Rail section point cloud filtering method and system based on multi-template matching

Abstract

This invention discloses a rail cross-section point cloud filtering method based on multi-template matching, comprising: acquiring the original point cloud of the rail cross-section; matching the original point cloud with a pre-constructed multi-template library, the multi-template library including a standard rail cross-section template and at least one typical defective rail cross-section template; during matching, a lightweight one-dimensional convolutional neural network is used to classify each point point by point to identify whether each point belongs to a rail category or a noise category, the rail categories including standard rails and defective rails; based on the classification results, points belonging to the rail category are retained, and points belonging to the noise category are removed to obtain a clean rail cross-section point cloud. This invention, through multi-template matching and lightweight one-dimensional convolutional neural network point-by-point classification, effectively removes noise from fasteners, ballast, etc., while fully preserving defect features such as side wear, vertical wear, and rail head spalling, providing a reliable clean rail cross-section point cloud data foundation for subsequent high-precision calculation of geometric parameters such as gauge, superelevation, and wear.

Description

Technical Field

[0001] This invention relates to the field of railway track inspection technology, and in particular to a method and system for filtering point clouds of rail cross sections based on multi-template matching. Background Technology

[0002] Track geometry parameters, including gauge, superelevation, level, alignment, elevation, and rail wear, are core indicators for evaluating track smoothness and train safety. The current mainstream method for dynamically measuring track geometry parameters is to use a line laser profilometer to acquire two-dimensional point clouds of rail cross-sections. However, due to the complexity of the actual railway environment, denoising of these point clouds is necessary. These point clouds contain not only the rail itself but also numerous noise points from fasteners, ballast, and other components, some of which have geometric shapes similar to the rail profile. Existing point cloud denoising methods mainly employ statistical filtering and radius filtering, which are insufficient for quickly removing noisy point clouds. For example, to remove fastener noise close to the rail, a large filtering radius is often required, which can lead to the accidental deletion of rail head edge points, affecting the accuracy of subsequent gauge calculations. In summary, the pursuit of improved noise reduction efficiency has led to the misclassification and removal of rail defects with morphologically similar characteristics to noise (such as lateral wear, vertical wear, and rail head spalling). This results in a significant decrease in the accuracy of subsequent calculations of core geometric parameters such as track gauge, superelevation, and wear, and may even render rail defects undetectable. Furthermore, while deep learning point cloud segmentation technology can improve classification accuracy, it requires a large number of manually labeled samples, resulting in extremely high labeling costs. Additionally, the large number of network parameters makes it difficult to run in real-time on vehicle-mounted embedded platforms, limiting its engineering practicality.

[0003] Therefore, there is an urgent need for a filtering method that can simultaneously identify standard rails and defective rails, effectively eliminating noise from fasteners, ballast, etc., while fully preserving defect characteristics such as wear and breakage, so that the geometric parameters of the entire track can be accurately calculated based on this. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for filtering rail cross-section point clouds based on multi-template matching, which can effectively overcome the above-mentioned defects.

[0005] Specifically, in a first aspect, the present invention provides a method for filtering point clouds of rail cross-sections based on multi-template matching, the method comprising the following steps: S1. Data Acquisition: Acquire the original point cloud of the rail cross-section; S2. Multi-template matching collaborative filtering: The original point cloud is matched with a pre-constructed multi-template library, which includes a standard rail cross-section template and at least one typical defective rail cross-section template; the matching includes: classifying each point in the original point cloud point by point using a lightweight one-dimensional convolutional neural network to identify whether each point belongs to the rail category or the noise category; wherein, the rail category includes standard rails and defective rails; S3. Retain Filtering Results: Based on the point-by-point classification results, retain the points belonging to the rail category and remove the points belonging to the noise category to obtain a clean rail cross-section point cloud.

[0006] By matching the original point cloud with a multi-template library containing standard and defect templates, and using a lightweight one-dimensional convolutional neural network to classify each point point by point, the system achieves accurate differentiation between noise and standard rails and defective rails. It improves the defect retention rate while achieving a high noise removal rate, solving the technical problem of traditional filtering losing defects in order to remove noise, and further making the final geometric parameter calculation more accurate.

[0007] Optionally, S2 further includes a coarse registration sub-step prior to the point-by-point classification, the coarse registration sub-step including: rigidly registering the original point cloud with the standard rail section template to eliminate the overall translational offset of the original point cloud.

[0008] By rigidly registering the original point cloud with a standard rail section template before point-by-point classification, the overall translational shift of the section caused by train vibration is eliminated, providing a unified spatial reference for subsequent point-by-point classification by a one-dimensional convolutional neural network and avoiding classification errors caused by coordinate shift.

[0009] Optionally, the input of the lightweight one-dimensional convolutional neural network is a one-dimensional point cloud sequence of a single rail cross-section after rigid registration, and the output is the probability that each point in the cross-section belongs to a predefined category; the predefined categories include four rail subcategories: rail standard, rail side wear, rail vertical wear, and rail spalling, as well as three noise subcategories: noisy fasteners, noisy ballast, and noisy other; the structure of the one-dimensional convolutional neural network includes, in sequence: multiple convolutional layers, a global average pooling layer, one or more fully connected layers, and an output layer.

[0010] By defining the specific structure of the one-dimensional convolutional neural network—multiple convolutional layers, a global average pooling layer, a fully connected layer, and an output layer—and predefined categories—four rail subcategories and three noise subcategories—a lightweight design was achieved, reducing single-section inference time and meeting the real-time processing requirements of the vehicle-mounted edge computing platform. Meanwhile, the refined rail subcategory labels provide a data foundation for subsequent adaptive fitting.

[0011] Optionally, after retaining points belonging to the rail category and removing points belonging to the noise category based on the point-by-point classification results, the method further includes a post-processing sub-step, which includes: performing radius filtering on the retained point cloud to remove isolated noise points.

[0012] By performing radius filtering after preserving the rail category point cloud, isolated noise points that may remain after point-by-point classification are eliminated, further improving the quality of the clean rail cross-section point cloud and providing cleaner data input for subsequent geometric parameter calculations.

[0013] Optionally, the step of pre-building the multi-template library includes: Multiple real, defect-free rail section samples were obtained, and the standard rail section template was generated statistically. Multiple real rail cross-section samples with specific defect types are obtained. After coordinate normalization of the multiple rail cross-section samples, typical defect rail cross-section templates corresponding to the specific defect types are statistically generated. The specific defect types include at least one of side wear, vertical wear, and rail head spalling.

[0014] By acquiring multiple real rail cross-section samples with and without defects and with typical defects, standard templates and defect templates are generated statistically after coordinate normalization. This enables the automatic construction of a multi-template library without manual annotation, avoiding the high cost problem of requiring a large number of labeled samples for deep learning methods.

[0015] Optionally, step S3 may be followed by a geometric parameter calculation step, which includes: calculating the geometric parameters of the track based on the point cloud of the clean rail cross-section.

[0016] By computing the geometric parameters of the track based on point cloud computing of pure rail cross-sections, the filtering results are transformed into quantitative indicators such as gauge, superelevation, and wear that can be used in engineering, thus realizing a value closed loop from point cloud filtering to engineering applications.

[0017] Optionally, the geometric parameter calculation step further includes: adaptively fitting the point cloud of different regions using the category labels attached to the point cloud of the clean rail cross section; The region includes the rail top region and the working edge region. When adaptively fitting the point cloud in the rail top region, points with rail vertical grinding or rail standard labels are preferred. When adaptively fitting the point cloud in the working edge region, points with rail side grinding labels are preferred.

[0018] By utilizing the category labels attached to the point cloud of the pure rail cross section, points with rail vertical wear or rail standard labels are prioritized for circular arc fitting in the rail top area, and points with rail side wear labels are prioritized for linear fitting in the working edge area. This achieves adaptive fitting for different areas, significantly reducing the deviation of track gauge measurement and side wear.

[0019] Optionally, the method further includes a result output step: comparing the geometric parameters with a preset safety threshold, and generating a fault alarm record when the geometric parameters exceed the limit; the fault alarm record includes at least one of mileage information, stock type, fault type, measured value, allowable limit, and recommended maintenance measures.

[0020] By comparing the calculated geometric parameters with preset safety thresholds and generating a fault alarm record containing mileage, track type, fault type, measured value, allowable limit and recommended maintenance measures when the limit is exceeded, the automatic judgment and output of the detection results are realized, providing a direct basis for railway maintenance and preventive maintenance.

[0021] In a second aspect, the present invention provides a rail cross-section point cloud filtering system based on multi-template matching, for implementing the rail cross-section point cloud filtering method based on multi-template matching as provided in the first aspect above, the system comprising: The data acquisition module is used to acquire the original point cloud of the rail cross-section; A multi-template matching collaborative filtering module is used to match the original point cloud with a pre-built multi-template library, which includes a standard rail cross-section template and at least one typical defective rail cross-section template. The matching includes: classifying each point in the original point cloud point by point using a lightweight one-dimensional convolutional neural network to identify whether each point belongs to a rail category or a noise category; wherein, the rail category includes standard rails and defective rails. The filter result retention module is used to retain points belonging to the rail category and remove points belonging to the noise category based on the point-by-point classification results, thereby obtaining a clean rail cross-section point cloud.

[0022] By constructing a data acquisition module, a multi-template matching collaborative filtering module, and a filtering result retention module corresponding to the method claims, a corresponding system protection is formed.

[0023] Optionally, the system further includes: The geometric parameter calculation module is used to calculate the geometric parameters of the track based on the point cloud of the clean rail cross-section. The result output module is used to compare the geometric parameters with preset safety thresholds. When the geometric parameters exceed the limits, a fault alarm record or maintenance guide sheet containing accurate mileage positioning information and recommended maintenance measures is generated.

[0024] By integrating a geometric parameter calculation module and a result output module into the system, geometric parameters are calculated based on pure point cloud data, and fault alarm records or maintenance guidance sheets containing accurate mileage positioning information and recommended maintenance measures are generated. This achieves end-to-end fully automated detection from raw point cloud input to maintenance guidance sheet output.

[0025] This invention provides a method for filtering rail cross-section point clouds based on multi-template matching. The method includes the following steps: S1, Data Acquisition: Acquire the original point cloud of the rail cross-section; S2, Multi-template Matching Collaborative Filtering: Match the original point cloud with a pre-constructed multi-template library, which includes a standard rail cross-section template and at least one typical defective rail cross-section template; the matching includes: classifying each point in the original point cloud point by point using a lightweight one-dimensional convolutional neural network to identify whether each point belongs to the rail category or the noise category; wherein, the rail category includes standard rails and defective rails; S3, Retaining Filtering Results: Based on the point-by-point classification results, retain the points belonging to the rail category and remove the points belonging to the noise category to obtain a clean rail cross-section point cloud. This invention matches the original point cloud with a multi-template library containing standard templates and at least one typical defect template, and uses a lightweight one-dimensional convolutional neural network to classify each point point by point to identify the rail category or noise category to which each point belongs. This effectively removes irrelevant noise, such as fastener and ballast noise, while fully preserving the required defect features such as side wear, vertical wear, and rail head breakage. This ensures that defects are not lost while removing noise, and provides a reliable and clean rail cross-section point cloud data foundation for high-precision calculation of subsequent geometric parameters such as gauge, superelevation, and wear. Attached Figure Description

[0026] To more clearly illustrate the technical solution of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A schematic diagram of the process for a rail cross-section point cloud filtering method based on multi-template matching provided by the present invention; Figure 2 A schematic diagram of a rail cross-section point cloud filtering system based on multi-template matching is provided for this invention. Figure 3 A functional framework diagram of a railway track geometry parameter calculation system provided by the present invention; Figure 4 A schematic diagram of a railway track geometry parameter calculation system provided by the present invention; Figure 5 An internal structural diagram of a computer device in one embodiment is shown. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0029] References to "an embodiment," "an exemplary embodiment," etc., in this invention indicate that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment must include that specific feature, structure, or characteristic. Furthermore, these phrases do not necessarily refer to the same embodiment. Additionally, when a specific feature, structure, or characteristic is described in connection with an embodiment, it is assumed that, whether explicitly described or not, the influence of such feature, structure, or characteristic on other embodiments is within the knowledge of those skilled in the art.

[0030] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "first", "second", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0031] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" or "several" means two or more, unless otherwise explicitly specified.

[0032] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0033] Based on the foregoing, current technologies for detecting track geometry parameters still have the following shortcomings: 1. Large data volume and limited processing efficiency: The number of cross-sectional point clouds collected over hundreds of kilometers of the entire line is enormous (approximately 2 million cross-sectional points per kilometer), and each cross-sectional point cloud itself contains a large number of noise points such as fasteners, ballast, and oil stains. Existing filtering methods (such as statistical filtering and radius filtering) rely on local density or distance distribution, which are ineffective in removing noise with similar shapes, such as fasteners, resulting in long post-processing times and making it difficult to achieve true real-time full-line detection. 2. Existing methods generally suffer from a technical bias: In order to improve denoising efficiency, a large radius filtering parameter is usually used, which leads to the misjudgment and removal of morphologically similar defect features (such as wear and chipping), making it impossible to effectively perform subsequent core defect detection tasks such as wear and corrugation. 3. Deep learning is used for cross-sectional point cloud classification, but it requires a large number of manually labeled samples, resulting in extremely high labeling costs, and the large number of network parameters makes it difficult to run in real time on an onboard embedded platform.

[0034] To address the shortcomings of existing technologies, this invention innovatively designs an intelligent filtering and geometric parameter calculation method for rail cross-section point clouds based on multi-template collaborative matching. This provides a system capable of quickly and accurately removing cross-section noise while fully preserving wear and deformation characteristics. The method constructs a multi-template library containing standard templates and various typical defect templates. A lightweight deep learning network is used to classify each point on the cross-section into multiple categories: rail standard, rail side wear, rail vertical wear, rail spalling, noise, etc. Points of all rail categories are retained. Based on this, geometric indicators such as gauge, superelevation, vertical wear, side wear, and corrugation parameters are accurately calculated. This aims to solve the problems of low calculation accuracy and high rate of missed / false detections of defects in current dynamic detection of track geometric parameters, providing reliable data support for refined maintenance of railway engineering. This invention is applicable to rail cross-section data preprocessing, noise removal from fasteners / ballast, and millimeter-level accuracy calculation of geometric parameters such as gauge, superelevation, and wear in high-speed railways, heavy-haul railways, and urban rail transit.

[0035] For details, please refer to Figure 1 This invention provides a flowchart of a rail cross-section point cloud filtering method based on multi-template matching. The method includes: S1, data acquisition: acquiring the original point cloud of the rail cross-section; S2, multi-template matching collaborative filtering: matching the original point cloud with a pre-constructed multi-template library, which includes a standard rail cross-section template and at least one typical defective rail cross-section template; the matching includes: classifying each point in the original point cloud point by point using a lightweight one-dimensional convolutional neural network to identify whether each point belongs to the rail category or the noise category; wherein, the rail category includes standard rails and defective rails; S3, retaining the filtering results: based on the point-by-point classification results, retaining the points belonging to the rail category and removing the points belonging to the noise category to obtain a clean rail cross-section point cloud.

[0036] In one feasible implementation, the data acquisition in step S1 is based on the acquisition by a line laser profilometer. Step S1 specifically includes: S101 Equipment Installation and Calibration Behind the bogie at the bottom of the operating train, two line laser profilers are symmetrically installed on the left and right sides via a high-precision integrated bracket (aiming at the left and right rails respectively).

[0037] To ensure that the cross-sections acquired by the left and right profilometers are located on the same cross section, the following measures are taken: Mechanical coplanar installation: The bracket is precision-machined to ensure that the laser lines of the left and right profilometers are coplanar. After installation, a laser collimator is used for calibration to ensure that the projections of the two laser lines on the top surface of the rail completely coincide. The fixed longitudinal offset Δx (mm) between the two is recorded and saved to the configuration file.

[0038] Hardware synchronous triggering: Based on the pulse signal output by the train odometer, two TTL trigger signals are generated simultaneously after frequency division by the FPGA and sent to the left and right profilers respectively, ensuring that the two devices collect cross-sections at the same time.

[0039] Mileage recording: During each data acquisition phase, the cumulative mileage value from the wheel encoder is recorded synchronously. left and mileage right The accuracy is better than ±1mm. To simplify the formula writing, mileage will be used in the following text. left Let it be S L mileage right Let it be S R .

[0040] S102 Data Acquisition Process The operating train travels along the track at normal speed. A line laser profilometer continuously acquires cross-sectional data of the rails at a set frequency, typically 1kHz to 2kHz. The specific frequency can be set according to the engineering requirements of those skilled in the art and is not limited here. Each cross-section outputs a set of two-dimensional point clouds. Left femoral section: S L ={(y i L Z i L )}, i=1…N; Right femur section: S R ={(y i R , z i R )}, i=1…N; Where y is the horizontal coordinate (perpendicular to the rail direction, with the zero point located at the center of the profilometer installation), and z is the vertical coordinate (relative to the sensor reference plane). The raw data is stored in the vehicle's solid-state drive in binary stream form, and key frames are simultaneously transmitted back to the ground monitoring center via 4G / 5G.

[0041] Based on the data obtained above, this invention constructs standard track section filtering and typical defect track section filtering. A convolutional neural network is used to perform feature matching on the acquired initial section point cloud image, achieving rapid noise point cloud removal. Specifically: In one feasible implementation, the construction steps of the pre-built multi-template library mentioned in step S2 include: acquiring multiple real, defect-free rail section samples and statistically generating standard rail section templates; acquiring multiple real rail section samples with specific defect types, performing coordinate normalization on the multiple rail section samples, and statistically generating typical defect rail section templates corresponding to the specific defect types; the specific defect types include at least one of lateral wear, vertical wear, and rail head spalling. This invention, by acquiring multiple real, defect-free and typical defect rail section samples, and statistically generating standard and defect templates after coordinate normalization, achieves automatic construction of a multi-template library without manual annotation, avoiding the high cost problem of requiring a large number of labeled samples in deep learning methods.

[0042] It is understandable that, in order to meet the requirements of multi-template collaborative filtering to remove noisy point clouds, the multi-template library is the basis for the collaborative filtering of this invention. Specifically, the multi-template collaborative filtering includes a standard template and several typical defect templates. The construction principle of the typical defect templates is: to face the most common, most harmful and easily obtained defect types in the railway field from actual detection data. Each template is generated by statistical synthesis of multiple real cross-sectional samples to ensure the representativeness, robustness and generalization ability of the template.

[0043] Furthermore, the standard rail section template T std The construction method includes: during the track maintenance window, using a high-precision static profilometer to scan a section of confirmed rail free of wear, pollution, and defects, continuously collecting 100 cross-sections, and taking the median of the z-value on the lateral coordinate y as the standard template point set. The standard template is stored as T. std ={(y i , z i )}, i=1…M. At the same time, extract its geometric features: the radius of the rail top arc, the slope of the working edge of the rail head, the width of the rail bottom, etc.

[0044] The construction method for typical rail section templates includes: based on statistical data and field experience from railway engineering departments, prioritizing the following three types of high-incidence and high-risk defects for template construction: 1. Side wear type: rail edges concave inward, rail head width reduced; 2. Vertical wear type: rail top lowered overall, rail head radius increased; 3. Rail head spalling type: localized material peeling off the rail top, forming depressions or gaps. For each defect type, a representative template is generated using a process of multi-case sampling, coordinate normalization, and statistical synthesis.

[0045] The specific methods for producing corresponding templates for various disease types are as follows: 1. Side wear template T side : Sampling: Select at least 20 sections confirmed to be lateral wear from historical track inspection vehicle data or on-site laser acquisition records. The sections must cover different wear levels and originate from different curve radius sections.

[0046] Preprocessing: Compare the point cloud of each cross-section with the standard rail cross-section template T. std ICP rigid registration (translation only, rotation not allowed) is performed to eliminate the overall offset caused by the installation posture. After registration, the transformation matrix (Δy, Δz) is obtained for each section.

[0047] Coordinate resampling: Interpolate all cross-sectional point clouds at equal intervals along the y-axis to the same sampling points (consistent with the y-coordinate of the standard template) to obtain the z value at each y-axis.

[0048] Statistical template generation: For a given wear grade group, assuming the group has K cross-sections, calculate the median z-value of all cross-sections at each y-coordinate:

[0049] The median is more robust to outliers (such as local noise in individual sections). All (y, z) template(y) Save the point set as the side wear template for this grade.

[0050] Hierarchical storage: Generate templates for different levels, for example: T side-light (Corresponding to wear of 2-4mm), T side-medium (4-6mm), T side-severe (6-10mm).

[0051] 2. Vertical wear template T vert : Sampling: Select more than 15 cross sections with obvious vertical wear, also from different sections.

[0052] Pre-processing and statistics: Similar to side wear, the cross-section is registered to the standard rail cross-section template T.std Then, the median of the z-value at each y-point is calculated in the rail top region (the region with the highest z-value). Since vertical wear is mainly manifested as a uniform overall decrease in the rail top, the template shape is similar to the standard rail cross-section template T. std Similar, but the top part of the track has shifted downwards as a whole.

[0053] Template generation: Template T is generated for typical values ​​of vertical wear. vert .

[0054] 3. Rail head falling block template T spall : Sampling: Collect more than 10 sections of the fallen block from the site. The fallen block area is usually located inside the rail top or near the working edge.

[0055] Processing: Compare each missing section with the standard rail section template T. std Registration, calculating the relationship between each point and the standard rail section template T std The z-value deviation. Only areas with a deviation less than -0.5mm (depression) are retained; the remaining areas are treated using a standard rail section template T. std Complete the gaps. Overlay the recessed areas of multiple templates and take the median to form the rail head breakage template T. spall .

[0056] Furthermore, after all the templates are constructed, the standard track section template and three typical defect track section templates are stored in the template library: T=(T std T side T vert T spall For cross sections with multiple disease characteristics (such as lateral wear + vertical wear), the single category with the highest matching degree should be selected for labeling.

[0057] In one feasible implementation, after building multiple template libraries, it is also necessary to align the mileage of the left and right cross sections. This is because mechanical installation and dynamic slippage may result in residual longitudinal offsets, so mileage interpolation and alignment of the left and right cross sections are required to ensure that the subsequent calculation results of geometric parameters such as gauge and superelevation are true, accurate, and meaningful. Specifically: for each cross section of the left track (S L ), search for the sequence of right femoral sections that matches S L The two most recent cross sections S R1 and S R2 (where S) R1 ≤S L ≤S R2 (As mentioned above, to simplify the formula writing, the mileage in the above text is omitted.) left Let it be S L mileage right Let it be S RFor each point (y) on the right femoral cross-section. R , z R Linear interpolation is performed to obtain the mileage S. L Point cloud of the right femoral section:

[0058]

[0059] If a calibration offset Δx exists, then S L Replace with S L After adding Δx, interpolation is performed. After alignment, the left and right sections strictly correspond to the same cross-section.

[0060] In this embodiment, the point clouds of the left and right rail sections after mileage alignment still contain a large number of noise points such as fasteners and ballast. This invention employs a multi-template collaborative matching method to judge each point in each section point cloud: if the local shape of the point matches the standard template or any defect template, it is determined to be a rail point and retained; otherwise, it is determined to be a noise point and discarded. The multi-template matching filtering process can sequentially include three sub-steps: coarse registration, used to eliminate overall offset; point-by-point multi-template matching; and post-processing, used to remove isolated points.

[0061] The coarse registration sub-step includes rigidly registering the original point cloud with the standard rail section template to eliminate the overall translational shift of the original point cloud. In this invention, by rigidly registering the original point cloud with the standard rail section template before point-by-point multi-template matching and classification, the overall translational shift of the section caused by train vibration is eliminated, providing a unified spatial reference for subsequent point-by-point classification by the one-dimensional convolutional neural network and avoiding classification errors caused by coordinate shifts. Specifically: Due to the vibrations caused by the operation of the testing vehicle, the cross-section undergoes slight translation (mainly lateral and vertical). It is necessary to first align the measured cross-section with the standard rail cross-section template T. std Rigid registration is performed to eliminate overall offset, ensuring the correspondence of local coordinates during subsequent point-by-point matching. For each cross-section, the measured point cloud S = {(y...} i , z i The ICP (Iterative Closest Point) algorithm and the standard rail section template T are used. std Registration is performed, allowing only translation (Δy, Δz), but not rotation. The registration objective function is:

[0062] After registration, the translation amounts (Δy, Δz) are obtained. The measured point cloud S of each cross section is translated as a whole by (Δy, Δz). After coarse registration, the measured cross section is roughly aligned with the template library, providing a correct spatial reference for subsequent point-by-point matching.

[0063] The point-by-point multi-template matching, as mentioned in step S2, involves using a lightweight one-dimensional convolutional neural network to classify each point in the original point cloud point by point to identify whether each point belongs to the rail category or the noise category. The input to the lightweight one-dimensional convolutional neural network is a one-dimensional point cloud sequence of a single rail cross-section after the aforementioned rigid registration. The output is the probability that each point in the cross-section belongs to a predefined category. The predefined categories include four rail subcategories: rail standard, rail side wear, rail vertical wear, and rail spalling; and three noise subcategories: noise fasteners, noise ballast, and other noise. The structure of the one-dimensional convolutional neural network includes, in sequence: multiple convolutional layers, a global average pooling layer, one or more fully connected layers, and an output layer.

[0064] In this invention, by defining the specific structure of the one-dimensional convolutional neural network (multiple convolutional layers, global average pooling layer, fully connected layer, and output layer) and predefined categories (four rail subcategories and three noise subcategories), a lightweight design with fewer than 50,000 parameters is achieved, reducing the single-section inference time and meeting the real-time processing requirements of the vehicle-mounted edge computing platform; at the same time, the refined rail subcategory labels provide a data foundation for subsequent adaptive fitting.

[0065] Understandably, in order to quickly and accurately determine whether each point is a rail (normal or defective) or noise, this invention prioritizes the use of a one-dimensional convolutional neural network (1D-CNN) to classify the cross-sectional point cloud point by point (other classification algorithms can also be used depending on the actual situation). A lightweight neural network is trained offline, and during online inference, the class probability is output for each point, retaining all points of rail category.

[0066] The input and output of this one-dimensional convolutional neural network (1D-CNN) are as follows: Each cross-section point cloud (already registered) is used as the one-dimensional sequence input, with an input dimension of (N, 2) (N being the number of cross-section points, including y and z coordinates). The output is the probability of each point belonging to a category, including: rail. 标准 Rails 侧磨 Rails 垂磨 Rails 掉块 ,noise 扣件 ,noise 道砟 ,noise 其他 The total number of categories is C=7. The network structure of this one-dimensional convolutional neural network (1D-CNN) can be three convolutional layers (kernel size 3, number of channels 32→64→128); global average pooling; fully connected layers (128→64); and an output layer (C=7 categories, including: rails). 标准 Rails 侧磨 Rails 垂磨 Rails 掉块 ,noise 扣件 ,noise 道砟 ,noise 其他Specifically, it includes the following layers in sequence: a first convolutional layer using 32 one-dimensional convolutional kernels of size 3 to output 32 feature maps; a second convolutional layer using 64 one-dimensional convolutional kernels of size 3 to output 64 feature maps; a third convolutional layer using 128 one-dimensional convolutional kernels of size 3 to output 128 feature maps; a global average pooling layer that compresses each feature map into a single value, outputting a 128-dimensional feature vector; a fully connected layer that reduces the 128-dimensional feature vector to 64 dimensions; and an output layer that maps the 64-dimensional feature vector to 7 categories, finally outputting the probability of each category via Softmax.

[0067] Furthermore, the training and inference method of this one-dimensional convolutional neural network (1D-CNN) is as follows: after the network is trained by samples, the measured cross-section is input into the network to obtain the class probability of each point. The probability threshold is set to 0.8. If the maximum probability is less than 0.8, it is corrected by neighborhood voting (after sorting the cross-section point cloud by y coordinate, for points with a probability lower than the threshold, the mode of the classes of the five adjacent points before and after it is taken as its final class).

[0068] Then proceed to step S3, retaining the filtering results: Based on the above point-by-point classification results, retain the points belonging to the rail category and remove the points belonging to the noise category, that is, retain all points with the label prefix "rail" (i.e., rail). 标准 Rails 侧磨 Rails 垂磨 Rails 掉块 It may also include steel rails. 波磨 Points can be selected based on specific engineering definitions (this embodiment does not impose any limitations), and all points with the prefix "noise" are removed to obtain a clean rail cross-section point cloud.

[0069] In one feasible implementation, in step S3, after retaining points belonging to the rail category and removing points belonging to the noise category based on the point-by-point classification results, the method further includes the aforementioned post-processing sub-step. This post-processing sub-step includes: performing radius filtering on the retained point cloud to remove isolated noise points. This invention, by performing radius filtering after retaining the rail category point cloud, removes potentially residual isolated noise points after point-by-point classification, further improving the quality of the clean rail cross-section point cloud and providing cleaner data input for subsequent geometric parameter calculations.

[0070] Specifically, radius filtering is performed on the retained point cloud: a radius of r = 2mm is preferably set, and if the number of neighboring points within the radius of a point is less than 3, it is discarded as isolated noise. The final output is a clean cross-sectional point cloud, with each point labeled with its category. After noise reduction through collaborative filtering using standard templates and typical defect templates, sporadic noise is further filtered out, finally generating a clean cross-sectional point cloud, and labeling the point cloud with its category.

[0071] Understandably, after the above multi-template collaborative filtering, the resulting clean rail cross-section point cloud has removed noise from fasteners, ballast, etc., while retaining the labels of various rail subcategories. Since the point cloud is still discrete, it needs to be fitted into continuous geometric features, such as rail top arcs and working edge straight lines, before calculating parameters such as gauge, superelevation, and wear based on these features. That is, step S3 includes a geometric parameter calculation step, which includes: calculating the track's geometric parameters based on the clean rail cross-section point cloud. Specifically: In this embodiment, the following fitting steps are performed on the left and right point clouds of each cross section. For simplicity, the following example uses a single-sided rail; the other side can be treated symmetrically.

[0072] For single-sided rails: the geometric parameter calculation steps further include: using the category labels attached to the point cloud of the pure rail cross section to adaptively fit the point cloud of different regions; wherein, the regions include the rail top region and the working edge region, when adaptively fitting the point cloud in the rail top region, points with rail vertical grinding or rail standard labels are given priority; when adaptively fitting the point cloud in the working edge region, points with rail side grinding labels are given priority.

[0073] Specifically, the filtered cross-sectional point cloud S clean ={(y i , z i label i (i=1…N) are sorted by y-coordinate from smallest to largest to obtain an ordered point sequence. Since the line laser profilometer outputs data in the scanning order, it is usually already ordered, but to ensure stability, it is explicitly sorted once. Based on the geometric shape of the point cloud, three key regions are automatically segmented: Rail top region: the top 20% of points with the highest z-values ​​(these points form the rail top arc). Working edge region: located inside the rail (the side with smaller y-values), and within 0–30 mm below the rail top, typically selected where y < -10 mm and z ∈ [z...]. max-30 , z max The points are labeled using the point cloud's category labels. i Auxiliary segmentation: For example, a label labeled "rail". 侧磨 Points labeled "" are prioritized for edge fitting; "rail" is the label for this purpose. 垂磨 Points marked with "" are preferentially used for track top fitting. Regarding the track bottom region: since this embodiment primarily utilizes the track top region and working edge region for geometric parameter calculation, the track bottom region can be used for other auxiliary calculations (such as track bottom slope detection). This invention does not limit this use and will not elaborate further.

[0074] Specifically, the fitting of the rail top circular arc in the rail top region is as follows: The least squares method is used to fit the circular arc. Let there be M points (y0, ...) in the track top region. j , zj The equation of the fitted circle is:

[0075] Among them (y) c , z c Let R be the coordinates of the center of the circle and R be the radius. Expand the equation and transform it into linear least squares form:

[0076] Let a = 2y c b=2z c c=R 2 -y c 2 -z c 2 Then for all points j, we have:

[0077] It can also be expressed as an AX=B matrix, where:

[0078]

[0079]

[0080] The highest point of the orbital apex is obtained as (y c , z c +R).

[0081] Specifically, the working edge line fitting in the working edge region is as follows: Select the working edge region point (y k , z k The least squares method is used to fit the straight line:

[0082] For K points, construct:

[0083]

[0084]

[0085] Solving for X, we get X = (A T A) -1 A T Y, we can obtain the slope k and intercept b.

[0086] In one feasible implementation, based on the above fitting results, the geometric parameters are calculated using the following formula. For the left and right rails, the highest point (y_t) of the left rail top is obtained respectively.cL , z cL +R L ), the highest point of the right stock rail (y) cR , z cR +R R ), and the left working side line y=k Lz +b L The right working side line y=k Rz +b R Based on the above information, calculate the track gauge, superelevation, vertical wear, and lateral wear respectively. Specifically: (1) Track gauge Track gauge is defined as the lateral distance between the working edges of the left and right rails at a point 16mm below the top of the rail. First, determine the z-coordinate corresponding to this measurement height: z g =z top -16, where z top Let z be the z-coordinate of the highest point of the rail top on this side. Since the highest points of the left and right rail tops may differ, calculate them separately: Left thigh height measurement: z gL =(z cL +R L )-16 Right thigh height measurement: z gR =(z cR +R R )-16 Substituting each into the equation of the working side line, we obtain the lateral position: Y l =k L ·z gL +b L y R =k R ·z gR +b R Measured track gauge: G=y R -y L Track gauge deviation (standard track gauge 1435mm, in millimeters): ΔG = G - 1435.

[0087] (2) Extremely high Superelevation is the elevation difference between the top surface of the outer rail and the top surface of the inner rail. Let the right rail be the outer rail (curved section), then: H=(z cR +R R )-(z cL +R L ) If the left rail is the outer rail, then take the opposite number. The superelevation deviation is: ΔH=HH design Where H design It is the design ultra-high value (which can be obtained from the line parameters).

[0088] (3) Vertical wear Vertical wear is the reduction in thickness of the rail top surface due to wheel-rail contact wear. (Based on a standard template T...) std The highest point of the orbit z std-top Based on: W vert =z std-top -(z c +R) Among them (z) c +R) represents the highest point measured at the top of the rail. W vert A positive value indicates wear. Note: For rails... 垂磨 The fitted rail top arc of the labeled points naturally reflects the shape after wear, so the fitting result can be used directly.

[0089] (4) Side wear Lateral wear refers to the width of the inward concavity of the working edge of the rail head due to wheel flange contact. Similarly, using a standard template, the lateral wear is measured at a position 16mm below the rail top. std Based on the reference, the lateral position y of the working edge at the same height was measured. meas : W side =|y std -y meas | Where y meas Based on the measured working side straight line at the measured height z g The place is worthwhile; y std Extracted from a standard template (or calculated by fitting the working edge straight line of the standard template). For rails 侧磨 The cross-section of the label, this value can be directly taken from the fitting result.

[0090] In one feasible implementation, the method further includes a result output step: comparing the geometric parameters with a preset safety threshold; when the geometric parameters exceed the limit, generating a fault alarm record; the fault alarm record includes at least one of the following: mileage information, branch, fault type, measured value, allowable limit, and recommended maintenance measures. Specifically: The calculated geometric parameters are compared with the allowable deviation thresholds for the corresponding speed class in the "Railway Line Repair Rules" (e.g., gauge ±2mm, horizontal ±4mm, side wear ≤5mm, etc.). For each section, if any parameter exceeds the limit, a defect alarm record is generated: "Mileage (accurate to meters) - Left and right tracks (if applicable) - Defect type (gauge exceeds limit, side wear exceeds limit, etc.) - Measured value allowable limit - Recommended maintenance measures (e.g., "rerouting", "grinding", etc.)".

[0091] Furthermore, to facilitate rapid location by on-site personnel, the following visualized mileage effect can be output: Two-dimensional waveform diagram: plots curves for parameters such as gauge deviation and superelevation deviation along the entire line. Sections exceeding limits are highlighted with a red background and marked with kilometer and 100-meter markers. Maintenance guidance sheet: For each point exceeding limits, a relative location description is automatically generated based on a pre-established landmark database (caten post number, level crossing, kilometer marker, etc.). For example: "Wear on the left side of the rail at K225+600 is 6.2mm. From catenary post #225-5 at K225+500, move 100 meters towards the greater mileage direction to the third sleeper."

[0092] All results can be output in the form of structured reports (Excel / CSV) and PDF documents, and can be automatically pushed to the engineering management system via the network.

[0093] In this invention, by comparing the calculated geometric parameters with preset safety thresholds, and generating a fault alarm record containing mileage, track type, fault type, measured value, allowable limit and recommended maintenance measures when the limit is exceeded, the automatic judgment and output of the detection results are realized, providing a direct basis for real-time condition maintenance and preventive maintenance of railway engineering.

[0094] Please see Figure 2 This invention provides a schematic diagram of a rail cross-section point cloud filtering system based on multi-template matching. The system includes: a data acquisition module 201 for acquiring the original point cloud of the rail cross-section; a multi-template matching collaborative filtering module 202 for matching the original point cloud with a pre-built multi-template library, which includes a standard rail cross-section template and at least one typical defective rail cross-section template; the matching includes: classifying each point in the original point cloud point by point using a lightweight one-dimensional convolutional neural network to identify whether each point belongs to the rail category or the noise category; wherein the rail category includes standard rails and defective rails; and a filtering result retention module 203 for retaining points belonging to the rail category and removing points belonging to the noise category based on the point-by-point classification results to obtain a clean rail cross-section point cloud.

[0095] Based on the above methods and systems, this invention also proposes a multi-template collaborative matching intelligent filtering and geometric parameter calculation system for rail cross-section point clouds. This system, based on synchronous data acquisition and centered on multi-template collaborative filtering and a one-dimensional convolutional neural network, constructs a fully automatic, real-time processing system from raw cross-section point cloud input to geometric parameter report output. This solves the engineering problems of incomplete noise removal, easy loss of defect features, and low computational efficiency in existing methods. The system mainly includes a data acquisition subsystem, a data preprocessing and filtering subsystem, and a parameter calculation and output subsystem. Its functional framework is as follows: Figure 3 As shown, the system structure diagram is as follows: Figure 4 As shown.

[0096] Please see Figure 3 This invention provides a functional framework diagram of a railway track geometry parameter calculation system. Specifically, the system includes a data acquisition subsystem 310, a preprocessing subsystem 320, and a result calculation subsystem 330. The data acquisition subsystem 310 includes a motion sensing module 311, a data transmission module 312, and a data storage module 313. The preprocessing subsystem 320 includes a filtering construction module 321, a point cloud noise reduction module 322, and a data alignment module 323. The result calculation subsystem 330 includes a feature fitting module 331, a parameter calculation module 332, and a result output module 333. Please continue reading... Figure 4This is a schematic diagram of a railway track geometry parameter calculation system provided by the present invention. The system includes: an acquisition layer 410, a data layer 420, a processing layer 430, a calculation layer 440, and an output layer 450. The acquisition layer 410 includes a motion sensing module 311, which specifically comprises a left-side laser profilometer 3111 and a right-side laser profilometer 3112. The data layer 420 includes a data transmission module 312 and a data storage module 313. The data transmission module 312 specifically comprises an onboard transfer unit 3121 and a 5G communication unit 3122, and the data storage module 313 specifically comprises a database 3131 and a log database 3132. The processing layer 430 includes a filtering construction module 321, a point cloud denoising module 322, and a data alignment module 323. The filtering construction module 321 is specifically composed of a standard module unit 3211 and a disease module unit 3212. The point cloud denoising module 322 is specifically composed of a neural network unit 3221, a feature matching unit 3222, and a filtering processing unit 3223. The data alignment module 323 is specifically composed of a left-right cross-section alignment unit 3231. The calculation layer 440 includes a parameter calculation module 332 and a feature fitting module 331. The parameter calculation module 332 is specifically composed of a parameter calculation unit 3321 and a wear calculation unit 3322. The feature fitting module 331 is specifically composed of a point cloud sorting unit 3311 and a track fitting unit 3312. The output layer 450 includes a result output module 333, which is specifically composed of a result output unit 3331 and a disease visualization unit 3332.

[0097] Combination Figure 3 and Figure 4 This is understandable: The mobile sensing module 311 consists of two high-precision line laser profilometers, one high-precision odometer encoder, and an embedded synchronization controller. The two profilometers are symmetrically mounted on the rear of the bogies at the bottom of the train via precision-machined rigid brackets and calibrated with a laser collimator to ensure that their laser lines are coplanar and perpendicular to the rail surface. Hardware synchronization is achieved by using the pulse signal from the odometer encoder as a reference, which is then frequency-divided by an FPGA to simultaneously trigger the left and right profilometers to acquire data, ensuring that the left and right cross-sections strictly correspond to the same odometer cross-section, with a synchronization accuracy better than 0.1mm (equivalent mileage). Data acquisition includes continuous cross-section acquisition by the profilometers at a frequency of 1kHz~2kHz during normal train operation, with each cross-section outputting a set of two-dimensional point clouds ((y,z) coordinates). The system simultaneously records the cumulative mileage value, timestamp, and vehicle speed for each cross-section. The final data storage method involves streaming the raw data in binary format to an onboard industrial-grade solid-state drive (capacity ≥1TB, supporting wide-temperature operation), generating a data block file every 10 minutes.

[0098] The data transmission module 312 temporarily stores the raw data (binary format) in streaming mode on the vehicle's solid-state drive (capacity ≥ 2TB, write speed ≥ 500MB / s) to prevent data loss due to network interruptions. Simultaneously, it transmits key frames (such as a low-resolution preview image per kilometer and device status logs) to the ground center in real time via a 4G / 5G communication module for monitoring data acquisition quality. After acquisition, the complete data is exported to the ground processing server via USB 3.0 or 10 Gigabit Ethernet.

[0099] The data storage module 313 adopts a hierarchical storage strategy: raw data is stored in folders according to timestamps (laser point clouds are stored as .las / .bin, and POS data is stored as .txt / .csv), while processed intermediate data (denoised point clouds) is stored in a solid-state cache and a mechanical hard drive archive. A metadata database is established to record acquisition time, line mileage, sensor parameters, processing logs, etc., supporting subsequent retrieval and reproduction.

[0100] The multi-template collaborative filtering module is the core processing module, running on an in-vehicle embedded GPU computing platform or a ground-based high-performance server. This module includes a data storage module 313, a filter construction module 321, a point cloud noise reduction module 322, and a data alignment module 323. The implementation process based on the internal structure of these modules is as follows: Left and right cross-section mileage alignment: Read the left and right cross-section data and mileage values, and use linear interpolation to align the left and right cross-sections according to the calibrated longitudinal offset Δx and real-time vehicle speed, and output the left and right cross-section point cloud pairs under the same mileage.

[0101] Multiple template library storage: Pre-stores standard templates T built offline. std Side wear template T side Vertical wear template T vert Rail head falling block template T spall The template is saved as a structured file, containing the (y, z) coordinates and geometric feature parameters of each template point.

[0102] Coarse registration: For each measured cross section, the ICP algorithm (translation only) and the standard template T are used. std Perform registration, calculate the overall lateral offset Δy and vertical offset Δz of the cross-section, and output the registered cross-section point cloud.

[0103] 1D-CNN Neural Network Inference: A pre-trained lightweight one-dimensional convolutional neural network model (parameter count <50k) is loaded. The registered cross-sectional point cloud (fixed number of points 1024, normalized coordinates) is input into the network, and parallel inference outputs a 7-dimensional probability vector for each point (corresponding to 7 categories). A probability threshold of 0.8 is set, and the classification results are corrected by neighborhood voting. This unit is optimized based on TensorRT or ONNXRuntime, with a single cross-sectional inference time of <1ms.

[0104] Filtering and Post-processing: Based on the classification results, retain the point clouds of all rail categories (standard, side-wear, vertical wear, and spalling) and remove the point clouds of all noise categories (fasteners, ballast, and others). Finally, perform radius filtering (this radius parameter can be configured according to the track grade and point cloud density; a value of 2mm is acceptable, and points with fewer than 3 neighboring points are removed), outputting clean cross-sectional point clouds with category labels.

[0105] The results calculation subsystem is used to calculate geometric parameters. This calculation runs on a ground server or a high-performance onboard industrial control computer, responsible for computing various track geometric parameters based on pure cross-section points and generating reports. Specifically, it includes: Feature fitting module 331: First, sort the filtered left and right cross-sectional point clouds, then perform least-squares fitting of the rail top circular arc and least-squares fitting of the working edge straight line respectively. This module utilizes the point cloud's category labels (such as "rail") 侧磨 Points marked with "" are prioritized for working edge fitting to improve fitting robustness.

[0106] Parameter calculation module 332: Based on the fitting results, calculate the following parameters according to the algorithm specified in the "Railway Line Repair Rules": track gauge (spacing of working sides 16mm below the rail top) and deviation, superelevation (height difference between the left and right rail tops) and deviation, vertical wear and lateral wear (relative to the standard template).

[0107] Result Output Module 333: This module includes an over-limit judgment and alarm generation unit and a result visualization and defect output unit. The over-limit judgment and alarm generation unit compares calculated values ​​with preset thresholds and generates alarm records for over-limit sections, including: mileage, track type, parameter name, measured value, allowable limit, and recommended maintenance measures. It automatically generates text descriptions that can be quickly located on-site, based on a pre-established landmark database (caten wire supports, level crossings, kilometer markers, etc.). The result visualization and defect output unit generates two-dimensional waveform diagrams (PNG / PDF) of parameters such as gauge deviation, superelevation deviation, and wear, highlighting over-limit sections; outputs structured data reports (Excel / CSV) containing all calculation results for each section; outputs maintenance guidance sheets (PDF) listing all over-limit points and their precise location descriptions; and supports automatic push of result data to the Railway Maintenance Management Information System (PWMIS) via 4G / 5G or internal network.

[0108] To enable those skilled in the art to better understand and implement the present invention, a specific embodiment is provided below for detailed description. It should be noted that this specific embodiment is only for better understanding of the present invention and is not intended to limit the specific engineering implementation of the present invention. This embodiment takes a section of a heavy-haul railway downline from K100+000 to K150+000 (50 km in length) as an example, and combines specific equipment models and parameters to provide a detailed description of the technical solution of the present invention. Specific Implementation Example 1: Example prerequisites: Work section: A heavy-haul railway down line, from K100+000 to K150+000, with a total length of 50 kilometers. This section includes straight sections, curves with a radius of 600m and 1000m, and a 12‰ gradient.

[0110] Operating vehicle: HXD1 electric locomotive, used for normal traction of freight trains at a speed of 50~70km / h.

[0111] Environmental conditions: Nighttime, dry, with a small amount of ballast splash and fastener reflections along the route.

[0112] Equipment configuration: left and right line laser profilometers (Gocator 2380, sampling rate 1kHz), odometer encoder (20,000 pulses / revolution), and onboard GPU computing platform (Jetson Orin 32GB).

[0113] Step S1: System calibration and initialization S11 Equipment Installation and Rigid Coplanar Calibration Behind the bottom bogie of the HXD1 locomotive, two Gocator 2380 profilers are mounted using an integrated precision bracket. A laser collimator is used for calibration to ensure the projections of the two laser lines onto the rail top surface are completely aligned. The fixed longitudinal offset Δx = 15mm between the two profilers (the left profiler is mounted further forward) is recorded. The offset is then written into the configuration file.

[0114] S12 Coordinate System Definition and Calibration Define the coordinate system for each profilometer: the origin is the laser emission center, the y-axis is horizontal (positive to the right), and the z-axis is vertical (positive upward). The profilometer is calibrated on-site using a calibration block (an isosceles trapezoidal block of known height) to obtain the scale factor and offset in the y and z directions, ensuring that the measurement error is <±0.1mm.

[0115] S13 Multi-template Library Loading The offline-built standard template T std Side wear template (T) side-light T side-medium T side-severe Vertical wear template T vert Rail head falling block template T spall Loaded into the storage unit of the onboard computing platform. The template data is in HDF5 format and contains the (y, z) coordinate sequence of each template point.

[0116] S14 System Self-Test Before the locomotive leaves the depot, the geometric parameter calculation system software is started. The software loads the configuration file and multiple template library, and performs self-checks on the communication status of the left and right profilers and the odometer encoder pulse signal. All statuses are confirmed to be "normal". The system time is synchronized with GNSS time.

[0117] Step S2: Dynamic Data Acquisition S21 data acquisition started The locomotive departs from point K100+000 and maintains a stable speed of 60 km / h. The operator clicks "Start Data Collection".

[0118] S22 Synchronous Trigger Acquisition Based on 20,000 pulses per meter output from the odometer encoder, the FPGA generates 1kHz TTL pulses after frequency division, simultaneously triggering the left and right profilers to acquire cross-sections. The cumulative pulse count for each cross-section is recorded (converted to odometer value). The locomotive travels 50 kilometers, acquiring approximately 50,000 cross-sections (sampling interval approximately 1 meter).

[0119] S23 Data Streaming Write The raw point cloud generated for each section (approximately 1500 points per profilometer) is written to the vehicle's solid-state drive in real time as a binary stream. The left and right data are stored separately, with filenames containing timestamps and starting mileage.

[0120] Step S3: Data Preprocessing and Multi-Template Collaborative Filtering S31 Mileage Alignment After data acquisition, the preprocessing program reads the data from the left and right stocks. For each mileage S of the left stock... L =K100+001.00, find mileage S in the right-hand data. R1 =K100+00.998 and S R2 =K100+001.002. Linear interpolation is performed on the point cloud of the right cross-section to obtain S. L The right section at that location. Simultaneously, based on Δx = 15mm, the left section mileage is corrected to S. L+0.015 Then perform matching. Output 50,000 pairs of point clouds that precisely correspond to the left and right cross-sections of the mileage.

[0121] S32 Multi-template Collaborative Filtering The filtering process is illustrated using a typical cross-section at K120+350 (with fastener noise and slight side wear). S321 coarse registration: The program will register the measured left femoral section point cloud S={(y i , z i ))} and standard template T std ICP registration was performed. The global translation was calculated (Δy=2.3mm, Δz=-1.1mm), indicating that the cross-section was translated 2.3mm to the right and 1.1mm downward due to vibration. The measured cross-section was then translated as a whole (-2.3mm, 1.1mm) to complete the coarse registration.

[0122] S322 Point-by-Point 1D-CNN Classification: The registered cross-section (fixed sampling of 1024 points) is input into a pre-trained 1D-CNN network. The network outputs a 7-dimensional probability for each point. Results show that approximately 120 points near the working edge of the rail head represent the "rail" classification. 侧磨 "Probability > 0.92; 80 points on the top of the rail" 标准 "Probability > 0.85; approximately 50 points of 'noise' in the fastener area" 扣件 "Probability > 0.95; approximately 30 points in the track bed area exhibit noise." 道砟 "Probability > 0.90. Post-processing: A point located on the edge of the working area has a probability of only 0.76, and its five preceding and following points are all 'rails'." 侧磨 "" was corrected to "rails" through neighborhood voting. 侧磨 ".

[0123] S323 Filtering and Post-processing: The program retains all tags labeled "rail". 标准 "rails" 侧磨 Points labeled "noise" (approximately 200 in total) were removed. 扣件 ","noise道砟 The points (approximately 80 in total) were then subjected to radius filtering (r=2mm): isolated noise points (nearest neighbor >2.5mm) were removed. The final result was a clean left-side cross-section point cloud of approximately 198 points, each with a category label. The entire process for processing a single cross-section took approximately 0.8ms.

[0124] S33 generates a clean cross-sectional point cloud database. Perform steps S31 to S32 above on all 50,000 cross sections to generate a database containing clean cross section point clouds and category labels for subsequent geometric parameter calculations.

[0125] Step S4: Geometric parameter calculation and output S41 Feature Fitting Taking the K120+350 left section processed in step S3 as an example: Rail top arc fitting: The program selects the label "rail". 标准 "and rails" 侧磨 "And the top 20% of points with the highest z-values ​​(40 points in total). The least squares method is used to fit the equation of the circular arc. The coordinates of the center (y) are obtained." c =-12.3mm, z c =-8.1mm), radius R=300.2mm. The coordinates of the highest point of the rail top are (y=-12.3mm, z=292.1mm).

[0126] Working edge straight line fitting: The program selects the label "rail". 侧磨 "And points located 0-30mm below the top of the rail (60 points in total). The least squares method was used to fit the straight line. The equation of the straight line was obtained as: y=-0.35*z-15.2."

[0127] S42 parameter calculation Track gauge: Calculate the highest point z of the left track top. Ltop =292.1mm, then the measured height z gL =292.1-16=276.1mm. Substituting this into the left working side straight line, we get y. L =-0.35*276.1-15.2=-111.8mm. Similarly, processing the right section (in this example, the right section is a standard section with no obvious wear), we obtain z. Rtop =292.5mm, y R =1435.0-111.8=1323.2mm (assuming standard track gauge is 1435mm). Actual track gauge G=y R -y L =1323.2-(-111.8)=1435.0mm. Track gauge deviation ΔG=0.0mm.

[0128] Side wear: from standard template T std The standard working edge position y is located 16mm below the top of the rail. std =-116.0mm. Measured position y meas =y L =-111.8mm. Side wear W side =|y std -y meas| =4.2mm.

[0129] S43 Over-limit detection and alarm The calculated lateral wear of 4.2mm was compared with the warning value (3mm) and clearance value (5mm) for heavy-haul railways in the "Railway Line Repair Rules". Since 4.2mm > 3mm but < 5mm, the system generated a yellow warning record: [Mileage K120+350][Left Stock][Side Wear Exceeds Warning][Measured Value 4.2mm][Warning Value 3mm][Recommendation to Strengthen Observation and Arrange a Grinding Plan]. Simultaneously, through the landmark database, it was found that this location is 50 meters towards the greater mileage direction from contact wire support #120-3 at K120+300.

[0130] S44 Results Visualization and Report Output Full-line waveform diagram: The system generates waveform diagrams for gauge deviation, superelevation, lateral wear, and vertical wear over a 50-kilometer section. In the section from K120+300 to K120+400, the lateral wear curve is displayed as a continuous bulge (peak value 4.5mm), with a bright yellow warning background.

[0131] Maintenance Guide Sheet: The system generates a PDF maintenance guide sheet listing all the over-limit points. The first item is: "Wearing on the left side of the rail at K120+350 is 4.2mm, warning. On-site location: Walk about 50 meters from the contact wire support #120-3 at K120+300 towards the greater mileage direction to find the 25th sleeper (calculated at 1.67 sleepers per meter)." Structured reports: The system exports Excel files containing all data for each section, including mileage, top elevation of left and right rails, gauge, superelevation, vertical wear, and lateral wear.

[0132] Step S5: Model Delivery and Application Engineers imported the generated reports and waveforms into the railway track maintenance management information system. The track maintenance section, based on the maintenance guidance sheet, scheduled the evening maintenance window to perform flaw detection and grinding on the rail at K120+350, and closed the loop on the warning record in the system.

[0133] This invention, by setting up a rail cross-section point cloud filtering method based on multi-template matching, brings the following significant technical benefits: 1. Precise filtering while preserving disease features: By combining standard templates with typical disease templates for collaborative matching and 1D-CNN point-by-point classification, noise such as fasteners and ballast is removed (removal rate >98%), while the point cloud of disease areas such as side wear, vertical wear, and chipping is completely preserved (preservation rate >99%), solving the problem that traditional filtering loses disease features in order to remove noise.

[0134] 2. Strong real-time onboard processing capability: The lightweight 1D-CNN network has less than 50,000 parameters, and the single-section inference on the edge computing platform is less than 1ms, with a processing speed of more than 1000 sections / second, which meets the real-time geometric parameter calculation requirements under normal train operation conditions.

[0135] 3. High parameter accuracy and fully automatic output: Adaptive fitting based on category labels ensures track gauge measurement deviation ≤ ±0.5mm and side wear deviation ≤ 0.2mm; the entire process requires no manual intervention and automatically generates over-limit reports and maintenance guide sheets with accurate mileage positioning, supporting railway track condition repair and preventive maintenance.

[0136] Figure 5 An internal structural diagram of a computer device according to one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 5 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program, which, when executed by the processor, causes the processor to implement the method described in this invention. The internal memory may also store a computer program, which, when executed by the processor, causes the processor to perform the method described in this invention. Those skilled in the art will understand that… Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0137] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the method of the present invention.

[0138] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0139] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0140] The above description is merely a specific implementation of this application and is not intended to limit the scope of protection of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for filtering point clouds of rail cross-sections based on multi-template matching, characterized in that, The method includes the following steps: S1. Data Acquisition: Acquire the original point cloud of the rail cross-section; S2. Multi-template matching collaborative filtering: The original point cloud is matched with a pre-constructed multi-template library, which includes a standard rail cross-section template and at least one typical defective rail cross-section template; the matching includes: classifying each point in the original point cloud point by point using a lightweight one-dimensional convolutional neural network to identify whether each point belongs to the rail category or the noise category; wherein, the rail category includes standard rails and defective rails; S3. Retain Filtering Results: Based on the point-by-point classification results, retain the points belonging to the rail category and remove the points belonging to the noise category to obtain a clean rail cross-section point cloud.

2. The rail cross-section point cloud filtering method based on multi-template matching according to claim 1, characterized in that, The S2 further includes a coarse registration sub-step prior to the point-by-point classification, which includes rigidly registering the original point cloud with the standard rail section template to eliminate the overall translational offset of the original point cloud.

3. The rail cross-section point cloud filtering method based on multi-template matching according to claim 2, characterized in that, The lightweight one-dimensional convolutional neural network takes as input a one-dimensional point cloud sequence of a single rail cross-section after rigid registration, and outputs as the probability that each point in the cross-section belongs to a predefined category. The predefined categories include four rail subcategories: rail standard, rail side wear, rail vertical wear, and rail spalling, as well as three noise subcategories: noise fasteners, noise ballast, and noise other. The structure of the one-dimensional convolutional neural network includes, in sequence: multiple convolutional layers, a global average pooling layer, one or more fully connected layers, and an output layer.

4. The rail cross-section point cloud filtering method based on multi-template matching according to claim 1, characterized in that, After retaining points belonging to the rail category and removing points belonging to the noise category based on the point-by-point classification results, the process further includes a post-processing sub-step, which includes: performing radius filtering on the retained point cloud to remove isolated noise points.

5. The rail cross-section point cloud filtering method based on multi-template matching according to claim 1, characterized in that, The steps for pre-building the multi-template library include: Multiple real, defect-free rail section samples were obtained, and standard rail section templates were generated statistically. Multiple real rail cross-section samples with specific defect types are obtained. After coordinate normalization of the multiple rail cross-section samples, typical defect rail cross-section templates corresponding to the specific defect types are statistically generated. The specific defect types include at least one of side wear, vertical wear, and rail head spalling.

6. The rail cross-section point cloud filtering method based on multi-template matching according to claim 1, characterized in that, Step S3 is followed by a geometric parameter calculation step, which includes: calculating the geometric parameters of the track based on the point cloud of the clean rail cross-section.

7. The rail cross-section point cloud filtering method based on multi-template matching according to claim 6, characterized in that, The geometric parameter calculation step further includes: using the category labels attached to the point cloud of the pure rail cross section to perform adaptive fitting of the point cloud of different regions; The region includes the rail top region and the working edge region. When adaptively fitting the point cloud in the rail top region, points with rail vertical grinding or rail standard labels are preferred. When adaptively fitting the point cloud in the working edge region, points with rail side grinding labels are preferred.

8. The rail cross-section point cloud filtering method based on multi-template matching according to claim 6 or 7, characterized in that, The method further includes a result output step: comparing the geometric parameters with a preset safety threshold, and generating a fault alarm record when the geometric parameters exceed the limit; the fault alarm record includes at least one of mileage information, stock type, fault type, measured value, allowable limit, and recommended maintenance measures.

9. A rail cross-section point cloud filtering system based on multi-template matching, used to implement the rail cross-section point cloud filtering method based on multi-template matching as described in any one of claims 1 to 7, characterized in that, The system includes: The data acquisition module is used to acquire the original point cloud of the rail cross-section; A multi-template matching collaborative filtering module is used to match the original point cloud with a pre-built multi-template library, which includes a standard rail cross-section template and at least one typical defective rail cross-section template. The matching includes: classifying each point in the original point cloud point by point using a lightweight one-dimensional convolutional neural network to identify whether each point belongs to a rail category or a noise category; wherein, the rail category includes standard rails and defective rails. The filter result retention module is used to retain points belonging to the rail category and remove points belonging to the noise category based on the point-by-point classification results, thereby obtaining a clean rail cross-section point cloud.

10. The rail cross-section point cloud filtering system based on multi-template matching according to claim 9, characterized in that, The system also includes: The geometric parameter calculation module is used to calculate the geometric parameters of the track based on the point cloud of the clean rail cross-section. The result output module is used to compare the geometric parameters with preset safety thresholds. When the geometric parameters exceed the limits, a fault alarm record or maintenance guide sheet containing accurate mileage positioning information and recommended maintenance measures is generated.

Images