Track geometry intelligent analysis and adjustment amount calculation method

Abstract

The present application belongs to the technical field of intelligent analysis and measurement, and discloses a kind of track geometric parameter intelligent analysis and adjustment quantity calculation method.The method synchronously collects laser radar point cloud and binocular sequence image, and constructs track three-dimensional real scene model by feature matching, beam adjustment and multi-modal registration fusion;Continuous mileage is converted into graph structure data, and parameter error is corrected by combining continuity constraint and Euler beam dynamics regularization with graph neural network;Based on finite element digital twinning and reinforcement learning agent, the optimal adjustment quantity is quickly solved by combining Gaussian process proxy model and active learning with geometric compliance and stress safety as constraints.The present application eliminates non-physical jump and stress rebound, improves detection accuracy and adjustment rationality, and is suitable for intelligent detection and fine maintenance operation of railway track.

Description

Technical Field

[0001] This invention belongs to the field of intelligent analysis and measurement technology, and discloses a method for intelligent analysis of track geometry parameters and calculation of adjustment quantities. Background Technology

[0002] Current track geometry parameter measurements primarily rely on a single lidar scanner mounted on a detection platform to scan the track surface and acquire point cloud data. In the data processing stage, conventional methods extract rail feature points from the point cloud and perform independent calculations for each cross-section or single point to derive geometric parameters such as gauge, level, and elevation. After determining the current track geometry, the adjustment amount calculation method typically uses geometric parameter deviation as a single indicator, combined with standard limits for linear interpolation or empirical formula derivation, directly outputting the required adjustment of shim thickness or track realignment amount.

[0003] The aforementioned existing technologies suffer from a core problem: conventional solutions sever the connection between the mechanical properties of the track as a continuous physical structure and geometric measurements, resulting in the calculated maintenance amount failing to maintain track smoothness after execution. Specifically, independently calculating single-section parameters disrupts the continuous topological constraints of the track across adjacent mileages, leading to non-physical jumps in the extracted geometric parameters. When directly deriving maintenance amounts based on deviation limits using these geometric parameters with jump errors, the stress redistribution and deformation response within the track caused by maintenance operations are not incorporated into the calculation system. This causes the calculation results to deviate from the actual stress boundaries of the track, resulting in localized stress concentrations and deformation rebound after maintenance. Summary of the Invention

[0004] The purpose of this invention is to provide a method for intelligent analysis of track geometry parameters and calculation of adjustment amounts, which can effectively solve the problems in the background art mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for intelligent analysis of track geometry parameters and calculation of adjustment quantities includes: Simultaneously acquire lidar point cloud data and binocular sequence image data of the track detection area; Feature matching and bundle adjustment are performed on sequential image data to generate a high-density textured mesh model. The textured mesh model is then registered with point cloud data and fused to construct a 3D reality model. Initial values ​​of geometric parameters of each section of the track are extracted based on a 3D real-scene model, and the correlation of continuous track mileage is transformed into graph structure data. By inputting graph structure data into a graph neural network topology feature learning model, the topological correlation features of geometric parameters of adjacent track mileage segments are learned, and the initial values ​​of geometric parameters are corrected to obtain accurate measurement values. The precise measured values ​​are input into the finite element digital twin model of the track structure to simulate the deformation response and stress state under different adjustment amounts. With the geometric parameters meeting the standards and the structural stress being safe as constraints, the optimal adjustment amount sequence for the entire mileage is obtained.

[0006] Preferably, the coordinate registration and data fusion of the texture mesh model and point cloud data includes: Extract multimodal local features from textured mesh models and geometric spatial features from point cloud data, and calculate the cross-modal mutual information matrix between multimodal local features and geometric spatial features; The maximum mutual information registration objective function is constructed based on the cross-modal mutual information matrix, and the optimal rigid body transformation matrix between the texture mesh model and the point cloud data is solved by the manifold space gradient descent algorithm. After coarse registration based on the optimal rigid body transformation matrix, a feature-level fusion metric based on the scatter matrix is ​​introduced to map the pixel-level grayscale features of the texture mesh and the spatial reflection intensity features of the point cloud to the joint probability space, thereby generating a high-precision 3D real-scene model that integrates multi-source heterogeneous features.

[0007] Preferably, the step of converting the continuous track mileage correlation into graph structure data includes: The initial values ​​of the geometric parameters of each section of the track are used as the initial feature vectors of the graph nodes, and the initial edge connections are constructed based on the spatial adjacency relationship between the sections. A time-series long short-term memory network is introduced to extract the dynamic evolution trend features of orbital geometric parameters in the time dimension, and the dynamic evolution trend features are embedded into the initial edge connections; Based on edge connections embedded with dynamic evolution trend features, a multi-head self-attention mechanism is used to calculate the spatial attention weights between nodes at different cross sections, remove topological noise edges caused by vibration of the detection platform, and output high-dimensional graph structure data with dynamic adaptive weights.

[0008] Preferably, the topological association features of the geometric parameters of adjacent mileage segments of the learning track include: During the training process of the graph neural network topology feature learning model, a dual-driving loss function is constructed, which includes orbital continuity constraint loss and physical dynamics regularization loss. Among them, the physical dynamics regularization loss is constructed based on the Euler beam differential equation in track structure mechanics. The hidden layer feature vector output by the graph neural network is substituted into the Euler beam differential equation to calculate the residual, and the residual is used as a penalty term to constrain the feature update direction. By backpropagating through a dual-drive loss function, the graph neural network is guided to extract topological correlation features while satisfying the physical deformation law of the orbit, so as to correct non-physical jump errors in the initial values ​​of geometric parameters.

[0009] Preferably, the step of obtaining the optimal repair quantity sequence for the entire mileage under the constraint that the geometric parameters after repair meet the standards and the structure is safe under stress includes: The decision-making process of track adjustment quantity sequence is modeled as a Markov decision process, with the state space being the current accurate measurement value, the action being the operation of applying the adjustment quantity, and the reward function combining geometric parameter deviation penalty and structural stress safety margin. Construct a reinforcement learning agent based on proximal policy optimization and conduct trial-and-error interaction in a virtual simulation environment built with a digital twin model; By leveraging the probability distribution of action policies output by reinforcement learning agents to guide the search direction of adjustment quantities, the traditional traversal solution is transformed into directional sampling in the policy space, accelerating convergence to the optimal adjustment quantity sequence for the entire mileage that satisfies the constraints.

[0010] Preferably, before extracting the multimodal local features of the textured mesh model and the geometric spatial features of the point cloud data, the method further includes: To address the local missing regions in point cloud data caused by track surface corrosion or debris occlusion, a point cloud generation network based on a variational autoencoder is constructed. The boundary point cloud of the locally missing region is used as a priori condition input into the point cloud generation network. Combined with the adversarial training mechanism, a probability distribution that conforms to the real surface morphology of the track is used to complete the point cloud. By incorporating the probability distribution-completed point cloud into the original point cloud data, the interference of missing regions on the calculation of cross-modal mutual information matrix is ​​eliminated, thereby improving the robustness of subsequent feature-level fusion metrics.

[0011] Preferably, before outputting the high-dimensional graph structure data with dynamically adaptive weights, the method further includes: The intermediate graph structure containing dynamic adaptive weights is projected onto the Riemannian manifold space to extract the intrinsic geometry of the graph structure data in the manifold space. Calculate the graph Laplacian matrix corresponding to the graph structure data, and construct manifold smoothing constraints based on graph Laplacian regularization; Under the constraint of manifold smoothing, the node features in the intermediate graph structure are subjected to diffusion smoothing to filter out high-frequency topological fluctuations caused by random noise from optical measurements, while retaining low-frequency topological manifold features that reflect the true smoothness of the track, thereby outputting high-dimensional graph structure data.

[0012] Preferably, the training process of the graph neural network topology feature learning model further includes: Multiple edge computing nodes deployed in different track segments use local graph structure data to train local copies of graph neural network topology feature learning models; After each edge computing node calculates the gradient of the dual-drive loss function locally, a differential privacy mechanism is introduced to add calibration noise to the gradient, generating an encrypted gradient. Each edge computing node uploads encrypted gradients to the central server for aggregation and updating. The central server then distributes the updated global model parameters to each edge computing node, enabling privacy-preserving transfer and collaborative learning of cross-track segment topological association feature knowledge.

[0013] Preferably, the trial-and-error interaction in the virtual simulation environment constructed by the digital twin model includes: In a virtual simulation environment, a small number of adjustment quantity sequence samples are extracted and input into a finite element digital twin model to obtain the corresponding deformation response and stress state simulation results. Using the adjustment quantity sequence sample as input and the simulation results as output, a surrogate model based on Gaussian process regression is trained to replace the computationally expensive finite element digital twin model. The acquisition function in Bayesian optimization is used to evaluate the prediction mean and uncertainty of the surrogate model. During each trial-and-error interaction of the reinforcement learning agent, the acquisition function is used to actively select the adjustment sequence that maximizes the information gain for simulation verification.

[0014] Preferably, the step of actively selecting the adjustment sequence that maximizes information gain through the acquisition function for simulation verification includes: Construct an information gain calculation function based on prediction entropy to quantify the uncertainty of the surrogate model in the predicted distribution of candidate adjustment quantity sequences; When the predicted entropy of the candidate adjustment quantity sequence is greater than the set dynamic confidence threshold, it is determined that the agent model has a cognitive blind spot in the local state space, and the active learning mechanism is triggered. After the active learning mechanism is triggered, the candidate adjustment quantity sequence is input into the finite element digital twin model to perform real simulation, and the simulation results are added to the training set as high-value labeled data. The Gaussian process hyperparameters of the surrogate model are fine-tuned and updated online until the prediction entropy of the surrogate model in the full state space is lower than the dynamic confidence threshold.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention incorporates the continuous physical properties of the track into the geometric parameter calculation process by constructing graph structure data and combining it with graph neural networks to learn topological correlation features, thus eliminating non-physical jump errors caused by independent calculation of single sections. The corrected geometric parameters are input into a finite element digital twin model, and the adjustment amount is solved under the constraint of structural stress safety. The stress redistribution and deformation response caused by the adjustment work are included in the calculation closed loop, avoiding local stress concentration and deformation rebound after adjustment. This ensures that the output adjustment amount sequence conforms to the actual stress boundary of the track, maintaining the long-term smoothness of the line.

[0016] 2. This invention utilizes a variational autoencoder to generate a probability distribution to complete the point cloud, and combines a cross-modal mutual information matrix with a feature-level fusion metric based on the scatter matrix to overcome the interference of missing data from single optical measurements. In the construction and training of the graph neural network, dynamic evolution trend features are extracted and embedded into edge connections using a time-series long short-term memory network. A physical dynamics regularization loss based on the Euler beam differential equation and diffusion smoothing processing based on Riemannian manifold space are introduced to filter out high-frequency topological fluctuations, ensuring that feature updates conform to the physical laws of orbital deformation. When solving for adjustment quantities, a reinforcement learning agent based on near-end policy optimization and a Gaussian process regression surrogate model are used, combined with an active learning mechanism based on prediction entropy, replacing the traditional traversal solution of the finite element model and reducing the computational cost of the digital twin simulation process. By combining edge computing nodes and differential privacy mechanisms in model aggregation, privacy-preserving transfer of cross-segment topological feature knowledge is achieved. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the intelligent analysis and adjustment calculation of track geometry parameters of the present invention; Figure 2 This is a flowchart of the point cloud missing data completion and multimodal data registration and fusion process of the present invention; Figure 3 This is a flowchart illustrating the construction process of the continuous track mileage map structure data of the present invention; Figure 4 This is a flowchart of the topological feature learning and precise geometric parameter correction process of the present invention; Figure 5 This is a flowchart of the optimal adjustment quantity sequence decision-making process based on reinforcement learning, as presented in this invention. Figure 6 This is a flowchart of the Gaussian process proxy model and active learning accelerated simulation of the present invention. Detailed Implementation

[0018] The intelligent analysis and adjustment calculation method for track geometry parameters described in this invention can be applied to a track inspection and maintenance system. This system includes a track inspection platform equipped with a LiDAR and a binocular industrial camera, multiple edge computing nodes, a central server, a finite element simulation module, and a data storage module. The track inspection platform operates at speeds ranging from 0 to 160 km / h. The LiDAR and binocular industrial camera achieve time synchronization via a hardware synchronization trigger unit, using a 1PPS+TOD time protocol as the synchronization benchmark. The time synchronization error is no greater than 1 ms, ensuring consistent mileage correspondence between point cloud data and image data at the same spatial location.

[0019] The track geometry parameters described in this invention include, but are not limited to, track gauge, horizontality, elevation, track alignment, and track bottom slope. Track gauge is defined as the horizontal distance 16mm below the inner top surface of the left and right rail heads. Horizontality is defined as the elevation difference between the working surfaces of the left and right rail tops at the same cross-section. Elevation is defined as the vertical deviation of the rail top elevation relative to the design centerline along the mileage direction. Track alignment is defined as the lateral deviation of the track gauge point relative to the design centerline. Track bottom slope is defined as the angle between the track bottom plane and the horizontal plane.

[0020] As a preferred embodiment, please refer to the appendix. Figure 1 Simultaneously, LiDAR point cloud data and binocular sequential image data of the track detection area are acquired. The LiDAR is a 32-line mechanical LiDAR with an angular resolution of 0.1°, a scanning frequency of 10Hz, and a point cloud sampling density of no less than 50 points / meter along the mileage direction. The binocular industrial camera is a global shutter camera with a baseline distance of 200mm, a resolution of 2048×1536, a frame rate of 20fps, a lens focal length of 8mm, and an overlap rate of no less than 70% between adjacent frames to ensure no motion blur in the track surface image. During the acquisition process, the point cloud data frames and image data frames corresponding to the same mileage marker are bound and stored using continuous track mileage as a unique index.

[0021] Feature matching and bundle adjustment are performed on the sequential image data to generate a high-density textured mesh model. The textured mesh model is then registered and fused with point cloud data to construct a 3D reality model. Key points are extracted from the stereo sequence images using the SIFT feature operator. Bidirectional matching is used to match features between the left and right eye images and adjacent frames. The RANSAC algorithm is used to remove mismatched points and solve for the fundamental matrix. Based on the matched feature points, a bundle adjustment objective function is constructed, which minimizes the reprojection error. The expression is: ; Where, x ij Let R be the pixel coordinates of the j-th feature point in the i-th frame of the image, P be the projection function of the pinhole camera, and R be the pixel coordinates of the j-th feature point in the i-th frame of the image. i t i Let X be the rotation matrix and translation vector of the camera in the i-th frame. j Let M be the world coordinates of the j-th feature point in 3D space, M be the total number of image frames, and N be the total number of feature points. The optimal camera extrinsic 3D point coordinates are obtained through iterative solving using the Levenberg-Marquardt algorithm, generating a sparse point cloud. A dense point cloud is generated using a multi-view stereo matching algorithm, and a high-density textured mesh model is generated using a Poisson surface reconstruction algorithm, with mesh element sizes no larger than 2mm. The textured mesh model is then registered with the LiDAR point cloud data. After coarse and fine registration, the texture features and spatial geometric features are fused to construct a 3D reality model.

[0022] Initial values ​​of geometric parameters for each section of the track are extracted based on a 3D reality model, transforming the continuous track mileage relationships into graph-structured data. Along the track mileage direction, a track section perpendicular to the design centerline is set every 0.25m. For each section, the 3D coordinates of the left and right rail top working surfaces, gauge points, and rail bottom edges are extracted from the 3D reality model. Based on the extracted 3D coordinates, the initial values ​​of the gauge, horizontal, vertical, track alignment, and rail bottom slope for the corresponding section are calculated, forming a set of initial values ​​for the geometric parameters of each section. Using each section as a basic unit, a graph structure is constructed based on the mileage adjacency relationships between sections, mapping the continuous track mileage relationships to the node and edge connection relationships of the graph.

[0023] Graph-structured data is input into a graph neural network topology feature learning model to learn the topological correlation features of geometric parameters of adjacent track segments, correcting the initial values ​​of the geometric parameters to obtain accurate measurement values. The graph neural network topology feature learning model employs a graph attention network structure with residual connections, containing three graph convolutional layers, each with eight attention heads. The input consists of node features and adjacency matrices from the graph-structured data, and the output is the corrected geometric parameters for each section. During model training, a loss function constrains the feature learning direction, enabling the model to learn the continuous topological correlation features of adjacent track segments, correcting non-physical jump errors and random measurement noise in the initial values, and outputting accurate measurement values ​​of the geometric parameters.

[0024] Precise measured values ​​are input into the finite element digital twin model of the track structure to simulate the deformation response and stress state under different adjustment amounts. With the constraint that the geometric parameters meet the standards after adjustment and the structure is structurally safe, the optimal adjustment amount sequence for the entire mileage is obtained. The finite element digital twin model of the track structure is constructed using Timoshenko beam elements, the rails are discretized using beam elements, the sleepers are using rigid body elements, the fasteners are using linear spring-damped elements, and the track bed is using continuous elastic foundation elements. The model element size is consistent with the cross-sectional mileage interval, which is 0.25m. The model boundary condition is simply supported at both ends of the track. Parameters such as fastener stiffness and track bed stiffness are assigned values ​​based on the measured track structure parameters on site. Different adjustment quantity sequences are input into the model to simulate the corresponding rail deformation response, bending stress, fastener stress, and other results. The constraint condition is that the geometric parameters after adjustment meet the limit requirements of TB10913-2019 "Railway Line Maintenance Rules", and the rail bending stress and fastener stress do not exceed the corresponding allowable values. With the optimization objective of minimizing the adjustment quantity and optimizing the line smoothness, the optimal adjustment quantity sequence for the entire mileage is obtained. The adjustment quantity sequence includes the lateral track shifting amount, vertical track lifting amount, and pad thickness adjustment amount for each mileage section.

[0025] As a preferred embodiment, refer to the appendix. Figure 2Before extracting the multimodal local features of the textured mesh model and the geometric spatial features of the point cloud data, the missing regions of the LiDAR point cloud data are first filled in. The specific process is as follows: For the original point cloud data, outliers are first removed by statistical filtering, and noise points are removed by radius filtering. Then, a region growing algorithm is used to identify locally missing regions. The criteria for identifying missing regions are: the local point density in the point cloud is less than 50 points / cm², and the continuous change of the region boundary normal vector conforms to the geometric characteristics of the rail surface. A point cloud generation network based on a variational autoencoder is constructed. The network includes an encoder, a decoder, and a discriminator. The encoder uses a 3-layer graph convolutional layer, the decoder uses a 3-layer transposed graph convolutional layer, and the discriminator uses a PointNet structure. The boundary point cloud of the locally missing regions is then processed. As a priori input to the encoder, N is the number of boundary points. The encoder encodes the boundary point cloud into a 128-dimensional latent space feature vector z; the decoder decodes the feature vector z into a completed point cloud. M represents the number of completion points; the discriminator is used to distinguish the generated completion point cloud from the real orbital surface point cloud.

[0026] The total loss function of the point cloud generation network is: ; Among them, L rec To reconstruct the loss, P gt L represents the true value of the point cloud on the orbital surface. KL For KL divergence loss, , Let p(z) be the posterior distribution of the latent features output by the encoder, and p(z) be the standard normal prior distribution; L adv To combat the losses, L adv =-log(D(P gen D is the output of the discriminator; , where represents the weight coefficients. After the network converges through iterative training, the boundary point cloud of the missing region is input into the trained network to generate a probability distribution-completed point cloud that conforms to the true surface morphology of the track. The completed point cloud is then fused into the original point cloud data to eliminate the interference of the missing region on subsequent registration and fusion.

[0027] After point cloud preprocessing is completed, coordinate registration and data fusion between the texture mesh model and the point cloud data are performed. The specific process is as follows: Multimodal local features of the textured mesh model and geometric spatial features of the point cloud data are extracted. For the textured mesh model, a fast point feature histogram operator is used to extract 33-dimensional geometric features of each mesh vertex, which are then combined with the corresponding pixel-level grayscale values ​​to form a 34-dimensional multimodal local feature vector. For the preprocessed point cloud data, a fast point feature histogram operator is used to extract 33-dimensional geometric features of each point cloud point, which are then combined with the spatial reflection intensity values ​​of the point cloud to form a 34-dimensional geometric spatial feature vector.

[0028] The cross-modal mutual information matrix between multimodal local features and geometric spatial features is calculated using the following formula: I(X;Y)=H(X)+H(Y)-H(X,Y); Where X is the multimodal local feature set of the texture mesh, Y is the geometric spatial feature set of the point cloud, H(X) is the edge entropy of X, H(Y) is the edge entropy of Y, and H(X,Y) is the joint entropy of X and Y. The elements M of the cross-modal mutual information matrix M... ij =I(X i ;Y j ), where X i Y is the multimodal feature of the i-th grid vertex. j Let be the geometric feature of the j-th point cloud point, with a matrix dimension of N×M, where N is the number of grid vertices and M is the number of point cloud points.

[0029] A maximum mutual information registration objective function is constructed based on the cross-modal mutual information matrix. The objective function is as follows: L reg =-I(T(X);Y); Where T is the rigid body transformation matrix, T(X) = R∙X + t, R is a 3×3 rotation matrix, t is a 3×1 translation vector, and the optimization objective is to minimize L. reg That is, to maximize the mutual information between the transformed mesh features and the point cloud features. The rotation matrix R is mapped to the Lie algebra SO(3) space corresponding to the Lie group SO(3), and the Lie algebra vector is iteratively updated by the gradient descent algorithm in manifold space, thereby updating the rotation matrix and the translation vector. The convergence condition of the iteration is that the change in the objective function is less than 10. -6 The optimal rigid body transformation matrix T is obtained by iterating no more than 100 times. opt .

[0030] After coarse registration based on the optimal rigid body transformation matrix, a feature-level fusion metric based on the scatter matrix is ​​introduced to map the pixel-level grayscale features of the texture mesh and the spatial reflection intensity features of the point cloud to the joint probability space. The intra-class scatter matrix S is then calculated. w Inter-class scatter matrix S b The expressions are as follows: ; ; Where X is the vector set of all features to be fused, X c Let c be the set of sample vectors for the c-th feature class, where c=1 corresponds to the image grayscale feature class and c=2 corresponds to the point cloud reflectance intensity feature class. Let N be the mean vector of the c-th class of features. c Let c be the number of samples for the c-th feature. The global mean vector of all features. A fusion metric is constructed based on the Fisher discriminant criterion, maximizing the objective function J=tr(S). b ) / tr(S w Solving for the optimal linear projection matrix W, we obtain W as... The eigenvector corresponding to the largest eigenvalue. The two types of features are mapped to the joint feature space using the projection matrix W, resulting in the fused features. , where F img F is the set of image grayscale features. pc This is a set of point cloud reflection intensity features. The fused features are combined with the registered geometric spatial information to generate a high-precision 3D reality model that integrates multi-source heterogeneous features.

[0031] As a preferred embodiment, refer to the appendix. Figure 3 Using the initial values ​​of the geometric parameters of each section of the track as the initial feature vectors of the graph nodes, a graph structure G=(V,E,A) is constructed. init ), where V is the set of nodes, E is the set of edges in the graph, and A init Let be the initial adjacency matrix of the graph; Each node v i The corresponding mileage is s i The track cross sections, i=1,2,...,N, where N is the total number of cross sections; node v i initial feature vector , where g i h is the initial value of the track gauge. i v is the initial horizontal value. i As the initial values ​​for high and low, a i Let s be the initial value of the orbital direction. i The feature vector represents the mileage value and has a 5-dimensional dimension. Initial edge connections are constructed based on the spatial adjacency relationships between cross-sections, with an initial adjacency matrix A. init Let A be an N×N matrix, where A init (i,j)=1. If and only if... In the formula S j Let L be the mileage value of the j-th section, and L be the adjacent mileage threshold. In this embodiment, L = 5m; otherwise, A init (i,j)=0.

[0032] A time-series Long Short-Term Memory (LSTM) network is introduced to extract the dynamic evolution trend features of orbital geometric parameters over time, where the time dimension consists of historical detection data from different detection periods for the same orbital segment. The LSTM network comprises two hidden layers, each with 64 hidden units. The input is a sequence of geometric parameters from the same cross-section over T historical detection periods, with an input dimension of T×5. The output is a 32-dimensional dynamic evolution trend feature for each cross-section. The dynamic evolution trend features are embedded into the initial edge connections, and edge e is... ij The initial weights are updated as follows: ; in, This represents the 32-dimensional dynamic evolution trend characteristics corresponding to the i-th orbital section. This represents the 32-dimensional dynamic evolution trend characteristics corresponding to the j-th orbital section. The width of the Gaussian kernel is used in this embodiment. =1, which embeds the evolution features of the time dimension into the edge connection weights.

[0033] Based on edge connections embedded with dynamic evolution trend features, a multi-head self-attention mechanism is used to calculate the spatial attention weights between nodes at different cross-sections. In this embodiment, the number of attention heads is 8. The calculation process for a single attention head is as follows: a linear transformation is performed on the node features to obtain the query vector. Key vector Value vector W q W k W v It is a linear transformation matrix with a dimension of 5×d. k d k =8; the weight of a single attention head is: ; Where h is the h-th attention head. The weights of the eight attention heads are concatenated and then subjected to a linear transformation to obtain the final spatial attention weight matrix. Topological noise edges caused by vibration of the detection platform are removed. The criteria for determining noise edges are as follows: < In this embodiment =0.01, reset the weight of the corresponding edge to 0, and obtain the intermediate adjacency matrix A containing the dynamically adaptive weights. mid Output intermediate graph structure data.

[0034] Before outputting the high-dimensional graph structure data with dynamically adaptive weights, the intermediate graph structure is smoothed using the Riemannian manifold space. The specific process is as follows: The intermediate graph structure is projected onto a Riemannian manifold space, which is a symmetric positive definite matrix manifold, and the eigenvector h of each node is represented. i Mapped to a symmetric positive definite matrix: ; in, =10 -6 Here, I is the regularization coefficient, and I is the identity matrix, ensuring that M... i It is a positive definite matrix. Calculate the graph Laplacian matrix L=DA corresponding to the graph structure data. mid Where D is the degree matrix, .

[0035] Construct manifold smoothing constraints based on graph Laplacian regularization, with the objective function as follows: ; Where H is an N×5 dimensional matrix composed of the features of all nodes. In this embodiment, the regularization coefficient is used. =0.1, the first term is the local smoothing term, and the second term is the graph Laplacian regularization term, which constrains the global smoothness of node features.

[0036] Under the constraint of manifold smoothing, the gradient descent algorithm is used to perform diffusion smoothing on the node features of the intermediate graph structure. The iterative convergence condition is L. smooth The change is less than 10 -6 The optimal node feature matrix H is obtained by iterating no more than 50 times. opt Through diffusion smoothing, high-frequency topological fluctuations caused by random noise from optical measurements are filtered out, while low-frequency topological manifold features reflecting the true smoothness of the orbit are retained, ultimately outputting high-dimensional graph structure data G. opt =(V,E opt A opt H opt ), where E opt For the optimized edge set, A opt This is the optimized adjacency matrix.

[0037] As a preferred embodiment, refer to the appendix. Figure 4 The graph neural network topology feature learning model adopts a graph attention network structure with residual connections, and the input is the high-dimensional graph structure data G output by the above embodiment. opt The input node feature is H opt The adjacency matrix is ​​A opt The model contains three graph convolutional layers, each with eight attention heads. Each layer has an output dimension of 64, and the final output layer has a dimension of 5, corresponding to the accurate measured values ​​of the corrected orbital geometric parameters.

[0038] During model training, a dual-driving loss function is constructed, comprising orbital continuity constraint loss and physical dynamics regularization loss. The specific process is as follows: Track continuity constraint loss L cont The expression used to constrain the continuity of geometric parameters between adjacent mileage segments is: ; in, The corrected geometric parameter vector for the i-th cross section in the model output, where N is the total number of cross sections. This loss function penalizes non-physical jumps in the geometric parameters of adjacent cross sections.

[0039] The physical dynamics regularization loss is constructed based on the Euler beam differential equation in track structure mechanics. The track structure is simplified to an Euler beam on an elastic foundation, and the Euler beam differential equation is: ; Where EI is the bending stiffness of the rail, w(x) is the vertical deflection of the rail at mileage x, k is the ballast bed coefficient, and q(x) is the vertical load borne by the rail. For the discrete track cross-section, the fourth derivative is approximated using a difference method, the difference scheme being: ; in, In this embodiment, the mileage interval between adjacent sections is... =0.25m. The corrected elevation parameters output by the graph neural network, i.e., the vertical deflection w. i Substitute into the Euler beam differential equation and calculate the residual: ; Where, q i Let be the vertical self-weight load of the i-th section, and be a constant. The physical dynamic regularization loss expression is: ; By using residuals as a penalty term to constrain the feature update direction, we can ensure that the geometric parameters output by the model conform to the laws of orbital physical deformation.

[0040] The total loss of the dual-drive loss function is: ; in, As the weighting coefficient, in this embodiment =0.5. The model weights are updated using the Adam optimizer via backpropagation through the total loss function, with a learning rate set to 10. -4 With a batch size of 32 and an iteration cycle of 100, the guided graph neural network extracts topological correlation features while satisfying the physical deformation law of the orbit, corrects non-physical jump errors in the initial values ​​of geometric parameters, and obtains accurate measurement values.

[0041] The model training process also includes a privacy-preserving collaborative learning process across orbital segments, specifically: Multiple edge computing nodes are deployed in different track sections. Each edge computing node corresponds to an independent railway maintenance section. Each edge computing node uses locally collected graph structure data to train a local copy of the graph neural network topology feature learning model. The local training process is the same as the process described above.

[0042] After each edge computing node locally computes the gradient of the dual-driving loss function, a differential privacy mechanism is introduced to add calibration noise to the gradient. First, gradient clipping is performed, reducing the L2 norm of the local gradient g to no more than a gradient threshold C. In this embodiment, C = 1.0. The clipped gradient is: ; Gaussian calibration noise is added to the clipped gradient, with a noise standard deviation of [value missing]. ,in For privacy budgeting of differential privacy, Generate cryptographic gradients to account for privacy failure probability: ; Where I is the identity matrix.

[0043] Each edge computing node uploads the encrypted gradient to the central server, which then uses a federated averaging algorithm to aggregate and update the gradient. The global gradient is: ; Where K is the number of edge computing nodes. This is the encrypted gradient uploaded by the k-th edge computing node. The central server updates the global model parameters based on the global gradient and distributes the updated global model parameters to each edge computing node. Each edge computing node updates its local model copy with the global model parameters, realizing privacy-preserving transfer and collaborative learning of topological association feature knowledge across track segments.

[0044] As a preferred embodiment, refer to the appendix. Figure 5 and attached Figure 6 The decision-making process for track adjustment quantity sequence is modeled as a Markov decision process, represented by a quintuple. ,in: State space S: each state This provides precise measurements of the track geometry and stress state for the current track segment. , where g t h is the accurate measured value of the track gauge for the t-th mileage segment. t v is the precise horizontal measurement value for the t-th mileage segment. t Let a be the precise elevation measurement value for the t-th mileage segment. tLet be the precise measurement value of the track orientation for the t-th mileage segment. For the bending stress of the rail, For the fastener under stress, the state dimension is 6-dimensional, and t is the mileage segment index; Action Space A: Each action This represents the maintenance amount for the current mileage segment. ,in This refers to the lateral clearance. For vertical track measurement, This is the adjustment amount for the pad thickness; the range of values ​​for this action is... , , ; State transition probability P: Indicates that in state s t Next, execute action a t Then transition to state s t+1 The probability was obtained by simulation using a finite element digital twin model of the orbital structure; Reward function R: R(s) t ,a t )=-(w1L dev +w2L strees ), where L dev This is a penalty term for geometric parameter deviation. , The adjusted geometric parameters, s limit For standard limits; L strees This is a penalty item for structural stress safety margin. , To adjust the stress on the rails and the load on the fasteners after the repair. These are the corresponding allowable values; w1=1.0 and w2=2.0 are weighting coefficients; Discount factor In this embodiment =0.95, used to balance immediate rewards and long-term rewards.

[0045] A reinforcement learning agent based on proximal policy optimization is constructed. The agent's policy network and value network adopt a shared fully connected network structure, containing three hidden layers with 128 hidden units per layer. The activation function is ReLU. The policy network outputs the Gaussian distribution mean and variance of the actions, while the value network outputs the value estimate of the state. Generalized dominance estimation is used as the advantage function, and the pruning objective function is: ; in, Let A be the probability ratio of the new strategy to the old strategy. t For the dominant function, The pruning coefficient is used. The agent interacts through trial and error in a virtual simulation environment constructed using a finite element digital twin model, collecting trajectory data and iteratively updating the policy network parameters through the pruning objective function until the policy converges. The action policy probability distribution output by the converged policy network guides the search direction for adjustment quantities, transforming the traditional traversal solution into directional sampling in the policy space, thus accelerating convergence to the optimal adjustment quantity sequence for the entire mileage that satisfies the constraints.

[0046] In the process of trial and error interaction of the intelligent agent, a Gaussian process regression surrogate model is constructed to replace the computationally expensive finite element digital twin model. The specific process is as follows: In the virtual simulation environment, the Latin hypercube sampling method is used to extract 100 initial adjustment quantity sequence samples in the action space. The samples are then input into the finite element digital twin model to obtain the corresponding deformation response and stress state simulation results, and an initial training set is constructed.

[0047] A surrogate model based on Gaussian process regression is trained. The model input is the adjustment quantity sequence sample 'a', and the output is the simulation result 'y'. The prior distribution adopts a zero-mean Gaussian distribution, and the covariance function adopts the squared exponential covariance function. ; in, Let l be the signal variance and l be the length scale. For noise variance, The Kronecker function is used. The model hyperparameters are solved by maximizing the marginal likelihood function. The surrogate model training is completed. After training, the surrogate model can quickly predict simulation results corresponding to any sequence of adjustment parameters, replacing the finite element digital twin model in the trial-and-error interaction process of reinforcement learning. The expected improvement acquisition function in Bayesian optimization is used to evaluate the prediction mean and uncertainty of the surrogate model. The acquisition function expression is: ; Among them, y min This represents the optimal target value in the current sample set. To obtain the predicted mean of the surrogate model at the adjustment quantity sequence a, during each trial-and-error interaction, the adjustment quantity sequence that maximizes the information gain is actively selected through the acquisition function for simulation verification.

[0048] The specific process of simulation verification by actively selecting the adjustment sequence that maximizes information gain through the acquisition function is as follows: Construct an information gain calculation function based on prediction entropy. The expression for prediction entropy is: ; in, H(a) represents the prediction variance of the surrogate model at the adjustment quantity sequence a. H(a) quantifies the uncertainty of the surrogate model in the predicted distribution of the candidate adjustment quantity sequence. The larger the entropy value, the higher the uncertainty.

[0049] Set dynamic confidence threshold ,in Predict the mean entropy for all candidate adjustment quantity sequences. Let H(a) be the standard deviation and k = 1.5 be the coefficient. When the predicted entropy H(a) of the candidate adjustment sequence > H... th When the agent model is found to have a cognitive blind spot in the local state space, the active learning mechanism is triggered.

[0050] After triggering the active learning mechanism, the candidate adjustment quantity sequence is input into the finite element digital twin model to perform a real simulation, obtaining the corresponding deformation response and stress state simulation results. The adjustment quantity sequence and simulation results are then added to the training set as high-value labeled data. The Gaussian process hyperparameters of the surrogate model are fine-tuned and updated online until the prediction entropy of the surrogate model in the full state space is lower than the dynamic confidence threshold H. th This ensures that the prediction accuracy of the proxy model meets the requirements.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent analysis of track geometry parameters and calculation of adjustment quantities, characterized in that, include: Simultaneously acquire lidar point cloud data and binocular sequence image data of the track detection area; Feature matching and bundle adjustment are performed on sequential image data to generate a high-density textured mesh model. The textured mesh model is then registered with point cloud data and fused to construct a 3D reality model. Initial values ​​of geometric parameters of each section of the track are extracted based on a 3D real-scene model, and the correlation of continuous track mileage is transformed into graph structure data. The process of converting the correlation between continuous track mileages into graph-structured data includes: The initial values ​​of the geometric parameters of each section of the track are used as the initial feature vectors of the graph nodes, and the initial edge connections are constructed based on the spatial adjacency relationship between the sections. A time-series long short-term memory network is introduced to extract the dynamic evolution trend features of orbital geometric parameters in the time dimension, and the dynamic evolution trend features are embedded into the initial edge connections; Based on the edge connection with embedded dynamic evolution trend features, the spatial attention weight between nodes at different cross sections is calculated using a multi-head self-attention mechanism. Topological noise edges caused by the vibration of the detection platform are removed, and high-dimensional graph structure data with dynamic adaptive weights are output. The graph structure data is input into a graph neural network topology feature learning model to learn the topological correlation features of geometric parameters of adjacent track mileage segments, and to correct the initial values ​​of the geometric parameters to obtain accurate measurement values; the learning of the topological correlation features of geometric parameters of adjacent track mileage segments includes: During the training process of the graph neural network topology feature learning model, a dual-driving loss function is constructed, which includes orbital continuity constraint loss and physical dynamics regularization loss. Among them, the physical dynamics regularization loss is constructed based on the Euler beam differential equation in track structure mechanics. The hidden layer feature vector output by the graph neural network is substituted into the Euler beam differential equation to calculate the residual, and the residual is used as a penalty term to constrain the feature update direction. By backpropagating through a dual-drive loss function, the graph neural network is guided to extract topological correlation features while satisfying the laws of orbital physical deformation, so as to correct non-physical jump errors in the initial values ​​of geometric parameters. The precise measured values ​​are input into the finite element digital twin model of the track structure to simulate the deformation response and stress state under different adjustment amounts. With the geometric parameters meeting the standards and the structural stress being safe as constraints, the optimal adjustment amount sequence for the entire mileage is obtained.

2. The method according to claim 1, characterized in that, The process of coordinate registration and data fusion between the texture mesh model and point cloud data includes: Extract multimodal local features from textured mesh models and geometric spatial features from point cloud data, and calculate the cross-modal mutual information matrix between multimodal local features and geometric spatial features; The maximum mutual information registration objective function is constructed based on the cross-modal mutual information matrix, and the optimal rigid body transformation matrix between the texture mesh model and the point cloud data is solved by the manifold space gradient descent algorithm. After coarse registration based on the optimal rigid body transformation matrix, a feature-level fusion metric based on the scatter matrix is ​​introduced to map the pixel-level grayscale features of the texture mesh and the spatial reflection intensity features of the point cloud to the joint probability space, thereby generating a high-precision 3D real-scene model that integrates multi-source heterogeneous features.

3. The method according to claim 1, characterized in that, The optimal repair quantity sequence for the entire mileage is obtained by solving under the constraint that the geometric parameters after repair meet the standards and the structure is safe under stress, including: The decision-making process of track adjustment quantity sequence is modeled as a Markov decision process, with the state space being the current accurate measurement value, the action being the operation of applying the adjustment quantity, and the reward function combining geometric parameter deviation penalty and structural stress safety margin. Construct a reinforcement learning agent based on proximal policy optimization and conduct trial-and-error interaction in a virtual simulation environment built with a digital twin model; By leveraging the probability distribution of action policies output by reinforcement learning agents to guide the search direction of adjustment quantities, the traditional traversal solution is transformed into directional sampling in the policy space, accelerating convergence to the optimal adjustment quantity sequence for the entire mileage that satisfies the constraints.

4. The method according to claim 2, characterized in that, Before extracting the multimodal local features of the textured mesh model and the geometric spatial features of the point cloud data, the method further includes: To address the local missing regions in point cloud data caused by track surface corrosion or debris occlusion, a point cloud generation network based on a variational autoencoder is constructed. The boundary point cloud of the locally missing region is used as a priori condition input into the point cloud generation network. Combined with the adversarial training mechanism, a probability distribution that conforms to the real surface morphology of the track is used to complete the point cloud. The probability distribution is used to complete the point cloud and fuse it into the original point cloud data.

5. The method according to claim 1, characterized in that, Before the output of high-dimensional graph structure data with dynamically adaptive weights, the following is also included: The intermediate graph structure containing dynamic adaptive weights is projected onto the Riemannian manifold space to extract the intrinsic geometry of the graph structure data in the manifold space. Calculate the graph Laplacian matrix corresponding to the graph structure data, and construct manifold smoothing constraints based on graph Laplacian regularization; Under the constraint of manifold smoothing, the node features in the intermediate graph structure are subjected to diffusion smoothing to filter out high-frequency topological fluctuations caused by random noise from optical measurements, while retaining low-frequency topological manifold features that reflect the true smoothness of the track, thereby outputting high-dimensional graph structure data.

6. The method according to claim 1, characterized in that, The training process of the graph neural network topology feature learning model also includes: Multiple edge computing nodes deployed in different track segments use local graph structure data to train local copies of graph neural network topology feature learning models; After each edge computing node calculates the gradient of the dual-drive loss function locally, a differential privacy mechanism is introduced to add calibration noise to the gradient, generating an encrypted gradient. Each edge computing node uploads the encrypted gradient to the central server for aggregation and update, and the central server then distributes the updated global model parameters to each edge computing node.

7. The method according to claim 3, characterized in that, The trial-and-error interaction in the virtual simulation environment constructed by the digital twin model includes: In a virtual simulation environment, a small number of adjustment quantity sequence samples are extracted and input into a finite element digital twin model to obtain the corresponding deformation response and stress state simulation results. Using the adjustment quantity sequence sample as input and the simulation results as output, a surrogate model based on Gaussian process regression is trained to replace the computationally expensive finite element digital twin model. The acquisition function in Bayesian optimization is used to evaluate the prediction mean and uncertainty of the surrogate model. During each trial-and-error interaction of the reinforcement learning agent, the acquisition function is used to actively select the adjustment sequence that maximizes the information gain for simulation verification.

8. The method according to claim 2, characterized in that, The simulation verification of the adjustment sequence that maximizes information gain by actively selecting the acquisition function includes: Construct an information gain calculation function based on prediction entropy to quantify the uncertainty of the surrogate model in the predicted distribution of candidate adjustment quantity sequences; When the predicted entropy of the candidate adjustment quantity sequence is greater than the set dynamic confidence threshold, it is determined that the agent model has a cognitive blind spot in the local state space, and the active learning mechanism is triggered. After the active learning mechanism is triggered, the candidate adjustment quantity sequence is input into the finite element digital twin model to perform real simulation, and the simulation results are added to the training set as high-value labeled data. The Gaussian process hyperparameters of the surrogate model are fine-tuned and updated online until the prediction entropy of the surrogate model in the full state space is lower than the dynamic confidence threshold.

Images