High-precision identification method for gap between ballastless track layers

Abstract

The invention relates to the technical field of ballastless tracks, and provides a ballastless track interlayer gap high precision identification method comprising the following steps: 1, collecting track structure gap damage; 2, a detection area, a gap adjacent area and a gap far-end area are arranged on the side face of the track structure, and the detection area is divided; step 3, performing data smooth sparse processing according to different partition attributes; step 4, carrying out space curved surface fitting on the detection area and the gap adjacent area which are subjected to smooth sparse processing through machine learning enhancement so as to obtain a benchmark reference curved surface; step 5, consistency detection; and step 6, parallel calculation is carried out on the gap adjacent areas in different detection subareas of each track plate, and the gap damage set size of each track plate is obtained. According to the invention, the gap between the ballastless track layers of the high-speed rail can be detected in a non-destructive manner, and accurate and reliable measurement of the submillimeter gap can be realized under the conditions of serious damage, foreign matter attachment and the like of the detection surface.

Description

Technical Field

[0001] The invention relates to the technical field of ballastless track, in particular to a high-precision identification method for inter-layer gaps of ballastless track. Background Art

[0002] Ballastless track is a typical layered structure. Under the influence of trains and complex environments, ballastless track will suffer from gap damage. The gap will gradually deteriorate the mechanical properties of the ballastless track. With the extension of service time, if these damages are not discovered in time, they will often quickly develop into larger damages, further weakening the durability and service life of the ballastless track bed. In order to ensure the safe operation of the line, the track status must be comprehensively and carefully inspected, which requires more accurate, advanced and faster detection technology.

[0003] The current gap detection method has the following shortcomings:

[0004] 1) The existing gap detection device mainly relies on manpower to move forward, and the detection speed and efficiency are low;

[0005] 2) When the camera of the existing detection device detects the gap at a specified position, there is an angle when the camera shoots, resulting in a blind area, which affects the accuracy of detecting the gap;

[0006] 3) In the process of detecting damage defects, the full-size point cloud line fitting calculation is relied on without data sparse reduction operation, resulting in low detection efficiency;

[0007] 4) The damage size is determined by searching the distance between damage boundary points, which is easily affected by the favorable points and lacks a reference surface;

[0008] 5) When the distance between the two points where the fitted straight line intersects is used as the characteristic value of the height difference of the track structure, the straight line fitting is prone to abnormal straight line angles;

[0009] 6) PCA data dimensionality reduction and filter sampling are prone to loss of local information. When data is missing, it is easily disturbed by noise points.

[0010] 7) The traditional least squares method fits the surface by minimizing the sum of squared errors, but it is usually only applicable to simple linear or polynomial surfaces and has difficulty in handling complex nonlinear data distributions. Summary of the invention

[0011] The present invention provides a high-precision identification method for ballastless track interlayer gaps, which can detect the interlayer gaps of high-speed railway ballastless track in a non-destructive manner, and can achieve accurate and reliable measurement of sub-millimeter gaps when there is severe damage or foreign matter attached to the detection surface.

[0012] A high-precision identification method for interlayer gaps of ballastless track according to the present invention comprises the following steps:

[0013] Step 1: Collect the damage of track structure cracks;

[0014] Step 2: setting a detection area, a gap adjacent area, and a gap distal area on the side of the track structure, and dividing the detection area;

[0015] Step 3: Perform data smoothing and sparse processing according to different partition attributes;

[0016] Step 4: Perform spatial surface fitting on the smoothed and sparsely processed detection area and the gap adjacent area through machine learning enhancement to obtain a reference surface;

[0017] Step 5: Consistency test;

[0018] Step 6: Perform parallel calculations on the gap adjacent areas in different inspection zones of each track slab to obtain the size of the gap damage set of each track slab.

[0019] Preferably, in step 1, the damage of the rail structure separation is collected by a wheel hub encoder, a line laser 3D stereo camera, and a dual-axis robotic arm, as follows:

[0020] 1.1) Unfold the dual-axis robotic arm and automatically adjust the line laser 3D stereo camera to the track plate and adjustment layer to vertically shoot the damage position according to the corresponding track type. When there is an interference on the side of the track, the robotic arm automatically lifts up and passes through the interference to continue collecting;

[0021] 1.2) Enable the wheel encoder, and use the external pulse drive mode to drive the laser 3D stereo camera into the acquisition working state;

[0022] 1.3) Start the vehicle and continuously collect and store the seam point cloud depth image data at high speed.

[0023] Preferably, in step 2, the detection area, the area adjacent to the gap, and the area distal to the gap are all rectangular areas, the detection area completely covers the gap, and the side of the track structure is set as the detection area; the horizontal center line of the area adjacent to the gap coincides with the horizontal center line of the interlayer gap; the ratio of the height of the area adjacent to the gap to the height of the detection area is n, and the value range of n is 10%-20%; in the detection area, after removing the area adjacent to the gap, the remaining area is the area distal to the gap.

[0024] Preferably, in step 3, specifically:

[0025] 3.1) For the adjacent area of ​​the gap, all data are retained, or every other 1 downsampling is performed. The data set of the adjacent area of ​​the gap of the i-th detection partition is recorded as D i , i=0,1,2,…m-1;

[0026] 3.2) In the far-end area from the seam, down-sampling is performed every a, where a is an integer between 3 and 7;

[0027] 3.3) The data set belonging to the i-th detection partition is recorded as P i , i=0,1,2,…m-1;

[0028] 3.4) The data set of the far-end area of ​​the i-th detection partition is recorded as P Li , i=0,1,2,…m-1;

[0029] 3.5) In P i In the data set, the point coordinate data format of the three-dimensional space is (x b ,y i ,z i ), then for m data, define the following piecewise function C(jΔy):

[0030]

[0031] Where j is 0, 1, 2...m-1, Δy is the discrete step length; y 0 ,y 1 ,y 2 ,y m-1 represents the y-direction coordinate, z 0 、z 1 、z m-1 Indicates the z-direction coordinate;

[0032] Set discrete step length:

[0033]

[0034] Define the control quantity e λ (j):

[0035] e λ (j)=C(jΔy)-Pos(j)+λ||Pos(j)|| 1

[0036] Where POS(j) is a smoothing function, λ>0 is the sparse regularization coefficient;

[0037] Sparse smoothing function iteration:

[0038] Pos(j+1)=Pos(j)+f λ (j)

[0039] In the formula, f λ (j) is a smoothing factor, which includes a combination of control amount and weight;

[0040]

[0041] In the formula, s 0 ,s1 ,s 2 is the coefficient for adjusting the degree of smoothness;

[0042] Define the initial conditions for the iteration:

[0043]

[0044] e(0)=e(-1)=0

[0045] After iterative calculation, the following data set is finally obtained:

[0046] (x b ,y 0 ,Pos(0)),(x b ,y 0 +Δy,Pos(1)),(x b ,y 0 +2Δy,Pos(2)),...,(x b ,y 0 +jΔy,Pos(j))

[0047] 3.6) Replace P with the smoothed sparse data obtained above i Data collection.

[0048] Preferably, in step 4, specifically:

[0049] 4.1) Perform spatial surface fitting on the data sets of the first detection partition, the middle detection partition, and the last detection partition of each track plate;

[0050] 4.2) Construct a feedforward neural network, where the network output z represents the surface function of the detection partition data set, i.e. z = f(x, y, θ);

[0051] 4.3) Optimize the target surface by combining the least squares error and regularization, use the gradient descent method to update the machine learning model parameters θ, and use the optimal solution of the least squares method as the initial weight;

[0052] 4.4) Initialize the model parameters θ, obtain the preliminary fitting results through the least squares method, use them as the initialization values ​​of the machine learning model, and use the training data for iterative optimization to minimize the loss function θ * , adjust the regularization coefficient λ according to the validation set error;

[0053]

[0054] In the formula, R(θ) represents regularization, and N represents the total number of discrete points;

[0055] 4.5) When the loss drops to the minimum, the fitted surface function is output.

[0056] Preferably, in step 5, specifically:

[0057] 5.1) In the spatial surface fitting function of the first detection partition, the middle detection partition, and the last detection partition of each track plate, the extracted surface coefficient vector is V 0 、V (l-1) / 2 and V l-1 ;

[0058] 5.2) Calculate V 0 -V (l-1) / 2 , V 0 -V l-1 , V (l-1) / 2 -V l-1 , and calculate the variance var of these 3 norms;

[0059] 5.3) If the variance var is less than or equal to the preset threshold th, where the threshold th ranges from 2 to 10, it is considered that the fitting surface has good consistency, that is, the fitting reference surface of the detection partition is highly coincident with the global fitting reference surface of each track side;

[0060] 5.4) The average value of the fitting surfaces of the three detection partitions is used as the global reference surface.

[0061] Preferably, in step 6, specifically:

[0062] 6.1) For the i-th detection partition, the reference surface is uniformly recorded as basep i ,i=0,1,2,3…m-1;

[0063] 6.2) Set the adjacent area data set P Li All in basep i The points below are regarded as points in the gap, and the points with the same x coordinates are recorded as col_low_i, i = 0, 1, 2, 3... Similarly, all points above and located at basep i The set of points on the surface is denoted as col_high_i, i=0,1,2,3…, and the points in the set are arranged in ascending order based on their y coordinates;

[0064] 6.3) Traverse each data set col_low_i and calculate the distance from each point to basep i The distance is recorded and the result is dis_i, i = 0, 1, 2, 3..., and the maximum distance value is found in dis_i and recorded as d_max_i;

[0065] 6.4) If d_max_i is greater than or equal to the preset gap detection depth threshold, the depth threshold value range is 5mm-10mm, then it is determined that there is gap damage, and in dis_i, the distance basep is taken i The average value of the nearest q points is the crack depth, and the value range of q is 3-5;

[0066] 6.5) Calculate the two points in the set of points that are the farthest from each other in the x direction. These two points are the edge points of the gap. The distance between them is the width of the gap. If col_low_i is an empty set, col_high_i is used to determine the width of the gap at this position.

[0067] 6.6) Count the number of positions k that are continuously determined to be gaps. Every time one position is detected, it indicates that there is actually a gap of c mm in length. k*c is the total length of the gaps in this data frame.

[0068] 6.7) Summarize and count the results of parallel calculation to obtain the aggregate size of the gap damage of each track slab.

[0069] The beneficial effects of the present invention are as follows:

[0070] (1) The technical solution of wheel encoder + dual-axis robotic arm + 3D stereo camera is used, which enables the camera to collect 3D data with minimal distortion perpendicular to the measured surface; when collecting data, the wheel encoder outputs sampling pulses based on the wheel rotation angle and is insensitive to the uniformity of vehicle speed; the sampling pulses are accumulated and counted to convert the vehicle's travel distance, and the data sampling geographic information can be located at the same time;

[0071] (2) In the processing method, the concepts of adjacent area, distal area and detection area are proposed, and the range and downsampling parameters of each area are limited. In the distal area, sparse sampling is performed to a large extent. With a small amount of data and relying on the breadth of its spatial distribution, the accuracy of the fitting surface is guaranteed. The above measures effectively reduce the time complexity and space complexity of the method.

[0072] (3) The surface model provided by the present invention is used as the fitting target of the reference surface. When the measured surface is deformed, the surface model can still be close to the real surface with high accuracy. At the same time, if the measured surface is a high-quality plane, the model can also be degenerated into a spatial plane model to meet the fitting accuracy requirements.

[0073] (4) The concept of parallel computing of detection partitions is proposed, and the number of detection partitions is set to an odd number. The size of the gap damage set is calculated in parallel in different detection partitions, which speeds up the calculation speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a flow chart of a high-precision identification method of inter-layer gaps of ballastless track in an embodiment;

[0075] Figure 2 It is a front view of the data collection process of the interlayer gap of the ballastless track in the embodiment;

[0076] Figure 3 It is a side view of the data collection process of the interlayer gap of the ballastless track in the embodiment;

[0077] Figure 4 The ballastless track layer gap adjacent area, gap distal area and detection area are divided in the embodiment;

[0078] Figure 5 The side detection partition number of each track plate in the embodiment. DETAILED DESCRIPTION

[0079] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0080] Example

[0081] like Figure 1 As shown, this embodiment provides a high-precision method for identifying gaps between ballastless track layers, which includes the following steps:

[0082] Step 1: Collect the damage of track structure cracks;

[0083] Step 2: setting a detection area, a gap adjacent area, and a gap distal area on the side of the track structure, and dividing the detection area;

[0084] Step 3: Perform data smoothing and sparse processing according to different partition attributes;

[0085] Step 4: Perform spatial surface fitting on the smoothed and sparsely processed detection area and the gap adjacent area through machine learning enhancement to obtain a reference surface;

[0086] Step 5: Consistency test;

[0087] Step 6: Perform parallel calculations on the gap adjacent areas in different detection zones of each track slab to obtain the size of the gap damage set of each track slab.

[0088] In step 1, the damage of the track structure gap 3 is collected through the wheel hub encoder 1, the line laser 3D stereo camera 2, and the dual-axis robot arm 4, such as Figure 2 and Figure 3 As shown, the details are as follows:

[0089] 1.1) Deploy the dual-axis robotic arm 4, and automatically adjust the line laser 3D stereo camera (2) to the track plate and the adjustment layer to vertically shoot the damage position according to the corresponding track type. When there is an interference on the side of the track, the robotic arm automatically lifts up and passes through the interference to continue collecting;

[0090] 1.2) Enable the wheel encoder 1, and the external pulse drive mode drives the laser 3D stereo camera 2 to enter the acquisition working state;

[0091] 1.3) Start the vehicle and continuously collect and store the seam point cloud depth image data at high speed.

[0092] In step 2, if Figure 4 As shown, the detection area 7, the gap adjacent area 5, and the gap distal area 6 are all rectangular areas. The detection area completely covers the gap (or the area where the gap may appear), and the side of the track structure is set as the detection area; the horizontal center line of the gap adjacent area coincides with the horizontal center line of the interlayer gap; the ratio of the height of the gap adjacent area to the height of the detection area is n, and the value range of n is 10%-20%; in the detection area, after removing the gap adjacent area, the remaining area is the gap distal area.

[0093] Each track plate is divided into l rectangular areas, where l must be an odd number, and 7 is recommended. Parallel calculations are performed in different detection areas. The specific form is as follows: Figure 5 As shown, the detection partitions are numbered 0, 1, 2, 3, etc. from left to right.

[0094] In step 3, specifically:

[0095] 3.1) For the adjacent area of ​​the gap, all data are retained, or every other 1 downsampling is performed. The data set of the adjacent area of ​​the gap of the i-th detection partition is recorded as D i , i=0,1,2,…m-1;

[0096] 3.2) In the far-end area from the seam, down-sampling is performed every a, where a is an integer between 3 and 7;

[0097] 3.3) The data set belonging to the i-th detection partition is recorded as P i , i=0,1,2,…m-1;

[0098] 3.4) The data set of the far-end area of ​​the i-th detection partition is recorded as P Li , i=0,1,2,…m-1;

[0099] 3.5) In P i In the data set, the point coordinate data format of the three-dimensional space is (x b ,y i ,z i ), then for m data, define the following piecewise function C(jΔy):

[0100]

[0101] Where j is 0, 1, 2...m-1, Δy is the discrete step length; y 0 ,y 1 ,y 2 ,y m-1 represents the y-direction coordinate, z 0 、z 1 、z m-1 Indicates the z-direction coordinate;

[0102] Set discrete step length:

[0103]

[0104] Define the control quantity e λ (j):

[0105] e λ (j)=C(jΔy)-Pos(j)+λ‖Pos(j)‖ 1

[0106] Where POS(j) is a smoothing function, λ>0 is the sparse regularization coefficient;

[0107] Sparse smoothing function iteration:

[0108] Pos(j+1)=Pos(j)+f λ (j)

[0109] In the formula, f λ (j) is a smoothing factor, which includes a combination of control amount and weight;

[0110]

[0111] In the formula, s 0 ,s 1 ,s 2 is the coefficient for adjusting the degree of smoothness;

[0112] Define the initial conditions for the iteration:

[0113]

[0114] e(0)=e(-1)=0

[0115] After iterative calculation, the following data set is finally obtained:

[0116] (x b ,y 0 ,Pos(0)),(x b ,y 0+Δy,Pos(1)),(x b ,y 0 +2Δy,Pos(2)),...,(x b ,y 0 +jΔy,Pos(j))

[0117] 3.6) Replace P with the smoothed sparse data obtained above i Data collection.

[0118] In step 4, specifically:

[0119] 4.1) Perform spatial surface fitting on the data sets of the first detection partition, the middle detection partition, and the last detection partition of each track plate;

[0120] 4.2) Construct a feedforward neural network (more than three layers of network structure), the network output z represents the surface function of the detection partition data set, that is, z = f(x, y, θ);

[0121] 4.3) Optimize the target surface by combining the least squares error and regularization, use the gradient descent method to update the machine learning model parameters θ, and use the optimal solution of the least squares method as the initial weight;

[0122] 4.4) Initialize the model parameters θ, obtain the preliminary fitting results through the least squares method, use them as the initialization values ​​of the machine learning model, and use the training data for iterative optimization to minimize the loss function θ * , adjust the regularization coefficient λ according to the validation set error;

[0123]

[0124] In the formula, R(θ) represents regularization, and N represents the total number of discrete points;

[0125] 4.5) When the loss drops to the minimum, the fitted surface function is output.

[0126] In step 5, specifically:

[0127] 5.1) In the spatial surface fitting function of the first detection partition, the middle detection partition, and the last detection partition of each track plate, the extracted surface coefficient vector is V 0 、V (l-1) / 2 and V l-1 ;

[0128] 5.2) Calculate V 0 -V (l-1) / 2 , V 0 -V l-1 , V (l-1) / 2 -V l-1The l2 norm of , and calculate the variance var of this 3 norms;

[0129] 5.3) If the variance var is less than or equal to the preset threshold th (th is recommended to be a value between 2 and 10), it is considered that the fitting surface has good consistency, that is, the fitting reference surface of the detection partition is highly coincident with the global fitting reference surface of each track side;

[0130] 5.4) The average value of the fitting surfaces of the three detection partitions is used as the global reference surface.

[0131] In step 6, specifically:

[0132] 6.1) For the i-th detection partition, the reference surface is uniformly recorded as basep i ,i=0,1,2,3…m-1;

[0133] 6.2) Set the adjacent area data set P Li All in basep i The points below are regarded as points in the gap, and the points with the same x coordinates are recorded as col_low_i, i = 0, 1, 2, 3... Similarly, all points above and located at basep i The set of points on the surface is denoted as col_high_i, i=0,1,2,3…, and the points in the set are arranged in ascending order based on their y coordinates;

[0134] 6.3) Traverse each data set col_low_i and calculate the distance from each point to basep i The distance is recorded and the result is dis_i, i = 0, 1, 2, 3..., and the maximum distance value is found in dis_i and recorded as d_max_i;

[0135] 6.4) If d_max_i is greater than or equal to the preset gap detection depth threshold (the depth threshold is recommended to be a value between 5mm and 10mm), it is determined that there is gap damage. In dis_i, the distance basep is taken i The average value of the nearest q points is the crack depth, and the value range of q is 3-5;

[0136] 6.5) Calculate the two points in the set of points that are the farthest from each other in the x direction. These two points are the edge points of the gap. The distance between them is the width of the gap. If col_low_i is an empty set, col_high_i is used to determine the width of the gap at this position.

[0137] 6.6) Count the number of positions k that are continuously determined to be gaps. Every time one position is detected, it indicates that there is actually a gap of c mm in length. k*c is the total length of the gaps in this data frame.

[0138] 6.7) Summarize and count the results of parallel calculation to obtain the aggregate size of the gap damage of each track slab.

[0139] This embodiment designs a gap acquisition device of wheel hub encoder + linear laser 3D stereo camera + dual-axis robotic arm, which can automatically adjust the laser camera to extend to the track plate and adjust the layer position according to the track type, and vertically collect the gap point cloud depth image to ensure the acquisition accuracy. The device can perform fast linear array high-precision acquisition with a resolution of 2560px and a scanning frame rate of 46khz, ensuring 0.1mm acquisition and recognition accuracy at a speed of 16.5km / h.

[0140] In order to increase detection efficiency and reduce recognition complexity, this embodiment proposes a sparse dimensionality reduction method for partitioning point cloud depth images containing seam damage, sets different partitions in the point cloud depth image, and reduces the data capacity of non-seam areas.

[0141] Due to differences in the construction or casting processes of the track plate and the adjustment layer, there is a slight curvature on the contact surface. To ensure high-precision identification, this embodiment proposes a high-order curved surface reference benchmark for gap identification to reduce the accuracy impact caused by outliers or slight curvatures.

[0142] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. A high-precision identification method for inter-layer gaps in ballastless track, characterized in that: The following steps are involved: Step 1: Collect the damage of track structure cracks; Step 2: setting a detection area, a gap adjacent area, and a gap distal area on the side of the track structure, and dividing the detection area; Step 3: Perform data smoothing and sparse processing according to different partition attributes; Step 4: Perform spatial surface fitting on the smoothed and sparsely processed detection area and the gap adjacent area through machine learning enhancement to obtain a reference surface; Step 5: Consistency test; Step 6: Perform parallel calculations on the gap adjacent areas in different detection zones of each track slab to obtain the size of the gap damage set of each track slab.

2. The high-precision identification method of inter-layer gaps of ballastless track according to claim 1 is characterized in that: In step 1, the damage of the track structure gap (3) is collected through the wheel hub encoder (1), the line laser 3D stereo camera (2), and the dual-axis robotic arm (4), as follows: 1.1) Deploy the dual-axis robotic arm (4) and automatically adjust the line laser 3D stereo camera (2) to the track plate and the adjustment layer to vertically shoot the damage position according to the corresponding track type. When there is an interference on the side of the track, the robotic arm automatically lifts up and passes through the interference to continue collecting; 1.2) Enable the wheel encoder (1), and use the external pulse drive mode to drive the laser 3D stereo camera (2) to enter the acquisition working state; 1.3) Start the vehicle and continuously collect and store the seam point cloud depth image data at high speed.

3. The high-precision identification method of inter-layer gaps of ballastless track according to claim 2 is characterized in that: In step 2, the detection area, the area adjacent to the gap, and the area distal to the gap are all rectangular areas. The detection area completely covers the gap, and the side of the track structure is set as the detection area; the horizontal center line of the area adjacent to the gap coincides with the horizontal center line of the interlayer gap; the ratio of the height of the area adjacent to the gap to the height of the detection area is n, and the value range of n is 10%-20%; in the detection area, after removing the area adjacent to the gap, the remaining area is the area distal to the gap.

4. The high-precision identification method of inter-layer gaps of ballastless track according to claim 3 is characterized by: In step 3, specifically: 3.1) For the adjacent area of ​​the gap, all data are retained, or every other 1 downsampling is performed. The data set of the adjacent area of ​​the gap of the i-th detection partition is recorded as D i , i=0,1,2,…m-1; 3.2) In the far-end area from the seam, down-sampling is performed every a, where a is an integer between 3 and 7; 3.3) The data set belonging to the i-th detection partition is recorded as P i , i=0,1,2,…m-1; 3.4) The data set of the far-end area of ​​the i-th detection partition is recorded as P Li , i=0,1,2,…m-1; 3.5) In P i In the data set, the point coordinate data format of the three-dimensional space is (x b ,y i ,z i ), then for m data, define the following piecewise function C(jΔy): Where j is 0, 1, 2…m-1, Δy is the discrete step length; y0, y1, y2, y m-1 Indicates the y-direction coordinate, z0, z1, z m-1 Indicates the z-direction coordinate; Set discrete step length: Define the control quantity e λ (j): e λ (j)=C(jΔy)-Pos(j)+λ||Pos(j)||1 Where POS(j) is a smoothing function, λ>0 is the sparse regularization coefficient; Sparse smoothing function iteration: Pos(j+1)=Pos(j)+f λ (j) In the formula, f λ (j) is a smoothing factor, which includes a combination of control amount and weight; In the formula, s0, s1, s2 are coefficients for adjusting the degree of smoothness; Define the initial conditions for the iteration: e(0)=e(-1)=0 After iterative calculation, the following data set is finally obtained: (x b ,y0,Pos(0)),(x b ,y0+Δy,Pos(1)),(x b ,y0+2Δy,Pos(2)),...,(x b ,y0+jΔy,Pos(j)) 3.6) Replace P with the smoothed sparse data obtained above i Data collection.

5. The high-precision identification method of inter-layer gaps of ballastless track according to claim 4 is characterized in that: In step 4, specifically: 4.1) Perform spatial surface fitting on the data sets of the first detection partition, the middle detection partition, and the last detection partition of each track plate; 4.2) Construct a feedforward neural network, where the network output z represents the surface function of the detection partition data set, i.e. z = f(x, y, θ); 4.3) Optimize the target surface by combining the least squares error and regularization, use the gradient descent method to update the machine learning model parameters θ, and use the optimal solution of the least squares method as the initial weight; 4.4) Initialize the model parameters θ, obtain the preliminary fitting results through the least squares method, use them as the initialization values ​​of the machine learning model, and use the training data for iterative optimization to minimize the loss function θ * , adjust the regularization coefficient λ according to the validation set error; In the formula, R(θ) represents regularization, and N represents the total number of discrete points; 4.5) When the loss drops to the minimum, the fitted surface function is output.

6. The high-precision identification method of inter-layer gaps of ballastless track according to claim 5 is characterized in that: In step 5, specifically: 5.1) In the spatial surface fitting function of the first detection partition, the middle detection partition, and the last detection partition of each track plate, the extracted surface coefficient vectors are V0, V (l-1) / 2 and V l-1 ; 5.2) Calculate V0-V (l-1) / 2 , V0-V l-1 , V (l-1) / 2 -V l-1 , and calculate the variance var of these 3 norms; 5.3) If the variance var is less than or equal to the preset threshold th, where the threshold th ranges from 2 to 10, it is considered that the fitting surface has good consistency, that is, the fitting reference surface of the detection partition is highly coincident with the global fitting reference surface of each track side; 5.4) The average value of the fitting surfaces of the three detection partitions is used as the global reference surface.

7. The high-precision identification method of inter-layer gaps of ballastless track according to claim 6 is characterized by: In step 6, specifically: 6.1) For the i-th detection partition, the reference surface is uniformly recorded as basep i ,i=0,1,2,3…m-1; 6.2) Set the adjacent area data set P Li All in basep i The points below are regarded as points in the gap, and the points with the same x coordinates are recorded as col_low_i, i = 0, 1, 2, 3... Similarly, all points above and located at basep i The set of points on the surface is denoted as col_high_i, i=0,1,2,3…, and the points in the set are arranged in ascending order based on their y coordinates; 6.3) Traverse each data set col_low_i and calculate the distance from each point to basep i The distance is recorded and the result is dis_i, i = 0, 1, 2, 3..., and the maximum distance value is found in dis_i and recorded as d_max_i; 6.4) If d_max_i is greater than or equal to the preset gap detection depth threshold, the depth threshold value range is 5mm-10mm, then it is determined that there is gap damage, and in dis_i, the distance basep is taken i The average value of the nearest q points is the crack depth, and the value range of q is 3-5; 6.5) Calculate the two points in the set of points that are the farthest from each other in the x direction. These two points are the edge points of the gap. The distance between them is the width of the gap. If col_low_i is an empty set, col_high_i is used to determine the width of the gap at this position. 6.6) Count the number of positions k that are continuously determined to be gaps. Every time one position is detected, it indicates that there is actually a gap of c mm in length. k*c is the total length of the gaps in this data frame. 6.7) Summarize and count the results of parallel calculation to obtain the aggregate size of the gap damage of each track slab.

Images