AI-based methods and systems for identifying railway line environments

Abstract

This application provides an artificial intelligence-based method and system for railway line environment identification. Specifically, it simultaneously collects laser point clouds of the ballast area, track inspection vehicle status parameters, and millimeter-wave radar data, fusing them to construct an enhanced point cloud with physical attributes. Using line-scanning lasers and millimeter-wave radar on the bottom of the track inspection vehicle, spiral and penetration scans are performed to generate a three-dimensional data matrix that integrates surface morphology and internal dielectric characteristics. A neural network analyzes this matrix to determine the curvature and gap distribution of the ballast contact surface, and enhances the weight of loose areas through spatiotemporal attention. An adversarial training-based generative model completes the occluded areas, outputting a complete virtual point cloud. Finally, the virtual point cloud and related features are quantitatively analyzed, and the ballast density and the coordinate range of the compacted area are output as key identification results. This application achieves accurate, dynamic, and non-contact quantitative assessment of railway ballast density and compacted areas through multi-source data fusion, deep learning, and point cloud generation technologies.

Description

Technical Field

[0001] This application relates to the field of intelligent railway detection technology, and in particular to a method and system for identifying railway line environment based on artificial intelligence. Background Technology

[0002] The density of railway ballast, i.e., the crushed stone layer under the sleepers, and whether it is loose or compacted, with the crushed stone bonded and hardened by mud, is a key factor directly affecting the geometric stability of the track and traffic safety. Traditional manual inspection is inefficient, subjective, incomplete in coverage, and poses safety risks. There is an urgent need for an automated, non-contact, high-precision intelligent detection technology that can penetrate the surface to achieve rapid, quantitative identification and location of ballast density and compacted areas along railway lines, replacing manual inspection and improving maintenance efficiency and safety.

[0003] Currently, a representative technical solution is point cloud acquisition and machine learning recognition based on vehicle-mounted 3D laser scanning. This solution utilizes a laser scanner installed on a track inspection vehicle to acquire high-density point cloud data of the track and ballast surfaces at high speed. The ballast area is extracted through point cloud processing algorithms, and further analysis is performed using machine learning models, surface roughness, or geometric features to attempt to distinguish between "normal," "loose," or "compacted" areas.

[0004] This scheme relies heavily on surface morphology information, which has significant limitations: First, it cannot perceive the internal structure of the ballast: lasers have difficulty penetrating gravel, making it impossible to obtain key internal physical properties such as the distribution of ballast particle gaps, changes in internal density, and the bonding of the slab. The judgment of the "loose" state is mainly based on the surface unevenness, which is easily affected by surface dirt and water accumulation, resulting in insufficient accuracy in identifying early or deep slab / loose conditions. Second, the occlusion problem is serious: surface gravel can obscure information from the underlying layers, leading to a large amount of missing point cloud data. Existing methods are unable to effectively reconstruct the true ballast accumulation state in the obscured areas. Third, the quantification capability is weak: it can usually only provide rough classification results, lacking precise numerical output of ballast density and precise definition of the spatial coordinate range of the boundaries and depths of slab areas, making it difficult to meet the needs of refined maintenance decisions. Summary of the Invention

[0005] This application provides a railway line environment identification method and system based on artificial intelligence, which solves the problems in the prior art that rely on surface topography data and cannot penetrate to perceive the internal structure of ballast, lack of obstructed areas, and weak quantification ability.

[0006] Firstly, this application provides a method for identifying railway line environments based on artificial intelligence, including:

[0007] Acquire laser point cloud data of the track ballast area, and simultaneously collect the operating status parameters of the track inspection vehicle and the echo data of the millimeter-wave radar;

[0008] Based on the operating state parameters, dynamic distortion compensation is performed on the laser point cloud data, and combined with the dielectric constant characteristics in the echo data, an enhanced point cloud model containing physical properties is constructed.

[0009] The surface of the ballast area is spirally scanned by the line-scanning laser module at the bottom of the track inspection vehicle, and the millimeter-wave radar is simultaneously triggered to perform a penetrating scan to generate a three-dimensional spatial data matrix that integrates the surface morphology and internal dielectric characteristics.

[0010] A neural network structure is used to analyze the curvature variation data of the contact surface of ballast particles in the three-dimensional spatial data matrix and identify the spatial distribution characteristics of the gaps between ballast particles. At the same time, the weight of loose areas is dynamically enhanced through a spatiotemporal attention mechanism.

[0011] A multi-scale point cloud generation model is constructed based on an adversarial training mechanism to geometrically complete the missing occluded areas in the spatial distribution features in order to output a virtual point cloud with a complete ballast stacking state.

[0012] The virtual point cloud is quantitatively analyzed. Combining the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature change law data, and the weight of loose regions, the ballast compaction value and the spatial coordinate range of the compacted area are output, which serve as key identification results of the ballast structure environmental status along the railway line.

[0013] Optionally, the multi-scale point cloud generation model based on the adversarial training mechanism, which geometrically completes the missing occluded regions in the spatial distribution features to output a virtual point cloud representing the complete ballast stacking state, includes:

[0014] A multi-scale point cloud generation model is constructed based on an adversarial training mechanism, which includes a generator unit and a discriminator unit. The generator unit is equipped with a coarse-grained generation channel and a fine-grained generation channel, and the discriminator unit is equipped with a geometric continuity verification module.

[0015] The occlusion region in the spatial distribution features is input into the generator unit. The basic geometry of the occlusion region is predicted through the coarse-grained generation channel, and the surface details of the ballast particles are added through the fine-grained generation channel to generate supplementary point cloud data of the occlusion region.

[0016] The supplementary point cloud data and the original spatial distribution features are input into the discriminator unit. The geometric continuity verification module detects the curvature connection state between the supplementary geometry and the adjacent region, and verifies the physical rationality of the particle gap transition, and outputs the discrimination result.

[0017] When the discrimination result fails the verification, the generator unit parameters are iteratively optimized to regenerate supplementary point cloud data. When the discrimination result passes the verification, the supplementary point cloud data is spliced ​​with the original spatial distribution features to form a complete three-dimensional point set data of the ballast stack state as a virtual point cloud output.

[0018] Optionally, the dynamic enhancement of loose region weights through a spatiotemporal attention mechanism includes:

[0019] A spatiotemporal attention mechanism is constructed based on the spatial distribution characteristics of ballast particle gaps;

[0020] The weight adjustment increment is increased for the ballast particle gap region with high frequency spatial location changes, and the time cumulative enhancement factor is applied to the ballast particle gap region that continues to expand over time.

[0021] The weight values ​​of loose regions in the spatial distribution features are increased based on the weight adjustment increment and time accumulation enhancement factor.

[0022] Optionally, the quantitative analysis of the virtual point cloud, combined with the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature variation data, and the weights of loose regions, outputs the ballast compaction value and the spatial coordinate range of the compacted area, which serves as a key identification result of the environmental status of the ballast structure along the railway line, including:

[0023] Quantitative analysis is performed on the virtual point cloud to calculate the point density distribution of the virtual point cloud within a unit volume spatial grid, and the median value of the point density distribution is used as the benchmark value for ballast compaction.

[0024] By combining the dielectric constant distribution characteristics, curvature variation data, and loose region weight values ​​in the enhanced point cloud model, a slab region determination condition is established. The slab region determination condition includes three independent determination conditions: the first determination condition is to mark the region in the enhanced point cloud model where the dielectric constant characteristic value is continuously higher than a set threshold; the second determination condition is to extract the flat region where the curvature variation is lower than the curvature threshold in the curvature variation data; and the third determination condition is to locate the region where the weight value is lower than the critical weight value in the loose region weight value distribution.

[0025] The region that simultaneously satisfies the first, second, and third judgment conditions is identified as the hardening region. The spatial inflection point coordinates of the outer contour of the hardening region are extracted, and the spatial inflection point coordinates are connected to form a three-dimensional polygon bounding box.

[0026] The ballast compaction benchmark value is output as the ballast compaction value, and the set of vertex coordinates of the three-dimensional polygon bounding box is output as the spatial coordinate range of the compacted area. The ballast compaction value and the spatial coordinate range of the compacted area together constitute the key identification result of the environmental status of the ballast structure along the railway.

[0027] Optionally, the spiral scanning of the ballast area surface by the line-scanning laser module at the bottom of the track inspection vehicle, simultaneously triggering the millimeter-wave radar to perform a penetrating scan, generates a three-dimensional spatial data matrix fusing surface morphology and internal dielectric characteristics, including:

[0028] The galvanometer deflection component of the line-scanning laser module at the bottom of the track inspection vehicle moves along an Archimedean spiral trajectory, driving the laser beam to cover the surface of the ballast area in a spiral path, and generating a set of three-dimensional coordinate points of the ballast surface morphology by receiving the reflected laser beam.

[0029] At the start of laser beam scanning, a trigger pulse is sent to the millimeter-wave radar controller to drive the millimeter-wave radar to emit a penetrating beam. After receiving the reflected signal inside the ballast layer, the dielectric constant characteristic values ​​of different depth positions of the ballast layer are extracted.

[0030] Based on the set of three-dimensional coordinate points on the surface, the surface coordinate points are vertically projected onto the underground coordinate system. According to the dielectric constant characteristic value of each depth layer below the projection point, the spatial mapping relationship between the three-dimensional coordinate points on the surface and the depth layer below the underground is established.

[0031] Based on the spatial mapping relationship, the data is integrated according to the spatial grid. The surface height value and the corresponding underground dielectric constant profile value are stored with the surface projection point as the center position. All spatial grid units are aggregated to generate a three-dimensional spatial data matrix.

[0032] Optionally, the step of using a neural network structure to analyze the curvature variation data of the ballast particle contact surface in the three-dimensional spatial data matrix and identify the spatial distribution characteristics of the ballast particle gaps includes:

[0033] The three-dimensional spatial data matrix is ​​processed by a neural network structure. The surface normal vector change sequence of the ballast particle contact surface area is extracted in the primary processing module of the neural network structure. The curvature change quantization value is calculated based on the surface normal vector change sequence to generate ballast particle contact surface curvature change law data.

[0034] In the advanced processing module of the neural network structure, a continuous spatial region with constant curvature quantization value is detected within the three-dimensional spatial data matrix. The continuous spatial region is identified as a set of ballast particle gaps. The three-dimensional spatial distribution parameters of the set of ballast particle gaps are measured to generate spatial distribution characteristics of ballast particle gaps.

[0035] Optionally, the step of dynamically compensating for distortion in the laser point cloud data based on the operating state parameters, and constructing an enhanced point cloud model containing physical properties by combining the dielectric constant characteristics in the echo data, includes:

[0036] Based on the longitudinal displacement, lateral offset and pitch angle values ​​in the operating status parameters, a position compensation vector is generated for each three-dimensional coordinate point in the laser point cloud data.

[0037] Based on the position compensation vector, the original coordinates of the laser point cloud data are translated and corrected point by point to obtain a set of corrected three-dimensional coordinate points;

[0038] The dielectric constant characteristic values ​​of the ballast layer at different depths are analyzed from the echo data of the millimeter-wave radar, and the dielectric constant characteristic values ​​are used as new physical attributes and mapped to the corresponding points in the set of corrected three-dimensional coordinate points according to their spatial locations.

[0039] For each corrected 3D coordinate point, the attribute dimension is expanded to generate a composite data unit that simultaneously contains spatial coordinates and dielectric constant values. All composite data units are aggregated to form an enhanced point cloud model.

[0040] Secondly, this application provides an artificial intelligence-based railway line environment identification system, including:

[0041] The acquisition module is used to acquire laser point cloud data of the track ballast area and simultaneously collect the operating status parameters of the track inspection vehicle and the echo data of the millimeter-wave radar.

[0042] A construction module is used to perform dynamic distortion compensation on the laser point cloud data based on the operating state parameters, and to construct an enhanced point cloud model containing physical properties by combining the dielectric constant characteristics in the echo data.

[0043] The generation module is used to perform a spiral scan on the surface of the ballast area using the line-scanning laser module at the bottom of the track inspection vehicle, and simultaneously trigger the millimeter-wave radar to perform a penetrating scan, so as to generate a three-dimensional spatial data matrix that integrates the surface morphology and internal dielectric characteristics.

[0044] The analysis module is used to analyze the curvature variation data of the contact surface of ballast particles in the three-dimensional spatial data matrix using a neural network structure and to identify the spatial distribution characteristics of the gaps between ballast particles. At the same time, it dynamically enhances the weight of loose areas through a spatiotemporal attention mechanism.

[0045] The completion module is used to construct a multi-scale point cloud generation model based on the adversarial training mechanism, and to geometrically complete the missing occluded areas in the spatial distribution features to output a virtual point cloud with a complete ballast stacking state.

[0046] The output module is used to perform quantitative analysis on the virtual point cloud, and combine the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature change law data and the weight of loose areas to output the ballast compaction value and the spatial coordinate range of the compacted area, which serves as a key identification result of the ballast structure environmental status along the railway line.

[0047] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement an artificial intelligence-based railway environment recognition method as described in the first aspect above.

[0048] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements an artificial intelligence-based method for identifying railway line environments as described in the first aspect.

[0049] In this application example, laser point cloud data of the track ballast area is acquired, and the operating status parameters of the track inspection vehicle and the echo data of the millimeter-wave radar are collected simultaneously. Based on the operating status parameters, dynamic distortion compensation is performed on the laser point cloud data, and combined with the dielectric constant characteristics in the echo data, an enhanced point cloud model containing physical properties is constructed. The surface of the ballast area is spirally scanned by a line-scanning laser module at the bottom of the track inspection vehicle, simultaneously triggering the millimeter-wave radar to perform a penetration scan, thereby generating a three-dimensional spatial data matrix that integrates surface morphology and internal dielectric characteristics. A neural network structure is used to analyze the ballast in the three-dimensional spatial data matrix. The data on the curvature variation of the particle contact surface is used to identify the spatial distribution characteristics of the gaps between ballast particles. At the same time, the weight of loose regions is dynamically enhanced through a spatiotemporal attention mechanism. A multi-scale point cloud generation model is constructed based on an adversarial training mechanism to geometrically complete the missing occluded regions in the spatial distribution characteristics to output a virtual point cloud of the complete ballast stacking state. The virtual point cloud is quantitatively analyzed, and combined with the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature variation data, and the weight of loose regions, the ballast compaction value and the spatial coordinate range of the compacted area are output as key identification results of the ballast structure environmental status along the railway line.

[0050] The technical solution of this application has the following beneficial effects:

[0051] This application simultaneously acquires laser point cloud data, track inspection vehicle status parameters, and millimeter-wave radar echo data. First, it dynamically compensates for point cloud distortion based on status parameters and integrates radar dielectric constant to construct a point cloud model with enhanced physical properties. Then, through the synchronous triggering of linear laser helical scanning and millimeter-wave radar penetration scanning, it generates a three-dimensional spatial data matrix that integrates surface morphology and internal dielectric characteristics. Next, it uses neural networks to analyze the curvature changes and gap distribution characteristics of ballast particle contact surfaces, and dynamically enhances the weight of loose areas using a spatiotemporal attention mechanism. Based on adversarial training, it constructs a multi-scale point cloud generation model, geometrically completes the obscured areas, and outputs a virtual point cloud of complete ballast accumulation. Finally, through quantitative analysis of the virtual point cloud, combined with dielectric constant distribution, curvature change law, and loose area weight, it accurately outputs the ballast density value and the spatial coordinate range of the compacted area, realizing non-contact, penetrating, and high-precision quantitative identification of the internal structural state of railway ballast, forming an automated intelligent detection closed loop that replaces manual inspection.

[0052] A multi-scale point cloud generation model is further constructed, consisting of a dual-channel generator with coarse-grained and fine-grained generation channels and a geometric continuity verification discriminant. The occluded region is input into the generator. First, the basic geometry is predicted through the coarse-grained channel, and then surface details of the ballast particles are added through the fine-grained channel to generate supplementary point clouds. The supplementary point clouds and the original data are input into the discriminant, and the geometric continuity verification module checks the connection status between the supplementary region and adjacent curvatures, as well as the physical rationality of particle gaps. If verification fails, the generator parameters are iteratively optimized and regenerated. After successful verification, the supplementary point clouds are stitched together with the original data to output a complete 3D virtual point cloud. This method achieves hierarchical completion of "basic geometric framework construction - granular detail restoration" through dual-scale generation channels. Combined with the dual constraints of curvature connection and gap physical rationality imposed by the geometric continuity verification module, it ensures that the completed point cloud of the occluded region conforms to the macroscopic topological structure of ballast stacking while maintaining the physical authenticity of granular microscopic features. This generates a complete ballast virtual point cloud with high geometric consistency and physical rationality, providing a reliable data foundation for subsequent quantitative analysis.

[0053] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1A flowchart of an artificial intelligence-based railway environment identification method provided in this application is shown;

[0056] Figure 2 The illustration shows a scenario diagram of a railway line environment recognition method based on artificial intelligence provided in this application;

[0057] Figure 3 A schematic diagram of the structure of a railway line environment recognition system based on artificial intelligence provided in this application is shown;

[0058] Figure 4 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0059] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0060] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0061] Research indicates that existing representative solutions based on vehicle-mounted 3D laser scanning and machine learning suffer from fundamental bottlenecks in the field of automated inspection of railway ballast condition. These solutions rely on single surface morphology information, making it impossible to penetrate and understand key internal physical properties such as the distribution of ballast particle gaps, changes in internal density, and the bonding state of slabs. Furthermore, severe surface occlusion leads to large-area gaps in point cloud data, making it difficult to reconstruct the true ballast stacking structure. Ultimately, the output results are limited to coarse qualitative classification, lacking precise numerical assessment of density and spatial coordinate definition of slab boundaries and depths, thus failing to meet the needs of refined and quantitative maintenance decisions. The core of these shortcomings lies in the triple predicament of existing technologies: the invisibility of the internal structure of ballast, the unknowability of occluded areas, and the inability to quantify state parameters.

[0062] To address the aforementioned issues, this application proposes an artificial intelligence-based method for railway line environment identification. Its core lies in simultaneously fusing laser point clouds, millimeter-wave radar echoes, and track inspection vehicle status parameters to construct an enhanced point cloud model incorporating physical attributes. A three-dimensional spatial data matrix is ​​generated through the synchronous triggering of linear laser helical scanning and millimeter-wave penetrating scanning. Furthermore, a neural network is used to analyze particle contact curvature and gap distribution characteristics, and a spatiotemporal attention mechanism is employed to dynamically focus on loose areas. Specifically, an adversarial training mechanism is applied to construct a multi-scale point cloud generation model, intelligently completing occluded areas under geometric-physical dual constraints, outputting a virtual point cloud of complete ballast accumulation. Finally, the virtual point cloud is quantitatively analyzed, and combined with dielectric constant, curvature patterns, and looseness weights, the ballast density value and the spatial coordinate range of the compacted area are accurately output. This method breaks through the triple dilemma of the background technology. It makes the internal structure "visible" by fusing millimeter wave penetration and dielectric constant, makes the occluded area "knowable" by using geometric-physical completion of the adversarial generative model, and outputs "measurable" density values ​​and precise spatial coordinates through a quantitative analysis engine that fuses multiple features. This provides a reliable technical closed loop for the automated, accurate, and quantitative detection of railway ballast condition, and solves the problems of insufficient identification accuracy and lack of quantification capability caused by the reliance on surface information in existing solutions.

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] Figure 1 A flowchart of an artificial intelligence-based railway environment identification method is provided as an embodiment of this application, such as... Figure 1 As shown, the method includes:

[0065] 101. Acquire laser point cloud data of the track ballast area, and simultaneously collect the operating status parameters of the track inspection vehicle and the echo data of the millimeter-wave radar;

[0066] In the above scheme, laser point cloud data refers to a dense set of three-dimensional coordinate points generated by emitting laser beams from the line-scanning laser module at the bottom of the track inspection vehicle and receiving reflected signals from the ballast surface. Each point contains spatial location (X, Y, Z) information, used to characterize the geometry of the ballast surface. Operating status parameters refer to the dynamic physical quantities of the track inspection vehicle during its movement, including vehicle speed, longitudinal acceleration, lateral acceleration, pitch angle, and roll angle, used to subsequently correct data distortion caused by vehicle movement during laser scanning. Millimeter-wave radar echo data refers to the sequence generated by the onboard millimeter-wave radar emitting electromagnetic waves towards the ballast and receiving reflected signals, including signal strength, phase difference, and propagation time. Analysis of this data allows for the extraction of the dielectric constant within the ballast.

[0067] In this embodiment, firstly, a laser beam at a frequency of 200Hz is emitted from the line-scanning laser module at the bottom of the track inspection vehicle to perform a high-speed line scan of the ballast surface on both sides of the track. When the laser beam hits a ballast particle, the receiver records the time difference and angle of the reflected light. Combined with the spatial positioning data from the built-in high-precision gyroscope, a single frame containing 50,000 to 80,000 three-dimensional coordinate points is generated as laser point cloud data. For example, when the laser beam scans a protruding gravel on the ballast surface, the distance between the point and the sensor is calculated based on the round-trip time difference of the beam. Combined with the scanning angle, its precise position (X=1.2m, Y=0.3m, Z=0.15m) is determined, forming a data point in the point cloud.

[0068] Secondly, the onboard IMU (Inertial Measurement Unit) collects the track inspection vehicle's operating parameters at a frequency of 100Hz. The vehicle speed, three-axis acceleration, pitch angle, and roll angle are recorded, and these parameters are encapsulated into time-stamped data packets. For example, when the track inspection vehicle passes a track joint and experiences a bump, the IMU records that at this moment, the longitudinal acceleration abruptly changes to +0.08g, and the pitch angle changes by +0.8°; this data packet is timestamped T=12.345s.

[0069] Finally, the millimeter-wave radar is triggered to emit a 77 GHz frequency-modulated continuous wave. After penetrating the ballast surface, the electromagnetic wave is reflected by different media. The receiver captures the echo signal, analyzes its amplitude attenuation and phase shift, and outputs the original echo sequence containing the time-intensity relationship. For example, when the radar wave penetrates the ballast, the echo amplitude attenuates by 20 dB in the gaps between dry gravel, while in the water-bearing compacted area, the amplitude attenuation is only 5 dB due to the increased dielectric constant, forming a characteristic signal that distinguishes the internal structure.

[0070] In a practical application, during routine inspections of a freight railway, track inspection vehicle A travels at 25 km / h to section B. The laser module under the vehicle scans the ballast area, generating a single frame of 72,000 surface point clouds. Simultaneously, the IMU unit records the vehicle's speed (25 km / h), lateral acceleration (0.05g), and pitch angle (-0.6°). A millimeter-wave radar synchronously emits electromagnetic waves, acquiring a 5ms echo signal. An abnormally high-amplitude echo is detected in region C (coordinates X=120m~125m), indicating a risk of slab compaction. All three types of data are marked with a unified timestamp "2023-05-10 14:30:25.123" for precise synchronization.

[0071] The aforementioned 101 overall solution, through multi-source synchronous acquisition, simultaneously obtains data on the fine geometric morphology of the ballast surface, the dynamic attitude of the vehicle, and the internal physical properties of the ballast in a single inspection operation, laying the foundation for subsequent fusion processing. Its core value lies in overcoming the limitations of a single sensor. Lasers can accurately depict the surface but cannot penetrate the interior, radar can perceive the interior but lacks geometric precision, and state parameters provide a basis for dynamic distortion compensation. The three work together to form a full-dimensional data support of "surface-interior-motion".

[0072] 102. Based on the operating state parameters, perform dynamic distortion compensation on the laser point cloud data, and combine the dielectric constant characteristics in the echo data to construct an enhanced point cloud model containing physical properties;

[0073] Optionally, step 102 may specifically include the following steps:

[0074] 1021. Based on the longitudinal displacement, lateral offset, and pitch angle values ​​in the operating status parameters, generate a position compensation vector for each three-dimensional coordinate point in the laser point cloud data;

[0075] 1022. Based on the position compensation vector, perform point-by-point translation correction on the original coordinates of the laser point cloud data to obtain a corrected three-dimensional coordinate point set;

[0076] 1023. Analyze the dielectric constant characteristic values ​​of different depth positions of the ballast layer from the echo data of the millimeter-wave radar, and map the dielectric constant characteristic values ​​as new physical attributes to the corresponding points in the set of corrected three-dimensional coordinate points according to their spatial positions.

[0077] 1024. Expand the attribute dimensions for each corrected 3D coordinate point to generate a composite data unit that simultaneously contains spatial coordinates and dielectric constant values. Aggregate all composite data units to form an enhanced point cloud model.

[0078] In the above scheme, the position compensation vector refers to the three-dimensional spatial offset (Δx, Δy, Δz) calculated based on the track inspection vehicle's motion deviation, namely longitudinal displacement, lateral offset, and pitch angle, used to correct the position error of each point in the laser point cloud. The corrected three-dimensional coordinate point set refers to the new point cloud generated after the original laser point cloud has been corrected by the position compensation vector, eliminating deformation caused by vehicle motion. The dielectric constant eigenvalue refers to a physical quantity resolved from millimeter-wave radar echoes, representing the electrical conductivity of the ballast material and used to distinguish internal structures. The composite data unit refers to the expanded data structure of each point cloud data point, containing both spatial coordinates (X, Y, Z) and dielectric constant attributes. The enhanced point cloud model refers to the point cloud set composed of all composite data units, possessing both geometric position and internal physical attribute information.

[0079] In this embodiment, firstly, step 1021 calculates the position compensation vector for each three-dimensional coordinate point in the laser point cloud based on the track inspection vehicle's operating state parameters, namely longitudinal displacement, lateral offset, and pitch angle, using a kinematic affine transformation model. The longitudinal displacement is obtained through vehicle speed integration; for example, when the vehicle speed is 30 km / h and the laser sampling interval is 5 ms, the longitudinal compensation Δx = 0.042 m. The lateral offset is derived from the second integral of IMU acceleration data; for example, a 0.1g lateral acceleration lasting 0.1 s produces Δy = 0.0049 m. The pitch angle compensation is calculated through trigonometric function projection; a laser installed at a height of 1.5 m at a pitch angle of +0.8° produces Δz = 0.021 m. Finally, the three-dimensional position compensation vector (Δx, Δy, Δz) for each point is output, providing a dynamic offset reference for subsequent correction.

[0080] Secondly, in step 1022, the original laser point cloud is corrected point-by-point using the position compensation vector generated in step 1021. First, the compensation vector is matched with the corresponding laser point at millisecond-level timestamps. Then, vector addition is performed on the original coordinates to obtain the corrected 3D coordinate point set. For example, the points... The correction point is obtained by superimposing the compensation amount (0.042, -0.005, 0.021) on (1.500, 0.200, 0.100). (1.542, 0.195, 0.121). This process eliminates spatial distortions caused by vehicle movement, such as Z-axis compression caused by the vehicle's head-up on uphill sections, thus underestimating the ballast height in the original point cloud by 2.1 cm, allowing the corrected point cloud to restore the true surface morphology.

[0081] Next, in step 1023, the millimeter-wave radar echo data is analyzed into characteristic values ​​of the dielectric constant of the ballast layer, and then mapped to the corrected three-dimensional coordinate point set output in step 1022 according to their spatial location. Depth layers are then defined using time-gating technology: based on the electromagnetic wave propagation formula... Inversely derived permittivity Where c is the speed of light, t is the echo time, and d is the depth. The relative permittivity reflects the dielectric properties of the ballast material; for example, dry crushed stone ≈ 3, and water-bearing compacted material > 12. The correction point is used as the reference point. Taking the region 20cm underground as an example (1.542, 0.195, 0.121), the measured echo time t = 1.35ns was obtained, and the calculated value was... This indicates the presence of water-bearing deposits at that location. Finally, the dielectric constant is used as a physical property and bound to the spatial coordinates.

[0082] Finally, by extending the attribute dimensions for each 3D coordinate point in the calibration 3D coordinate point set in step 1024, a composite data unit containing both spatial coordinates and dielectric constant values ​​is created. For example, the point... (1.542, 0.195, 0.121) and dielectric constant =12.8 is fused into a four-dimensional data structure (1.542, 0.195, 0.121, 12.8). All composite units are aggregated to form an enhanced point cloud model. Its spatial distribution inherits the millimeter-level geometric accuracy of laser point clouds, capable of resolving 2mm ballast particle undulations. The physical dimension carries dielectric characteristics obtained through radar penetration, such as the dielectric constant of dry gravel. ≈3.0, dielectric constant of the saturated junction region >12.0. This model achieves an integrated characterization of ballast's "surface morphology-internal properties," providing a multimodal data foundation for subsequent intelligent identification.

[0083] In practical applications, during inspection operations on the E railway, when the track inspection vehicle passed through the F curve section at 28 km / h, the IMU recorded a longitudinal displacement of +0.8 m, a lateral offset of -0.05 m, and a pitch angle of -1.2°. For a given original laser point (5.214, -1.032, 0.208), a compensation vector of (+0.011, -0.003, -0.028) was calculated, and the corrected coordinates were updated to (5.225, -1.035, 0.180). Simultaneous millimeter-wave radar echo analysis revealed the dielectric constant of a layer 20 cm below this point. =14.7, corresponding to water-bearing compacted material, finally generating a composite data unit (5.225, -1.035, 0.180, 14.7), which together with 3,200 similar units in the surrounding area constitutes the enhanced point cloud model of this section.

[0084] The overall solution described above (102) eliminates vehicle motion distortion through dynamic compensation, ensuring the geometric accuracy of the point cloud. By fusing dielectric constant characteristics, the point cloud is upgraded from a pure geometric model to a physical model that integrates both shape and material properties. The resulting enhanced point cloud model accurately reflects the true morphology of the ballast surface and reveals the internal material properties, providing a high-information-density data foundation for subsequent intelligent recognition.

[0085] 103. The surface of the ballast area is spirally scanned by the line-scanning laser module at the bottom of the track inspection vehicle, and the millimeter-wave radar is simultaneously triggered to perform a penetrating scan to generate a three-dimensional spatial data matrix that integrates the surface morphology and internal dielectric characteristics.

[0086] Optionally, step 103 may specifically include the following steps:

[0087] 1031. The galvanometer deflection component of the line-scanning laser module at the bottom of the track inspection vehicle moves along the Archimedean spiral trajectory, driving the laser beam to cover the surface of the ballast area in a spiral path, and generating a set of three-dimensional coordinate points of the ballast surface morphology by receiving the reflected laser beam.

[0088] 1032. At the start of laser beam scanning, a trigger pulse is sent to the millimeter-wave radar controller to drive the millimeter-wave radar to emit a penetrating beam. After receiving the reflected signal inside the ballast layer, the dielectric constant characteristic values ​​of different depth positions of the ballast layer are extracted.

[0089] 1033. Based on the set of three-dimensional coordinate points on the surface, the surface coordinate points are vertically projected onto the underground coordinate system. According to the dielectric constant characteristic values ​​of each depth layer below the projection point, the spatial mapping relationship between the three-dimensional coordinate points on the surface and the depth layers below is established.

[0090] 1034. Based on the spatial mapping relationship, integrate the data according to the spatial grid, store the surface height value and the corresponding underground dielectric constant profile value with the surface projection point as the center position, and aggregate all spatial grid units to generate a three-dimensional spatial data matrix.

[0091] In the above scheme, the galvanometer deflection assembly is the core optical actuator that controls the scanning direction of the laser beam, used to precisely guide the laser beam to the target surface. The Archimedes spiral trajectory refers to the spatial scanning path formed by the laser beam on the ballast surface, mathematically characterized by a linear expansion of radial distance with rotation angle, achieving comprehensive coverage from the center outwards. The surface three-dimensional coordinate point set refers to the spatial location dataset of the ballast surface generated by laser ranging, containing the horizontal coordinates (x, y) and vertical height value z of each scanning point, used to characterize the geometric distribution of ballast particles. The penetrating beam refers to the electromagnetic signal emitted by millimeter-wave radar, whose physical characteristic is that it can penetrate the ballast gravel layer and reflect at different medium interfaces, used to obtain the dielectric properties of the internal structure of the ballast. The underground coordinate system refers to a three-dimensional spatial reference system extending vertically downwards from the surface reference plane, with its depth axis φ having the surface as its zero point, used to locate the spatial position of underground dielectric characteristics. Dielectric constant profile values ​​refer to the sequence of dielectric constant data corresponding to different depth layers below the same surface projection point, reflecting the variation of the physical properties of the ballast layer along the depth direction. A three-dimensional spatial data matrix refers to a structured data container organized according to a regular spatial grid, where each grid cell stores the surface height value and the associated underground dielectric constant profile.

[0092] In this embodiment, step 1031 first performs a surface scan using a line-scanning laser module at the bottom of the track inspection vehicle. The galvanometer deflection assembly is controlled to move along an Archimedean spiral trajectory, driven by a preset angular velocity and radial expansion velocity. The coordinates of the laser landing point are then calculated. When the laser beam covers the ballast surface in a helical path, the receiver records the time difference and angle of the reflected signal. Combined with high-precision inertial navigation data, a set of three-dimensional coordinate points on the ballast surface is generated. For example, at t=0.5s, the coordinates of the scanned points are calculated. The height of this point was measured to be z=0.15m, forming the surface point (-0.1,0.0,0.15).

[0093] Secondly, in step 1032, a hardware-level synchronization trigger pulse is sent to the millimeter-wave radar controller at the start of laser scanning, driving the radar to emit a 77GHz frequency-modulated continuous wave that penetrates the ballast layer. When the electromagnetic wave encounters different medium interfaces and is reflected, such as gaps between gravel or slabs, the dielectric constant characteristic values ​​of each depth layer are extracted using a time-domain analytical algorithm after the receiver captures the echo signal. For example, at the surface point (-0.1, 0.0), the radar detects an echo time delay of 4.2ns at a depth of 30cm, according to the formula... The calculated dielectric constant is 15.3, at which point c = 3e8m / s and d = 0.3m, characterizing the core region of the slab.

[0094] Next, in step 1033, the surface three-dimensional coordinate point set generated in step 1031 is vertically projected to the underground coordinate system: using surface points Taking (8.732, 1.204, 0.21) as an example, its projection reference point is set as... (8.732, 1.204, 0). With Divide the image downwards into depth layers, with each layer measuring 10 cm. Bind the dielectric constant characteristic values ​​obtained from the 1032 diode to the corresponding layer according to depth. For example... The depth sequence below the point [0-10cm, 10-20cm, 20-30cm] is associated with the dielectric value [3.1, 5.2, 15.3], forming a vertical mapping chain of "surface location - underground properties".

[0095] Finally, in step 1034, the data is gridded based on spatial mapping relationships, dividing the detection area into 0.1m × 0.1m surface grid units. Each unit integrates the average height of all surface points within it and the associated underground dielectric constant profile value. For example, grid (87,12) contains 3 surface points, with a calculated average height of 0.22m and bound dielectric profile [3.1,5.2,15.3]. Finally, 200 × 150 grid units are aggregated to generate a three-dimensional spatial data matrix, where grid (90,15) stores data {height: 0.18m, dielectric profile: [3.0,4.1,8.9,15.2]}, intuitively revealing the distribution of the 1.2m deep compaction zone.

[0096] In practical applications, when the track inspection vehicle travels at a speed of 20 km / h, it first drives a laser beam along an Archimedean spiral trajectory via a galvanometer deflection assembly, generating a surface point (2.874, -0.532, 0.18) at t=0.6s with an angular velocity of 8π rad / s and a radial velocity of 0.15 m / s. Simultaneously, a pulse is triggered to activate the millimeter-wave radar, which detects an echo delay of 3.8 ns at a depth of 25 cm underground at this point, and calculates the dielectric constant. =14.2; then the surface point is vertically projected onto the grid (28,-5), and associated with the dielectric profile values ​​[3.2,4.5,14.2,12.1] at a depth of 0-40cm below it; finally, 500 grids in the region are integrated to generate a three-dimensional data matrix, in which the grid (30,-6) stores the average surface height of 0.21m and the dielectric profile [3.0,4.0,8.5,15.0], accurately locating the core area of ​​the compaction at a depth of 35cm.

[0097] The aforementioned overall scheme 103 achieves seamless coverage of the ballast surface through spiral scanning, simultaneously acquiring internal dielectric characteristics through penetrating scanning, and establishing a precise surface-to-subsurface spatial mapping relationship. The resulting three-dimensional spatial data matrix integrates surface morphology and internal physical properties within a unified spatial framework, forming an integrated "surface-to-subsurface" multi-dimensional data base, providing multi-source information for spatial alignment for subsequent intelligent identification.

[0098] 104. A neural network structure is used to analyze the curvature variation data of the contact surface of ballast particles in the three-dimensional spatial data matrix and identify the spatial distribution characteristics of the gaps between ballast particles. At the same time, the weight of loose areas is dynamically enhanced through a spatiotemporal attention mechanism.

[0099] Optionally, step 104 may specifically include the following steps:

[0100] 1041. The three-dimensional spatial data matrix is ​​processed through a neural network structure. The surface normal vector change sequence of the ballast particle contact surface area is extracted in the primary processing module of the neural network structure. The curvature change quantization value is calculated based on the surface normal vector change sequence to generate ballast particle contact surface curvature change law data.

[0101] 1042. In the advanced processing module of the neural network structure, a continuous spatial region with constant curvature quantization value is detected in the three-dimensional spatial data matrix, the continuous spatial region is identified as a set of ballast particle gaps, the three-dimensional spatial distribution parameters of the set of ballast particle gaps are measured, and spatial distribution characteristics of ballast particle gaps are generated.

[0102] 1043. Constructing a spatiotemporal attention mechanism based on the spatial distribution characteristics of ballast particle gaps;

[0103] 1044. Increase the weight adjustment increment for the ballast particle gap region with high frequency spatial location changes, and apply the time cumulative enhancement factor to the ballast particle gap region that continues to expand over time.

[0104] 1045. Based on the weight adjustment increment and time accumulation enhancement factor, increase the weight value of loose regions in the spatial distribution characteristics.

[0105] In the above scheme, the surface normal vector change sequence refers to dynamic data describing the evolution of the geometric morphology of the ballast particle contact surface, including the angular difference sequence of the vertical direction vectors of adjacent surface points, which can be used to quantify the surface unevenness characteristics. The curvature change quantification value refers to a numerical index characterizing the sharpness of the ballast particle edges, which can be used to identify stress concentration areas on the particle contact surface. The ballast particle gap set refers to the three-dimensional continuous spatial region identified as unfilled with gravel, which can be used to assess the structural stability of the ballast layer. The three-dimensional spatial distribution parameters refer to a set of quantification indicators describing the geometric characteristics of the particle gaps, which can be used to determine the gap morphology type. The spatiotemporal attention mechanism refers to an intelligent algorithm that dynamically adjusts data weights, including a spatial position offset calculation module and a temporal evolution trend analysis module, which can be used to focus on high-risk loose areas. The weight adjustment increment refers to the enhancement coefficient applied to areas of high-frequency spatial change, which can be used to increase the detection priority of abrupt change areas. The time-cumulative enhancement factor refers to the progressive enhancement coefficient applied to continuously expanding areas, which can be used to provide early warning of progressive loosening and deterioration.

[0106] In this embodiment, the three-dimensional spatial data matrix is ​​first input into the primary processing module of the 3D convolutional neural network in step 1041. The surface elevation data is then scanned using a 5×5×5 convolutional kernel to calculate the normal vector of each point, i.e., the direction vector perpendicular to the surface. Taking point (1.2, 0.5, 0.18) as an example, the normal vector sequence within its 3cm neighborhood [(0.12, 0.85, 0.51), (0.15, 0.82, 0.54)] is extracted, and the rate of change of the angle between adjacent normal vectors is calculated. ,in For curvature quantization value, The normal vectors of adjacent points, Given the distance between adjacent points on the surface, a matrix of curvature variation data is generated by traversing all contact surface regions. In this matrix, the curvature of a certain corner region suddenly increases to 0.31 mm⁻¹, while the curvature of the normal region is <0.05 mm⁻¹.

[0107] Secondly, in step 1042, the curvature data is input into the graph neural network advanced processing module to automatically detect continuous regions with curvature values ​​< 0.01 mm⁻¹. These continuous spatial regions are identified as sets of ballast particle gaps and determined as gaps. A three-dimensional region growing algorithm is executed with coordinates (120.5, 8.7) as the center to identify a connected space with a volume of 3.2 m³. The volume parameters are calculated by converting voxel counts and resolution. The surface area is calculated as 15.6 m² by integrating the surface triangular patches. The depth is measured as 3.2 m and the width as 1.5 m, resulting in a depth-to-width ratio of 2.1. Finally, the feature vector of this gap [position (120.5, 8.7), volume 3.2, surface area 15.6, depth-to-width ratio 2.1] is output as the spatial distribution feature of the ballast particle gaps.

[0108] Next, a dual-channel data structure is constructed based on the spatial distribution characteristics of the ballast particle gaps in step 1043: the spatial channel loads the coordinates of the center of all current gaps, for example, establishing a location set L={(120.5,8.7,0.2),(122.1,9.3,0.3)}; the temporal channel is associated with the historical database, extracting the volume sequence T={2.8,3.0,3.2} m³ of the three past detections at that location. A spatiotemporal encoder binds the location coordinates to historical trends, forming a timestamped feature map, establishing a data foundation for subsequent dynamic analysis.

[0109] Then, in step 1044, an enhancement strategy is implemented for the two types of risk areas: First, the spatial location offset is calculated, and when the Euclidean distance between the current gap center and the position of the previous week exceeds the threshold of 0.2m, the spatial weight increment is triggered. Secondly, analyzing the time trend, when the volume increases by 6.7% and 7.1% for three consecutive detections, all exceeding the 5% threshold, a time accumulation factor is applied exponentially based on the number of cycles. For example, the Euclidean distance between the current gap center (120.5, 8.7) and the position from last week (120.2, 8.9) is 0.28m, exceeding the 0.2m threshold, thus triggering a spatial weight increment. When the volume increases by 6.7% and 7.1% for three consecutive detections, and all exceed the 5% threshold, a time accumulation factor is applied exponentially based on the number of cycles. At this point, the gap simultaneously satisfies the conditions of spatial abrupt change and temporal continuous expansion.

[0110] Finally, the final risk weight value is synthesized through step 1045: taking the base weight. Multiplied by spatial weight increment With time accumulation factor The weight values ​​of the loose region are obtained. When the weight value is greater than 2.0, it is automatically marked as a high-risk loose area, and the gap is given a red warning indicator. Simultaneously, the latest volume is added to the historical sequence for trend analysis in the next period, completing the dynamic monitoring loop. For example, the basic weight of the gap (120.5, 8.7) is... At this time, the spatial weight increment is =1.5, time accumulation factor is =1.728, the weight value of the loose region is calculated. At this point, the weight value is greater than 2.0, and it is automatically marked as a high-risk loose area and given a red warning sign. The latest volume of 3.2m³ is added to the historical sequence and updated to T={3.0,3.2,3.2}.

[0111] In practical applications, during the inspection of section N of railway M, the neural network processes the three-dimensional data matrix: the primary module calculates the normal vector sequence [(0.10,0.88,0.46),(0.08,0.90,0.42)] for the neighborhood of point (120.5,8.7,0.23), the dot product of 0.978 inverse cosine yields 12.1°, and combined with the surface distance of 5.8mm, the curvature is 0.035mm. - ¹, while the curvature of the adjacent corner region jumps to 0.33 mm. - ¹; The advanced module identifies a constant curvature region of 0.008 mm at (120.5, 8.7). - ¹, the gap volume was measured to be 3.2 m³ and the surface area to be 15.6 m² after 3D segmentation; the spatiotemporal mechanism was used to load the historical volume sequence [2.8, 3.0, 3.2] m³ of this point and the position of the previous week (120.2, 8.9), and the spatial offset of 0.28 m was calculated to trigger the increment. =1.5, continuous volume growth (7.1%, 6.7%) triggering time factor =1.2³=1.728; the final weight value of the loose area is 1.0×1.5×1.728=2.592>2.0, so it is marked as a high-risk loose area and the historical database is updated.

[0112] The aforementioned overall scheme 104 uses neural networks to accurately quantify the curvature changes and spatial distribution of gaps in the contact surface of ballast particles, breaking through the limitations of traditional manual experience; it combines a spatiotemporal attention mechanism to dynamically track positional shifts and continuously expanding high-risk areas, assigning weight increments and cumulative factors; and finally, it automatically marks loose risk areas based on synthetic weights, realizing a closed loop from microscopic feature extraction to macroscopic risk warning, providing data-driven dynamic decision-making basis for ballast structure stability assessment.

[0113] 105. Construct a multi-scale point cloud generation model based on an adversarial training mechanism, and perform geometric completion on the missing occluded areas in the spatial distribution features to output a virtual point cloud of complete ballast stacking state.

[0114] Optionally, step 105 may specifically include the following steps:

[0115] 1051. Construct a multi-scale point cloud generation model based on an adversarial training mechanism, which includes a generator unit and a discriminator unit. The generator unit is configured with a coarse-grained generation channel and a fine-grained generation channel, and the discriminator unit is configured with a geometric continuity verification module.

[0116] 1052. Input the occluded area in the spatial distribution features into the generator unit, predict the basic geometry of the occluded area through the coarse-grained generation channel, and add the surface details of the ballast particles through the fine-grained generation channel to generate supplementary point cloud data of the occluded area.

[0117] 1053. Input the supplementary point cloud data and the original spatial distribution features into the discriminator unit, and use the geometric continuity verification module to detect the curvature connection state between the supplementary geometry and the adjacent region, while verifying the physical rationality of the particle gap transition, and output the discrimination result.

[0118] 1054. When the discrimination result fails the verification, the generator unit parameters are iteratively optimized to regenerate supplementary point cloud data. When the discrimination result passes the verification, the supplementary point cloud data is spliced ​​with the original spatial distribution features to form a complete three-dimensional point set data of the ballast stack state as the virtual point cloud output.

[0119] In the above scheme, the generator unit refers to the occlusion region prediction module based on an adversarial training mechanism. It includes a dual-channel collaborative structure for coarse-grained geometric framework generation and fine-grained surface detail addition, which can be used to output physically reasonable supplementary point cloud data. The coarse-grained generation channel refers to the neural network component that generates the basic geometric contour, which can be used to construct the macroscopic three-dimensional framework of the occluded region. The fine-grained generation channel refers to the neural network component that restores the microscopic features of the ballast, which can be used to add granular texture undulations. The discriminator unit is an intelligent evaluation module that verifies the reasonableness of the generated results. It includes dual detection logic for geometric continuity and physical laws, which can be used to ensure the engineering reliability of the supplementary point cloud. The geometric continuity verification module is an algorithm unit that detects the quality of morphological connection, which can be used to eliminate geometric discontinuities. The physical reasonableness verification is a logical unit for constraining mechanical compliance, including gap rate threshold determination, which can be used to ensure that the generated results conform to the mechanical specifications of the ballast structure. The supplementary point cloud data refers to the virtual three-dimensional point set generated for the occluded region, including coarse-grained basic framework points and fine-grained surface detail points, which can be used to reconstruct the complete ballast stacking state.

[0120] In this embodiment, a dual-module structure is first built based on the adversarial training framework in step 1051: the generator unit adopts the PointNet++ network architecture, where the coarse-grained generation channel is designed as a fully connected layer outputting 256 basic points to construct the macroscopic geometric framework, and the fine-grained generation channel uses a 3D transposed convolutional layer to output 2048 detail points to add granular texture; the discriminator unit integrates a geometric continuity verification module and a physical rationality verification module. The network weights are initialized based on the ballast engineering specifications, the coarse-grained channel learns the ballast pile slope characteristics such as a 25° inclination angle, and the fine-grained channel learns the standard deviation of the crushed stone surface, such as ±2cm undulation.

[0121] Secondly, in step 1052, the set of boundary points of the occluded region is input into the generator: the coarse-grained channel uses the least squares method to fit the plane equation and generate the basic points of the grid; the fine-grained channel adds random perturbations based on the generated coordinates. This simulates the particle characteristics of ballast. For example, given an input occlusion region boundary of {(10.2,3.1,0.18),(10.3,3.2,0.20)}, the coarse-grained channel generates a 20×20 grid of base points based on the plane equation z=0.25x+0.1y, while the fine-grained channel adds random perturbations. The surface texture details are formed by simulating the characteristics of ballast particles.

[0122] Next, after merging the supplemented point cloud with the original point cloud in step 1053, the result is input into the discriminator: the geometric continuity module calculates the curvature of the edge points of the supplemented region. curvature of adjacent original points Find the difference ,when If the value is less than the threshold of 0.02, the verification is passed; subsequently, the total volume of the generated area is calculated using the physical rationality module. interstitial volume The gap ratio is calculated and compared with a threshold of 40% to determine whether it meets the ballast mechanical specifications. The calculation formula is as follows: For example, in the decision maker, the geometric continuity module calculates the curvature of the edge points of the completed region. curvature of adjacent original points The difference If the value is less than the threshold of 0.02, the verification is passed; at the same time, the physical rationality module calculates the total volume of the generated area. interstitial volume The calculated gap ratio is 35% < threshold 40%, which is considered to meet the ballast mechanical specifications.

[0123] Finally, iterative optimization and stitching are performed in step 1054. If the verification passes, the supplementary points are directly stitched with the original points; if the verification fails, the generator parameters are adjusted using the backpropagation algorithm: the learning rate of the coarse-grained channel is reduced to 0.0001 to smooth the geometric framework, and the noise amplitude of the fine-grained channel is increased to ±3cm to enhance detail diversity. The supplementary point cloud is regenerated until the verification passes, and finally, a complete virtual point cloud is output.

[0124] In practical applications, during the detection of the S section of the R railway, the input boundary point set for the occluded area is {(15.2,4.3,0.22),(15.3,4.4,0.25)}; the generator uses a coarse-grained channel to fit the plane equation z=0.28x+0.12y to generate an 18×18 basic mesh, and a fine-grained channel to superimpose perturbation functions. The base point (15.25, 4.35, 0.255) is converted to a fine-grained point (15.25, 4.35, 0.268); the discriminator detects and completes the edge curvature of 0.072mm. - ¹ 0.069 mm from the original region - The difference between the two points is 0.003 < the threshold of 0.02, and the gap ratio of 36.2% (0.38 m³ / total volume 1.05 m³) is less than the 40% threshold. Finally, the 1,650 supplementary points generated are stitched together with the original 7,800 points to output a complete virtual point cloud. When the initial gap ratio reaches 42%, the discriminator triggers parameter optimization. After adjusting the fine-grained noise amplitude, the second generation gap ratio of 37.1% passes verification.

[0125] The above-mentioned overall scheme of 105 achieves intelligent completion of occluded areas through adversarial training: the coarse-grained channel ensures the rationality of macroscopic geometry, and the fine-grained channel restores the microscopic particle features; the discriminator's dual verification ensures that the completion result is both consistent with morphology and meets engineering mechanical constraints, and the final output of the complete virtual point cloud provides a reliable data foundation for the quantitative analysis of ballast condition.

[0126] 106. Perform quantitative analysis on the virtual point cloud, and combine the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature change law data, and the weight of loose areas to output the ballast compaction value and the spatial coordinate range of the compacted area, which will serve as the key identification result of the ballast structure environmental status along the railway line.

[0127] Optionally, step 106 may specifically include the following steps:

[0128] 1061. Perform quantitative analysis on the virtual point cloud, calculate the point density distribution of the virtual point cloud in the unit volume spatial grid, and use the median value of the point density distribution as the benchmark value of ballast compaction.

[0129] 1062. Based on the dielectric constant distribution characteristics, curvature change law data and loose region weight values ​​in the enhanced point cloud model, establish the slab region determination conditions. The slab region determination conditions include three independent determination conditions. The first determination condition is to mark the region in the enhanced point cloud model where the dielectric constant characteristic value is continuously higher than a set threshold. The second determination condition is to extract the flat region where the curvature change amount is lower than the curvature threshold in the curvature change law data. The third determination condition is to locate the region where the weight value is lower than the critical weight value in the loose region weight value distribution.

[0130] 1063. Determine the region that simultaneously satisfies the first, second, and third judgment conditions as the hardening region, extract the spatial turning point coordinates of the outer contour of the hardening region, and connect the spatial turning point coordinates to form a three-dimensional polygon bounding box.

[0131] 1064. The ballast compaction benchmark value is output as the ballast compaction value, and the set of vertex coordinates of the three-dimensional polygon bounding box is output as the spatial coordinate range of the compacted area. The ballast compaction value and the spatial coordinate range of the compacted area together constitute the key identification result of the environmental status of the ballast structure along the railway line.

[0132] In the above scheme, point density distribution refers to a quantitative indicator reflecting the compactness of ballast, including the statistical characteristics of the number of virtual point clouds within a unit volume spatial grid, which can be used to map the benchmark value of ballast compactness. The benchmark value of ballast compactness refers to a normalized value characterizing the overall compactness of ballast, including a scalar value linearly transformed by the median of point density, which can be used to assess the compactness of the ballast layer structure. The slab compaction region determination criteria refer to the triple logical constraints for identifying ballast slab compaction characteristics, including the synergistic mechanism of dielectric constant threshold, curvature flatness threshold, and weighted critical values, which can be used to accurately locate slab compaction regions. The three-dimensional polygonal bounding box refers to the geometric structure describing the spatial extent of the slab compaction region, including the set of convex polyhedron vertices generated by connecting the spatial inflection points of the outer contour, which can be used to define the machine-readable boundary of the slab compaction region. The coordinates of the spatial inflection points refer to the key location points of the slab compaction region boundary, including the corner points of the horizontally projected polygon and the depth extension extreme points, which can be used to construct the geometric framework of the three-dimensional bounding box.

[0133] In this embodiment, the complete virtual point cloud is first divided into spatial meshes in step 1061, with 0.1m³ cube units covering the detection area, and the number of point clouds contained in each unit is counted. The formula for calculating the point density per unit volume is: Where N is the number of points, and V = 0.001 m³. The median of all collected grid density values ​​is taken after sorting. The ballast is converted into a density benchmark value through linear mapping to characterize the overall density of the ballast. The conversion formula is as follows: ,in Points / m³ is the theoretical maximum value. For example, a grid (50,60) contains 158 points, and the calculated density is 158,000 points / m³; the median density of 500 grids is 152,000 points / m³, so the density baseline value is 0.76.

[0134] Secondly, a logic AND gate criterion is established by coordinating the three types of data in step 1062: First, the dielectric distribution of the enhanced point cloud is scanned, and the dielectric constant of three or more consecutive grid cells is marked. Regions with curvature greater than 12; secondly, extract curvature data with curvature changes less than 0.05mm. - ¹ Flat regions; finally, low-risk cells with a loose weight value < 1.0 are located. A candidate nodal region is marked only if the spatial location simultaneously meets three independent conditions. For example, at coordinates (120.5, 8.7), the dielectric value is 14.2 > 12, and the curvature change is 0.04 mm. - ¹<0.05, loose weight value 0.9<1.0, is marked as a candidate slab unit.

[0135] Next, in step 1063, three-dimensional boundary extraction is performed on the candidate slab region: the outer corner points of the horizontal projection of the region are identified by the edge detection algorithm, and the depth range is determined by combining the dielectric anomaly peak value. The convex hull algorithm is used to connect spatial turning points: horizontal planes connect corner points to form polygons, which are then extended vertically to depth extrema to generate a minimum volume 3D bounding box. For example, corner points (120.3,8.6)(120.7,8.6)(120.7,8.9)(120.3,8.9) and depths of 0.3-0.6m generate an 8-vertex bounding box {(120.3,8.6,0.3)...(120.3,8.9,0.6)}.

[0136] Finally, the analysis results are integrated through step 1064. The compaction benchmark value obtained in step 1061 is used as the ballast compaction value; the set of vertex coordinates of the three-dimensional bounding box generated in step 1063 is used as the spatial range of the compacted area. These are encapsulated into key-value pairs in a machine-readable format to form a machine-parseable railway ballast condition diagnostic report, providing data support for maintenance decisions. For example, the key identification results are encapsulated in JSON format: {"Compaction":0.76,"Compacted Area":[[120.3,8.6,0.3],[120.7,8.6,0.3],...]}.

[0137] In practical applications, during the inspection of section W of the V railway, the virtual point cloud was divided into 0.1m³ grid cells. A grid (60, 70) yielded 162 points with a density of 162,000 points / m³. The median density of 500 grid cells was 154,000 points / m³. Linear mapping yielded a compaction baseline value of 0.77. The area at coordinates (130.2, 9.5) simultaneously met three conditions: dielectric value 15.1 > 12, curvature 0.03 mm⁻¹ < 0.05, and weight 0.85 < 1.0, and was thus marked as a candidate area for compaction. Edge detection was used to extract its outer corner point (130.0). The data points are defined as follows: (130.0, 9.3), (130.4, 9.3), (130.4, 9.7), (130.0, 9.7) and a depth of 0.4-0.7m. An 8-vertex bounding box {(130.0, 9.3, 0.4)..(130.0, 9.7, 0.7)} is generated using the convex hull algorithm. The final output is a structured result {"Density": 0.77,"Condensed Area": ​​[[130.0, 9.3, 0.4], [130.4, 9.3, 0.4], ... [130.0, 9.7, 0.7]]}, completing the full diagnosis from data quantization to spatial positioning.

[0138] The above-mentioned overall scheme 106 quantifies the overall compaction status of ballast through point density distribution, breaking through the limitations of traditional qualitative assessment; based on multi-condition collaborative judgment of dielectric characteristics, curvature flatness and weight values, it accurately identifies the spatial range of compaction areas; the final output structured data provides a locationable and quantifiable decision basis for railway maintenance.

[0139] The following is a complete example for steps 101-106, such as Figure 2As shown, when the track inspection vehicle enters section Y at a speed of 25 km / h, it simultaneously activates laser scanning and millimeter-wave radar: the line-scanning laser module collects point clouds of the ballast surface at a frequency of 200 Hz, generating 12,500 three-dimensional coordinate points per frame; at the same time, the IMU records the vehicle speed of 25 km / h and the pitch angle of -0.8°; the millimeter-wave radar detects the dielectric constant at a position 20 cm below the coordinates (85.32, 4.15). =14.7 is an abnormally high value. The three types of data are precisely aligned using the unified timestamp "2023-09-15 10:30:25.456".

[0140] Subsequently, dynamic compensation of the laser point cloud was performed based on IMU state parameters: For the point (85.32, 4.15, 0.21), elevation compensation ΔZ = 0.018m was calculated, and the corrected coordinates were (85.32, 4.15, 0.192); the radar dielectric value was then... =14.7 Bind this point to generate a composite data unit (85.32,4.15,0.192,14.7). 8,200 enhanced point cloud units for the entire region have been constructed, among which point (86.10,4.22,0.180,3.2) represents dry gravel.

[0141] Next, the spiral scan and penetration scan fusion is initiated: the laser beam scans the surface along the Archimedean spiral trajectory at an angular velocity of 6πrad / s, generating a point (85.32, 4.15, 0.192) at t=0.7s; simultaneously, the radar is triggered to detect the dielectric profile [3.1, 4.5, 14.7] (0-30cm) below this point; the data is projected onto the grid (853, 415) to store the data {height 0.192m, dielectric profile [3.1, 4.5, 14.7]}, forming a three-dimensional spatial matrix of 200×150 grid.

[0142] Then, the neural network analyzes the structural features: the primary module calculates the curvature of the neighborhood of point (85.32, 4.15) as 0.35 mm⁻¹; the advanced module identifies a set of gaps with a volume of 2.8 m³ at (85.32, 4.15); the spatiotemporal attention mechanism monitors that the gap has shifted by 0.22 m compared to the previous week and its volume has continuously increased by 11%, and applies a spatial increment of 1.5 and a time factor of 1.2³ = 1.728, increasing the overall weight to 2.592 to mark the high-risk loose area.

[0143] Generative adversarial model to complete occluded regions: Input occlusion boundary points The coarse-grained channels generate a 15×15 basic mesh; the fine-grained channels add perturbations. The point (85.32, 4.12, 0.19) is converted to (85.32, 4.12, 0.198); the discriminator verifies that the curvature difference is 0.012 < 0.02 threshold and the gap rate is 37% < 40%, and outputs 1,500 supplementary points to stitch together with the original 9,800 points to form a complete virtual point cloud.

[0144] Finally, the quantitative analysis and results are as follows: The virtual point cloud is meshed, with a median density of 154,000 points / m³, resulting in a density of 0.77; at (85.32, 4.15), the following conditions are simultaneously met: =14.7>12, curvature 0.04<0.05, weight 0.85<1.0 three conditions; extract corner points (85.30,4.10)(85.35,4.10)(85.35,4.20)(85.30,4.20) and depth 0.3-0.6m to generate an 8-vertex bounding box. The final output structured result is: {"Density":0.77,"Condensed region":[[85.30,4.10,0.3],[85.35,4.10,0.3],…[85.30,4.20,0.6]]}.

[0145] Figure 3 This application provides a schematic diagram of the structure of an artificial intelligence-based railway environment recognition system, as shown in the embodiments. Figure 3 As shown, the system includes:

[0146] The acquisition module 31 is used to acquire laser point cloud data of the track ballast area and simultaneously collect the operating status parameters of the track inspection vehicle and the echo data of the millimeter-wave radar.

[0147] The construction module 32 is used to perform dynamic distortion compensation on the laser point cloud data based on the operating state parameters, and to construct an enhanced point cloud model containing physical properties by combining the dielectric constant characteristics in the echo data.

[0148] The generation module 33 is used to perform a spiral scan on the surface of the ballast area by the line scanning laser module at the bottom of the track inspection vehicle, and simultaneously trigger the millimeter-wave radar to perform a penetrating scan, so as to generate a three-dimensional spatial data matrix that integrates the surface morphology and internal dielectric characteristics.

[0149] The analysis module 34 is used to analyze the curvature variation data of the contact surface of ballast particles in the three-dimensional spatial data matrix using a neural network structure and to identify the spatial distribution characteristics of the gaps between ballast particles. At the same time, it dynamically enhances the weight of loose areas through a spatiotemporal attention mechanism.

[0150] The completion module 35 is used to construct a multi-scale point cloud generation model based on the adversarial training mechanism, and to geometrically complete the missing occluded areas in the spatial distribution features to output a virtual point cloud with a complete ballast stacking state.

[0151] Output module 36 is used to perform quantitative analysis on the virtual point cloud, and combine the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature change law data and the weight of loose area to output the ballast compaction value and the spatial coordinate range of the compacted area, which serves as a key identification result of the ballast structure environmental status along the railway line.

[0152] Figure 3 The aforementioned AI-based railway line environment recognition system can perform... Figure 1 The implementation principle and technical effects of the AI-based railway environment identification method described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the AI-based railway environment identification system in the above embodiments perform operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0153] In one possible design, Figure 3 The illustrated embodiment of an artificial intelligence-based railway line environment recognition system can be implemented as a computing device, such as... Figure 4 As shown, the computing device may include a storage component 41 and a processing component 42;

[0154] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 42.

[0155] The processing component 42 is used for the above Figure 1 The embodiment describes an artificial intelligence-based method for identifying the environment along a railway line.

[0156] The processing component 42 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0157] Storage component 41 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0158] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0159] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0160] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0161] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0162] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown is an artificial intelligence-based method for identifying the environment along a railway line.

[0163] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0164] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0165] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying railway line environments based on artificial intelligence, characterized in that, include: Acquire laser point cloud data of the track ballast area, and simultaneously collect the operating status parameters of the track inspection vehicle and the echo data of the millimeter-wave radar; Based on the operating state parameters, dynamic distortion compensation is performed on the laser point cloud data, and combined with the dielectric constant characteristics in the echo data, an enhanced point cloud model containing physical properties is constructed. The surface of the ballast area is spirally scanned by the line-scanning laser module at the bottom of the track inspection vehicle, and the millimeter-wave radar is simultaneously triggered to perform a penetrating scan to generate a three-dimensional spatial data matrix that integrates the surface morphology and internal dielectric characteristics. A neural network structure is used to analyze the curvature variation data of the contact surface of ballast particles in the three-dimensional spatial data matrix and identify the spatial distribution characteristics of the gaps between ballast particles. At the same time, the weight of loose areas is dynamically enhanced through a spatiotemporal attention mechanism. A multi-scale point cloud generation model is constructed based on an adversarial training mechanism to geometrically complete the missing occluded areas in the spatial distribution features in order to output a virtual point cloud with a complete ballast stacking state. The virtual point cloud is quantitatively analyzed. Combining the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature change law data, and the weight of loose regions, the ballast compaction value and the spatial coordinate range of the compacted area are output, which serve as key identification results of the ballast structure environmental status along the railway line.

2. The method according to claim 1, characterized in that, The multi-scale point cloud generation model based on the adversarial training mechanism performs geometric completion on the missing occluded areas in the spatial distribution features to output a virtual point cloud representing the complete ballast stacking state, including: A multi-scale point cloud generation model is constructed based on an adversarial training mechanism, which includes a generator unit and a discriminator unit. The generator unit is equipped with a coarse-grained generation channel and a fine-grained generation channel, and the discriminator unit is equipped with a geometric continuity verification module. The occlusion region in the spatial distribution features is input into the generator unit. The basic geometry of the occlusion region is predicted through the coarse-grained generation channel, and the surface details of the ballast particles are added through the fine-grained generation channel to generate supplementary point cloud data of the occlusion region. The supplementary point cloud data and the original spatial distribution features are input into the discriminator unit. The geometric continuity verification module detects the curvature connection state between the supplementary geometry and the adjacent region, and verifies the physical rationality of the particle gap transition, and outputs the discrimination result. When the discrimination result fails the verification, the generator unit parameters are iteratively optimized to regenerate supplementary point cloud data. When the discrimination result passes the verification, the supplementary point cloud data is spliced ​​with the original spatial distribution features to form a complete three-dimensional point set data of the ballast stack state as a virtual point cloud output.

3. The method according to claim 1, characterized in that, The method of dynamically enhancing the weights of loose regions through a spatiotemporal attention mechanism includes: A spatiotemporal attention mechanism is constructed based on the spatial distribution characteristics of ballast particle gaps; The weight adjustment increment is increased for the ballast particle gap region with high frequency spatial location changes, and the time cumulative enhancement factor is applied to the ballast particle gap region that continues to expand over time. The weight values ​​of loose regions in the spatial distribution features are increased based on the weight adjustment increment and time accumulation enhancement factor.

4. The method according to claim 1, characterized in that, The virtual point cloud is quantitatively analyzed, and combined with the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature variation data, and the weight of loose regions, the ballast compaction value and the spatial coordinate range of the compacted area are output. These outputs serve as key identification results for the environmental condition of the ballast structure along the railway line, including: Quantitative analysis is performed on the virtual point cloud to calculate the point density distribution of the virtual point cloud within a unit volume spatial grid, and the median value of the point density distribution is used as the benchmark value for ballast compaction. By combining the dielectric constant distribution characteristics, curvature variation data, and loose region weight values ​​in the enhanced point cloud model, a slab region determination condition is established. The slab region determination condition includes three independent determination conditions: the first determination condition is to mark the region in the enhanced point cloud model where the dielectric constant characteristic value is continuously higher than a set threshold; the second determination condition is to extract the flat region where the curvature variation is lower than the curvature threshold in the curvature variation data; and the third determination condition is to locate the region where the weight value is lower than the critical weight value in the loose region weight value distribution. The region that simultaneously satisfies the first, second, and third judgment conditions is identified as the hardening region. The spatial inflection point coordinates of the outer contour of the hardening region are extracted, and the spatial inflection point coordinates are connected to form a three-dimensional polygon bounding box. The ballast compaction benchmark value is output as the ballast compaction value, and the set of vertex coordinates of the three-dimensional polygon bounding box is output as the spatial coordinate range of the compacted area. The ballast compaction value and the spatial coordinate range of the compacted area together constitute the key identification result of the environmental status of the ballast structure along the railway.

5. The method according to claim 1, characterized in that, The method involves using a linear scanning laser module at the bottom of the track inspection vehicle to perform a spiral scan of the ballast area, simultaneously triggering a millimeter-wave radar to perform a penetrating scan, thereby generating a three-dimensional spatial data matrix that fuses surface morphology and internal dielectric characteristics, including: The galvanometer deflection component of the line-scanning laser module at the bottom of the track inspection vehicle moves along an Archimedean spiral trajectory, driving the laser beam to cover the surface of the ballast area in a spiral path, and generating a set of three-dimensional coordinate points of the ballast surface morphology by receiving the reflected laser beam. At the start of laser beam scanning, a trigger pulse is sent to the millimeter-wave radar controller to drive the millimeter-wave radar to emit a penetrating beam. After receiving the reflected signal inside the ballast layer, the dielectric constant characteristic values ​​of different depth positions of the ballast layer are extracted. Based on the set of three-dimensional coordinate points on the surface, the surface coordinate points are vertically projected onto the underground coordinate system. According to the dielectric constant characteristic value of each depth layer below the projection point, the spatial mapping relationship between the three-dimensional coordinate points on the surface and the depth layer below the underground is established. Based on the spatial mapping relationship, the data is integrated according to the spatial grid. The surface height value and the corresponding underground dielectric constant profile value are stored with the surface projection point as the center position. All spatial grid units are aggregated to generate a three-dimensional spatial data matrix.

6. The method according to claim 1, characterized in that, The method of using a neural network structure to analyze the curvature variation data of the ballast particle contact surface in the three-dimensional spatial data matrix and identify the spatial distribution characteristics of the ballast particle gaps includes: The three-dimensional spatial data matrix is ​​processed by a neural network structure. The surface normal vector change sequence of the ballast particle contact surface area is extracted in the primary processing module of the neural network structure. The curvature change quantization value is calculated based on the surface normal vector change sequence to generate ballast particle contact surface curvature change law data. In the advanced processing module of the neural network structure, a continuous spatial region with constant curvature quantization value is detected within the three-dimensional spatial data matrix. The continuous spatial region is identified as a set of ballast particle gaps. The three-dimensional spatial distribution parameters of the set of ballast particle gaps are measured to generate spatial distribution characteristics of ballast particle gaps.

7. The method according to claim 1, characterized in that, The process of dynamically compensating for distortion in the laser point cloud data based on the operating state parameters, and constructing an enhanced point cloud model containing physical properties by combining the dielectric constant characteristics in the echo data, includes: Based on the longitudinal displacement, lateral offset and pitch angle values ​​in the operating status parameters, a position compensation vector is generated for each three-dimensional coordinate point in the laser point cloud data. Based on the position compensation vector, the original coordinates of the laser point cloud data are translated and corrected point by point to obtain a set of corrected three-dimensional coordinate points; The dielectric constant characteristic values ​​of the ballast layer at different depths are analyzed from the echo data of the millimeter-wave radar, and the dielectric constant characteristic values ​​are used as new physical attributes and mapped to the corresponding points in the set of corrected three-dimensional coordinate points according to their spatial locations. For each corrected 3D coordinate point, the attribute dimension is expanded to generate a composite data unit that simultaneously contains spatial coordinates and dielectric constant values. All composite data units are aggregated to form an enhanced point cloud model.

8. A railway line environment recognition system based on artificial intelligence, characterized in that, include: The acquisition module is used to acquire laser point cloud data of the track ballast area and simultaneously collect the operating status parameters of the track inspection vehicle and the echo data of the millimeter-wave radar. A construction module is used to perform dynamic distortion compensation on the laser point cloud data based on the operating state parameters, and to construct an enhanced point cloud model containing physical properties by combining the dielectric constant characteristics in the echo data. The generation module is used to perform a spiral scan on the surface of the ballast area using the line-scanning laser module at the bottom of the track inspection vehicle, and simultaneously trigger the millimeter-wave radar to perform a penetrating scan, so as to generate a three-dimensional spatial data matrix that integrates the surface morphology and internal dielectric characteristics. The analysis module is used to analyze the curvature variation data of the contact surface of ballast particles in the three-dimensional spatial data matrix using a neural network structure and to identify the spatial distribution characteristics of the gaps between ballast particles. At the same time, it dynamically enhances the weight of loose areas through a spatiotemporal attention mechanism. The completion module is used to construct a multi-scale point cloud generation model based on the adversarial training mechanism, and to geometrically complete the missing occluded areas in the spatial distribution features to output a virtual point cloud with a complete ballast stacking state. The output module is used to perform quantitative analysis on the virtual point cloud, and combine the dielectric constant distribution characteristics in the enhanced point cloud model, the curvature change law data and the weight of loose areas to output the ballast compaction value and the spatial coordinate range of the compacted area, which serves as a key identification result of the ballast structure environmental status along the railway line.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement an artificial intelligence-based railway environment identification method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The system contains a computer program that, when executed by a computer, implements an artificial intelligence-based method for identifying railway line environments as described in any one of claims 1 to 7.

Images