Construction and Application Methods of Local Descriptors for Segmenting Leakage Inside Tunnels

Abstract

This invention discloses a method for constructing and applying local descriptors for segmenting leakage within tunnel pipes, belonging to the field of intelligent monitoring technology for subway tunnels. It solves the technical problems of single intensity features and insufficient segmentation accuracy in existing leakage segmentation methods. The technical solution includes the following steps: S1, scanning the subway tunnel using a mobile laser scanning system to acquire high-precision point cloud data; S2, preprocessing the tunnel point cloud, extracting the lining point cloud and segmenting it into individual pipe segments; S3, constructing neighborhoods for points within each individual pipe segment and generating local intensity descriptors; S4, combining the local intensity descriptors and a CatBoost classifier to segment leakage within the tunnel pipes. The beneficial effects of this invention are that by effectively expanding the dimensions of intensity features, it can accurately characterize the local intensity distribution of leakage points and non-leakage points, significantly improving leakage segmentation accuracy and providing reliable technical support for intelligent monitoring and safety maintenance of subway tunnels.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for subway tunnels, and in particular to the construction and application method of local descriptors for segmenting leakage within tunnel pipes. Background Technology

[0002] Leakage, a typical defect in subway tunnels, seriously affects tunnel durability and subway operational safety. In recent years, mobile laser scanning systems have been widely used for tunnel structure and defect monitoring due to their high precision and efficiency. Laser reflection intensity can reflect the physical characteristics of the tunnel surface, providing a key basis for leakage segmentation in three-dimensional space. However, traditional segmentation methods rely only on the reflection intensity characteristics in a single dimension, making it difficult to characterize the local characteristics of leakage points and non-leakage points, thus affecting segmentation accuracy. Summary of the Invention

[0003] The purpose of this invention is to overcome the problems in the prior art and provide a method for constructing and applying local descriptors for segmenting leakage inside tunnel pipes. By effectively expanding the dimensions of intensity features, it can accurately characterize the local intensity distribution of leakage points and non-leakage points, significantly improve the leakage segmentation accuracy, and provide reliable technical support for intelligent monitoring and safety maintenance of subway tunnels.

[0004] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for constructing and applying a local descriptor for segmenting leakage within a tunnel pipe, comprising the following steps:

[0005] S1 uses a mobile laser scanning system to scan the subway tunnel and acquire high-precision point cloud data;

[0006] S2, preprocess the tunnel point cloud, extract the lining point cloud and segment it into individual segments;

[0007] S3: Construct a neighborhood for points within a single tunnel segment and generate a local intensity descriptor;

[0008] S4, combining local intensity descriptors and CatBoost classifiers, segments leakage within tunnel pipes.

[0009] Further, in step S2, the preprocessing of the tunnel point cloud, extracting the lining point cloud and segmenting it into individual segments includes the following steps:

[0010] S2.1, the point cloud of the subway tunnel is P(x i ,y i ,z i ) i＝1,2,…,n Set the slice thickness and divide the tunnel point cloud into multiple slices using the following formula:

[0011]

[0012] in

[0013] nslice is the number of slices.

[0014] ceil(·) means rounding up.

[0015] y max and y min y i The maximum and minimum values, and the coordinates of the slice point are. k = 1, 2, ..., nslice;

[0016] S2.2, Project the slice points onto the XOZ plane to obtain the cross-sectional points. By fitting an ellipse to the cross-sectional points, the standard equation of the ellipse is obtained, as shown in the following equation:

[0017]

[0018] in,

[0019] (x center ,z center () represents the coordinates of the center of the circle.

[0020] 'a' is the major semi-axis.

[0021] b is the shorter semi-axis;

[0022] S2.3, Calculate the distance from the cross-section point to the fitted ellipse. Set distance thresholds δ1 and δ2. If the distance from a point to the fitted ellipse satisfies δ1 < d < δ2, then the point is a tunnel lining point; otherwise, the point is a non-lining point. The obtained lining point cloud is P'(x j ,y j ,z j ) j＝1,2,…,m ;

[0023] S2.4, Based on the design width of the tunnel segment, divide the lining point cloud into n segments using the following formula:

[0024]

[0025] in,

[0026] nSegment represents the number of pipe segments.

[0027] ceil(·) means rounding up.

[0028] and y j The maximum and minimum values,

[0029] width refers to the design width of the pipe segment.

[0030] Furthermore, in step S3, the process of constructing a neighborhood for a point within a single segment and generating a local intensity descriptor includes the following steps:

[0031] S3.1 Project a single segment onto the XOZ plane to obtain the cross-sectional point (x j ,z j Ellipse fitting is performed on the cross-sectional points to obtain the center (x). o ',z o The cross-section is translated, and the center of the circle is moved to the origin. The translation process is shown in the following formula:

[0032]

[0033] in,

[0034] (x j ',z j ') indicates the cross-section point after translation;

[0035] S3.2, Calculate the query point p in a single segment. o (x o ,y o ,z o The central angle θ of a circle is calculated using the following formula:

[0036]

[0037] in,

[0038] Q1 represents the first quadrant.

[0039] Q2 represents the second quadrant.

[0040] Q3 represents the third quadrant.

[0041] Q4 represents the fourth quadrant.

[0042] arctan() is the arctangent function;

[0043] S3.3, For the current query point, set the step size Δy o And Δθ, using the mileage range [y o -Δy o ,y o +Δy o Construct a neighborhood using the mileage value y1 and the angular range [θ-Δθ, θ+Δθ]. If the mileage value y1 and the central angle θ1 of a point satisfy y o -Δy o ≤y1≤y o +Δy oIf the angle range is θ-Δθ≤θ1≤θ+Δθ, then the point is a neighboring point of the query point; when Δy o When Δy is small and Δθ is large, the neighborhood exhibits a vertical distribution; when Δy is small... o When Δθ is large and Δy is small, the neighborhood exhibits a horizontal distribution; by adjusting the two step size parameters Δy... o Given Δθ, construct a vertically distributed neighborhood N1 and a horizontally distributed neighborhood N2, and extract the corresponding two neighborhood point sets N1(p o ) and N2(p o ); Calculate the mean I of the reflection intensity of the neighborhood point set. mean and standard deviation I std If the reflection intensity I at the current query point o Satisfy I o <I mean -3I std If the condition is met, the point is preliminarily identified as a leak point; otherwise, it is preliminarily identified as a non-leak point. The proportions of leak points in the two neighboring point sets, prob1 and prob2, are calculated respectively, and the neighboring point set of the current query point is determined by the following formula:

[0044]

[0045] S3.4, After determining the neighborhood point set, press [I i -min(I i )] / [max(I i )-min(I i The reflection intensity of the neighborhood point set is normalized, and the normalized reflection intensity value ranges from [0,1]. With ω as the width of the intensity interval, [0,1] is divided into K intensity intervals, and the number n points in the neighborhood point set that fall into each intensity interval is counted. K Calculate the probability density value p for each intensity interval using the following formula. K Obtain the local intensity descriptor;

[0046]

[0047] Where N is the total number of points in the neighborhood point set.

[0048] Further, in step S4, the segmentation of tunnel leakage by combining the local intensity descriptor and the CatBoost classifier includes the following steps:

[0049] S4.1, construct num decision trees, set the maximum number of iterations, learning rate η, decision tree depth depth, loss function parameter loss_function, and initialize the CatBoost model;

[0050] S4.2, using local descriptors as training features, fitting the features with the constructed decision tree, and calculating the value of the loss function according to the following formula:

[0051]

[0052] in,

[0053] n is the number of samples.

[0054] x i It is a feature of the i-th sample.

[0055] y i It is the true label of the i-th sample.

[0056] f(x i ) is the model's predicted label.

[0057] l(y i ,f(x i )) represents the loss of the i-th sample;

[0058] S4.3, calculate the prediction function f in each iteration t using the following formula. (t) (x i gradient of )

[0059]

[0060] in,

[0061] g i (t) Let be the gradient of the i-th sample in the t-th iteration, representing the prediction error of the sample.

[0062] This represents taking the partial derivative of the function;

[0063] S4.4, based on the gradient calculated in the t-th iteration, refit the training features to obtain a new predicted label h. t (x i The model is optimized by progressively minimizing the loss function using the following formula:

[0064] f (t+1) (x i )=f (t) (x i )+η·h t (x i (10)

[0065] in,

[0066] η represents the learning rate, a hyperparameter used to control the step size of model updates in each iteration;

[0067] S4.5 Repeat steps S4.2, S4.3 and S4.4 until the maximum number of iterations is reached, terminate training and obtain the optimal model. Then, use the model to infer the test data and obtain the predicted labels of each point within a single segment.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] 1. This invention overcomes the limitations of traditional neighborhoods, which lack directionality and cannot be applied to the construction of local neighborhoods for leakage points. By constructing a new type of neighborhood, it can effectively fit the distribution direction of leakage and improve the ability to accurately describe the distribution direction of leakage.

[0070] 2. This invention breaks through the limitation of the single dimension of intensity features in traditional tunnel leakage segmentation methods. By constructing a local intensity descriptor, the dimension of intensity features is effectively expanded, thereby effectively characterizing the local intensity distribution of leakage points and non-leakage points.

[0071] 3. This invention enables accurate segmentation of leakage within a single tunnel segment, providing important technical support for intelligent monitoring and safety maintenance of subway tunnels.

[0072] 4. This invention focuses on the innovative construction and application of local strength descriptors, breaking through the limitations of traditional leakage segmentation technology in terms of strength feature dimensions. It can effectively expand the dimensions of strength features and accurately characterize the local characteristics of leakage points and non-leakage points, achieving high-precision segmentation of leakage within a single tunnel segment. This significantly improves the accuracy of leakage segmentation and provides reliable technical support and guarantee for intelligent monitoring and safe maintenance of subway tunnels. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the process of the present invention.

[0074] Figure 2 This is the test data for the present invention.

[0075] Figure 3 This is a scatter plot of the tunnel point cloud after point cloud preprocessing according to the present invention.

[0076] Figure 4 This is a schematic diagram illustrating the neighborhood construction implemented in this invention.

[0077] Figure 5 This is a schematic diagram illustrating the local intensity descriptor extraction method of the present invention.

[0078] Figure 6 This is a schematic diagram illustrating how the global intensity descriptor is combined with the CatBoost classifier to achieve leakage segmentation in this invention. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0080] The invention will now be further described with reference to the accompanying drawings.

[0081] Example

[0082] like Figure 1 As shown, this embodiment provides a method for constructing and applying a local descriptor for segmenting leakage within a tunnel pipe, including the following steps:

[0083] Step 1: As Figure 2 As shown, a mobile laser scanning system is used to scan a subway tunnel to obtain a high-precision point cloud P(x). i ,y i ,z i ) i＝1,2,…,n , where n is the number of points in the data collection cloud, and the observed values ​​are the three-dimensional coordinates of points on the inner surface of the tunnel.

[0084] Step 2: Preprocess the tunnel point cloud, extract the lining point cloud and segment it into individual segments, as follows:

[0085] (1) Set the slice thickness to 0.2m, and divide the tunnel point cloud into multiple slices using the following formula:

[0086]

[0087] Where nslice is the number of slices, ceil(·) represents rounding up, and y max and y min y i The maximum and minimum values, and the coordinates of the slice point are. k = 1, 2, ..., nslice;

[0088] (2) Project the slice points onto the XOZ plane to obtain the cross-section points. By fitting an ellipse to the cross-sectional points, the standard equation of the ellipse is obtained, as shown in the following equation:

[0089]

[0090] Among them, (x center ,z center ( ) represents the coordinates of the center of the circle, a is the major semi-axis, and b is the minor semi-axis;

[0091] (3) Calculate the distance from the cross-section point to the fitted ellipse. Set distance thresholds δ1 to -0.01m and δ2 to 0.01m. If the distance from a point to the fitted ellipse satisfies δ1 < d < δ2, then the point is a tunnel lining point; otherwise, the point is a non-lining point. The obtained lining point cloud is P'(x j ,y j ,z j ) j＝1,2,…,m ;

[0092] (4) The design width of the tunnel segment is 1.2m, and the lining point cloud is divided into 20 tunnel segments, such as Figure 3 As shown, the process of segmenting a single tunnel segment is illustrated by the following formula:

[0093]

[0094] Where nSegment is the number of segments, and ceil(·) represents rounding up. and y j The maximum and minimum values ​​are given, and width is the design width of the pipe segment.

[0095] Step 3: Construct a neighborhood for each point within a single tunnel segment and generate a local intensity descriptor, as detailed below:

[0096] (1) Project a single segment onto the XOZ plane to obtain the cross-sectional point (x j ,z j Ellipse fitting is performed on the cross-sectional points to obtain the center (x). o ',z o The cross-section is translated, and the center of the circle is moved to the origin. The translation process is shown in the following formula:

[0097]

[0098] Among them, (x j ',z j ') indicates the cross-section point after translation;

[0099] (2) Calculate the query point p in a single segment. o (x o ,y o ,z o The central angle θ of a circle is calculated using the following formula:

[0100]

[0101] Where Q1 represents the first quadrant, Q2 represents the second quadrant, Q3 represents the third quadrant, Q4 represents the fourth quadrant, and arctan() is the arctangent function;

[0102] (3) Figure 4As shown, for the current query point (white dot), the step size Δy is set. o And Δθ, using the mileage range S=[y o -Δy o ,y o +Δy o and angular range θ s = [θ-Δθ, θ+Δθ] constructs a neighborhood: if the mileage value y1 and the central angle θ1 of a point satisfy y o -Δy o ≤y1≤y o +Δy o If the angle range is θ-Δθ≤θ1≤θ+Δθ, then the point is a neighboring point of the query point; when Δy o When Δy is small and Δθ is large, the neighborhood exhibits a vertical distribution; when Δy is small... o When Δθ is large and Δy is small, the neighborhood exhibits a horizontal distribution; by adjusting the two step size parameters Δy... o Given Δθ, construct a vertically distributed neighborhood N1 and a horizontally distributed neighborhood N2, and extract the corresponding two neighborhood point sets N1(p o ) and N2(p o ); Calculate the mean I of the reflection intensity of the neighborhood point set. mean and standard deviation I std If the reflection intensity I at the current query point o Satisfy I o <I mean -3I std If the condition is met, the point is preliminarily identified as a leak point; otherwise, it is preliminarily identified as a non-leak point. The proportions of leak points in the two neighboring point sets, prob1 and prob2, are calculated respectively, and the neighboring point set of the current query point is determined by the following formula:

[0103]

[0104] (4) After determining the neighborhood point set, press [I i -min(I i )] / [max(I i )-min(I i The reflection intensity of the neighborhood point set is normalized. The normalized reflection intensity value ranges from [0,1]. The intensity interval width ω is set to 0.05, so [0,1] is divided into 20 intensity intervals. The number n points in the neighborhood point set that fall into each intensity interval is counted. K The probability density value p for each intensity interval is calculated using the following formula. K Obtain a 20-dimensional local intensity descriptor, such as Figure 5 As shown,

[0105]

[0106] Where N is the total number of points in the neighborhood point set.

[0107] Step Four: As Figure 6 As shown, by combining the local intensity descriptor and the CatBoost classifier, leakage segmentation within the tunnel pipe is achieved, as detailed below:

[0108] (1) Construct 1000 decision trees, set the maximum number of iterations to 1000, the learning rate η to 0.01, the decision tree depth to 10, the loss function loss_function to Logloss function, and initialize the CatBoost model;

[0109] (2) Using local descriptors as training features, fit the features using the constructed decision tree, and calculate the value of the loss function according to the following formula:

[0110]

[0111] Where n is the number of samples, x i y is a feature of the i-th sample. i f(x) is the true label of the i-th sample. i ) is the model's predicted label, l(y) i ,f(x i )) represents the loss of the i-th sample;

[0112] (3) Calculate the prediction function f in each iteration t using the following formula. (t) (x i gradient of )

[0113]

[0114] in,

[0115] g i (t) It is the gradient of the i-th sample in the t-th iteration, representing the prediction error of the sample;

[0116] This represents taking the partial derivative of the function;

[0117] (4) Based on the gradient calculated in the t-th iteration, refit the training features to obtain a new predicted label h. t (x i The model is optimized by progressively minimizing the loss function using the following formula:

[0118] f (t+1) (x i )=f (t) (x i )+η·h t (xi (10)

[0119] in,

[0120] η represents the learning rate, a hyperparameter used to control the step size of model updates in each iteration;

[0121] (5) Repeat steps (2), (3), and (4) in step four until the maximum number of iterations is reached, then terminate the training and obtain the optimal model; subsequently, use this model to infer the test data and obtain the predicted labels for each point within a single segment, such as... Figure 6 As shown, based on the prediction results, the points marked as leaks are extracted.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing and applying a local descriptor for segmenting leaks within a tunnel pipe, characterized in that, The method comprises the following steps: S1, scanning the subway tunnel by using a mobile laser scanning system to obtain high-precision point cloud data; S2, preprocessing the tunnel point cloud, extracting the lining point cloud and dividing it into single segments; S3, constructing a neighborhood for the points in a single segment and generating a local intensity descriptor; S4, combining the local intensity descriptor and the CatBoost classifier to segment the tunnel pipe leakage; In step S3, the step of constructing a neighborhood for the points in a single segment and generating a local intensity descriptor comprises the following steps: S3.1, project the single tube piece to XOZ plane, get section points , ellipse fitting to the section points, get the center of the circle , translate the section, move the center of the circle to the origin, the translation process is shown as follows: (4) wherein, represents the translated section point; S3.2, in a single segment, calculate the query point the central angle of the circle The central angle calculation formula is as follows: (5) wherein, Indicates the first quadrant, Indicates the second quadrant. Indicates the third quadrant. represents the fourth quadrant; is the arctangent function; S3.3, set step size for current query point and , construct neighborhood with range of mileage and range of angle , if the mileage value and the central angle of a point satisfy and range of angle , respectively, then the point is the neighborhood point of the query point; when is smaller, is larger, the neighborhood is vertically distributed; when is larger, is smaller, the neighborhood is horizontally distributed; by adjusting two step size parameters and , construct vertically distributed neighborhood and horizontally distributed neighborhood , extract the corresponding two neighborhood point sets and ; calculate the mean value and the standard deviation of the reflectance intensity of the neighborhood point set, if the reflectance intensity of the current query point satisfies , then the point is preliminarily determined as a leakage point, otherwise, the point is preliminarily determined as a non-leakage point; calculate the proportion of the number of leakage points in the two neighborhood point sets and , respectively, and determine the neighborhood point set of the current query point according to the following formula: (6) S3.4, after determining the neighborhood point set, according to The reflection intensity of the neighborhood point set is normalized, and the normalized reflection intensity value ranges from 0 to 1, so that For the intensity interval width, divide [0, 1] into Intensity intervals, and count the number of points in each intensity interval The probability density value of each intensity interval is calculated according to the following formula , and the local intensity descriptor is obtained; (7) wherein, is the total number of points of the neighborhood point set.

2. The method for constructing and applying local descriptors for segmenting leaks within a tunnel pipe according to claim 1, wherein, In step S2, the step of preprocessing the tunnel point cloud, extracting the lining point cloud and dividing it into single segments comprises the following steps: S2.1, the subway tunnel point cloud is , set the slice thickness , divide the tunnel point cloud into multiple slices according to the following formula: (1) wherein, for the number of slices, denotes the ceiling function, and are respectively the maximum and minimum values of the slice point coordinates ;​​ S2.2, project the slice point to XOZ plane to obtain the section point ellipse fitting is performed on the section point to obtain an ellipse standard equation, as shown in the following formula: (2) wherein, is the center of the circle, is the major semi-axis, is the short semi-axis; S2.3, calculate the distance from the section point to the fitted ellipse , set the distance threshold and , if the distance from a point to the fitted ellipse meets , the point is a tunnel lining point, otherwise, the point is a non-lining point, and the obtained lining point cloud is ; S2.4, according to the designed width of the segment, segment the lining point cloud into segments: (3) wherein, for the number of segments, denotes the ceiling function, and are respectively the maximum and minimum values of Design width of the segment.

3. The method for constructing and applying local descriptors for segmenting leaks within a tunnel pipe according to claim 1, wherein, In step S4, the step of combining the local intensity descriptor and the CatBoost classifier to segment the tunnel pipe leakage comprises the following steps: S4.1, build a decision tree, set the maximum number of iterations , learning rate , depth of decision tree , loss function parameters, initialize the CatBoost model; S4.2, taking the local descriptor as a training feature, fitting the feature by using the constructed decision tree, and calculating the value of the loss function according to the following formula: (8) wherein, for the number of samples, is the feature of the th sample, is the true label of the th sample, is the predicted label of the model, represents the loss for the th sample; S4.3, calculate for each iteration round the gradient of the prediction function in the middle (9) wherein, is the gradient of the sample in the iteration, representing the prediction error of the sample, denotes the partial derivative of the function; S4.4, based on the first re-fitting the training features with the newly computed gradients, resulting in new predicted labels minimizing the loss function step by step, thereby optimizing the model: (10) wherein, denotes the learning rate, which is a hyperparameter, used to control the step size of the model update in each iteration; S4.5, repeating steps S4.2, S4.3 and S4.4 until the maximum number of iterations is reached, terminating the training and obtaining an optimal model, and then using the model to infer the test data to obtain the predicted label of each point in a single segment.

Images