Measurable turnout apparent disease intelligent identification method

Abstract

The embodiment of the invention discloses a measurable turnout apparent disease intelligent identification method, which comprises the following steps of: acquiring top images and oblique images of a stock rail and a switch rail by using a rail detection vehicle, and inputting the top images into a top width analysis model which comprises a switch rail identification module and a top width identification module, the switch rail recognition module is used for recognizing the outline of the switch rail, the top width recognition module determines the position of the switch rail with the preset width according to the outline of the switch rail, the height difference analysis model comprises an image conversion module and a height difference calculation module, and the image conversion module converts an oblique image into a lateral image; and the height difference calculation module determines the positions of the switch rail and the stock rail at a preset height difference, compares the output position data of the top width analysis model with the output position data of the height difference analysis model, and outputs a disease identification result when the two position data are inconsistent.

Description

Technical Field

[0001] The invention relates to the field of turnout apparent disease identification, and in particular to a measurable turnout apparent disease intelligent identification method. Background Art

[0002] Turnouts are key line branching devices in railway track systems, designed to solve the problem of train path selection in multi-track networks. They have the characteristics of complex structure, short service life, large number of components, and limited train speed. As the weak link of the track, its performance directly affects the line's capacity, wheel-rail relationship, and maintenance cost.

[0003] The point rail is an important part of the turnout switch. It is the key to the turnout. When the train passes through the point rail, it will generate a large lateral impact force and a longitudinal impact force. Therefore, the point rail often suffers from diseases, such as wear and deformation of the point rail, the point rail is not close to the base rail, and the point rail is not close to the slide bed. These problems will accelerate the wear and scrapping of the point rail, and may cause excessive dynamic response when the train is running, resulting in dynamic unevenness, affecting the safety, stability and comfort of the train when passing the turnout.

[0004] The point rail needs to have a certain reduction value compared to the base rail. When the train passes through the switch, the reduction value of the tip of the point rail (for example, 23mm) is to ensure that the wheel can gradually increase the contact area with the point rail, thereby smoothly transferring the load. If the tip is not properly reduced, the wheel will have a severe impact at the tip of the point rail. Where the point rail has a certain width (for example, 20mm), the base rail and the point rail jointly bear the vertical load of the wheel. Since the wheel tread has a slope, the point rail should have a certain reduction value (for example, 2mm) at this section. At this width section, the point rail has better deformation resistance. If the width of the point rail does not meet expectations and it bears the wheel load together with the base rail, it will cause the point rail to be overloaded, bringing the hidden danger of point rail disease.

[0005] The existing detection of the reduction value of the point rail mainly relies on manual contact measurement using feeler gauges, calipers or simple measuring instruments, which has the defects of low efficiency, poor accuracy and strong subjectivity. The traditional method is also affected by ambient light and oil pollution on the surface of the rail, and is prone to miss local wear or cracks, resulting in discontinuous data and unable to meet the high-precision and real-time maintenance needs of high-speed railways. At the same time, it is not enough to only measure the reduction value of the point rail. The reduction value of the point rail is not combined with the cross-sectional width of the point rail, and it is also impossible to reflect whether the stress condition of the point rail is good.

[0006] Therefore, designing an accurate, fast and measurable intelligent identification method for turnout surface defects is an urgent problem to be solved by technical personnel in this field. Summary of the invention

[0007] In view of this, an embodiment of the present invention provides a method for intelligent identification of measurable apparent diseases of turnouts, which uses a track inspection vehicle to obtain the top images and diagonal images of the stock rail and the switch rail, determines the position of the switch rail with a preset width through the top image, determines the position of the switch rail with a preset height difference through the perspective transformation of the diagonal image, and compares whether the position of the switch rail with the preset width is consistent with the position of the switch rail with the preset height difference to identify diseases, improving the convenience and accuracy of the identification of apparent diseases of turnouts and reducing the identification cost.

[0008] The present application discloses a method for intelligent identification of measurable apparent diseases of turnouts, including the following steps: Using a camera located directly above the stock rail and the switch rail in the track inspection vehicle to obtain the top images of the stock rail and the switch rail; Using a camera located obliquely above the stock rail and the switch rail in the track inspection vehicle to obtain the diagonal images of the stock rail and the switch rail; Inputting the top image into a top width analysis model, the top width analysis model includes a switch rail identification module and a top width identification module, the switch rail identification module is used to identify the switch rail contour, and the top width identification module determines the position of the switch rail with a preset width according to the switch rail contour; Inputting the diagonal image into a height difference analysis model, the height difference analysis model includes an image transformation module and a height difference calculation module, the image transformation module transforms the diagonal image into a lateral image in the horizontal direction, and the height difference calculation module is used to calculate the height difference between the switch rail and the stock rail according to the lateral image and determine the positions of the switch rail and the stock rail at a preset height difference; Comparing the output position data of the top width analysis model and the height difference analysis model, and when the two position data are inconsistent, outputting a disease identification result.

[0009] Preferably, the switch rail identification module includes a feature extraction unit and a contour prediction unit. The feature extraction unit adopts a U-Net structure. The output features of each layer of the encoder are transmitted to the decoder of the same layer and cross-layer transmitted to the decoders of different layers. The decoder outputs contour features after passing the input features through an edge-aware attention module, a transposed convolutional layer, a bilinear interpolation layer, and a refinement convolutional layer in sequence. The edge-aware attention module performs Sobel edge extraction on the features from the previous layer of the decoder to output edge features, performs content feature extraction on the features from the encoder to output content features, calculates attention weights according to the edge features and the content features, and outputs enhanced features according to the attention weights for feature weighting.

[0010] Preferably, the profile feature is input into the profile prediction unit. The profile prediction unit includes a parallel semantic segmentation path and an edge enhancement path. The profile feature passes through the semantic segmentation path and the edge enhancement path respectively and then is fused to output the switch rail profile. The profile feature passes through staged multiple convolution operations in the semantic segmentation path and then outputs the switch rail profile probability. The profile feature passes through depthwise separable convolutions, residual blocks, and depthwise separable convolutions in the edge enhancement path and then outputs a distance field map.

[0011] Preferably, the top width recognition module includes a profile parameterization unit, a width calculation unit, and a width matching unit. The profile parameterization unit is used to convert the switch rail profile into a switch rail profile point sequence. The width calculation unit determines the width array of each point of the profile according to the switch rail profile point sequence. The width matching unit determines the profile points with a preset width according to the width array.

[0012] Preferably, the profile parameterization unit takes the switch rail tip as the origin and the switch rail direction as the x-axis to establish a coordinate system, determines the coordinates of each point of the switch rail profile, and establishes a switch rail profile point sequence. The width calculation unit calculates the normal direction point by point along the switch rail profile point sequence, projects a ray to the side of the switch rail close to the stock rail, determines the intersection point of the ray and the side, calculates the Euclidean distance between each point of the switch rail profile point sequence and the intersection point, and establishes a width array. The width matching unit searches for the profile point closest to the preset width in the width array.

[0013] Preferably, the image transformation module includes a track feature detection unit, a coordinate system construction unit, a perspective transformation unit, and a remapping unit. The track feature detection unit detects the switch rail tip and the stock rail working surface and determines the feature points of the switch rail and the stock rail; The coordinate system construction unit takes the switch rail tip as the origin of the coordinate system, uses the direction of the straight line fitted parallel to the stock rail working surface as the x-axis, and uses the direction perpendicular to the x-axis as the y-axis to establish a coordinate system; The perspective transformation unit constructs and solves a perspective transformation matrix by matching the feature points in the oblique image with the feature points in the reference image; The remapping unit compensates for the longitudinal offset of the image caused by vehicle vibration through the pitch angle deviation value measured by the camera in real time, corrects the perspective transformation matrix, and uses the corrected perspective transformation matrix to map any point in the oblique image to the corresponding position in the lateral image to realize the conversion of the image perspective and obtain the lateral image.

[0014] Preferably, the height difference calculation module includes an edge extraction unit, a reference plane unit, a height difference calculation unit, and a positioning unit. The edge extraction unit detects the rail top edge point sets of the switch rail and the stock rail according to the lateral image. The reference plane unit models the rail tops of the switch rail and the stock rail according to the rail top edge point sets of the switch rail and the stock rail, and determines the edge point sequences of the switch rail and the stock rail. The height difference calculation unit calculates the distance between the switch rail and the stock rail point by point in the vertical direction according to the edge point sequences of the switch rail and the stock rail, forming a height difference array. The positioning unit searches for the contour point closest to the preset height difference in the height difference array.

[0015] Preferably, determine the horizontal distance between the contour point output by the top width analysis model and the tip of the switch rail, determine the horizontal distance between the contour point output by the height difference analysis model and the tip of the switch rail, compare the two horizontal distances, and if the deviation is greater than the preset threshold, output the disease identification result.

[0016] The present application proposes a measurable intelligent identification method for turnout apparent diseases. The method uses a track inspection vehicle to obtain switch rail images, does not rely on manual measurement, reduces the detection cost, and improves the identification accuracy. After performing perspective transformation on the oblique image to measure the height difference, it solves the limitation that the track inspection vehicle cannot directly obtain lateral images, and improves the applicability of the track inspection vehicle in the scenario of switch rail detection. Through the design of the switch rail recognition module, the contour of the switch rail is accurately determined, providing a solid data basis for the measurement of the switch rail width. By identifying and comparing the preset width position and the preset height difference position, the joint detection of the switch rail reduction value and the switch rail width is realized, and the disease hidden danger of the switch rail can be more accurately discovered. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Through the following description of the embodiments of the present invention with reference to the accompanying drawings, the above and other objects, features, and advantages of the present invention will become more apparent. In the drawings: Figure 1 is a flowchart of a measurable intelligent identification method for turnout apparent diseases provided by an exemplary embodiment of the present invention; Figure 2 is a schematic structural diagram of a top width analysis model provided by an exemplary embodiment of the present invention; Figure 3 is a schematic structural diagram of a height difference analysis model provided by an exemplary embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The present application will be described based on embodiments, but the present application is not limited to these embodiments only. In the following detailed description of the present application, some specific details are described in detail. Those skilled in the art can fully understand the present application without the description of these details. In order to avoid obscuring the essence of the present application, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0019] In addition, those of ordinary skill in the art should understand that the accompanying drawings provided herein are for illustrative purposes only, and the drawings are not necessarily drawn to scale.

[0020] Unless the context clearly requires otherwise, words such as "including" and "comprising" in the entire application document should be interpreted as having an inclusive meaning rather than an exclusive or exhaustive meaning; that is, the meaning of "including but not limited to".

[0021] In the description of the present application, it should be understood that terms such as "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. In addition, in the description of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0022] This embodiment provides a method for intelligent identification of measurable apparent diseases of turnouts, as Figure 1 shown, the method provided in this embodiment includes the following steps: Step 1: Use the cameras located directly above the stock rail and switch rail in the track inspection vehicle to obtain the top images of the stock rail and switch rail.

[0023] The existing measurement of the switch rail lowering value generally relies on manual use of special measuring instruments, using high-precision sensors for contact measurement, or using a switch rail lowering value measuring ruler to measure the distances of each section of the switch rail. It still mainly relies on manual on-site measurement, which requires a huge amount of labor costs and measurement work. Moreover, since the switch rail is attached to the inner side of the stock rail, installing image acquisition equipment or laser measurement equipment in the middle of the track will inevitably interfere with the operation of the train, and installing equipment beside the track will be affected when the train passes. Therefore, using the track inspection vehicle to obtain the switch rail image and using the image data as the data basis for disease identification is an efficient and inexpensive method, which avoids the cumbersome and high cost of manual contact measurement and can also improve the identification accuracy.

[0024] Use the cameras located obliquely above the stock rail and switch rail in the track inspection vehicle to obtain the oblique images of the stock rail and switch rail.

[0025] The cameras of the track inspection vehicle are generally distributed directly below the vehicle body to vertically photograph the rail surface for detecting planar defects such as wear and cracks. They are also distributed on both sides of the vehicle body bottom, obliquely aligned with the working surface and the web of the rail, with a shooting angle of 30 to 45 degrees, so as to simultaneously take into account the rail head profile and the rail bottom state. In order to measure the lowering value of the switch rail, the best image is necessarily the side image of the switch rail and the stock rail in the horizontal direction. However, in order to improve the applicability of the track inspection vehicle, the oblique image is directly used as the data basis for switch rail measurement, and the accuracy of the measurement is ensured through subsequent algorithm design.

[0026] Step 2: Input the top image into the top width analysis model 100. The top width analysis model 100 includes a switch rail recognition module 110 and a top width recognition module 120. The switch rail recognition module 110 is used to recognize the switch rail profile, and the top width recognition module 120 determines the position of the switch rail with a preset width according to the switch rail profile.

[0027] Step 3: Input the oblique image into the height difference analysis model 200. The height difference analysis model 200 includes an image transformation module 210 and a height difference calculation module 220. The image transformation module 210 transforms the oblique image into a lateral image in the horizontal direction, and the height difference calculation module 220 is used to calculate the height difference between the switch rail and the stock rail according to the lateral image and determine the positions of the switch rail and the stock rail at a preset height difference. Step 4: Compare the output position data of the top width analysis model and the height difference analysis model. When the two position data are inconsistent, output the disease recognition result.

[0028] In step 2, as Figure 2 shown, the switch rail recognition module 110 includes a feature extraction unit 111 and a contour prediction unit 112. The feature extraction unit 111 adopts a U-Net structure, and the output features of each layer of the encoder are transmitted to the decoder of the same layer and cross-layer transmitted to the decoders of different layers.

[0029] Among them, U-Net is a classic convolutional neural network architecture. Through the collaborative work of the encoder and the decoder, it can effectively extract image features. In the U-Net structure, each layer of the encoder will output a feature map. These feature maps not only contain the features extracted by the current layer, but also retain certain local information. These feature maps will be directly transmitted to the corresponding layers in the decoder. For example, the feature map output by the first layer of the encoder will be transmitted to the first layer of the decoder, and the output feature map of the second layer of the encoder will be transmitted to the second layer of the decoder, and so on.

[0030] In this embodiment, in addition to intra-level transmission, the U-Net structure also cross-level transmits the feature maps of the encoder to different layers of the decoder. The role of the skip connection is to directly transfer the features extracted in the encoder to deeper layers in the decoder, thereby enhancing feature transfer and fusion. For example, the feature map output by the first layer of the encoder may be transmitted to the second or third layer of the decoder, and the feature map output by the second layer of the encoder may be transmitted to the third or fourth layer of the decoder, and so on. Since feature maps of different levels contain feature information of different scales, through cross-level transmission, features of different scales can be fused together, thereby better capturing multi-scale information in the image.

[0031] The decoder sequentially passes the input features through an edge-aware attention module, a transposed convolutional layer, a bilinear interpolation layer, and a refinement convolutional layer, and then outputs the contour features. The edge-aware attention module performs Sobel edge extraction on the features from the previous layer of the decoder to output edge features, performs content feature extraction on the features from the encoder to output content features, calculates attention weights based on the edge features and the content features, and outputs enhanced features by weighting the features according to the attention weights.

[0032] Among them, the edge-aware attention module uses the Sobel operator to extract the edge features output by the previous layer of the decoder as the basis for attention generation, simultaneously captures features of different thicknesses output by the encoder from the same layer and cross-layers, and automatically adjusts the attention weights according to the feature content, enhancing the model's perception ability of the switch rail contour.

[0033] The transposed convolutional layer uses a minimized convolutional kernel (3×3) for transposition, and bilinear interpolation is performed after the transposed convolution, rather than the traditional preposition. The purpose of this design is to enable the transposed convolution to first establish semantic rationality, and the bilinear interpolation is responsible for geometric smoothing. Finally, the refinement convolutional layer uses a 1×1 convolution to correct the local distortions missed in the previous two stages and outputs the contour features.

[0034] The contour features are input to the contour prediction unit 112. The contour prediction unit 112 includes a parallel semantic segmentation path and an edge enhancement path. The contour features are respectively fused after passing through the semantic segmentation path and the edge enhancement path to output the switch rail contour. The contour features sequentially pass through staged multiple convolution operations in the semantic segmentation path and then output the switch rail contour probability. The contour features sequentially pass through depthwise separable convolutions, residual blocks, and depthwise separable convolutions in the edge enhancement path and then output the distance field map.

[0035] Among them, the main function of the semantic segmentation path is to generate a contour probability map, that is, to predict the probability that each pixel belongs to the contour. By gradually restoring the spatial resolution of the feature map in stages, it is made consistent with the resolution of the input image, providing semantic-level contour information for subsequent contour synthesis. At the same time, the convolutional operations in stages learn different features and patterns at each stage, thereby enhancing the generalization ability of the model. By gradually extracting and refining features, the model can better adapt to the input data and task requirements.

[0036] The staged multiple convolutional operations of the semantic segmentation path sequentially include transposed convolution, 3×3 convolution and batch normalization, transposed convolution, 3×3 convolution and batch normalization, and output the switch rail contour probability. The transposed convolution is used to expand the spatial size of the feature map while reducing the number of channels. This step helps to restore the spatial resolution of the feature map and reduce the computational burden; the 3×3 convolution is used to further process the feature map, increasing non-linear features, and stabilizing the training process through batch normalization.

[0037] The main function of the edge enhancement path is to generate sub-pixel edges, that is, to extract edge information through a distance field regression network. This path focuses on the precise extraction and enhancement of edges, providing high-precision edge information for subsequent contour synthesis.

[0038] The edge enhancement path first uses depthwise separable convolution to perform depthwise separable convolution processing on the input feature map, extracting edge features while reducing the computational amount; then introduces residual learning to enhance the feature extraction ability and avoid the problem of gradient disappearance; finally, the feature map is further processed through depthwise separable convolution again to generate a distance field map. The value of each pixel in the distance field map represents the distance from the pixel to the nearest edge, as follows: d is the pixel distance to the nearest edge.

[0039] Finally, the semantic segmentation path provides semantic-level contour information, that is, the probability that each pixel belongs to the contour, and the edge enhancement path provides high-precision edge information, that is, the distance from each pixel to the nearest edge or the probability of the edge. By fusing the outputs of these two paths, semantic information and edge information can be fully utilized to generate more accurate and robust contour prediction results.

[0040] In step 2, the top width recognition module 120 is used to accurately locate the cross-section position where the top width of the switch rail reaches a preset value (such as 20 mm). As Figure 3As shown in the figure, the top width recognition module 120 includes a contour parameterization unit 121, a width calculation unit 122, and a width matching unit 123. The contour parameterization unit 121 is used to convert the switch rail contour into a sequence of switch rail contour points. The width calculation unit 122 determines the width array of each point of the contour according to the sequence of switch rail contour points. The width matching unit 123 determines the contour points with a preset width according to the width array.

[0041] The contour parameterization unit 121 takes the tip of the switch rail as the origin and the direction of the switch rail as the X-axis to establish a coordinate system, determines the coordinates of each point of the switch rail contour, and establishes a sequence of switch rail contour points. The width calculation unit 122 calculates the normal direction point by point along the sequence of switch rail contour points, projects a ray towards the side of the switch rail close to the stock rail, determines the intersection point of the ray and the side, calculates the Euclidean distance between each point of the sequence of switch rail contour points and the intersection point, and establishes a width array. The width matching unit 123 searches for the contour points closest to the preset width in the width array.

[0042] In step 3, the image transformation module 210 uses pure vision geometric transformation technology to convert the obliquely captured side image of the turnout into a standard horizontal view, eliminate perspective distortion, and provide a standardized input for height difference calculation.

[0043] The image transformation module 210 includes an orbital feature detection unit 211, a coordinate system construction unit 212, a perspective transformation unit 213, and a remapping unit 214. The orbital feature detection unit 211 uses a depthwise separable convolutional network to detect the tip of the switch rail and the working surface of the stock rail, and determines the feature points of the switch rail and the stock rail. Among them, the feature points of the switch rail and the stock rail are the points with uniqueness and stability in the switch rail image, which can be the tip of the switch rail, the corner point of the slide table, and the inner point of the stock rail. By finding the corresponding feature points in the source image and the target image, the geometric transformation relationship between the two images can be determined. The corresponding relationship of the feature point pairs is the basis for calculating the perspective transformation matrix.

[0044] The coordinate system construction unit 212 takes the tip of the switch rail as the origin of the coordinate system, uses the direction of the straight line fitted parallel to the working surface of the stock rail as the X-axis, and the direction perpendicular to the X-axis as the Y-axis to establish a coordinate system.

[0045] The perspective transformation unit 213 constructs and solves the perspective transformation matrix by matching the feature points in the oblique image with the feature points in the reference image.

[0046] Among them, the oblique image is the source image, and the reference image is the lateral reference image. Detect the feature points in the oblique image, find multiple pairs of matching feature points between the reference images, and use multiple pairs of matching feature points to calculate the perspective transformation matrix of each image relative to the reference image as follows.

[0047] Among them, (x2, y2) are the coordinates of the feature points in the reference image, (x1, y1) are the coordinates of the feature points in the source image, and H is the perspective transformation matrix.

[0048] The remapping unit 214 compensates for the longitudinal offset of the image caused by vehicle vibration through the pitch angle deviation value measured by the camera in real time, corrects the perspective transformation matrix, and uses the corrected perspective transformation matrix to map any point in the oblique image to the corresponding position in the lateral image, realizing the conversion of the image perspective to obtain the lateral image.

[0049] Among them, the remapping unit 214 is responsible for finely adjusting the image after perspective transformation to ensure that the output image meets the requirements of high-precision height difference calculation.

[0050] In order to eliminate the longitudinal offset of the image caused by vehicle vibration, the remapping unit 214 obtains the pitch angle deviation Δθ (unit: degree) measured by the camera in real time, and corrects the third row parameters of the perspective transformation matrix: Among them, k is the calibration coefficient, Δθ is the pitch angle deviation of the camera, h 31 and h 32 are the first two elements of the third row of the original perspective transformation matrix, and h' 31 and h' 32 are the first two elements of the third row of the corrected perspective transformation matrix.

[0051] h 31 and h 32 are the first two elements of the third row of the perspective transformation matrix, which affect the scaling and position of the points. By adjusting h 31 and h 32 the perspective effect of the image can be controlled.

[0052] Finally, use the corrected perspective transformation matrix to map any point in the oblique image to the corresponding point in the lateral image, so as to obtain the transformed lateral image, as follows.

[0053] Among them, (x i , y i ) are the coordinates of each point in the oblique image, (x j , y j ) are the coordinates of each point in the transformed lateral image, is the corrected perspective transformation matrix, and ω is the scaling factor.

[0054] In step 3, the height difference calculation module 220 includes an edge extraction unit 221, a reference plane unit 222, a height difference calculation unit 223, and a positioning unit 224. The edge extraction unit 221 detects the rail top edge point sets of the switch rail and the stock rail according to the lateral image. The reference plane unit 222 models the rail tops of the switch rail and the stock rail according to the rail top edge point sets of the switch rail and the stock rail, determines the edge point sequences of the switch rail and the stock rail. The height difference calculation unit 223 calculates the distance between the switch rail and the stock rail point by point in the vertical direction according to the edge point sequences of the switch rail and the stock rail, forms a height difference array. The positioning unit 224 searches for the contour point closest to the preset height difference in the height difference array.

[0055] Among them, the reference plane unit 222 determines the working surfaces of the switch rail and the stock rail, which is the reference for height difference calculation. By accurately modeling the rail top shape of the stock rail, it provides a spatial reference system for the switch rail height difference calculation. Through modeling, the influence of fasteners, rust spots and other interfering objects can be excluded, and the smooth constraint of the rail surface geometry can be maintained.

[0056] The reference plane unit 222 models the rail tops of the switch rail and the stock rail according to the rail top edge point sets of the switch rail and the stock rail, including modeling the stock rail and modeling the switch rail. The stock rail modeling includes: 1. Selecting the edge points in the range from the working surface of the stock rail to the center line of the rail head; 2. Using a cubic polynomial model as follows: Where, is the surface height of the stock rail at the x position, x is the position coordinate along the stock rail direction, a0, a1, a2, a3 are the coefficients of the polynomial, used to define the shape of the polynomial, and ε is the error term, representing the difference between the model predicted value and the actual value; 3. Through the RANSAC algorithm for robust fitting, output the fitting coefficients a0, a1, a2, a3, obtain the cubic polynomial model, and determine the edge point sequence of the stock rail; The switch rail modeling uses piecewise spline interpolation fitting, including: 1. Setting a control point every 10 mm along the working surface of the switch rail in the switch rail edge point set, and setting a control point every 2 mm at the tip of the switch rail and in the turnout area; 2. Setting a curve function for the spline between two adjacent control points as follows: Where, S i (x) is the surface height of the switch rail at the x position between the i-th and the i + 1-th control points, x is the position coordinate along the switch rail direction, i is the serial number of the control point, a i , b i , c i , d i are the coefficients of the spline curve function; 3. To avoid excessive curvature changes in the curve, maintain the smoothness and continuity of the curve, limit the curvature change rate, and ensure that the absolute value of the curvature change rate is less than 0.01 mm -2 , as follows: where is the curvature of each spline segment; 4. Use the least squares fitting algorithm to construct the objective function, as follows: where m is the number of control points, is the data fitting term, representing the sum of the squared errors between the spline curve s(x) and the control points (x j , y j ), is the regularization term, used to control the smoothness of the spline curve, and λ is the smoothing factor; Solve the objective function to determine the coefficients a i , b i , c i , d i of the spline curve when the objective function takes the minimum value, and obtain each spline curve function to determine the edge point sequence of the basic rail.

[0057] In step 4, determine the horizontal distance between the contour points output by the top width analysis model 100 and the tip of the switch rail, determine the horizontal distance between the contour points output by the height difference analysis model 200 and the tip of the switch rail, compare the two horizontal distances, and if the deviation is greater than the preset threshold, output the disease identification result.

[0058] Among them, the position data output by the top width analysis model 100 is the coordinate position in the coordinate system established in the top surface image, while the position data output by the height difference analysis model 200 is the coordinate position in the coordinate system established in the side image after perspective transformation. The two coordinate systems are different, but both coordinate systems use the tip of the switch rail as the origin. Then, the x-axis coordinates of the two coordinate positions can both represent the distances between the two contour points and the tip of the switch rail. On the switch rail, whether viewed from the top surface or the side surface, the same distance from the tip of the switch rail represents the same cross-section position. Therefore, in step 4, compare the x-axis coordinates of the contour points output by the top width analysis model 100 with the x-axis coordinates of the contour points output by the height difference analysis model 200. If the deviation between the two is greater than the preset threshold, it indicates that the position of the switch rail with the preset width is inconsistent with the position of the switch rail with the preset height difference, which also means that the switch rail does not bear the required wheel load at the position of the preset width, and there is a potential disease hazard. At this time, it is necessary to output the disease identification result.

[0059] The above are only the preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various modifications and changes can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. An intelligent recognition method for measurable apparent diseases of turnouts, characterized in that, Including the following steps: Using a camera located directly above the stock rail and the switch rail in the track inspection vehicle to obtain top images of the stock rail and the switch rail; Using a camera located obliquely above the stock rail and the switch rail in the track inspection vehicle to obtain oblique images of the stock rail and the switch rail; Inputting the top image into a top width analysis model, the top width analysis model includes a switch rail recognition module and a top width recognition module, the switch rail recognition module is used to recognize the switch rail contour, and the top width recognition module determines the position of the switch rail with a preset width according to the switch rail contour; Inputting the oblique image into a height difference analysis model, the height difference analysis model includes an image transformation module and a height difference calculation module, the image transformation module transforms the oblique image into a lateral image in the horizontal direction, and the height difference calculation module is used to calculate the height difference between the switch rail and the stock rail according to the lateral image and determine the positions of the switch rail and the stock rail at a preset height difference; Comparing the output position data of the top width analysis model and the height difference analysis model, and when the two position data are inconsistent, outputting a disease recognition result.

2. The method according to claim 1, characterized in that, The switch rail recognition module includes a feature extraction unit and a contour prediction unit. The feature extraction unit adopts a U-Net structure. The output features of each layer of the encoder are transmitted to the decoder of the same layer and cross-layer transmitted to the decoders of different layers. The decoder outputs contour features after sequentially passing the input features through an edge-aware attention module, a transposed convolutional layer, a bilinear interpolation layer, and a refinement convolutional layer. The edge-aware attention module performs Sobel edge extraction on the features from the previous layer of the decoder to output edge features, performs content feature extraction on the features from the encoder to output content features, calculates attention weights according to the edge features and the content features, and outputs enhanced features according to the attention weights for feature weighting.

3. The method according to claim 2, wherein The contour features are input into the contour prediction unit. The contour prediction unit includes a parallel semantic segmentation path and an edge enhancement path. The contour features are respectively passed through the semantic segmentation path and the edge enhancement path and then fused to output the switch rail contour. The contour features output the probability of the switch rail contour after sequentially passing through staged multiple convolutional operations in the semantic segmentation path. The contour features output a distance field map after sequentially passing through depthwise separable convolutions, residual blocks, and depthwise separable convolutions in the edge enhancement path.

4. The method according to claim 1, characterized in that, The top width recognition module includes a contour parameterization unit, a width calculation unit, and a width matching unit. The contour parameterization unit is used to convert the switch rail contour into a sequence of switch rail contour points. The width calculation unit determines an array of widths of each point of the contour according to the sequence of switch rail contour points. The width matching unit determines the contour points with a preset width according to the width array.

5. The method according to claim 4, characterized in that, The contour parameterization unit establishes a coordinate system with the tip of the switch rail as the origin and the direction of the switch rail as the x-axis, determines the coordinates of each point on the switch rail contour, and establishes a sequence of switch rail contour points. The width calculation unit calculates the normal direction point by point along the sequence of switch rail contour points, projects a ray towards the side of the switch rail close to the stock rail, determines the intersection point of the ray and the side, calculates the Euclidean distance between each point in the sequence of switch rail contour points and the intersection point, and establishes a width array. The width matching unit searches for the contour point closest to the preset width in the width array.

6. The method according to claim 1, wherein The image transformation module includes an orbit feature detection unit, a coordinate system construction unit, a perspective transformation unit, and a remapping unit. The orbit feature detection unit detects the tip of the switch rail and the working surface of the stock rail, and determines the feature points of the switch rail and the stock rail. The coordinate system construction unit uses the tip of the switch rail as the origin of the coordinate system, uses the direction of the straight line fitted parallel to the working surface of the stock rail as the x-axis, and uses the direction perpendicular to the x-axis as the y-axis to establish a coordinate system. The perspective transformation unit constructs and solves a perspective transformation matrix by matching the feature points in the oblique image with the feature points in the reference image. The remapping unit compensates for the longitudinal offset of the image caused by vehicle vibration through the pitch angle deviation value measured by the camera in real time, corrects the perspective transformation matrix, and uses the corrected perspective transformation matrix to map any point in the oblique image to the corresponding position in the lateral image to realize the conversion of the image view and obtain the lateral image.

7. The method according to claim 1, wherein The height difference calculation module includes an edge extraction unit, a reference plane unit, a height difference calculation unit, and a positioning unit. The edge extraction unit detects the set of rail top edge points of the switch rail and the stock rail according to the lateral image. The reference plane unit models the rail tops of the switch rail and the stock rail according to the set of rail top edge points of the switch rail and the stock rail, and determines the sequence of edge points of the switch rail and the stock rail. The height difference calculation unit calculates the distance between the switch rail and the stock rail point by point in the vertical direction according to the sequence of edge points of the switch rail and the stock rail to form a height difference array. The positioning unit searches for the contour point closest to the preset height difference in the height difference array.

8. The method according to claim 1, wherein Determine the horizontal distance between the contour point output by the top width analysis model and the tip of the switch rail, determine the horizontal distance between the contour point output by the height difference analysis model and the tip of the switch rail, compare the two horizontal distances, and if the deviation is greater than the preset threshold, output the disease identification result.

Images