Railway turnout curve-based inspection anomaly early warning method and system

Abstract

This invention discloses a method and system for early warning of inspection anomalies based on railway turnout curves, relating to the field of track inspection and early warning technology. The method includes acquiring three-dimensional point cloud data of the railway turnout area track; performing voxelization and outlier filtering on the three-dimensional point cloud data to obtain a filter set; based on the filter set, generating a neighborhood point set by a preset neighborhood radius and performing convolution operations to output a reference track point cloud feature matrix; obtaining the current track point cloud feature matrix; using the reference and current track point cloud feature matrices as input, executing a parallel filtering mechanism to determine the final reference point; based on the final reference point, solving for track geometric offset errors and performing clustering; using cubic B-spline curves to generate the track centerline for turnout curve curvature analysis; and constructing a turnout curve parameter vector. This invention can monitor track change trends in real time and achieve efficient and accurate inspection of railway turnout curves.

Description

Technical Field

[0001] This invention relates to the field of track inspection and early warning technology, and in particular to a method and system for early warning of inspection anomalies based on railway turnout curves. Background Technology

[0002] In recent years, railway turnout inspection has gradually shifted from traditional manual inspection to intelligent inspection. Traditional manual inspection methods rely on human experience and visual inspection, resulting in low efficiency, large errors, and difficulty in timely detection of potential hazards. With the continuous advancement of information and automation technologies, more and more sensors and intelligent detection technologies are being introduced into railway turnout inspection. The application of lidar, point cloud technology, and image processing technology has greatly improved the detection accuracy of railway turnouts, enabling automated inspection and remote monitoring, thus increasing work efficiency and reducing human error.

[0003] However, existing technologies still have the following shortcomings:

[0004] 1. It relies heavily on single indicators such as track geometric error or curvature, and cannot simultaneously consider the geometric offset and curve shape changes in the turnout area. For example, although it can detect track gauge changes or track elevation errors in the turnout area, it lacks a comprehensive assessment of the turnout curve shape and is difficult to identify complex issues such as sharp curves and deformations.

[0005] 2. The method of coarse matching is often used, which ignores the special characteristics of the turnout curve area and is prone to mismatch or omission. Especially when the curve shape of the turnout area changes greatly, the accuracy of the registration result cannot be effectively guaranteed. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] Therefore, this invention provides a method and system for early warning of inspection anomalies based on railway turnout curves, which solves the problem that the existing technology lacks a comprehensive evaluation of the turnout curve morphology and the accuracy of the registration results cannot be effectively guaranteed.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a method for early warning of inspection anomalies based on railway turnout curves, comprising,

[0010] The three-dimensional point cloud data of the railway turnout area is obtained, and the three-dimensional point cloud data is processed by voxelization and outlier filtering to obtain a filter set.

[0011] Based on the filtering set, a neighborhood point set is generated by pre-setting the neighborhood radius, and convolution operation is performed to output the reference orbit point cloud feature matrix. The current orbit point cloud feature matrix is ​​obtained. Using the reference and current orbit point cloud feature matrices as input, a parallel filtering mechanism is executed to determine the final reference point.

[0012] Based on the final reference point, after solving the track geometric offset error and performing clustering, the track centerline is generated using cubic B-spline curves to perform turnout curve curvature analysis and construct the turnout curve parameter vector.

[0013] Based on the turnout curve parameter vector, future track state prediction and fusion are performed to generate abnormal risk indicators for inspection anomaly early warning.

[0014] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves described in this invention, wherein: the three-dimensional point cloud data of the railway turnout area track is acquired, and the three-dimensional point cloud data is subjected to voxelization processing and outlier filtering to obtain a filter set, including:

[0015] A multi-line lidar is installed on a railway inspection vehicle to perform a three-dimensional scan of the railway turnout area, obtaining a set of original track point cloud data. The original track point cloud data set is then voxelized using a preset voxel division side length to generate a set of voxel centers.

[0016] Perform statistical outlier filtering on the voxelized points in the voxel center set to obtain the filtered set.

[0017] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves described in this invention, the step of generating a set of neighboring points based on a filtered set, performing convolution operations through a preset neighborhood radius, outputting a reference track point cloud feature matrix, and obtaining the current track point cloud feature matrix includes:

[0018] By setting a preset neighborhood radius, a set of neighborhood points is generated for each voxel center point in the selected set. Convolution operations are then performed within this neighborhood to generate orbital structure feature vectors for each voxel center point. These feature vectors are combined to form a reference orbital point cloud feature matrix. The original point cloud data at the current time is then acquired, and the same operations used to generate the reference orbital point cloud feature matrix are repeated to obtain the current orbital point cloud feature matrix.

[0019] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves described in this invention, the parallel screening mechanism is composed of an attention mechanism and a KD tree.

[0020] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves described in this invention, the parallel screening mechanism includes:

[0021] By using a preset learnable mapping matrix, the feature matrix of the current orbit point cloud and the feature matrix of the reference orbit point cloud are mapped and transformed, and the coordinates of the voxel center points are predicted by combining the attention mechanism to generate attention coordinates.

[0022] Based on the feature matrix of the reference orbit point cloud, the coordinates of the center point of each voxel are extracted to form a reference coordinate set. Then, the KDTree construction function in the SciPy library is used to recursively partition the reference coordinate set to generate a KD tree.

[0023] Based on the current orbit point cloud feature matrix, the coordinates of each current voxel center point are extracted. Nearest neighbor search is used to search for the voxel center point closest to the current voxel center point coordinates in the KD tree. After calculating the spatial distance using Euclidean distance, it is determined whether the current voxel center point coordinates match the voxel center point, and the matching result is generated.

[0024] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves described in this invention, wherein: determining the final reference point includes:

[0025] Based on the matching results and attention coordinates, the consistency error is calculated to filter the consistency of the orbital position and generate the final reference point.

[0026] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves described in this invention, the step of clustering based on the final reference point to solve the track geometric offset error includes:

[0027] Based on the final reference point, the orbital geometric offset error is calculated, and K-means clustering is performed on the coordinates of all current voxel center points to generate a set of center point coordinates.

[0028] Curve parameters are generated for each center point based on the set of center point coordinates.

[0029] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves described in this invention, the step of using cubic B-spline curves to generate track centerlines for turnout curve curvature analysis and constructing turnout curve parameter vectors includes:

[0030] Based on the curve parameters, the track centerline is generated by fitting a cubic B-spline curve, the curvature of the turnout curve is calculated to generate a curvature sequence, the average curvature is generated by solving the curvature sequence, and the turnout curve parameter vector is obtained by combining the track geometric offset error and the turnout curve curvature.

[0031] As a preferred embodiment of the inspection anomaly early warning method based on railway turnout curves according to the present invention, the step of performing future track state prediction and fusion based on the turnout curve parameter vector to generate anomaly risk indicators for inspection anomaly early warning includes:

[0032] Based on the turnout curve parameter vector, a time series is constructed, and a prediction vector is obtained through a Transformer encoder.

[0033] By using preset weights, including weights for predicting orbital geometric offset errors. Weights for predicting the curvature of turnout curves Weights for predicting mean curvature A weighted fusion strategy is used to fuse the prediction vectors to generate anomaly risk indicators;

[0034] Based on abnormal risk indicators and according to preset early warning thresholds, including early warning thresholds and It outputs the inspection anomaly warning level of railway turnout curves, including no risk, risk present, and high risk.

[0035] Secondly, this invention provides an early warning system for inspection anomalies based on railway turnout curves, including:

[0036] The preprocessing generation module is used to acquire three-dimensional point cloud data of the railway turnout area track, perform voxelization processing and outlier filtering on the three-dimensional point cloud data, and obtain a filter set.

[0037] The convolution and parallel filtering module is used to generate a set of neighborhood points based on the filtering set and a preset neighborhood radius, perform convolution operations, output the feature matrix of the reference orbit point cloud, obtain the feature matrix of the current orbit point cloud, and use the reference and current orbit point cloud feature matrices as inputs to execute a parallel filtering mechanism to determine the final reference point.

[0038] The clustering analysis module is used to solve the track geometric offset error based on the final reference point, perform clustering, generate the track centerline using cubic B-spline curves, perform turnout curve curvature analysis, and construct the turnout curve parameter vector.

[0039] The early warning generation module is used to predict and fuse future track conditions based on turnout curve parameter vectors, and generate abnormal risk indicators for inspection anomaly early warning.

[0040] The beneficial effects of this invention are as follows: This invention collects three-dimensional point cloud data of railway turnout areas, performs voxelization processing and outlier filtering, and extracts track features of the turnout areas using convolution operations. It comprehensively considers track geometric offset errors and curvature changes of turnout curves to comprehensively evaluate the state of the turnout areas. Furthermore, this invention uses an attention mechanism combined with a KD tree structure for parallel filtering and matching. By adaptively adjusting the attention weights, it accurately matches the current track point cloud and the reference track point cloud, thereby significantly improving registration accuracy. Secondly, through a bidirectional attention mechanism, prediction model, abnormal risk indicators, and early warning thresholds, it can monitor track change trends in real time and achieve efficient and accurate inspection of railway turnout curves. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0042] Figure 1 This is a flowchart of the inspection anomaly early warning method based on railway turnout curves in Example 1.

[0043] Figure 2 This is a structural diagram of the inspection anomaly early warning system based on railway turnout curves in Example 1.

[0044] Figure 3 This is a flowchart of the parallel filtering mechanism implemented in Example 1.

[0045] Figure 4 This is a flowchart for determining the final reference point in Example 1. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Example 1, referring to Figures 1-4 This is the first embodiment of the present invention, which provides a method for early warning of inspection anomalies based on railway turnout curves, including the following steps:

[0050] S1. Obtain the three-dimensional point cloud data of the railway turnout area track, perform voxelization and outlier filtering on the three-dimensional point cloud data, and obtain the filter set.

[0051] S1.1 Install a multi-line lidar on the railway inspection vehicle to perform three-dimensional scanning of the railway turnout area to obtain the original track point cloud data set. Perform voxelization processing on the original track point cloud data set with the preset voxel division side length to generate a voxel center set.

[0052] Specifically, the railway turnout area is scanned in three dimensions using lidar to obtain raw point cloud data, which forms the raw track point cloud data set.

[0053] Subsequently, the voxel partitioning side length was set to... The original point cloud data is spatially voxelized by dividing the voxel side length to generate several voxel meshes.

[0054] The original point cloud points that fall within the same voxel grid are used to form voxel units;

[0055] Calculate the geometric center of all original point cloud points within the voxel unit, and use this geometric center as the voxel center point.

[0056] Repeat the operation to generate several voxel centers, forming a set of voxel centers.

[0057] The expression for calculating the geometric center of all original point cloud points within a voxel unit is:

[0058]

[0059] In the formula, Indicates the first The coordinates of the voxel center point of the voxel unit. Indicates the first The total number of point cloud points within an individual pixel unit. Indicates the first The first in the individual element unit The coordinates of a point cloud point;

[0060] It should be noted that the lidar is a rotating three-dimensional laser scanning device with a scanning frequency of 10Hz to 20Hz, a horizontal scanning angle of 360°, and a vertical field of view of 20° to 40°.

[0061] It should also be noted that the width of railway rail heads is typically 65mm to 75mm, and the track gauge is 1435mm. To ensure that the voxel dimensions can accurately describe the geometry of the rail head and to avoid oversampling, the voxel division side lengths are... It should be less than 1 / 3 to 1 / 5 of the rail head width; therefore, the voxel size range is determined as follows: When the voxel side length is less than 5 mm, the excessive number of point clouds will significantly increase the computational complexity. When the voxel side length is greater than 20 mm, the geometric features of the rail edge cannot be effectively preserved. Therefore, a voxel side length of 0.01 m is preferred, which is 1 / 7 of the rail head width of about 70 mm. This can effectively reduce the point cloud density while maintaining the details of the track structure.

[0062] S1.2 Perform statistical outlier filtering on the voxelized points in the voxel center set to obtain the filtered set;

[0063] Specifically, in the set of voxel centers, for each voxel center, search its... The nearest neighbor voxel centers are used to form a neighborhood point set. Then, the Euclidean distance between each voxel center and its neighboring points in the neighborhood point set is calculated, and the average distance is obtained to obtain the average distance of each voxel center.

[0064] Calculate the mean and standard deviation of all average distances;

[0065] Depend on In principle, the exclusion threshold is calculated by combining the mean and standard deviation;

[0066] If the average distance is greater than the rejection threshold, the voxel center point is identified as an outlier and deleted; otherwise, it is retained, and a filter set is generated.

[0067] Specifically, the Euclidean distances from each voxel center point's neighborhood points to all its neighboring points in its neighborhood set are calculated, and the average distance is obtained for each voxel center point. The formula is as follows:

[0068]

[0069] In the formula, Indicates the first The average distance between the centers of individual elements. This represents the total number of neighboring points. Represents Euclidean distance. Indicates the first The coordinates of the center point of the individual element, Indicates the first Neighborhood points of the individual element center The coordinates;

[0070] The formula for calculating the mean and standard deviation of all average distances is:

[0071]

[0072]

[0073] In the formula, The mean of the average distance. This represents the total number of average distances. The standard deviation of the average distance;

[0074] Depend on In principle, the exclusion threshold is calculated by combining the mean and standard deviation, using the following formula:

[0075]

[0076] In the formula, Indicates the rejection threshold;

[0077] It should be noted that the scanning density of railway track point clouds is usually 2000-5000 points / m². When the number of neighboring points is less than 10, the average distance calculation is easily affected by local noise. When the number of neighboring points is greater than 30, it will span different track structure areas. Therefore, the number of neighboring points can be limited to 10-30. In order to make the neighborhood coverage area approximately 0.005-0.01 m², which can completely cover the local area of ​​the rail head, while avoiding the inclusion of track fasteners or ballast structure points, 20 can be taken as the default value for example.

[0078] S2. Based on the filtering set, generate a set of neighborhood points by pre-setting the neighborhood radius, perform convolution operation, output the feature matrix of the reference orbit point cloud, obtain the feature matrix of the current orbit point cloud, and use the reference and current orbit point cloud feature matrices as input to execute a parallel filtering mechanism to determine the final reference point.

[0079] The parallel selection mechanism consists of an attention mechanism and a KD tree;

[0080] S2.1. By setting a preset neighborhood radius, a neighborhood point set is generated for each voxel center point in the selection set, and a convolution operation is performed within the neighborhood range to generate the orbital structure feature vector of each voxel center point. These feature vectors are then combined to form the reference orbital point cloud feature matrix and the current orbital point cloud feature matrix.

[0081] Specifically, set the neighborhood radius. Based on the neighborhood radius, a spherical neighborhood search method is used to obtain the neighborhood point set of each voxel center point in the filter set;

[0082] Convolution operations are performed in the neighborhood of each voxel center using a convolutional network (KPConv);

[0083] The track structure feature vector of each voxel center point is obtained, including track elevation change features, rail side profile features, and track gauge spatial structure features. The coordinates of the voxel center points are recorded and encapsulated together with the track structure feature vectors. All track structure feature vectors are combined to form a reference track point cloud feature matrix.

[0084] Obtain the raw point cloud data at the current moment, and repeat the same operation as when generating the reference orbit point cloud feature matrix to obtain the current orbit point cloud feature matrix.

[0085] Specifically, convolution operations are performed in the neighborhood of each voxel center point using a convolutional network (KPConv), as shown in the formula:

[0086]

[0087] In the formula, Indicates the first The orbital structure feature vector of the individual element center point. The number of convolution kernels (can be randomly sampled and generated within a spherical space of the radius of the neighborhood point set). Indicates the first Each convolutional kernel weight parameter (generated using the Xavier initialization method) Represents the kernel function. Indicates the first Each convolution kernel coordinate (selected from the number of convolution kernels, with the corresponding voxel center point coordinates as the convolution kernel coordinates). Representing neighborhood points Orbital point cloud data (generated when outlier deletion is completed);

[0088] Among them, kernel function The Gaussian weighting function is used, and the formula is as follows:

[0089]

[0090] In the formula, The base of the natural logarithm. This represents the Gaussian kernel bandwidth parameter. The kernel function bandwidth typically satisfies a certain condition with respect to the neighborhood radius. When the neighborhood radius When it is 0.05m, then To ensure that the convolution kernel effectively covers the neighborhood and the weight distribution is smooth, the Gaussian kernel bandwidth parameter can be set to 0.02m for example.

[0091] It should be noted that: neighborhood radius The settings are based on the following: the geometric dimensions of railway turnout areas are generally standard gauge 1435 mm, rail head width 65–75 mm, rail web thickness 16–20 mm, and rail height 150–180 mm, while the laser point cloud scanning density is typically... The corresponding average point spacing is To ensure that the neighborhood contains at least 20 to 40 valid points, the neighborhood radius must be greater than or equal to [a certain value]. Therefore, the lower limit can be set to greater than or equal to 0.03m. However, if the neighborhood radius is too large, it will cause the neighborhood points to span different track structures, such as rail heads, fasteners, and sleepers, with distances between these structures approximately [missing information]. Therefore, the upper limit can be set to less than or equal to 0.08m for example. However, to ensure that the entire rail surface is covered, that it does not cross sleepers or fasteners, and that the convolution feature can be calculated stably, 0.05m can be taken as the default value for example.

[0092] It should also be noted that when using the convolutional network (KPConv), it should be built and trained first. The construction content is as follows: the convolutional network is built through the point cloud convolutional network KPConv (Kernel Point Convolution Network) model architecture, including the input layer, neighborhood search layer, kernel point convolutional layer, non-linear activation layer and feature output layer.

[0093] The mean squared error is used as the loss function, and the Adam optimizer is used to train the model parameters. During the training process, when the decrease value of the loss function no longer decreases significantly, the iteration stops and the trained convolutional network is output.

[0094] S2.2. By using a preset learnable mapping matrix, the feature matrix of the current orbit point cloud and the feature matrix of the reference orbit point cloud are mapped and transformed, and the coordinates of the voxel center points are predicted by combining the attention mechanism to generate attention coordinates.

[0095] Specifically, a learnable mapping matrix is ​​generated by random initialization, including query, key, and value mapping matrices;

[0096] The current orbit point cloud feature matrix is ​​transformed by querying the mapping matrix to generate a query matrix, and the reference orbit point cloud feature matrix is ​​transformed by key-value mapping matrix to generate a key-value matrix.

[0097] The attention mechanism is used to calculate the attention weight matrix by combining the query, key, and value matrices.

[0098] Based on the attention weight matrix, the predicted coordinates of the voxel center points are calculated and used as attention coordinates.

[0099] Specifically, through an attention mechanism, combining the query, key, and value matrices, the attention weight matrix is ​​calculated using the following formula:

[0100]

[0101] In the formula, Represents the attention weight matrix. This represents the softmax function. Represents the query matrix. Represents the key matrix. This indicates the transpose operation. Represents a value matrix, Indicates the feature dimension of the query or key matrix;

[0102] Based on the attention weight matrix, the predicted coordinates of the voxel center points are calculated and used as the attention coordinates, using the following formula:

[0103]

[0104] In the formula, Indicates the current voxel center point Predicted coordinates This represents the total number of voxel center points. Represents the relationship between the attention weight matrix and the first... The element value corresponding to the center point of an individual element.

[0105] S2.3. Based on the feature matrix of the reference orbit point cloud, extract the coordinates of the center point of each voxel to form a reference coordinate set, and use the KDTree construction function in the SciPy library to recursively partition the reference coordinate set to generate a KD tree.

[0106] Specifically, among the coordinates of all voxel center points, a coordinate dimension is randomly selected, the coordinates in that dimension are sorted in ascending order, the median point of the coordinate values ​​of the current coordinate dimension is found from the ascending list, and the median point of the coordinate values ​​is set as the root node of the tree.

[0107] The tree is divided based on the coordinate value of that dimension using the root node. If the coordinate value of that dimension is less than the median of the coordinate values, it is divided into the left subtree set; otherwise, it is divided into the right subtree set.

[0108] Recursively partition the left and right subtree sets until there are no more points in the left and right subtree sets, then the KD tree construction is complete.

[0109] It should be noted that after sorting, the median point is determined based on the parity of the number of voxel center points;

[0110] For example, when the number of voxel center points is odd, the voxel center point in the middle position of the ascending sorted list can be directly selected as the median point.

[0111] When the number of voxel centers is even, the lower median is used as the median point, that is, the point with the smaller coordinate value among the two intermediate points is selected. For example:

[0112] The list sorted in ascending order is 1.2, 1.5, 1.8, 2.1, 2.3, 2.7;

[0113] The total number is 6, and the medians are 1.8 and 2.1. According to the lower median rule, 1.8 is selected as the median point.

[0114] S2.4 Based on the current orbit point cloud feature matrix, extract the coordinates of each current voxel center point, use nearest neighbor search to search for the voxel center point closest to the current voxel center point coordinates in the KD tree, and use Euclidean distance to calculate the spatial distance, then determine whether the current voxel center point coordinates match the voxel center point, and generate the matching result.

[0115] Specifically, a filtering threshold is set. If the spatial distance is greater than the filtering threshold, it is considered that the current voxel center point and the voxel center point do not correspond and are marked as mismatched. Otherwise, it is considered that the current voxel center point and the voxel center point correspond and are marked as matched.

[0116] Based on the current voxel center point and voxel center point of the marker matching, extract their corresponding coordinates as the matching result.

[0117] It should be noted that in this invention, the preferred voxel side length is 0.01m. If the center points of two voxels come from the same orbital structure, their maximum spatial offset will generally not exceed 5×0.01m. Therefore, the filtering threshold can be exemplarily set to 0.05m, which can both ensure matching accuracy and avoid mismatches.

[0118] S2.5. Based on the matching results and attention coordinates, calculate the consistency error to perform consistency screening of the orbital position and generate the final reference point.

[0119] Specifically, based on the matching results, the coordinates of the voxel center points are extracted from the matching results, and the consistency error is obtained by solving the coordinate error between the coordinates of the voxel center points and the attention coordinates.

[0120] Set a test threshold. If the consistency error is less than or equal to the test threshold, the consistency test of the orbit position is passed, and the arithmetic mean of the coordinates of the voxel center point and the attention coordinates is taken as the final reference coordinates. Otherwise, it fails, and the coordinates of the voxel center point in the matching result are directly used as the final reference coordinates.

[0121] It should be noted that the inspection threshold is determined based on the comprehensive error during the acquisition and processing of 3D point cloud data of the railway turnout area, mainly including the following three sources of error:

[0122] (1) LiDAR measurement error: The typical measurement accuracy of lidar in railway inspection scenarios is about 2~8mm. Considering the changes in scanning angle and surface reflection error, the upper limit of its spatial measurement error is about 0.01m.

[0123] (2) Voxelization sampling error: In this invention, the preferred voxel partitioning side length is 0.01m. The maximum deviation between the voxel center point and the true point cloud geometric center is usually no more than half the length of the voxel diagonal. The upper bound of the error can be expressed as: the voxelization sampling error is less than or equal to ;

[0124] (3) Attention mechanism coordinate prediction error. Since the attention mechanism is affected by feature extraction error and model prediction error when calculating the correspondence of voxel center points, its prediction deviation usually does not exceed 0.015m.

[0125] Therefore, to ensure that the consistency check can cover the cumulative effect of errors, the verification threshold should be greater than or equal to the sum of the lidar measurement error, voxel sampling error, and attention mechanism coordinate prediction error. To adapt to different lidar accuracies and point cloud density conditions, the selectable range of the verification threshold can be set to 0.015m~0.05m for example. This range can cover the comprehensive error generated during point cloud acquisition and prediction, and can also avoid track structure mismatch due to an excessively large threshold, thereby improving the positioning accuracy of railway turnout curve reference points. Under the superposition of errors, the verification threshold can be set to 0.03m for example.

[0126] Secondly, the filtering threshold is used for coarse matching screening, and the verification threshold is used for fine matching verification. Therefore, the verification threshold must be less than the filtering threshold. Under this condition, the present invention can achieve the effects of eliminating mismatch points, improving the positioning accuracy of reference points, and enhancing the stability of track geometric offset error calculation.

[0127] S3. Based on the final reference point, after solving the track geometric offset error and performing clustering, the track centerline is generated using cubic B-spline curves to perform turnout curve curvature analysis and construct the turnout curve parameter vector.

[0128] S3.1 Based on the final reference point, calculate the orbital geometric offset error, and use K-means clustering to perform clustering based on the coordinates of all current voxel center points to generate a set of center point coordinates;

[0129] Specifically, extract the coordinates of the current voxel center point from the matching results;

[0130] The coordinate error between the current voxel center point coordinates and the final reference coordinates is calculated as the orbital geometric offset error;

[0131] Set the number of clusters to 2 (i.e., left and right track points). Based on the coordinates of all current voxel center points, use K-means clustering to cluster the left and right track points of the current voxel center points, and obtain the set of left and right track points of the current voxel center points.

[0132] Extract the coordinates of the left and right orbit points from the set of left and right orbit points, and calculate the arithmetic mean of the coordinates of the left and right orbit points to generate the center coordinates, thus forming the set of center point coordinates.

[0133] S3.2 Generate curve parameters for each center point based on the set of center point coordinates;

[0134] Specifically, based on the set of center point coordinates, the spatial distance between each center point coordinate and the previous center point coordinate is calculated using Euclidean distance. Then, the chord length parameterization method is used in conjunction with the spatial distance to generate curve parameters for each center point.

[0135] It should be noted that the formula for generating curve parameters for each center point is:

[0136]

[0137] In the formula, Indicates the first Curve parameters at each center point Indicates from the first center point to the... The cumulative length of the spatial distance between all adjacent points of a center point. Indicates the first The spatial distance between each center point and the previous center point This represents the total number of center points.

[0138] S3.3 Based on curve parameters, the track centerline is generated by fitting a cubic B-spline curve, the curvature of the turnout curve is calculated to generate a curvature sequence, the average curvature is generated by solving the curvature sequence, and the track geometric offset error and the turnout curve curvature are combined to obtain the turnout curve parameter vector.

[0139] Specifically, based on the curve parameters and the set of center point coordinates, the orbital centerline is obtained by fitting a cubic B-spline curve.

[0140] Subsequently, at the center line of the track and Calculate the curvature of the turnout curve on the plane of direction;

[0141] Repeat the operation to generate a curvature sequence;

[0142] The maximum curvature is selected from the curvature sequence by maximizing the operation as the curvature of the turnout curve. The average curvature of the curvature sequence is obtained by using the mean formula. The track geometric offset error, the curvature of the turnout curve, and the average curvature are then concatenated into the turnout curve parameter vector.

[0143] It should be noted that: based on the curve parameters and the set of center point coordinates, the orbital centerline is obtained by fitting a cubic B-spline curve, and the formula is:

[0144]

[0145] In the formula, Indicates the centerline of the track in the curve parameters Spatial coordinates at the location, Indicates the first Cubic B-spline basis functions at center points Indicates the first The coordinates of the center point;

[0146] The recursive definition of the cubic B-spline basis function is as follows:

[0147]

[0148] In the formula, Indicates the first The central point is The B-spline basis function values ​​at order of 1. Indicates the center point index. This represents the order index of the cubic B-spline basis function. Indicates curve parameters, Indicates that the index is The node value, Indicates that the index is The node value, Indicates the first The central point is The B-spline basis function values ​​at order of 1. Indicates that the index is The node value, Indicates the first The central point is The B-spline basis function values ​​at order;

[0149] The node values ​​are generated based on the total number of center points and the order of the cubic B-spline basis function, expressed as:

[0150]

[0151] It should also be noted that the initial fitting of a cubic B-spline curve should begin with a zero-order function, as shown in the formula:

[0152]

[0153] In the formula, Indicates the first The center points are of order [order]. The B-spline basis function value at time, Indicates otherwise;

[0154] Subsequently, at the center line of the track and The formula for calculating the curvature of the turnout curve on the plane of direction is:

[0155]

[0156] In the formula, Indicates the curvature of the turnout curve. Indicates the center line of the track Direction versus curve parameters The first derivative, Indicates the center line of the track Direction versus curve parameters The first derivative, and They represent the center lines of the orbits respectively. and Direction versus curve parameters The second derivative;

[0157] The expression for the first derivative is:

[0158]

[0159]

[0160] In the formula, Indicates the centerline of the track Direction versus curve parameters Differentiation operation, Indicates the centerline of the track Direction versus curve parameters Differentiation operation, Indicates the centerline of the track in the curve parameters place Direction coordinates, Indicates the centerline of the track in the curve parameters place Direction coordinates;

[0161] Track centerline in curve parameters place and The expression for the direction coordinate value is:

[0162]

[0163]

[0164] In the formula, Indicates the first Each center point Direction coordinates Indicates the first Each center point Direction coordinates;

[0165] The expression for the second derivative is:

[0166]

[0167]

[0168] In the formula, Indicates the centerline of the track Direction versus curve parameters Second derivative operations, Indicates the centerline of the track Direction versus curve parameters Second derivative operations.

[0169] S4. Based on the turnout curve parameter vector, perform future track state prediction and fusion to generate abnormal risk indicators for inspection anomaly early warning.

[0170] S4.1. Based on the turnout curve parameter vector, construct a time series and obtain the prediction vector through a Transformer encoder;

[0171] Specifically, a time series is constructed based on the turnout curve parameter vector;

[0172] The time series is input into a standard Transformer encoder, and the output is the encoded hidden state sequence, including the hidden state vector;

[0173] The attention mechanism is used to calculate the attention weights between the current hidden state vector and other hidden state vectors in the hidden state sequence in the time series dimension.

[0174] Based on attention weights, the hidden state vector is used to predict the future orbital state and generate a prediction vector.

[0175] It should be noted that: the attention weights between the current hidden state vector and other hidden state vectors in the hidden state sequence are calculated along the time series dimension using the following formula:

[0176]

[0177] In the formula, Indicates time Regarding time Attention weights Represents an exponential function. Indicates time The hidden state vector, Indicates time The hidden state vector, This indicates the transpose operation. Indicates the length of the time series. Indicates time The hidden state vector, This represents the dimension of the hidden state vector;

[0178] Based on attention weights, the hidden state vector is used to predict the future orbital state, generating a prediction vector, as shown in the formula:

[0179]

[0180] In the formula, Indicates time The predicted vector at that time, Represents the weight matrix. Represents the bias vector;

[0181] It should be noted that the Transformer encoder in this invention adopts the standard Transformer encoder architecture in the prior art and uses an end-to-end approach to iteratively update the model parameters.

[0182] S4.2, Using preset weights, including weights for predicting orbital geometric offset errors. Weights for predicting the curvature of turnout curves Weights for predicting mean curvature A weighted fusion strategy is used to fuse the prediction vectors to generate anomaly risk indicators;

[0183] Specifically, through weighted fusion, the track geometric offset error, turnout curve curvature, and average curvature in the prediction vector are fused to calculate the abnormal risk index.

[0184] The formula for calculating the abnormal risk index is as follows:

[0185]

[0186] In the formula, An anomaly risk index representing the prediction vector, This represents the weight of the predicted orbital geometric offset error. Indicates the first The predicted orbital geometric offset error at time [time]. This represents the weight used to predict the curvature of the turnout curve. Indicates the first Predicted maximum turnout curve curvature at any given time. The weights representing the predicted mean curvature Indicates the first The predicted mean curvature at time;

[0187] It should be noted that the weights of the predicted orbital geometric offset error... The weighting is set based on the degree of impact of abnormal railway turnout structure on safety. The predicted track geometric offset error can directly reflect track gauge abnormality, track alignment deviation, and track surface height difference. These factors are the most direct factors affecting railway safety, so their weights are set to the maximum.

[0188] The maximum curvature of a turnout curve originates from a curvature sequence. This maximum curvature primarily reflects local sharp bends, rail wear, and turnout switch rail deformation. Because it is a local extreme value indicator, it is sensitive to anomalies but has low stability. Therefore, the weighting... Slightly lower than the weight of the predicted orbital geometric offset error ;

[0189] The average curvature reflects the overall track bending trend and long-term deformation. Because its changes are relatively gradual, it is less sensitive to sudden anomalies, therefore its weight is relatively low. Set to minimum;

[0190] To ensure that the abnormal risk indicators have a uniform scale and meet the weight normalization constraint, the weight of the predicted orbital geometric offset error can be used as an example. The weight of the predicted turnout curve curvature is 0.5. The weight of the predicted mean curvature is 0.3. It is 0.2;

[0191] Secondly, abnormal risk indicators It is a comprehensive indicator used to assess whether there are abnormal risks in the railway turnout area. In this invention, the indicator integrates track geometric errors (such as track gauge changes and track bed height changes) and curve curvature (such as the degree of sharpness of the turnout curve). Therefore, it can effectively reflect the changing trend of track geometry and improve the accuracy and stability of turnout curve anomaly identification. Combined with subsequent early warning thresholds, it can effectively ensure effective intervention before problems occur in the railway turnout curve.

[0192] S4.3 Based on abnormal risk indicators and according to preset early warning thresholds, including early warning thresholds and Output the inspection anomaly warning level of railway turnout curves, including no risk, risk present, and high risk;

[0193] Specifically, set early warning thresholds. and ,and ;

[0194] If the abnormal risk indicator is less than the warning threshold If so, it means that the inspection status of the railway turnout curve is risk-free;

[0195] If the abnormal risk indicator is greater than or equal to the warning threshold And less than the warning threshold If so, it indicates that the inspection status of the railway turnout curve is at risk.

[0196] If the abnormal risk indicator is greater than the warning threshold This indicates that the inspection status of the railway turnout curve is high-risk.

[0197] When the inspection status of a railway turnout curve is deemed to be at risk or high risk, the current inspection status is sent to the inspection personnel via wireless transmission technology to provide early warning of inspection anomalies.

[0198] It should be noted that in the turnout maintenance standards, conventional railways generally use the following: High-speed railways are generally The curvature of a conventional speed turnout is approximately Small radius turnouts can reach By performing normalization on this basis, the warning threshold can be obtained. and The exemplary value range is 0 and 1. However, in order to identify track degradation trends in advance and avoid alarms only when approaching safety limits, the present invention can, for example, set the warning threshold to... and Values , .

[0199] This embodiment also provides an inspection anomaly early warning system based on railway turnout curves, including:

[0200] The preprocessing generation module is used to acquire three-dimensional point cloud data of the railway turnout area track, perform voxelization processing and outlier filtering on the three-dimensional point cloud data, and obtain a filter set.

[0201] The convolution and parallel filtering module is used to generate a set of neighborhood points based on the filtering set and a preset neighborhood radius, perform convolution operations, output the feature matrix of the reference orbit point cloud, obtain the feature matrix of the current orbit point cloud, and use the reference and current orbit point cloud feature matrices as inputs to execute a parallel filtering mechanism to determine the final reference point.

[0202] The clustering analysis module is used to solve the track geometric offset error based on the final reference point, perform clustering, generate the track centerline using cubic B-spline curves, perform turnout curve curvature analysis, and construct the turnout curve parameter vector.

[0203] The early warning generation module is used to predict and fuse future track conditions based on turnout curve parameter vectors, and generate abnormal risk indicators for inspection anomaly early warning.

[0204] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for early warning of inspection anomalies based on railway turnout curves, characterized in that: include, The three-dimensional point cloud data of the railway turnout area is obtained, and the three-dimensional point cloud data is processed by voxelization and outlier filtering to obtain a filter set. Based on the filtering set, a neighborhood point set is generated by pre-setting the neighborhood radius, and convolution operation is performed to output the reference orbit point cloud feature matrix. The current orbit point cloud feature matrix is ​​obtained. Using the reference and current orbit point cloud feature matrices as input, a parallel filtering mechanism is executed to determine the final reference point. Based on the final reference point, after solving the track geometric offset error and performing clustering, the track centerline is generated using cubic B-spline curves to perform turnout curve curvature analysis and construct the turnout curve parameter vector. Based on the turnout curve parameter vector, future track state prediction and fusion are performed to generate abnormal risk indicators for inspection anomaly early warning.

2. The inspection anomaly early warning method based on railway turnout curves as described in claim 1, characterized in that: The process involves acquiring 3D point cloud data of the railway turnout area track, performing voxelization and outlier filtering on the 3D point cloud data to obtain a filtered set, including: A multi-line lidar is installed on a railway inspection vehicle to perform a three-dimensional scan of the railway turnout area, obtaining a set of original track point cloud data. The original track point cloud data set is then voxelized using a preset voxel division side length to generate a set of voxel centers. Perform statistical outlier filtering on the voxelized points in the voxel center set to obtain the filtered set.

3. The inspection anomaly early warning method based on railway turnout curves as described in claim 2, characterized in that: The process of generating a neighborhood point set based on a filtered set, performing convolution operations using a preset neighborhood radius, outputting a reference orbit point cloud feature matrix, and obtaining the current orbit point cloud feature matrix includes: By setting a preset neighborhood radius, a neighborhood point set is generated for each voxel center point in the selected set. Convolution operation is performed within the neighborhood range to generate the orbital structure feature vector of each voxel center point. These vectors are then combined to form a reference orbital point cloud feature matrix. The original point cloud data at the current time is obtained, and the same operation as when generating the reference orbital point cloud feature matrix is ​​repeated to obtain the current orbital point cloud feature matrix.

4. The inspection anomaly early warning method based on railway turnout curves as described in claim 1, characterized in that: The parallel filtering mechanism consists of an attention mechanism and a KD tree.

5. The inspection anomaly early warning method based on railway turnout curves as described in claim 1, characterized in that: The parallel filtering mechanism includes: By using a preset learnable mapping matrix, the feature matrix of the current orbit point cloud and the feature matrix of the reference orbit point cloud are mapped and transformed, and the coordinates of the voxel center points are predicted by combining the attention mechanism to generate attention coordinates. Based on the feature matrix of the reference orbit point cloud, the coordinates of the center point of each voxel are extracted to form a reference coordinate set. Then, the KDTree construction function in the SciPy library is used to recursively partition the reference coordinate set to generate a KD tree. Based on the current orbit point cloud feature matrix, the coordinates of each current voxel center point are extracted. Nearest neighbor search is used to search for the voxel center point closest to the current voxel center point coordinates in the KD tree. After calculating the spatial distance using Euclidean distance, it is determined whether the current voxel center point coordinates match the voxel center point, and the matching result is generated.

6. The inspection anomaly early warning method based on railway turnout curves as described in claim 5, characterized in that: The determination of the final reference point includes: Based on the matching results and attention coordinates, the consistency error is calculated to filter the consistency of the orbital position and generate the final reference point.

7. The inspection anomaly early warning method based on railway turnout curves as described in claim 6, characterized in that: The clustering of orbital geometric offset errors based on the final reference point includes: Based on the final reference point, the orbital geometric offset error is calculated, and K-means clustering is performed on the coordinates of all current voxel center points to generate a set of center point coordinates. Curve parameters are generated for each center point based on the set of center point coordinates.

8. The inspection anomaly early warning method based on railway turnout curves as described in claim 7, characterized in that: The method of using cubic B-spline curves to generate track centerlines for turnout curve curvature analysis and constructing turnout curve parameter vectors includes: Based on the curve parameters, the track centerline is generated by fitting a cubic B-spline curve, the curvature of the turnout curve is calculated to generate a curvature sequence, the average curvature is generated by solving the curvature sequence, and the turnout curve parameter vector is obtained by combining the track geometric offset error and the turnout curve curvature.

9. The inspection anomaly early warning method based on railway turnout curves as described in claim 8, characterized in that: The process of predicting and fusing future track conditions based on turnout curve parameter vectors to generate abnormal risk indicators for early warning of inspection anomalies includes: Based on the turnout curve parameter vector, a time series is constructed, and a prediction vector is obtained through a Transformer encoder. By using preset weights, including weights for predicting orbital geometric offset errors. Weights for predicting the curvature of turnout curves Weights for predicting mean curvature A weighted fusion strategy is used to fuse the prediction vectors to generate anomaly risk indicators; Based on abnormal risk indicators and according to preset early warning thresholds, including early warning thresholds and It outputs the inspection anomaly warning level of railway turnout curves, including no risk, risk present, and high risk.

10. An inspection anomaly early warning system based on railway turnout curves, based on the inspection anomaly early warning method based on railway turnout curves as described in any one of claims 1 to 9, characterized in that: include, The preprocessing generation module is used to acquire three-dimensional point cloud data of the railway turnout area track, perform voxelization processing and outlier filtering on the three-dimensional point cloud data, and obtain a filter set. The convolution and parallel filtering module is used to generate a set of neighborhood points based on the filtering set and a preset neighborhood radius, perform convolution operations, output the feature matrix of the reference orbit point cloud, obtain the feature matrix of the current orbit point cloud, and use the reference and current orbit point cloud feature matrices as inputs to execute a parallel filtering mechanism to determine the final reference point. The clustering analysis module is used to solve the track geometric offset error based on the final reference point, perform clustering, generate the track centerline using cubic B-spline curves, perform turnout curve curvature analysis, and construct the turnout curve parameter vector. The early warning generation module is used to predict and fuse future track conditions based on turnout curve parameter vectors, and generate abnormal risk indicators for inspection anomaly early warning.

Images