Point switch fault identification method and system based on data analysis

Abstract

The invention discloses a point switch fault identification method and system based on data analysis, and relates to the technical field of point switch fault identification, and the method comprises the following steps: before three-dimensional scanning, carrying out the space division of an overall curved surface of a point rail heel and point switch connecting region, and dividing the overall curved surface into a plurality of sub-regions with the same specification; and after the sub-region division is completed, carrying out one-time omnibearing initial scanning on each sub-region by adopting a preset reference point cloud space distribution density so as to obtain basic curved surface structure data in the sub-region. According to the method, the curved surface complexity of the sub-regions is automatically predicted through the convolutional neural network, high-complexity curved surface regions such as the root, the rotating interface and the orifice of the switch rail can be effectively identified, point cloud encryption is realized in the high-complexity regions, full expression of local geometric details is ensured, and the accuracy of point cloud encryption is improved. And the capability of identifying typical faults such as micro cracks, orifice deformation and eccentric abrasion is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of switch machine fault identification, and particularly relates to a method and system for switch machine fault identification based on data analysis. Background Art

[0002] A switch machine (also known as a turnout operating machine) is a core device in the railway signal system. It is mainly used to drive and lock the turnout part of the switch (commonly known as the "railway switch"), enabling a train to safely and smoothly transfer from one track to another. Its basic principle is to control the rotation of the switch rail through an electric, hydraulic, or electro-hydraulic hybrid drive device, ensuring that the switch rail is in close contact with or separated from the stock rail, completing the track conversion. At the same time, a locking mechanism is used to ensure that the switch is firmly locked after being positioned, preventing the switch rail from shifting due to external forces or train impacts and ensuring train operation safety. Modern switch machines are also equipped with electrical detection, remote control, and status feedback functions, which can be linked with the dispatching command center or the signal interlocking system to achieve automatic control and real-time monitoring, greatly improving the safety of railway transportation and the dispatching efficiency.

[0003] When identifying switch machine faults based on data analysis, image technology is often combined to visually monitor and diagnose the key components and working states of the switch machine. Common image technologies include video surveillance, infrared imaging, structured light or laser scanning, image recognition, and deep learning object detection technologies, which are mainly used to identify common faults in parts such as the switch rail, stock rail, switch machine housing, drive rod, and locking device. For example, high-definition video and image recognition technologies can detect appearance faults such as insufficient contact between the switch rail and the stock rail, switch rail jamming, foreign object intrusion, and damage to the switch machine box; infrared imaging can monitor abnormal temperature rises in components such as the drive motor and locking electromagnet of the switch machine to assist in identifying potential electrical or mechanical overload faults; three-dimensional scanning or structured light technology can be used to monitor structural defects such as wear, deformation, and misalignment of the switch and switch machine components.

[0004] Fault identification of the curved surface of the connection area between the heel of the switch rail and the switch rod through three-dimensional scanning technology mainly aims to accurately detect structural defects such as deformation, wear, misalignment, and microcracks in this key connection part. The heel of the switch rail is the direct connection point of the switch rail with the drive systems such as the switch rod and the locking device. Its curved surface structure is complex, with various special-shaped interfaces and transition surfaces. It is the area where the most force is concentrated during the long-term operation of the switch machine and is prone to problems such as fatigue bending, hole wear, hole mouth cracks, welding deformation, and local misalignment. Through high-precision three-dimensional scanning, the true geometric shape of this area can be completely obtained. Combining point cloud fitting with comparison with the reference model can effectively identify potential hazards such as minor deformations, abnormal gaps, and connection instability at the interface between the heel of the switch rail and the switch rod, providing a reliable geometric diagnosis basis for preventing faults such as switch rail movement failure, abnormal locking, or poor switch operation.

[0005] The prior art has the following deficiencies:

[0006] When the prior art performs three-dimensional scanning fault identification on the curved surface of the connection area between the heel of the switch rail and the switch rod, it usually uniformly assigns the same point cloud spatial distribution density to the curved surface of this area based on experience. However, there are significant differences in the complexity of the curved surface structure in the connection area between the heel of the switch rail and the switch rod. Especially in local areas such as the root of the switch rail, the rotating interface, and the orifice, its geometric structure mostly shows small-scale features and complex curved surface combinations. Since the uniform assignment of point cloud density fails to make differential adjustments for complex curved surfaces, it leads to insufficient point cloud distribution in high-complexity areas, making it difficult to effectively obtain local geometric details, and thus unable to accurately identify typical defects including orifice wear, cracks, micro-deformations, abnormal gaps, etc., easily causing missed detections of faults such as micro-cracks at the root of the switch rail, ovalization of the transmission orifice, and eccentric wear, and seriously threatening the safe and reliable operation of the switch machine when severe.

[0007] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure. Therefore, it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0008] The purpose of the present invention is to provide a method and system for switch machine fault identification based on data analysis. Through the automatic prediction of the complexity of sub-region curved surfaces by a convolutional neural network, it can effectively identify high-complexity curved surface areas such as the root of the switch rail, the rotating interface, and the orifice, and achieve point cloud encryption in high-complexity areas to ensure the full expression of local geometric details and improve the identification ability of typical faults such as micro-cracks, orifice deformation, and eccentric wear. At the same time, the reference point cloud density is maintained in low-complexity areas to avoid redundant sampling, reduce data redundancy and the subsequent processing burden, and solve the problems of insufficient point cloud in high-complexity areas and waste of scanning resources in low-complexity areas existing in the prior art, so as to solve the problems in the above background art.

[0009] To achieve the above purpose, the present invention provides the following technical solution: A method for switch machine fault identification based on data analysis, including the following steps:

[0010] Before three-dimensional scanning, spatially divide the overall curved surface of the connection area between the heel of the switch rail and the switch rod, and divide the overall curved surface into several sub-regions with the same specifications;

[0011] After completing the sub-region division, for each sub-region, perform a full-range initial scan using a preset reference point cloud spatial distribution density to obtain the basic curved surface structure data within the sub-region;

[0012] For the acquired point cloud data of each sub-region, key features that can reflect the complexity of the surface structure are extracted. After feature engineering processing of the extracted key features, the geometric change characteristics of the internal surface of each sub-region are mined to provide a quantitative description for the subsequent judgment of the complexity of the surface.

[0013] The processed features are input as feature vectors into the pre-trained convolutional neural network model, and the surface structure complexity of each sub-region is regressed and predicted through the neural network to achieve automatic and intelligent determination of the surface complexity of the sub-region;

[0014] According to the complexity results of each sub-region predicted by the convolutional neural network model, the spatial distribution density of the point cloud of the corresponding sub-region is dynamically adjusted. If a sub-region is identified as a highly complex structure surface, the spatial distribution density of the original reference point cloud is increased based on its complexity prediction result; if a sub-region is identified as a low-complexity structure surface, the original spatial distribution density of the reference point cloud is maintained;

[0015] According to the adjusted density allocation strategy, all sub-areas of the connection area between the heel of the entire point rail and the switch pole are re-scanned in all directions.

[0016] Preferably, a regular grid division method is used to spatially divide the entire curved surface of the connection area between the heel of the point rail and the switch rod, and the specific steps are as follows:

[0017] First, a three-dimensional spatial coordinate system is established for the connection area between the heel of the point rail and the switch rod. According to the design benchmark or preliminary point cloud data of the switch machine, the boundary range and coordinate axis direction of the area to be divided are determined to provide a spatial reference for subsequent mesh division.

[0018] Secondly, according to the set grid division parameters (such as grid size and unit side length), the whole area is divided regularly and evenly along the three directions of X, Y and Z, and the surface space is divided into several cubic or rectangular sub-areas of equal specifications. Each sub-area has a unique space number and position coordinates.

[0019] Finally, the intersection analysis between mesh and surface is used to screen out the valid mesh sub-regions that intersect or contain the actual surface of the heel of the point rail and the switch rod connection area, and the invalid blank meshes are eliminated, and only the valid sub-regions (the sub-regions involved in the subsequent feature extraction and point cloud density adjustment) are retained.

[0020] Preferably, for the obtained point cloud data of each sub-region, key features that can reflect the complexity of the surface structure are extracted. Among them, the extracted features include the degree of fluctuation of the normal vectors of adjacent points within the sub-region and the density of contour lines within the unit area of the sub-region. After performing feature engineering on the extracted key features, a normal vector fluctuation reference value and a contour line density reference value are generated respectively. The geometric change characteristics of the surface within each sub-region are quantified through the normal vector fluctuation reference value and the contour line density reference value, providing a basis for subsequent judgment of the surface complexity.

[0021] Preferably, the processed normal vector fluctuation reference value and contour line density reference value are used as feature vectors and input into a pre-trained convolutional neural network model. Through the neural network, a complexity score is generated, and based on the complexity score, a regression prediction is made on the surface structure complexity of each sub-region, realizing the automatic and intelligent determination of the sub-region surface complexity.

[0022] Preferably, according to the complexity results of each sub-region predicted by the convolutional neural network model, the point cloud spatial distribution density of the corresponding sub-region is dynamically adjusted. The specific steps are as follows:

[0023] According to the complexity prediction result of the sub-region by the convolutional neural network model, the complexity score of this sub-region is obtained, and a complexity score reference threshold is introduced as the criterion for distinguishing high-complexity surfaces and low-complexity surfaces. To achieve the continuity of density adjustment, a point cloud density adjustment factor is introduced to determine the multiple of the point cloud density increment in this region. The calculation formula of the point cloud density adjustment factor is:

[0024]

[0025] , where: λ i is the point cloud density adjustment factor of the sub-region, determining the multiple of the point cloud density increment in this sub-region, α y is the density enhancement coefficient, controlling the density increase amplitude of the high-complexity region, β y is the non-linear control parameter, used to control the curve growth of the point cloud density adjustment factor, Complexity score is the complexity score of the sub-region output by the convolutional neural network model, T c is the reference threshold of the complexity score. When Complexity score >T c , it is determined as a high-complexity structure surface. When Complexity score ≤T c , it is determined as a low-complexity structure surface.

[0026] Preferably, after obtaining the point cloud density adjustment factor λ of the sub-region iFinally, based on the original reference point cloud spatial distribution density, the actual point cloud spatial distribution density of each sub-area is adaptively adjusted. The adjustment formula is as follows:

[0027] ρ i =ρ0·λ i

[0028] , where: i is the spatial distribution density of the point cloud after adaptive adjustment in the sub-region, that is, the adjusted sampling density, and ρ0 is the spatial distribution density of the reference point cloud.

[0029] Preferably, the specific steps of generating a normal vector fluctuation reference value after performing feature engineering processing on the fluctuation degree of the normal vectors of adjacent points in the sub-region are as follows:

[0030] For each point cloud in the sub-region, it is recorded as point set P, P = {p i}, for each pair of adjacent points in Represents point p i For the points in the neighborhood of , calculate the direction cosine value of the normal vector of the neighboring point. The calculation expression is:

[0031] C ij =|n i ·n j |

[0032] , where: n i and n j Point p i and point p j The unit normal vector, C ij Represents point p i and point p j The cosine of the unit normal angle, C ij ∈[0, 1], the closer the value is to 1, the closer the directions are, the closer the value is to 0, the directions differ by 90 degrees, and the smaller the value is, the more dramatic the change of the normal vector is. |·| means that we only care about the size of the unit normal vector angle and not the direction.

[0033] Next, construct the direction cosine set C of all adjacent point pairs in the sub-region, C = {C ij};

[0034] The normal vector fluctuation reference value is calculated based on the direction cosine set C. The calculation expression is:

[0035]

[0036] , where: N vi is the normal vector fluctuation reference value, (1-C ij) represents the fluctuation intensity of the normal vectors of adjacent points. The larger the value, the stronger the fluctuation of the normal vectors of the adjacent point pair. (1 - C ij ) ∈ [0, 1], w ij is the weight of the adjacent edge of the adjacent point pair and can be taken where d ij is the distance between adjacent points, used to emphasize that the contribution of local adjacent point pairs with fluctuations is greater. П represents the product of all adjacent point pairs.

[0037] Preferably, the specific steps for generating the contour density reference value after performing feature engineering on the density of contour lines within a unit area are as follows:

[0038] First, based on the point cloud data within the sub-region, project the surface of the sub-region onto a contour surface (isosurface) along the direction perpendicular to the normal of the main surface (or a custom scanning direction) to generate a set of contour lines, denoted as where N c is the total number of contour lines generated within the sub-region. After the contour lines are generated, calculate the total curvature of the contour lines as an index of the geometric complexity of the contour lines within the sub-region. The calculation expression is:

[0039]

[0040] where: TCC represents the total curvature of the contour lines, C u is the u-th contour line, κ(s) is the curvature value at the arc length parameter s on the curve, and ds is the arc length element along the contour line;

[0041] To comprehensively reflect the quantity and bending characteristics of the contour lines, define the contour density reference value as follows:

[0042]

[0043] where: C di is the contour density reference value, A is the actual projected area of the sub-region on the surface, γ is the area scaling factor, taking 0 < γ < 1 (generally, 0.5 - 0.8 can be taken), used to avoid the overestimation phenomenon caused by the natural increase in the contour line length in sub-regions with larger areas.

[0044] A switch machine fault identification system based on data analysis includes a region division module, an initial scanning module, a feature extraction and feature engineering module, a complexity prediction module, a point cloud density adaptive adjustment module, and an adaptive scanning execution module;

[0045] The region division module, before three-dimensional scanning, spatially divides the overall surface of the connecting area between the heel of the switch rail and the switch rod into several sub-regions with the same specifications;

[0046] The initial scanning module, after completing the sub-region division, for each sub-region, performs a full-range initial scan using a preset benchmark point cloud spatial distribution density to obtain the basic surface structure data within the sub-region;

[0047] The feature extraction and feature engineering module extracts key features that can reflect the complexity of the surface structure from the point cloud data of each sub-region that has been obtained. After performing feature engineering processing on the extracted key features, it mines the geometric change characteristics of the internal surface of each sub-region, providing a quantitative description for subsequent judgment of the surface complexity;

[0048] The complexity prediction module inputs the processed features as feature vectors into a pre-trained convolutional neural network model, and through the neural network, performs regression prediction on the surface structure complexity of each sub-region, realizing the automated and intelligent determination of the surface complexity of the sub-region;

[0049] The point cloud density adaptive adjustment module dynamically adjusts the point cloud spatial distribution density of the corresponding sub-region according to the complexity results of each sub-region predicted by the convolutional neural network model. If a certain sub-region is identified as a high-complexity structure surface, based on its complexity prediction result, the original benchmark point cloud spatial distribution density is increased; if a certain sub-region is identified as a low-complexity structure surface, the original benchmark point cloud spatial distribution density is maintained;

[0050] The adaptive scanning execution module re-performs a full-range scan on all sub-regions of the entire connection area between the heel of the switch rail and the switch rod according to the adjusted density allocation strategy.

[0051] In the above technical solution, the technical effects and advantages provided by the present invention:

[0052] Through the automated prediction of the surface complexity of the sub-region by the convolutional neural network of the present invention, it can effectively identify high-complexity surface areas such as the root of the switch rail, the rotating interface, and the orifice, and achieve point cloud encryption in the high-complexity area to ensure the full expression of local geometric details and improve the recognition ability of typical faults such as micro-cracks, orifice deformation, and eccentric wear; at the same time, maintain the benchmark point cloud density in the low-complexity area, avoid redundant sampling, reduce data redundancy and the subsequent processing burden, solve the problems of insufficient point cloud in the high-complexity area and waste of scanning resources in the low-complexity area existing in the prior art, and provide an effective technical guarantee for the high-precision and intelligent three-dimensional scanning and fault recognition of the switch machine. Description of the Drawings

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0054] Figure 1 This is the method flow chart of a switch machine fault identification method based on data analysis according to the present invention.

[0055] Figure 2 This is the module schematic diagram of a switch machine fault identification system based on data analysis according to the present invention. Detailed implementation manners

[0056] Now, the exemplary embodiments will be described more fully with reference to the accompanying drawings. However, the exemplary embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these exemplary embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art.

[0057] The present invention provides a switch machine fault identification method based on data analysis as shown in Figure 1 the following, which includes the following steps:

[0058] Before three-dimensional scanning, spatially divide the overall curved surface of the connection area between the heel of the switch rail and the switch rod, and divide the overall curved surface into several sub-regions with the same specifications.

[0059] The specific division can be based on the CAD geometric model or the preliminary point cloud model of the switch machine, and the complex free-form surface is divided into several small-scale sub-regions by using the regular grid partitioning or voxelization method. Each sub-region serves as the basic unit for subsequent surface complexity analysis and point cloud density adjustment. The function of this step is to refine the scanning units of the target area, so that subsequent differential point cloud density control can be implemented according to the structural complexity of different sub-regions, avoiding a one-size-fits-all unified scanning strategy for the overall curved surface.

[0060] Use the regular grid partitioning method to spatially divide the overall curved surface of the connection area between the heel of the switch rail and the switch rod. The specific steps are as follows:

[0061] First, establish a three-dimensional space coordinate system for the connection area between the heel of the switch rail and the switch rod, and determine the boundary range and axis directions of the area to be divided according to the design reference of the switch machine or the preliminary point cloud data, providing a spatial reference for subsequent grid partitioning.

[0062] Secondly, according to the set grid partitioning parameters (such as grid size, unit side length), regularly and uniformly divide the overall area along the X, Y, and Z directions, and divide the curved surface space into several cube or rectangular sub-regions with the same specifications. Each sub-region has a unique spatial number and position coordinate.

[0063] Finally, the intersection analysis between mesh and surface is used to screen out the valid mesh sub-regions that intersect or contain the actual surface of the heel of the point rail and the switch rod connection area, and the invalid blank meshes are eliminated, and only the valid sub-regions (the sub-regions involved in the subsequent feature extraction and point cloud density adjustment) are retained.

[0064] Through the above steps, the spatial division of the complex curved surface of the connection area between the heel of the point rail and the switch rod can be achieved, providing a basis for subsequent regional feature extraction and adaptive point cloud distribution.

[0065] After completing the sub-region division, for each sub-region, a full-scale initial scan is performed using the preset reference point cloud spatial distribution density to obtain the basic surface structure data in the sub-region;

[0066] The density of the reference point cloud is usually determined based on the equipment performance and detection accuracy requirements, such as the initial point spacing is set to 1mm or finer. During the omnidirectional scanning process, the angle and position of the scanning device are adjusted to ensure that each sub-area is scanned from multiple perspectives without blind spots. The core function of this step is to quickly and comprehensively obtain the original point cloud data of the curved surface of the connection area between the heel of the point rail and the switch rod, providing the necessary three-dimensional spatial information for subsequent feature extraction and complexity analysis.

[0067] For the acquired point cloud data of each sub-region, key features that can reflect the complexity of the surface structure are extracted. After feature engineering processing of the extracted key features, the geometric change characteristics of the internal surface of each sub-region are mined to provide a quantitative description for the subsequent judgment of the complexity of the surface.

[0068] For the point cloud data of each sub-region that has been acquired, key features that can reflect the complexity of the surface structure are extracted from them. The extracted features include the fluctuation degree of the normal vectors of adjacent points in the sub-region and the density of the contour lines within the unit area of the sub-region. After feature engineering processing of the extracted key features, normal vector fluctuation reference values and contour line density reference values are generated respectively. The normal vector fluctuation reference values and contour line density reference values are used to quantify the geometric change characteristics of the internal surface of each sub-region, providing a basis for the subsequent judgment of the complexity of the surface.

[0069] For each sub-region, if the normal vectors of adjacent points in the sub-region fluctuate to a high degree, it usually indicates that the surface of the sub-region has dramatic geometric changes in space and has high structural complexity. Specifically, the point cloud normal vector reflects the local orientation of the surface at each point. When the normal vector directions of adjacent points change frequently and the differences are significant, it indicates that the curvature distribution of the sub-region is uneven, and there may be complex geometric structures such as sharp edges, concave and convex mutations, twists and bends, or multi-scale micro-features. Especially in mechanical connection parts, welding transition zones, or areas of concentrated wear and deformation, normal vector fluctuations often become a key indicator to reveal the high complexity of the surface.

[0070] The specific steps of generating the normal vector fluctuation reference value after feature engineering processing of the fluctuation degree of the normal vectors of adjacent points in the sub-region are as follows:

[0071] For each point cloud in the sub-region, it is recorded as point set P, P = {p i}, for each pair of adjacent points in Represents point p i For the points in the neighborhood of , calculate the direction cosine value of the normal vector of the neighboring point. The calculation expression is:

[0072] C ij =|n i ·n j |

[0073] , where: n i and n j Point p i and point p j The unit normal vector, C ij Represents point p i and point p j The cosine of the unit normal angle, C ij ∈[0, 1], the closer the value is to 1, the closer the direction is, the closer the value is to 0, the direction is 90 degrees different, and the smaller the value is, the more dramatic the change of the normal vector is. |·| means that only the size of the unit normal vector angle is concerned, not the direction (that is, the internal and external unit normal vectors are not distinguished);

[0074] Next, construct the direction cosine set C of all adjacent point pairs in the sub-region, C = {C ij};

[0075] The purpose of this step is to construct a basic data set of fluctuations of the local normal vector, which provides a basis for the subsequent direct quantification of the degree of fluctuation. At the same time, it avoids the traditional mean or variance processing method and retains the directional information of the fluctuation.

[0076] The normal vector fluctuation reference value is calculated based on the direction cosine set C. The calculation expression is:

[0077]

[0078] , where: N vi is the normal vector fluctuation reference value, (1-C ij ) represents the fluctuation intensity of the normal vector of the adjacent points. The larger the value, the stronger the fluctuation of the normal vector of the adjacent points. (1-C ij )∈[0,1],w ij is the weight of the adjacent edge of the adjacent point, which can be taken as where d ijis the distance between adjacent point pairs, which is used to emphasize that the contribution of local adjacent point pair fluctuations is greater, and П represents the product of all adjacent point pairs;

[0079] When there are a large number of normal vector mutations in the sub-region (i.e., C ij is small), then (1 - C ij ) tends to 1, resulting in the normal vector fluctuation reference value approaching 1; while when the normal vectors of most point pairs are similar, (1 - C ij ) is close to 0, and the normal vector fluctuation reference value rapidly decays close to 0. In this way, the normal vector fluctuation reference value can highlight the local strong fluctuation phenomenon rather than the global average change, and is particularly sensitive to surfaces with obvious edges, concavities and convexities, and small-scale complex features. The role of this step is to directly exponentiate and quantify the normal vector changes with significant local fluctuations, which is used for the accurate identification of highly complex surfaces.

[0080] From the normal vector fluctuation reference value, the larger the performance value of the normal vector fluctuation reference value generated by performing feature engineering on the fluctuation degree of the normal vectors of adjacent points in the sub-region, the more complex the geometric change characteristics of the surface in the sub-region. The reason is that: the normal vector fluctuation reference value is essentially the cumulative quantification of the degree of direction change between the normal vectors of adjacent points. When there are a large number of curvature mutations, edges, corners, concavities and convexities, or complex local details on the surface in the sub-region, the direction difference of the normal vectors of adjacent points will increase significantly, manifested as the angle between the normal vectors of adjacent points tending to increase, which in turn causes the direction cosine value of the normal vector to decrease, resulting in an increase in the value of (1 - C ij ) used in the exponential calculation, and finally increasing the overall normal vector fluctuation reference value. Therefore, the level of the normal vector fluctuation reference value is positively correlated with the severity of the local normal vector change of the surface. The larger the value, the more severe the undulation, the richer the details, and the more complex the structure of the surface in the sub-region; conversely, if the surface is smooth and the change is gentle, the difference between the normal vectors of adjacent points is small, and the normal vector fluctuation reference value decreases accordingly, reflecting that the surface is relatively simple and the structural change is not obvious.

[0081] The density of contour lines in the unit region is closely related to the geometric change characteristics of the surface. If the contour lines are densely distributed, it usually indicates that there are significant height fluctuations, concavity and convexity changes, or curvature mutations in the surface in the sub-region, that is, the geometric change characteristics of the surface are relatively complex. Contour lines are essentially the isosurfaces of the surface in a certain direction. When the surface undulates gently, the distance between the contour lines is large. On the contrary, when there are sharp turns, grooves, protrusions, steep curvature changes, or dense distributions of small-scale structures locally on the surface, the contour lines will show a dense, distorted or even overlapping state. Therefore, the contour line density can be used as an important geometric feature to measure the surface complexity, and can effectively reflect whether there are complex surface areas with severe deformation or rich details in the connection area between the heel of the switch rail and the switch rod.

[0082] The specific steps for generating the contour density reference value after feature engineering on the density of contour lines within a unit area are as follows:

[0083] First, based on the point cloud data within the sub-region, project the sub-region surface onto an equipotential surface (isosurface) along the direction perpendicular to the normal of the main surface (or a custom scanning direction) to generate a set of contour lines, denoted as where N c is the total number of contour lines generated within the sub-region. After the contour lines are generated, calculate the total curvature of the contour lines as an index of the geometric complexity of the contour lines within the sub-region. The calculation expression is:

[0084]

[0085] , where: TCC represents the total curvature of the contour lines, C u is the u-th contour line, κ(s) is the curvature value at the arc length parameter s on the curve, and ds is the arc length element along the contour line;

[0086] This step directly measures based on the shape characteristics of the contour lines, using the curvature change of the contour lines in space. If the surface has drastic undulations and rich concavities and convexities, the contour lines are dense and have a high curvature, and the total curvature of the contour lines naturally increases, reflecting the complexity of the geometric changes of the sub-region surface.

[0087] To comprehensively reflect the quantity and bending characteristics of the contour lines, define the contour density reference value as follows:

[0088]

[0089] , where: C di is the contour density reference value, A is the actual projected area of the sub-region on the surface, γ is the area scaling factor, taking 0 < γ < 1 (generally, 0.5 - 0.8 can be taken) to avoid the phenomenon of false high values brought by the natural increase in the contour line length in sub-regions with a large area;

[0090] The contour density reference value not only considers the morphological complexity (curvature bending) of the contour lines, but also realizes area normalization through A γ to avoid confusion in the index between large-area simple surfaces and small-area highly complex surfaces. A high contour density reference value indicates that there are many and curved contour lines per unit area, strongly indicating that the surface is a highly complex structure; a low contour density reference value indicates that the contour lines are sparse and have a gentle shape, indicating that the surface is relatively simple.

[0091] From the contour density reference value, it can be seen that the larger the performance value of the contour density reference value generated after feature engineering processing of the density of contours within the unit area of the sub-region, the more complex the geometric variation characteristics of the surface within the sub-region. Conversely, it indicates that the geometric variation characteristics of the surface within the sub-region are less complex. The fundamental reason is that the contour density reference value synthesizes the number of contours within the unit area and the degree of bending of the contours themselves. If the geometric variation of the surface within the sub-region is drastic, it is often accompanied by features such as undulation, concavity and convexity, broken lines, and mutations, manifested as the contours being closely distributed in space and the curve bending degree increasing significantly, resulting in a significant increase in the value of the contour density reference value after integral accumulation; while when the surface within the sub-region is relatively flat or the variation is slow, the contours are sparsely distributed and the curvature changes gently, leading to a relatively small value of the contour density reference value. Therefore, the contour density reference value can objectively reflect the intensity of the local geometric variation of the surface. The larger the performance value, the more complex the surface features and the finer the structure.

[0092] Input the processed features as feature vectors into a pre-trained convolutional neural network (CNN) model, and through the neural network, perform regression prediction on the complexity of the surface structure of each sub-region to achieve automated and intelligent determination of the complexity of the sub-region surface;

[0093] Input the processed normal vector fluctuation reference value and contour density reference value as feature vectors into a pre-trained convolutional neural network model, generate a complexity score through the neural network, and perform regression prediction on the complexity of the surface structure of each sub-region based on the complexity score to achieve automated and intelligent determination of the complexity of the sub-region surface.

[0094] The pre-trained convolutional neural network model refers to the convolutional neural network (CNN) adopted in the complexity determination method of the present invention. Instead of being trained temporarily after the sub-region feature extraction is completed, it is an in-depth neural network model that has already possessed strong surface complexity feature recognition and regression prediction capabilities through a systematic training process based on a large number of surface point cloud data of the connection area between the heel of the switch rail and the switch rod with known complexity annotations in advance. The design of this convolutional neural network usually revolves around the surface complexity determination task. Structurally, it includes basic modules such as an input layer, several groups of convolutional layers, pooling layers, feature fusion layers, fully connected layers, and output layers, and can perform feature extraction, feature combination, and non-linear mapping learning for the input feature vectors (i.e., the normal vector fluctuation reference value and the contour density reference value). The core role of the convolutional neural network is to establish an implicit relationship model between the feature index (input) and the surface complexity (output) by utilizing its capabilities in processing spatial correlation, graphic features, and local feature combination. In this technical solution, the CNN can learn the inherent non-linear mapping law between the normal vector fluctuation and the contour density and the actual surface complexity, and can effectively handle the situation where the feature distribution is complex and the feature space highly overlaps, and complete the regression prediction of the complexity. The meaning of pre-training refers to that this neural network model has completed repeated training, optimization, and verification based on historical data before the deployment or implementation of the present invention, and has stable prediction capabilities, and can be directly deployed in the complexity prediction process to achieve fast inference from feature input to complexity prediction output.

[0095] Furthermore, a pre-trained convolutional neural network model is usually built on a large number of point cloud datasets of the connection area between the heel of the switch rail and the switch rod collected in reality or generated by simulation. These training datasets not only contain rich real surface samples with high complexity, low complexity, multi-scale, and multi-morphology, but also, for each surface area, are manually or semi-automatically labeled by professional engineers or according to three-dimensional surface morphology standards, and corresponding surface complexity labels are assigned (for example: complexity grading, complexity score). In the process of building the model, first, each surface sample is divided into several sub-regions consistent with the actual detection process, and the normal vector fluctuation reference value and the contour density reference value are calculated for each sub-region to form a feature input set; subsequently, through the end-to-end learning of the convolutional neural network, the model gradually fits the corresponding relationship between the input features and the target complexity labels. During the training process, the mean square error (MSE) or the mean absolute error (MAE) is usually used as the loss function, combined with optimizers such as Adam and SGD, to continuously optimize the network parameters until the model reaches an ideal prediction accuracy and generalization ability on the validation set. The neural network model after training is regarded as a pre-trained model, which can be directly called in the actual fault detection and point cloud processing process to automatically determine the complexity of unknown sub-regions, without the need for re-training or parameter adjustment in actual applications, thus significantly improving the response speed and determination accuracy of the detection system.

[0096] The pre-trained convolutional neural network model not only plays the core role of an intelligent surface complexity determiner, but also has many technical advantages. First of all, using pre-training can effectively solve the non-linear and multi-scale coupling problems between features and outputs in surface complexity determination. Taking the normal vector fluctuation reference value and the contour density reference value as examples, the relationship between them and surface complexity is not a simple linear relationship, but often shows various complex non-linear and multi-feature combination effects. For example, there are some surfaces at the root of the switch rail. Even if the normal vector fluctuation reference value is low, due to the abnormal increase in the contour density reference value, they still show the characteristics of complex surfaces. Conventional linear models or threshold determination are difficult to accurately model such implicit relationships. The convolutional neural network relies on the mechanism of convolutional kernel to extract spatial features + multi-layer non-linear activation + feature fusion, and can automatically learn and extract the deep relationship between them and complexity, effectively solving the problem of misjudgment or missed judgment caused by the complex distribution of surface features. Secondly, pre-training helps to decouple the complex model training process from the actual detection process, so that the detection process only needs to perform forward inference to quickly output the complexity prediction result of the sub-region, which has the advantages of high efficiency, stability, and real-time, and greatly improves the closed-loop efficiency of point cloud acquisition - complexity recognition - adaptive scanning.

[0097] The convolutional neural network model is not specifically limited herein, as long as it can implement the comprehensive analysis of the normal vector fluctuation reference value N vi and the contour density reference value C di to generate the complex score Complexity score is acceptable. To implement the technical solution of the present invention, the present invention provides a specific implementation; the complex score Complexity score is generated by the following expression: Complexity score = h w *N vi + h l *C di , where h w , h l are respectively the preset proportionality coefficients of the normal vector fluctuation reference value N vi and the contour density reference value C di , and both h w , h l are greater than 0.

[0098] The preset proportionality coefficients (i.e., h w and h l ) refer to the coefficients that assign weights to different input feature indicators during the calculation process of generating the complex score. These coefficients are constant values set in advance according to experience, data statistics, or optimization results during the model training or system design stage, and are used to balance the influence degrees of the normal vector fluctuation reference value and the contour density reference value on the final complex score. Specifically, the role of the preset proportionality coefficients is to control the contribution weights of the two indicators to the scoring result. For example, if the normal vector fluctuation reference value can better reflect the surface complexity in a certain scenario, h w can be set larger to enhance its influence; conversely, if the contour density is a more critical judgment basis, the weight of h l can be appropriately increased. Since both of these coefficients are "greater than 0" and the values are adjustable, the system can select an appropriate combination of proportionality coefficients through experimental optimization, making the generated complex score more in line with the actual structural complexity distribution and improving the accuracy and adaptability of intelligent discrimination.

[0099] From the complex score, it can be seen that the larger the value of the normal vector fluctuation reference value generated after feature engineering processing of the fluctuation degree of the normal vectors of adjacent points in the sub-region, and the larger the value of the contour density reference value generated after feature engineering processing of the contour density in the unit region of the sub-region, that is, the larger the value of the complex score generated when the surface structure complexity of each sub-region is regressively predicted by the pre-trained convolutional neural network model, it indicates that the surface geometric change characteristics in this sub-region are more complex; on the contrary, it indicates that the surface geometric change characteristics in this sub-region are less complex.

[0100] According to the complexity results of each sub-region predicted by the convolutional neural network (CNN) model, dynamically adjust the point cloud spatial distribution density of the corresponding sub-region. If a certain sub-region is identified as a high-complexity structural surface, based on its complexity prediction result, improve the original reference point cloud spatial distribution density to enhance the capture ability of surface details; if a certain sub-region is identified as a low-complexity structural surface, while maintaining the original reference point cloud spatial distribution density to ensure detection accuracy, avoid global over-encryption, optimize the scanning resource allocation, achieve point cloud encryption in high-complexity areas, keep the reference density in low-complexity areas, and balance accuracy and efficiency;

[0101] According to the complexity results of each sub-region predicted by the convolutional neural network model, dynamically adjust the point cloud spatial distribution density of the corresponding sub-region. The specific steps are as follows:

[0102] According to the complexity prediction result of the convolutional neural network model for the sub-region, obtain the complexity score of this sub-region, and introduce a complexity score reference threshold as the criterion for distinguishing high-complexity surfaces and low-complexity surfaces. To achieve the continuity of density adjustment, introduce a point cloud density adjustment factor to determine the multiple of the point cloud density increment in this region. The calculation formula of the point cloud density adjustment factor is:

[0103]

[0104] , where: λ i is the point cloud density adjustment factor of the sub-region, which determines the multiple of the point cloud density increment in this sub-region, α y is the density enhancement coefficient, which controls the density increase amplitude in high-complexity regions, β y is the non-linear control parameter, which is used to control the curve growth of the point cloud density adjustment factor, Complexity score is the complexity score of the sub-region output by the convolutional neural network model, T c is the reference threshold of the complexity score. When Complexity score >T c , it is determined as a high-complexity structural surface. When Complexity score ≤T c , it is determined as a low-complexity structural surface;

[0105] The function of this step is: by designing a smooth and controllable point cloud density adjustment factor, realize the non-linear mapping between complexity and point cloud density, so that the higher the complexity, the more significant the point cloud encryption, while the low-complexity region remains unchanged, avoiding the sudden encryption problem caused by hard thresholds, and enhancing the flexibility and intelligence of density adjustment.

[0106] After obtaining the point cloud density adjustment factor λ of the sub-region iAfter that, based on the original spatial distribution density of the reference point cloud, the actual spatial distribution density of each sub-region is adaptively adjusted, and the adjustment formula is as follows:

[0107] ρ i = ρ0·λ i

[0108] , where: ρ i is the spatial distribution density of the point cloud after adaptive adjustment of the sub-region, that is, the adjusted sampling density, and ρ0 is the spatial distribution density of the reference point cloud;

[0109] The function of this step is to directly increase the point cloud density in the high-complexity region according to the sub-region complexity score and its relative relationship with the reference threshold, that is, multiply by λ i > 1 encryption coefficient on the basis of the original reference density; for low-complexity regions, when Complexity score ≤T c , λ i = 1, so that its reference density remains unchanged, so as to realize point cloud encryption in high-complexity regions and maintain the reference density in low-complexity regions, effectively balancing the detail capture ability and the scanning resource utilization efficiency.

[0110] According to the adjusted density allocation strategy, all sub-regions of the entire connection area between the heel of the switch rail and the switch rod are scanned again in all directions;

[0111] At this stage, the high-complexity region will be sampled with a higher point cloud density, and the low-complexity region will maintain the reference density, so as to obtain a set of point cloud data with variable accuracy, local encryption, and global efficiency. The finally obtained point cloud data can completely and accurately reflect the tiny geometric features of complex regions such as the root of the switch rail, the rotating interface, and the orifice, effectively improving the recognition accuracy of typical faults such as cracks, wear, orifice deformation, and eccentricity of the root of the switch rail. Complete the adaptive scanning and modeling process for complexity, and generate a high-quality point cloud data set that meets the subsequent fault recognition requirements.

[0112] Through the above switch machine fault identification method based on data analysis, it is possible to achieve precise perception of the complexity difference of the curved surface in the connection area between the heel of the switch rail and the switch rod and adaptive point cloud density regulation, significantly improving the quality of three-dimensional scan data and the reliability of fault identification. Specifically, through the automatic prediction of the complexity of the sub-region curved surface by the convolutional neural network, this method can effectively identify high-complexity curved surface regions such as the root of the switch rail, the rotating interface, and the orifice, and encrypt the point cloud in the high-complexity region to ensure that local geometric details are fully expressed, improving the identification ability for typical faults such as micro-cracks, orifice deformation, and eccentric wear. At the same time, the reference point cloud density is maintained in the low-complexity region to avoid redundant sampling, reducing data redundancy and the subsequent processing burden. This solution takes into account both detection accuracy and system efficiency, solves the problems of insufficient point cloud in high-complexity regions and waste of scanning resources in low-complexity regions existing in the prior art, and provides an effective technical guarantee for high-precision and intelligent three-dimensional scanning and fault identification of switch machines.

[0113] The present invention provides a Figure 2 switch machine fault identification system based on data analysis as shown, including a region division module, an initial scan module, a feature extraction and feature engineering module, a complexity prediction module, a point cloud density adaptive adjustment module, and an adaptive scan execution module;

[0114] The region division module, before three-dimensional scanning, spatially divides the overall curved surface of the connection area between the heel of the switch rail and the switch rod into several sub-regions of the same specification;

[0115] The initial scan module, after completing the sub-region division, for each sub-region, performs a full-range initial scan using a preset reference point cloud spatial distribution density to obtain the basic curved surface structure data within the sub-region;

[0116] The feature extraction and feature engineering module extracts key features that can reflect the complexity of the curved surface structure from the obtained point cloud data of each sub-region, and after performing feature engineering processing on the extracted key features, mines the geometric change characteristics of the internal curved surface of each sub-region to provide a quantitative description for subsequent judgment of the complexity of the curved surface;

[0117] The complexity prediction module inputs the processed features as feature vectors into a pre-trained convolutional neural network model, and through the neural network, performs regression prediction on the complexity of the curved surface structure of each sub-region to achieve automatic and intelligent determination of the complexity of the sub-region curved surface;

[0118] The point cloud density adaptive adjustment module dynamically adjusts the point cloud spatial distribution density of the corresponding sub-regions according to the complexity results of each sub-region predicted by the convolutional neural network model. If a certain sub-region is identified as a high-complexity structure surface, the original reference point cloud spatial distribution density is increased based on its complexity prediction result; if a certain sub-region is identified as a low-complexity structure surface, the original reference point cloud spatial distribution density is maintained.

[0119] The adaptive scanning execution module re-scans all sub-regions of the entire connecting area between the heel of the switch rail and the switch rod in all directions according to the adjusted density allocation strategy.

[0120] A method for identifying switch machine faults based on data analysis provided by an embodiment of the present invention is implemented through the above-mentioned system for identifying switch machine faults based on data analysis. The specific methods and processes of the system for identifying switch machine faults based on data analysis are detailed in the embodiments of the above-mentioned method for identifying switch machine faults based on data analysis, and will not be elaborated here.

[0121] The above formulas are all dimensionless and take their numerical calculations. The formulas are obtained by collecting a large amount of data for software simulation to obtain a formula that is closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0122] As described above, only the specific implementation manners of the present application are provided, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present application, and all should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0123] Only some exemplary embodiments of the present invention have been described by way of illustration above. Undoubtedly, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

Claims

1. A method for identifying switch machine faults based on data analysis, characterized in that, The steps include the following: Before 3D scanning, spatially divide the overall surface of the connection area between the heel of the switch rail and the switch rod, and divide the overall surface into several sub-regions with the same specifications; After completing the sub-region division, for each sub-region, conduct a full-round initial scan using the preset spatial distribution density of the reference point cloud to obtain the surface structure data within the sub-region; From the obtained point cloud data of each sub-region, extract the key features that can reflect the complexity of the surface structure. After performing feature engineering processing on the extracted key features, explore the geometric change characteristics of the internal surface of each sub-region to provide a quantitative description for subsequent judgment of the surface complexity; Input the processed features as feature vectors into a pre-trained convolutional neural network model, and use the neural network to perform regression prediction on the surface structure complexity of each sub-region to achieve automated and intelligent determination of the sub-region surface complexity; According to the complexity results of each sub-region predicted by the convolutional neural network model, dynamically adjust the spatial distribution density of the point cloud in the corresponding sub-region. If a certain sub-region is identified as a high-complexity structure surface, based on its complexity prediction result, increase the original spatial distribution density of the reference point cloud; If a certain sub-region is identified as a low-complexity structure surface, maintain the original spatial distribution density of the reference point cloud; According to the adjusted density allocation strategy, re-conduct a full-round scan of all sub-regions in the entire connection area between the heel of the switch rail and the switch rod; 2. The method for identifying a switch machine fault based on data analysis according to claim 1, wherein Use the regular grid division method to spatially divide the overall surface of the connection area between the heel of the switch rail and the switch rod. The specific steps are as follows: Establish a three-dimensional space coordinate system for the connection area between the heel of the switch rail and the switch rod. According to the preliminary point cloud data, determine the boundary range and axis directions of the area to be divided to provide a spatial reference for subsequent grid division; According to the set grid division parameters, conduct regular and uniform division of the overall area along the X, Y, and Z directions, divide the surface space into several cubes with the same specifications, and each sub-region has a unique spatial number and position coordinate; Use the grid and surface intersection analysis to screen out the effective grid sub-regions that intersect or contain the actual surface of the heel of the switch rail and the switch rod connection area, eliminate the invalid blank grids, and only retain the effective sub-regions; 3. A method for identifying switch machine faults based on data analysis according to claim 1, characterized in that, From the obtained point cloud data of each sub-region, extract the key features that can reflect the complexity of the surface structure. Among them, the extracted features include the fluctuation degree of the normal vectors of adjacent points within the sub-region and the density of contour lines within the unit area of the sub-region. After performing feature engineering processing on the extracted key features, generate the normal vector fluctuation reference value and the contour line density reference value respectively, and quantify the geometric change characteristics of the internal surface of each sub-region through the normal vector fluctuation reference value and the contour line density reference value; 4. A method for identifying switch machine faults based on data analysis according to claim 3, characterized in that Input the processed normal vector fluctuation reference value and contour line density reference value as feature vectors into a pre-trained convolutional neural network model, generate a complexity score through the neural network, and perform regression prediction on the surface structure complexity of each sub-region based on the complexity score to achieve automated and intelligent determination of the sub-region surface complexity; 5. A method for identifying turnout machine faults based on data analysis according to claim 4, characterized in that According to the complexity results of each sub-region predicted by the convolutional neural network model, the spatial distribution density of the point cloud of the corresponding sub-region is dynamically adjusted. The specific steps are as follows: According to the complexity prediction results of the sub-region by the convolutional neural network model, the complexity score of the sub-region is obtained, and the complexity score reference threshold is introduced as the criterion for distinguishing high-complexity surfaces from low-complexity surfaces. The point cloud density adjustment factor is introduced to determine the point cloud density increment multiple of the area. The calculation formula of the point cloud density adjustment factor is: ， Among them: λ i is the point cloud density adjustment factor of the sub-region, α y is the density enhancement coefficient, which controls the density increase amplitude in the high-complexity region, β y is the non-linear control parameter, which is used to control the curve growth of the point cloud density adjustment factor, Complexity score is the complexity score of the sub-region output by the convolutional neural network model, T c is the reference threshold of the complexity score. When Complexity score >T c it is determined as a high-complexity structure surface. When Complexity score ≤T c it is determined as a low-complexity structure surface.

6. The method for identifying a switch machine fault based on data analysis according to claim 5, wherein After obtaining the point cloud density adjustment factor λ of the sub-region i Based on the original reference point cloud spatial distribution density, the actual point cloud spatial distribution density of each sub-region is adaptively adjusted, and the adjustment formula is as follows: ρ i = ρ0·λ i , Where: ρ i is the point cloud spatial distribution density after sub-region adaptive adjustment, that is, the adjusted sampling density, and ρ0 is the reference point cloud spatial distribution density.

7. A method for identifying switch machine faults based on data analysis according to claim 3, characterized in that, The specific steps of generating the normal vector fluctuation reference value after feature engineering processing of the fluctuation degree of the normal vectors of adjacent points in the sub-region are as follows: For the point cloud within each sub-region, denoted as point set P, P = {p i}, for each pair of adjacent points where represents the points within the neighborhood of point p i , calculate the direction cosine value of the normal vectors of this pair of adjacent points. The calculation expression is as follows: C ij = |n i ·n j |, where: n i and n j are the unit normal vectors of point p i and point p j respectively, and C ij represents the cosine value of the included angle between the unit normal vectors of point p i and point p j , and C ij ∈[0, 1]; Construct the set C of direction cosines for all adjacent point pairs within this sub-region, C = {C ij}; The normal vector fluctuation reference value is calculated based on the direction cosine set C. The calculation expression is: ， Where: N vi is the reference value of the normal vector fluctuation, and (1 - C ij ) represents the fluctuation intensity of the normal vectors of adjacent points. (1 - C ij ) ∈ [0, 1], w ij is the weight of the adjacent edge of this adjacent point pair, and П represents the product of all adjacent point pairs.

8. A method for identifying switch machine faults based on data analysis according to claim 3, characterized in that The specific steps for generating the density reference value of contour lines after feature engineering processing of the density of contour lines in the unit area are as follows: Based on the point cloud data within the sub-region, perform an isometric surface projection on the sub-region surface along the direction perpendicular to the normal of the main surface to generate a set of contour lines, denoted as where N c is the total number of contour lines generated within the sub-region. After the contour lines are generated, calculate the total curvature of the contour lines as an index of the geometric complexity of the contour lines within the sub-region. The calculation expression is: ， Where: TCC represents the total curvature of the contour line, C u is the u-th contour line, k(s) is the curvature value at the arc length parameter s on the curve, and ds is the arc length element along the contour line; The density reference value of contour lines comprehensively reflects the number and curvature characteristics of contour lines. The calculation formula of the density reference value of contour lines is as follows: ， Where: C di is the reference value of dense contour lines, A is the actual projected area of the sub-region on the surface, and γ is the area scaling factor, where 0 < γ < 1.

9. A switch machine fault identification system based on data analysis, which is used to implement the switch machine fault identification method based on data analysis described in any one of the above claims 1-8, and is characterized in that, It includes a region division module, an initial scanning module, a feature extraction and feature engineering module, a complexity prediction module, a point cloud density adaptive adjustment module, and an adaptive scanning execution module; The area division module divides the whole curved surface of the connection area between the heel of the point rail and the switch rod into several sub-areas with the same specifications before the three-dimensional scanning; The initial scanning module, after completing the sub-region division, performs an all-round initial scan for each sub-region using the preset reference point cloud spatial distribution density to obtain the surface structure data within the sub-region; The feature extraction and feature engineering module extracts key features that can reflect the complexity of the surface structure from the acquired point cloud data of each sub-region. After feature engineering processing of the extracted key features, the geometric change characteristics of the internal surface of each sub-region are mined to provide a quantitative description for the subsequent judgment of the complexity of the surface. The complexity prediction module inputs the processed features as feature vectors into the pre-trained convolutional neural network model, and uses the neural network to perform regression prediction on the surface structure complexity of each sub-region, thus realizing the automatic and intelligent determination of the surface complexity of the sub-region; The point cloud density adaptive adjustment module dynamically adjusts the point cloud spatial distribution density of the corresponding sub-region according to the complexity results of each sub-region predicted by the convolutional neural network model. If a sub-region is identified as a highly complex structure surface, the original reference point cloud spatial distribution density is improved based on its complexity prediction result. If a sub-region is identified as a low-complexity structure surface, the original reference point cloud spatial distribution density is maintained; The adaptive scanning execution module re-performs a full-scale scan of all sub-areas of the connection area between the heel of the point rail and the switch pole according to the adjusted density allocation strategy.

Images