Fault detection method and system for track structure data based on knowledge graph

Abstract

The application discloses a kind of based on knowledge graph's track structure data's fault detection method and system, and the present application relates to the technical field of knowledge graph, obtains the stereoscopic model of track structure, and determines multiple initial nodes and corresponding initial associated edge in combination with the entity set and the topological relationship set between entity of track structure, further constructs corresponding initial knowledge graph, and determines corresponding space-time dynamic knowledge graph in combination with the state transition matrix of track structure, improve the accuracy of space-time dynamic knowledge graph.The space-time topological characteristics are deduced, and determine multiple fault detection events of track structure in combination with the past fault content of track structure;According to the identification of each fault detection event, determine multiple fault areas, and mark corresponding fault content, determine corresponding fault path in combination with causal mechanism, determine corresponding fault maintenance strategy according to the fault path and the current use state of track structure, improve the accuracy of fault maintenance strategy.

Description

Technical Field

[0001] This invention relates to the technical field of knowledge graphs, and in particular to a fault detection method and system based on knowledge graph-based track structure data. Background Technology

[0002] With the continuous expansion of rail transit networks and the increasing speed of trains, track structures are subjected to extremely complex coupling effects of train dynamic loads and environmental temperature loads during long-term use. Inevitably, their internal components will experience geometric deformation, material degradation, and fatigue damage. Therefore, how to accurately and efficiently detect track structure faults and make maintenance decisions is a core technical challenge in ensuring the safe operation of rail transit.

[0003] Existing methods often lack a unified and accurate spatial benchmark, making it difficult to perform in-depth spatial registration and cross-modal entity alignment of multimodal data from different detection perspectives and sampling frequencies. At the same time, when constructing the initial knowledge graph of the track, most existing methods remain at the level of simple mapping of static topological relationships, failing to fully extract the complex set of topological relationships between track components, such as spatial adjacency, functional dependence, and mechanical transmission. They also lack the ability to characterize the time-varying mechanical properties and coupling degradation trends of the track, affecting the accuracy of the spatiotemporal dynamic knowledge graph and thus the accuracy of the track structure's fault maintenance strategy. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a fault detection method and system based on knowledge graph-based track structure data.

[0005] This invention provides a fault detection method based on knowledge graph-based track structure data, comprising:

[0006] Track structure data is determined based on track structure detection. Spatial registration and cross-modal entity alignment of the track structure data are performed based on the absolute track mileage coordinate system, thereby determining the entity set of the track structure and the set of topological relationships between entities.

[0007] Obtain a three-dimensional model of the track structure, and determine multiple initial nodes and corresponding initial associated edges by combining the entity set of the track structure and the topological relationship set between entities. Further construct the corresponding initial knowledge graph, and determine the corresponding spatiotemporal dynamic knowledge graph by combining the state transition matrix of the track structure.

[0008] The node features in the spatiotemporal dynamic knowledge graph are labeled, and the node features are aggregated in the neighborhood and updated in terms of state evolution. The spatiotemporal topological features of each node are extracted, the spatiotemporal topological features are deduced, and multiple fault detection events of the track structure are determined by combining the past fault content of the track structure.

[0009] Multiple fault areas are identified based on the identification of each fault detection event, and the corresponding fault content is marked. Furthermore, the corresponding fault path is determined by combining the causal mechanism, and the corresponding fault maintenance strategy is determined based on the fault path and the current usage status of the track structure.

[0010] This invention provides a fault detection system based on knowledge graph-based track structure data, which is applied to the aforementioned fault detection method based on knowledge graph-based track structure data. The fault detection system based on knowledge graph-based track structure data includes:

[0011] The entity module is used to determine track structure data based on track structure detection. It performs spatial registration and cross-modal entity alignment on the track structure data based on the absolute track mileage coordinate system, thereby determining the entity set of the track structure and the set of topological relationships between entities.

[0012] The knowledge graph module is used to obtain a three-dimensional model of the track structure, and combine the entity set of the track structure and the topological relationship set between the entities to determine multiple initial nodes and corresponding initial associated edges, further construct the corresponding initial knowledge graph, and combine the state transition matrix of the track structure to determine the corresponding spatiotemporal dynamic knowledge graph;

[0013] The fault detection module is used to label the node features in the spatiotemporal dynamic knowledge graph, perform neighborhood aggregation and state evolution update on the node features, extract the spatiotemporal topological features of each node, deduce the spatiotemporal topological features, and determine multiple fault detection events of the track structure in combination with the past fault content of the track structure.

[0014] The fault maintenance module is used to identify multiple fault areas based on the identification of each fault detection event, mark the corresponding fault content, further combine the causal mechanism to determine the corresponding fault path, and determine the corresponding fault maintenance strategy based on the fault path and the current usage status of the track structure.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] (1) Based on the detection of the track structure, the track structure data is determined. The track structure data is spatially registered and cross-modal entity aligned based on the absolute mileage coordinate system of the track, thereby determining the entity set of the track structure and the set of topological relationships between entities. The three-dimensional model of the track structure is obtained, and multiple initial nodes and corresponding initial associated edges are determined by combining the entity set of the track structure and the set of topological relationships between entities. The corresponding initial knowledge graph is further constructed, and the corresponding spatiotemporal dynamic knowledge graph is determined by combining the state transition matrix of the track structure. The entity set of the track structure and the set of topological relationships between entities are introduced, and the initial knowledge graph and the state transition matrix of the track structure are further controlled, which improves the accuracy of the spatiotemporal dynamic knowledge graph.

[0017] (2) Mark the node features in the spatiotemporal dynamic knowledge graph, perform neighborhood aggregation and state evolution update on the node features, extract the spatiotemporal topological features of each node, deduce the spatiotemporal topological features, and determine multiple fault detection events of the track structure in combination with the past fault content of the track structure; determine multiple fault regions based on the identification of each fault detection event, mark the corresponding fault content, further combine the causal mechanism to determine the corresponding fault path, determine the corresponding fault maintenance strategy based on the fault path and the current use status of the track structure, further control the fault detection events, fully consider the fault path and the current use status of the track structure, and improve the accuracy of the fault maintenance strategy. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the fault detection method based on knowledge graph-based track structure data in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the structural composition of a fault detection system based on knowledge graph track structure data in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] Please see Figure 1 A fault detection method based on knowledge graph-based track structure data is proposed and applied to knowledge graph scenarios. The fault detection method based on knowledge graph-based track structure data includes:

[0022] Step S11: Determine track structure data based on track structure detection, and perform spatial registration and cross-modal entity alignment on the track structure data based on the absolute track mileage coordinate system, thereby determining the entity set of the track structure and the set of topological relationships between entities;

[0023] Step S12: Obtain the three-dimensional model of the track structure, and determine multiple initial nodes and corresponding initial associated edges by combining the entity set of the track structure and the topological relationship set between entities. Further construct the corresponding initial knowledge graph, and determine the corresponding spatiotemporal dynamic knowledge graph by combining the state transition matrix of the track structure.

[0024] Step S13: Mark the node features in the spatiotemporal dynamic knowledge graph, perform neighborhood aggregation and state evolution update on the node features, extract the spatiotemporal topological features of each node, deduce the spatiotemporal topological features, and determine multiple fault detection events of the track structure in combination with the past fault content of the track structure.

[0025] Step S14: Based on the identification of each fault detection event, multiple fault areas are identified and the corresponding fault content is marked. Further, the corresponding fault path is determined by combining the causal mechanism, and the corresponding fault maintenance strategy is determined based on the fault path and the current usage status of the track structure.

[0026] In step S11, the specific steps are as follows:

[0027] S111: The track structure is dynamically inspected by a comprehensive inspection train to obtain multimodal inspection data of the track structure. The multimodal inspection data is projected onto the absolute track mileage coordinate system and multi-scale spatial registration is performed. During the registration process, the geometric morphological characteristics of the track components are determined by combining a cross-modal entity alignment mechanism.

[0028] S112: Mark the semantic features of track components and perform feature-level fusion with the corresponding geometric features to eliminate spatial ambiguity under different detection perspectives and sampling frequencies, thereby outputting an optimized data combination. Based on this data combination, the track components of the track structure are mapped as graph entities. The spatial adjacency relationship, functional dependency relationship and mechanical transmission relationship between entities are extracted by combining track design specifications and line engineering rules, thereby determining the entity set of the track structure and the set of topological relationships between entities.

[0029] In step S111, the specific steps are as follows:

[0030] At this time, during the dynamic travel along the track, the integrated inspection train simultaneously acquires multi-modal detection data of the track structure through the multi-source heterogeneous sensor array it is equipped with. In this process, in order to eliminate the attitude jitter and large error of mileage accumulation caused by the dynamic operation of the train, a high-precision absolute mileage sequence is generated based on the "onboard integrated navigation system: INS / odometer fusion".

[0031] The collected "multimodal data, including but not limited to three-dimensional laser point clouds, two-dimensional high-resolution surface images, and vehicle vibration acceleration time-series signals" are uniformly projected into the "track absolute mileage coordinate system". This coordinate system takes the track design centerline as the reference axis, the absolute mileage along the line as the X-axis, the horizontal direction of the track surface as the Y-axis, and the vertical direction as the Z-axis, so that each frame of multimodal data is assigned a unique three-dimensional absolute spatial coordinate and timestamp, thereby providing a unified spatial reference for subsequent cross-modal calculations.

[0032] Under a unified coordinate system, multi-scale spatial registration is implemented to address the issues of perspective differences and uneven sampling density in multimodal data. Specifically, at the "macroscale: 100-meter-level segments," rigid registration based on "track design alignment: horizontal and vertical curves" is used to eliminate long-wavelength geometric position errors. At the "microscale: millimeter-level components," a non-rigid fine registration is performed using the Iterative Closest Point (ICP) algorithm combined with local point cloud normal vector features. During the registration process, a cross-modal entity alignment mechanism is introduced: semantic segmentation is performed based on the 2D image to extract the 2D semantic mask boundary of the component; using the camera's intrinsic and extrinsic parameter matrices, this 2D semantic mask is back-projected into the registered 3D point cloud space. Through spatial intersection operations and nearest neighbor search, the point cloud clusters are precisely anchored and fused with the same physical component in the image, thereby achieving cross-modal entity alignment at the physical component level in the absolute mileage coordinate system and eliminating spatial ambiguity caused by different detection perspectives.

[0033] Optionally, for the cross-modal entity alignment mechanism, based on the two-dimensional detection image, a semantic segmentation network is used to extract the two-dimensional semantic mask region of the track component; according to the pre-calibrated laser camera extrinsic and intrinsic parameter matrices, the two-dimensional semantic mask is back-projected onto the registered three-dimensional point cloud space to form a three-dimensional semantic projection region; then, nearest neighbor search is performed in this three-dimensional region, and the point cloud clusters belonging to the same physical component are spatially intersected with the image mask. If the proportion of the intersection volume to the total volume of the point cloud cluster exceeds a preset matching threshold (e.g., 70%), it is determined that the two are anchored as the same entity; the three-dimensional geometric coordinates, image texture features, and semantic labels of the entity are bound and stored as input for subsequent feature-level fusion.

[0034] After completing cross-modal entity alignment, the geometric features of track components are extracted based on the aligned multimodal fusion data. For the point cloud modality, the Alpha-shape algorithm or region growing method is used to extract the three-dimensional envelope mesh of the component surface, and then the macro-geometric parameters of the component, such as the gauge, orientation, and elevation of the rail, and the flatness of the track slab, as well as the micro-geometric parameters, such as the gap width of the fastener spring clips and the slope of the rail bottom, are calculated. For the image modality, surface texture and edge topology features are extracted to help define the geometric boundaries. The point cloud geometric feature vectors of the same physical component and the image edge feature vectors are concatenated at the feature level to output a high-precision multi-dimensional geometric feature matrix of the component in the absolute mileage coordinate system, which serves as the core static spatial attribute for the subsequent construction of knowledge graph nodes.

[0035] In step S112, the specific steps are as follows:

[0036] At this point, after completing cross-modal entity alignment and geometric feature extraction, the system performs semantic feature labeling on the track components. These semantic features include not only category labels defined based on the ontology library, such as rails, fasteners, and track slabs, but also surface state labels based on visual modality recognition, such as microcracks, corrosion, and peeling.

[0037] A feature-level fusion strategy is adopted, which maps low-dimensional semantic discrete features into high-dimensional continuous vectors through one-hot encoding or embedding layers. These vectors are then concatenated with the "continuous geometric feature matrix, such as curvature, gap width, and coordinate tensor" extracted from S111 in the channel dimension. To eliminate scale differences caused by different detection perspectives and temporal misalignments caused by different sampling frequencies, a multi-head cross-attention mechanism is further introduced. Geometric features are used as queries, and semantic features are used as keys and values ​​for feature interaction calculation. This allows the network to adaptively focus on key geometric abrupt regions that are sensitive to spatial ambiguity. Dimension alignment and dimensionality reduction are performed through fully connected layers, outputting an optimized data combination containing deep multimodal information, i.e., a high-dimensional fused feature vector. This data combination completely eliminates spatial ambiguity between multi-source data and becomes the only digital mirror representing the true physical state of the track components.

[0038] Based on the optimized data combination of the above output, the system performs a mapping operation from physical space to graph topology space; wherein, each independent physical component in the track structure, such as a single rail segment, a single fastening system, or a single track slab, is defined as a unique graph entity node; the high-dimensional fusion feature vector obtained above is directly used as the initial node attribute matrix of the graph entity for assignment.

[0039] For the multi-head cross-attention mechanism, the continuous geometric feature matrix of the track components extracted in step S111 is used as the query vector. The labeled discrete semantic features are mapped to high-dimensional continuous vectors as key and value vectors after one-hot encoding. To eliminate scale differences and temporal misalignments caused by different detection perspectives and sampling frequencies, the number of attention heads is set to 8, and the feature dimension of each head is one-eighth of the original feature dimension. Each attention head independently performs the following scaled dot product attention operation: calculate the dot product of the query vector and the key vector, divide by the square root of the feature dimension for scaling, normalize by Softmax to obtain the attention weight distribution, and then sum the weight distribution with the value vector to obtain the weighted output feature of that head.

[0040] The output features of the eight attention heads are concatenated along the channel dimension, and then passed through a fully connected linear transformation layer for dimension alignment and feature compression, resulting in a fused feature vector with the same dimension as the input features. In this attention mechanism, the weight and bias parameters of all linear transformation layers are obtained through end-to-end backpropagation training. The loss function used during training is a multi-task joint loss function including geometric reconstruction loss and semantic classification loss. An adaptive moment estimation optimizer is used, with an initial learning rate of 0.001 and 200 training epochs until the loss function converges.

[0041] Meanwhile, to prevent the attention mechanism from overfitting noisy data, a random deactivation layer with a dropout rate of 0.1 is introduced at the output of each attention head, and a batch normalization layer is used after the fully connected transformation layer to accelerate training convergence. Through the above-mentioned multi-head number setting, scaling dot product attention calculation, dimension concatenation and linear transformation, training strategy and regularization measures, those skilled in the art can fully realize the deep interactive fusion of geometric features and semantic features without relying on additional creative labor, thereby effectively eliminating spatial ambiguity between multi-source data and outputting optimized data combinations.

[0042] Meanwhile, the absolute mileage coordinates of the components, the detection timestamp, and "design parameters, such as material stiffness and design life" are incorporated into the node attribute set as auxiliary attributes. Through this mapping process, the continuously distributed track structure in the physical world is discretized into a set of entity nodes in the graph data structure, thus completing the construction of the underlying data foundation of the knowledge graph.

[0043] After generating the graph entity set, the system combines "track design specifications, such as the design standard for ballastless track of high-speed railway" and line engineering rules to extract multi-dimensional topological relationships between entities in order to construct a complete set of topological relationships between entities. This extraction process is specifically subdivided into three types of targeted extraction of relationships: spatial adjacency relationship extraction: collision detection and distance calculation are performed based on the absolute mileage coordinates of entity nodes and the three-dimensional spatial bounding box; when the Euclidean distance or mileage difference between two entities in space is less than the preset engineering threshold, spatial adjacency edges are established, such as "sleeper_adjacent_sleeper" and "rail_connected_rail", and the relative position vector is assigned as the edge weight.

[0044] Functional dependency extraction: Based on the track engineering rules and assembly logic tree, rule reasoning is performed. By matching the preset assembly ontology rule library, the fastening, support and constraint relationships between components are identified, such as "fastener_fastening_rail" and "track slab_support_self-compacting concrete layer". Such relationships reflect the structural integrity of the track system.

[0045] Mechanical transmission relationship extraction: Based on the train-track coupling dynamics mechanism in the track design specifications, the mechanical transmission relationship is extracted. According to the transmission path of train load in the track structure, wheel load is transmitted from top to bottom and lateral force is dispersed through fasteners. Mechanical transmission edges between entities are established, such as "rail_load transmission_fastener" and "track slab_stress diffusion_base plate". The theoretical stiffness ratio and stress attenuation coefficient are used as the initial properties of the mechanical edges.

[0046] All the extracted edges are deduplicated and normalized to form a set of topological relationships between entities, which together with the graph entity set constitute the initial knowledge graph topology network of the track structure.

[0047] In step S12, the specific steps are as follows:

[0048] S121: Retrieve the 3D model of the track structure, map the entity set to the spatial grid of the 3D model of the track structure, and use it as multiple initial nodes. Assign corresponding weights between the corresponding initial nodes according to the set of topological relationships, thereby generating the corresponding initial associated edges, and construct the corresponding initial knowledge graph in combination with the working history of the track structure.

[0049] S122: Obtain the track structure state transition matrix, which dynamically changes with train load and temperature load. Sequentially expand the node attributes in the initial knowledge graph along the time dimension, and use the spatiotemporal graph convolution operator to update the weights of the initial associated edges in a temporal evolution and nonlinear feature mapping, so that the initial knowledge graph is updated with time series into a spatiotemporal dynamic knowledge graph that can characterize the time-varying mechanical properties and coupling degradation trend of the track structure.

[0050] In step S121, the specific steps are as follows:

[0051] At this point, the system retrieves the 3D model of the track structure from the BIM (Building Information Modeling) database or digital twin platform. This 3D model contains a refined 3D geometric mesh of the components and the assembly hierarchy topology. The "entity set, i.e., the track components carrying fusion feature vectors" determined in S112 is accurately mapped to the corresponding 3D spatial mesh of the 3D model through the absolute mileage coordinate system and the spatial coordinate transformation matrix.

[0052] During the mapping process, a collision detection algorithm based on spatial voxels or bounding boxes is used to verify the matching degree between physical entities and the 3D model mesh, and to eliminate floating or penetration errors caused by detection noise. After the mapping is completed, each spatial mesh cell filled with entity data in the 3D model is defined as an "initial node" in a graph computing network, and the multidimensional fusion feature vector output by the aforementioned S112 is used as the initial state attribute of the initial node, thereby realizing the transformation from a 3D geometric model to a graph topology network node.

[0053] After generating the initial nodes, the system constructs connecting edges between the corresponding initial nodes based on the set of topological relationships between entities extracted in S112, including spatial adjacency, functional dependency, and mechanical transmission relationships, which are defined as "initial associated edges". In order to enable the graph to represent the strength of real physical interactions, the system assigns quantitative weights to different types of topological relationships.

[0054] Specifically, for spatial adjacency relationships, distance weights are assigned based on the inverse attenuation calculation of the three-dimensional Euclidean distance or mileage difference between nodes; for functional dependency relationships, discrete state weights are assigned based on assembly tolerances and tightness levels; for mechanical transmission relationships, mechanical transmission probabilities are calculated based on the theoretical stiffness ratio, stress diffusion angle, or load distribution coefficient in the track design specifications, and these probabilities are used as the initial weights of the mechanical edges. Through the above weight assignment mechanism, the system generates an initial set of associated edges that contains multiple physical semantics and has continuous weight attributes.

[0055] After completing the construction of nodes and associated edges, the system further integrates the track structure's operational history data to improve the dynamic attribute base of the graph. The operational history data includes, but is not limited to: cumulative total weight (TraficGrossTonnage, TGT), historical highest / lowest ambient temperature, existing "maintenance records, such as grinding and component replacement history", and train speed graph. The system converts the above-mentioned time-dimensional operational history data into time-series statistical feature vectors, and updates the attribute matrix of the corresponding initial node by using feature splicing or linear weighted fusion.

[0056] Meanwhile, the initial associated edge weights are dynamically corrected based on the work history data. For example, the weights of the mechanical transmission edges are reduced according to the increase of the cumulative total weight, using a nonlinear decay function to simulate stiffness degradation. The network structure, which integrates multimodal detection data, spatial three-dimensional mesh, multidimensional topological relationships, and work history data, is output as an "initial knowledge graph," providing a complete data foundation for the subsequent introduction of a state transition matrix to achieve spatiotemporal dynamics.

[0057] In step S122, the specific steps are as follows:

[0058] At this point, the system acquires external environmental input parameters of the track structure, including time-varying "train load spectrum: covering axle load, train speed, and passing frequency" and "temperature load spectrum: covering daily average temperature amplitude and rail temperature gradient changes." Based on track structure dynamics theory, the system constructs a state transition matrix A(t) that is dynamically updated with time step t. This state transition matrix A(t) aims to describe the transition probability and evolution law of the track components' "physical state, such as stiffness degradation and plastic cumulative deformation" from the current time t to the next time t+Δt under different load conditions. The element values ​​of the state transition matrix A(t) are obtained by jointly solving the "fatigue damage accumulation function caused by train load, such as the stress ratio parameter based on Miner's linear cumulative damage theory" and the thermal expansion and contraction deformation function caused by temperature load, thus enabling it to characterize the nonlinear coupling effect between external excitation and the internal mechanism of the track.

[0059] For the state transition matrix A(t), a physical state vector of the track structure at any time t is defined. This vector consists of the key physical parameters of all track components, specifically including the rail gauge deviation, track alignment irregularity amplitude, fastener spring gap, fastener stiffness coefficient, track slab flatness deviation, track slab concrete crack width, and base plate settlement displacement. Based on this, the state transition matrix A(t) is constructed as an N×N square matrix, where N equals the total dimension of the aforementioned physical state vector.

[0060] Each off-diagonal element of matrix A(t) represents the coupling influence coefficient of the j-th physical parameter on the i-th physical parameter, while the diagonal elements represent the persistence degradation coefficient of each physical parameter itself. All matrix elements are obtained by jointly solving the fatigue damage accumulation function caused by train load and the thermal expansion and contraction deformation function caused by temperature load, as detailed below:

[0061] For the train load contribution, based on Miner's linear cumulative damage theory, the cumulative fatigue damage degree of each component at the current moment is calculated according to the cumulative total weight passed, train axle load and passing frequency. This damage degree is converted into a degradation rate coefficient in the interval of 0 to 1 through a Sigmoid nonlinear mapping function, and then multiplied by a preset basic coupling coefficient to obtain the sub-matrix elements of the train load contribution.

[0062] For the temperature load contribution, the thermal expansion and contraction deformation of each component is calculated based on the current rail temperature gradient and the daily average ambient temperature amplitude. This deformation is then mapped to a temperature influence coefficient through a hyperbolic tangent function, and multiplied by a preset temperature-mechanical coupling coefficient to obtain the sub-matrix elements of the temperature load contribution.

[0063] The train load contribution submatrix and the temperature load contribution submatrix are added element by element, and then multiplied by the current time step of the combined load action (e.g., 1 hour) to obtain the complete state transition matrix A(t) for the current time step. This matrix is ​​updated once with each time step, and the current train operation data and ambient temperature data are reread during the update, thereby realizing a dynamic response to external time-varying loads.

[0064] The fatigue damage accumulation function is implemented as follows: The stress amplitude within each time step is substituted into the standard SN curve to calculate the fatigue consumption within that step, and then accumulated to the total fatigue consumption to obtain the current accumulated fatigue damage degree. The "thermal expansion and contraction deformation function" is implemented as follows: Based on the product of the linear expansion coefficient of the track component and the current temperature change, the free deformation of the component within the current time step is calculated, and then multiplied by the component constraint coefficient to correct the confined deformation under actual constraint conditions.

[0065] After obtaining the dynamic state transition matrix, the system serializes and expands the "initial knowledge graph: containing static node attributes and initial associated edge weights" generated by S121 along the time dimension; whereby, for each node in the initial knowledge graph, its fusion feature vector at the initial time t0 is defined as the initial state h(t0).

[0066] The node states are expanded along the discrete time axis t1, t2, ..., tn. Within each time step, the node state attributes of the previous time step and the external load parameters of the current time are input into the state transition matrix A(t) to calculate the predicted node state attributes at the current time. Through this process, the originally static graph node attributes are transformed into a time series tensor sequence containing historical state dependencies, enabling the graph nodes to remember and represent their own degradation process over time.

[0067] Based on the temporalization of node attributes, the system introduces a spatiotemporal graph convolution operator to perform temporal evolution updates and nonlinear feature mapping on the weights of the initial associated edges. This spatiotemporal graph convolution operator adopts graph convolution operations based on the spectral domain or spatial domain in the spatial dimension. Based on the state attributes of the current node and its neighboring nodes, it recalculates the mechanical coupling strength between nodes and outputs the dynamic adjustment coefficients of the edge weights.

[0068] Optionally, the spatiotemporal graph convolution operator consists of two concatenated parts: a spatial graph convolution module and a time series modeling module. In the spatial dimension, a spectral domain-based graph convolution operation is employed, specifically implemented using the Chebyshev polynomial approximation method. First, based on the initial set of nodes and initial set of associated edges in the initial knowledge graph, an adjacency matrix is ​​constructed. The rows and columns of this matrix correspond to each node, and the matrix elements are the weights of the associated edges assigned in step S121.

[0069] Obtain the degree matrix of the adjacency matrix, which is a diagonal matrix whose diagonal elements are the sum of the elements in each row of the adjacency matrix. Based on the adjacency matrix and the degree matrix, calculate the normalized Laplacian matrix, which is calculated as follows: the normalized Laplacian matrix equals the identity matrix minus the negative half power of the degree matrix multiplied by the adjacency matrix multiplied by the negative half power of the degree matrix. Perform a K-order approximate expansion of the normalized Laplacian matrix using Chebyshev polynomials, where K is 3, to aggregate the feature information of first-order, second-order, and third-order neighbors. The forward propagation process of each spatial graph convolutional layer is as follows: multiply the input node feature matrix by each term of the Chebyshev polynomial, then sum the results with the corresponding learnable weight parameters, and after a nonlinear transformation using the ReLU activation function, output the spatially aggregated node feature matrix.

[0070] In the time dimension, gated recurrent units or long short-term memory networks are used to extract the temporal dependency features of the edge weight change sequence. The system performs nonlinear feature mapping between the dynamic edge weight adjustment coefficient at the current moment and the initial associated edge weight in S121, such as nonlinear composition through activation functions, and outputs the dynamic associated edge weight updated over time. Through the temporal update of the above node attributes and the dynamic evolution of the associated edge weight, the initial static knowledge graph is transformed into a "spatiotemporal dynamic knowledge graph" that can accurately characterize the "variable mechanical properties of the track structure, such as stress redistribution caused by stiffness decay" and the degradation trend of multi-component coupling.

[0071] In step S13, the specific steps are as follows:

[0072] S131: In this spatiotemporal dynamic knowledge graph, the corresponding node features are marked according to the identification of the spatiotemporal dynamic knowledge graph. The node features include structural geometric deviation, vibration response and material degradation. The dynamic graph attention mechanism is used to perform multi-hop neighborhood aggregation on the node features, and the temporal degradation trajectory of the node features is determined by combining the state evolution update mechanism, so as to extract the spatiotemporal topological features of each node. The spatiotemporal topological features integrate local deformation and global mechanical transmission.

[0073] S132: Map the spatiotemporal topological features to a high-dimensional manifold space, and perform manifold learning and pattern reconstruction in the high-dimensional manifold space to output the reconstructed feature combination. At the same time, obtain the past fault content of the track structure, and perform multi-level iterations in combination with the reconstructed feature combination. During the iteration process, determine the abnormal deviation degree and fault evolution content of the current spatiotemporal topological features, and then output multiple fault detection events of the track structure at different degradation stages.

[0074] In step S131, the specific steps are as follows:

[0075] At this point, based on the spatiotemporal dynamic knowledge graph generated by S122, the system performs multi-dimensional feature identification and labeling for each graph node. The node features are specifically divided into three orthogonal dimensions: structural "geometric deviation, such as spatial morphological parameters like gauge deviation, track alignment irregularities, and relative displacement between components", "vibration response, such as dynamic stiffness parameters like power spectral density, dominant frequency drift, and vibration level extracted from onboard / trackside acceleration sensors", and "material degradation, such as surface crack width based on visual recognition, corrosion area ratio, and internal voiding rate based on ultrasonic / radar detection". The system normalizes the above continuous physical quantities and transforms them into high-dimensional feature vectors through multi-source heterogeneous sensor data analysis, assigning each node in the graph an initial multi-dimensional attribute with clear physical meaning.

[0076] To capture the coupling degradation effect between track components, the system employs a dynamic graph attention mechanism (DynamicGAT) to perform multi-hop neighborhood aggregation of node features. Unlike static graphs, this dynamic graph attention mechanism, when calculating the attention weight between node i and its neighbor node j, not only inputs the current feature vectors of both nodes but also introduces the "dynamic associated edge weights, i.e., the instantaneous mechanical transmission coefficients" generated in S122 as a prior bias for attention calculation. The system calculates the dynamic correlation between node pairs through a learnable nonlinear mapping function and uses the Softmax function for normalization to obtain attention weight coefficients. By stacking multiple layers of graph attention networks, the system achieves "multi-hop" aggregation, that is, it not only aggregates the features of directly adjacent nodes, such as rails aggregating adjacent fasteners, but also aggregates the features of "N-hop topological neighborhoods, such as rails aggregating the features of the underlying track slab and base foundation." In this process, the attention weights can adaptively focus on neighboring nodes with abnormal mechanical transmission or severe degradation, thereby suppressing the interference of redundant information.

[0077] Optionally, this dynamic graph attention mechanism consists of an input transformation layer, a dynamic attention coefficient calculation layer, a multi-hop neighborhood aggregation layer, and an output transformation layer connected in series. In the input transformation layer, the initial multidimensional feature vectors of each node in the spatiotemporal dynamic knowledge graph are mapped to high-dimensional hidden layer feature vectors of the same dimension through a learnable linear transformation weight matrix.

[0078] In the dynamic attention coefficient calculation layer, for each pair of nodes i and j connected by an associated edge, the following dynamic attention coefficients are calculated: the high-dimensional hidden feature vector of node i, the high-dimensional hidden feature vector of node j, and the current dynamic weight value of the associated edge generated in step S122 are concatenated into a joint vector. This joint vector is input into a single-layer feedforward neural network. The output of this neural network is non-linearly mapped using the LeakyReLU activation function to obtain unnormalized attention coefficients. For all first-order neighbor nodes of node i, the above unnormalized attention coefficients are calculated separately. Then, the Softmax function is used to normalize all unnormalized coefficients in the set, so that the sum of the attention coefficients from node i to each of its neighbor nodes is equal to 1, thereby obtaining the normalized dynamic attention weights of node i to each of its neighbor nodes.

[0079] In the process of calculating the attention coefficient, the weight of the dynamically associated edge is used as a prior bias in the splicing. This means that when the mechanical transmission coefficient between a neighboring node and node i decreases due to stiffness degradation or temperature influence, the neighboring node will be assigned a lower weight in the attention calculation, and vice versa. This achieves adaptive adjustment of the attention distribution to the time-varying mechanical properties of the track structure.

[0080] In the multi-hop neighborhood aggregation layer, the first-hop aggregation uses the normalized dynamic attention weights calculated above to perform a weighted summation of the high-dimensional hidden feature vectors of all first-order neighbor nodes of node i. After a nonlinear transformation by the Sigmoid activation function, the first-order aggregated feature vector of node i is output. The second-hop aggregation uses the same attention coefficient calculation method, taking the first-order aggregated feature vector as the feature input of the current node, and repeating the above attention coefficient calculation and weighted summation process to obtain the second-order aggregated feature vector of node i. This vector integrates the feature information of all second-order neighbor nodes that can be reached from node i through two associated edge steps.

[0081] After each hop aggregation is completed, a multi-head attention mechanism is used to enhance the stability of feature representation. Specifically, the number of attention heads is set to 4. Each head independently performs the above attention coefficient calculation and neighborhood aggregation process. Then, the output feature vectors of the 4 heads are concatenated in the channel dimension and compressed back to the original feature dimension through a linear transformation layer to obtain the final output of the hop aggregation.

[0082] The dynamic graph attention mechanism is set to two layers, completing the first and second hop aggregation operations mentioned above. The total number of hops corresponds to the number of hops in the multi-hop aggregation. A residual connection is set after each layer, which adds the input feature vector and the output feature vector of the current layer element-wise to alleviate the gradient vanishing problem in deep network training.

[0083] After completing the spatial dimension neighborhood aggregation, the system introduces a state evolution update mechanism to integrate the temporal dimension degradation information. Specifically, the spatial feature vector sequence obtained by each node after multi-hop aggregation is input into the state evolution operator, such as a temporal update module based on neural differential equations or gated recurrent units. This operator recursively updates the hidden state of the node by remembering historical degradation information through its internal hidden state memory and combining it with the external load input at the current moment. The system outputs the spatiotemporal topological features of each node. This feature vector is a highly compressed digital twin representation, which not only includes the node's own local deformation and material degradation, but also deeply integrates the global structural time-varying mechanical properties and multi-component coupling degradation trends transmitted through multi-hop mechanical transmission paths.

[0084] In step S132, the specific steps are as follows:

[0085] At this point, the system projects the "spatiotemporal topological feature vector: containing local deformation and global coupling degradation information" extracted by S131 onto a high-dimensional manifold space through a "nonlinear mapping function, such as a multilayer perceptron or autoencoder part". This high-dimensional manifold space aims to unfold the implicit degradation state of the orbital structure under the influence of complex multiphysics fields into a concrete geometric manifold structure. Within the manifold space, the system uses "manifold learning algorithms, such as variants of isometric mapping or local linear embedding" to calculate the geodesic distance and local tangent space coordinates between feature points to capture the nonlinear manifold structure in the state evolution process. Based on the baseline manifold distribution under the orbital health state, the system performs pattern reconstruction on the current spatiotemporal topological features. Among them, through manifold orthogonal decomposition or autoencoder decoding operations, high-frequency random observation noise in the features is filtered out, and a reconstructed feature combination reflecting the principal components of the orbital's true physical degradation is reconstructed. This reconstructed feature combination has smooth dynamic trajectory characteristics in the low-dimensional manifold tangent space, providing a pure baseline pattern for subsequent anomaly quantification.

[0086] After outputting the reconstructed feature combination, the system obtains the historical fault database of the track structure, extracts past fault content, including feature templates of typical fault types, fault degradation curves and failure thresholds; the system inputs the reconstructed feature combination and past fault content into a multi-level iterative calculation framework; in each iteration: calculate the Mahalanobis distance or reconstructed error of the reconstructed feature relative to the healthy baseline distribution in the manifold space, and determine the abnormal deviation of the current spatiotemporal topological features.

[0087] Secondly, the reconstructed features are dynamically time-warped and matched with historical fault degradation curves along the time axis to extract the evolution direction and rate of the features, and to determine the content of fault evolution, such as continuous decrease in stiffness and accelerated crack propagation. In the multi-level iterative process, the system adaptively adjusts the bandwidth parameter of the manifold kernel function based on the feedback of deviation gradient and evolution rate until the deviation converges or the preset maximum number of iterations is reached. Through this multi-level iterative process, the system not only quantifies the degree of deviation of the current state from the normal state, but also clarifies which known fault mode it is evolving along.

[0088] Based on the abnormal deviation and fault evolution content output by multi-level iteration, the system divides the entire life cycle of the track structure into stages. The system presets multi-level degradation threshold boundaries and defines the manifold state points where the deviation crosses a specific threshold and the evolution direction matches a specific fault mode as corresponding fault detection events. Among them, when the deviation is low but the evolution trend shows a negative divergence, it is marked as an early degradation event; when the deviation rises sharply and matches a known fault feature template, it is marked as an explicit fault initiation event; when the deviation exceeds the critical safety threshold, it is marked as an emergency fault event. The system outputs multiple discrete fault detection events of the track structure at different degradation stages, realizing the leap from continuous physical features to discrete semantic fault events.

[0089] In step S14, the specific steps are as follows:

[0090] S141: Dynamically identify each fault detection event, and combine spatiotemporal dynamic knowledge graph to perform spatial reverse tracing and cluster analysis during the identification process to determine multiple fault regions, and mark the corresponding fault category and severity content in each fault region;

[0091] S142: By intervening in multiple candidate nodes within the fault area through counterfactual reasoning, the false associations and confounding factors between candidate nodes are decoupled, thereby determining the actual root cause node that caused the fault and the corresponding fault propagation path. Combined with the current usage status of the track structure, multi-objective optimization is performed to achieve Pareto optimization between track operation safety risks and maintenance resource costs, so as to output a fault maintenance strategy for the fault path.

[0092] In step S141, the specific steps are as follows:

[0093] At this time, the system receives multiple fault detection events output by S132 and performs dynamic identification on them, that is, real-time analysis of the degradation stage to which the event belongs and the associated component type; the system uses the fault detection event as the trigger source to start the spatial reverse tracing mechanism in the spatiotemporal dynamic knowledge graph constructed by S122; in this process, the system anchors the fault event to the corresponding abnormal node in the graph, and uses the abnormal node as the starting point to trace back along the reverse direction of the "mechanical transmission relationship" and "functional dependency relationship" in the graph, for example, from "rail" to "fastener", and then back to "track slab", using reverse breadth-first search or reverse random walk algorithm to traverse the path.

[0094] During the traversal, the system combines the dynamic associated edge weights generated by the temporal evolution update in S122, namely the mechanical coupling strength and stiffness degradation coefficient at the current moment, to calculate the mechanical contribution and state coupling correlation of upstream and downstream nodes to the fault event. By filtering upstream and downstream associated nodes whose contribution exceeds the preset threshold, the system constructs the complete associated node sub-link that caused the fault event, and completes the spatial reverse tracing at the physical transmission topology level.

[0095] After completing the spatial reverse tracing of multiple fault detection events and obtaining the corresponding associated node sub-links, the system performs cluster analysis on all sub-links. Since track structure faults often exhibit local clustering characteristics, such as multiple fasteners failing together due to foundation settlement, the system adopts graph clustering algorithms based on topology compactness, such as the Louvain algorithm or spectral clustering variants.

[0096] During the clustering process, the system uses the "spatial adjacency relationship" and "dynamic mechanical transmission edge weight" between associated nodes as the basis for graph cuts, calculates the modularity density and mechanical transmission compactness between nodes, and maximizes the intra-cluster compactness and minimizes the inter-cluster coupling. The system aggregates nodes that are physically adjacent, highly coupled in mechanical transmission, and similar in degradation state into connected subgraphs. Each independent and closed connected subgraph is defined and divided into a "fault region", thereby realizing the mapping and definition from discrete abnormal node links to continuous physical space regions.

[0097] After delineating multiple fault regions, the system performs multi-level semantic content labeling for each fault region. The system extracts the abnormal deviation degree of all nodes in the fault region and the decay rate of the dynamic associated edge weights for comprehensive weighted calculation, thereby quantifying the degradation depth of the region. Based on the quantification results, the system labels the corresponding "fault category" at the region level, such as: geometric position exceedance, component fracture failure, foundation settlement and joint separation, determined according to the type of dominant fault event in the region" and "severity content, such as: divided into "early warning", "moderate degradation", "severe failure" etc. At the same time, the system converts the fault category and severity content into semantic tags and attaches them to the corresponding fault region subgraph of the graph, so that the abstract graph topology structure has intuitive engineering semantics to guide on-site engineering practice.

[0098] In step S142, the specific steps are as follows:

[0099] At this point, after delineating the fault area with semantic labels, the system extracts all the map nodes in the area as a set of candidate nodes; due to the strong mechanical coupling and temporal co-occurrence between the track structure components, there will inevitably be false correlations and confounding factors in the observation data, such as: temperature drop and fastener loosening occur at the same time, but temperature is not the direct physical cause of the fracture.

[0100] To decouple such confounding effects, the system introduces a counterfactual reasoning mechanism based on a structural causal model. Within the graph computation framework, the system executes a "Do operator" intervention operation. For each candidate root cause node, it hypothetically severs its observational association with the external environment and forces its state to be set to "normal, no degradation." The system calculates the reduction ratio of the overall abnormal deviation of the target fault area after the intervention operation propagates downstream along the dynamic association edges. If the reduction ratio is lower than a preset threshold, the candidate node is determined to be either an affected representation node or a false association node disturbed by confounding factors.

[0101] If the reduction ratio is significantly higher than the threshold, it proves that the state change of the node is a necessary condition for causing regional failure. By traversing all candidate nodes and comparing the differences in the graph state before and after counterfactual intervention, the system effectively eliminates co-occurring but non-causal false associations and accurately locates the "actual root cause node" that causes the failure.

[0102] After identifying the actual root cause node, the system generates a fault propagation path based on the topology of the spatiotemporal dynamic knowledge graph. Starting from the actual root cause node, the system extracts the forward mechanics transfer edge and functional dependency edge that contribute the most to reducing downstream anomalies during counterfactual intervention. Using a directed acyclic graph shortest path search algorithm, such as the improved Dijkstra algorithm, with dynamic mechanics transfer weights as path costs, the system traces the connected topology link from the root cause node to the final "fault manifestation node, such as the broken fastener node". In this process, the system combines the temporal degradation trajectory extracted in S13 to assign a time dimension attribute to the topology link, thereby reconstructing a complete "fault propagation path". This path not only clearly defines the spread of the fault in the spatial physical structure, but also depicts the temporal evolution process of the hidden danger from its inception, development to its manifestation.

[0103] After obtaining the fault propagation path, the system accesses the current usage status data of the track structure, including but not limited to: recent train operation density, allowable maintenance window time, spare parts inventory status, and environmental and climatic conditions. The system transforms these usage statuses into constraints and constructs a decision model centered on "fault path maintenance." This model sets two mutually interacting optimization objectives:

[0104] One is to maximize the elimination rate of track operation safety risks, namely the expected stiffness recovery and life extension of root cause nodes and critical propagation nodes after maintenance.

[0105] Secondly, it minimizes maintenance resource costs, including labor hours, material consumption, and indirect economic losses caused by speed limits or shutdowns. The system employs a multi-objective evolutionary algorithm, such as the non-dominated sorting genetic algorithm NSGA-II with an elite retention strategy, to perform Pareto optimization on the maintenance scheme space under constraints, such as combinations of strategies like "replacing only the apparent fasteners," "grouting and reinforcing the root cause track slab," and "replacing the entire section of components." Through iterative calculations, the system outputs a Pareto front composed of a series of non-dominated solutions and recommends the optimal fault maintenance strategy based on the current line's "operational priority, such as prioritizing safety during the Spring Festival travel rush and prioritizing cost during weekdays."

[0106] Specific application scenarios and corresponding implementation methods of this application:

[0107] In step S11, in the fault detection scenario of "high-speed railway ballastless track structure including WJ-8 type fastener system and CRTSIII type track slab", the comprehensive inspection train enters the track section at a speed of 350 km / h; the linear array laser camera mounted on the bottom of the train acquires two-dimensional grayscale images of the rails and fasteners at a frequency of 2kHz, while the three-dimensional laser scanner acquires high-density three-dimensional point clouds of the rail surface and sleeper area at a frequency of 1MHz.

[0108] The detection system combines each frame of laser point cloud and image data with data from the onboard high-precision gyroscope and projects it onto the absolute mileage coordinate system of the track. For example, the system simultaneously anchors the image frame and point cloud frame acquired at a certain moment onto the absolute mileage coordinate "K100+020.350", so that the multimodal data achieves initial alignment in absolute space.

[0109] Due to the lateral serpentine motion of the train traveling at high speed, there is a slight offset of a few millimeters between the center of the fastener in the image and the center of the fastener in the point cloud at the microscopic scale. At the macroscopic scale, the system performs overall rotation and translation correction on the data based on the design horizontal curve parameters of the K100+020 segment. Then, at the microscopic scale, the semantic mask of the "elastic bar" of the WJ-8 type fastener is extracted from the image. Using a pre-calibrated laser camera extrinsic matrix, the mask is back-projected onto the three-dimensional point cloud space. Through spatial intersection operation, the system accurately aligns the discrete point set representing the "WJ-8 elastic bar" in the point cloud with the "elastic bar" region in the image, thus realizing the cross-modal entity anchoring of this specific physical component.

[0110] For the aligned WJ-8 fastener entity at K100+020.350, the system extracts its 3D point cloud envelope and calculates the actual gap width between the elastic clip toe and the rail base as 8.5mm, while the design standard value is 10mm. At the same time, the rail base slope at this location is extracted as 1 / 38. These high-precision geometric features, together with the absolute mileage of the component and the semantic tag "WJ-8 fastener", constitute a complete data combination and are mapped to the subsequent knowledge graph nodes as entity attributes.

[0111] For the WJ-8 type fastener at K100+020.350 after alignment with S111, the system labels its semantic features as category label [WJ-8 fastener] and visual status label [slight corrosion of elastic bar]. At the same time, combined with the geometric features extracted by S111, the elastic bar gap is 8.5mm and the rail bottom slope is 1 / 38. The system performs feature fusion through a cross-attention mechanism.

[0112] Since the visual sensor has a sampling frequency of 2kHz while the laser sampling frequency is 1MHz, the fusion network automatically reduces the weight of visual edge features and strengthens the gap abruptness features extracted by the laser point cloud. Finally, it outputs the optimized 256-dimensional fusion feature vector of the fastener, which completely eliminates the spatial ambiguity caused by the perspective tilt and sampling misalignment caused by the train's serpentine motion.

[0113] The system defines the physical WJ-8 fastener at K100+020.350 as a graph entity node Nodefastener_C, and uses the aforementioned 256-dimensional feature vector along with the absolute mileage information K100+020.350 as the initial node attribute of Nodefastener_C. Similarly, the rail section and CRTSIII type track slab at this mileage are mapped to nodes Noderail_C and Nodeslab_A, respectively, forming a local entity set.

[0114] Based on spatial coordinate calculations, the bounding boxes of Nodefastener_C and Noderail_C intersect in three-dimensional space, establishing a spatial adjacency relationship. According to the ballastless track assembly rule library, it is deduced that the function of the WJ-8 fastener is to provide the clamping force between the rail and the sleeper or track slab, thus establishing a functional dependency relationship between Nodefastener_C and Noderail_C for fastening.

[0115] Based on the train load transfer mechanism, the system establishes a mechanical transfer relationship between Noderail_C and Nodefastener_C, transferring loads and assigning an initial stress attenuation coefficient attribute to the edge based on the design stiffness of the WJ-8 fastener. Through the above operations, the local component at K100+020.350 is completely transformed into a knowledge graph subnet containing entities and multidimensional topological relationships.

[0116] In step S12, the system retrieves the BIM 3D model of the high-speed railway ballastless track. For the WJ-8 fastener entity at K100+020.350 extracted in S112, the system locates it using absolute mileage coordinates and highlights the corresponding "3D mesh unit: a mesh cluster that typically contains fine components such as elastic bars, gauge baffles, and flat washers" in the BIM model. Through a coordinate system transformation matrix, the 256-dimensional multimodal fusion "feature vector: containing information such as a gap of 8.5mm and a state of [slight corrosion]" of the WJ-8 fastener is assigned to the mesh unit, making it officially become an initial node Nodefastener_C in the initial knowledge graph. Similarly, the rail section and CRTSIII type track slab at this mileage are also mapped to initial nodes Noderail_C and Nodeslab_A.

[0117] The system constructs initial associated edges and assigns weights based on the topological relationships determined in S112. Between Noderail_C and Nodefastener_C, the system identifies a "mechanical transfer relationship". According to the design specifications for high-speed railway ballastless track, the distribution coefficient for the train wheel load transferred to the WJ-8 fastener via the rail is typically 0.45 to 0.50. Accordingly, the system assigns an initial mechanical transfer weight of 0.48 to the directed edge from Noderail_C to Nodefastener_C. Simultaneously, a "functional dependency relationship" fastening edge is established between Nodefastener_C and Nodeslab_A, and an initial fastening stiffness weight value is assigned based on the design range of the WJ-8 fastener. Through this process, locally discrete nodes are assigned connection weights that conform to mechanical mechanisms.

[0118] The system reads the operational history data of this track section: the section has been in operation for 55 years, with a cumulative total weight of 300 million tons, and rail grinding was performed just last month. The system adds and integrates feature vectors such as [cumulative total weight: 300MGT] and [recent maintenance: rail grinding] into the attributes of initial nodes such as Noderail_C. At the same time, considering the fatigue cumulative effect of 300 million tons, the system dynamically corrects the mechanical transfer weight from Noderail_C to Nodefastener_C from the initial 0.48 to 0.43 through a preset nonlinear attenuation function to reflect the decline in the fatigue resistance of the fastening system under long-term use. After the above processing, an initial knowledge graph of the track integrating BIM geometric features, inspection data, and usage history is completed.

[0119] The system acquired nearly a month's worth of operational data for this track section: an average of 80 high-speed trains passed through daily, train load spectrum, axle load of approximately 17 tons, a cold wave occurred during the month, with ambient temperature fluctuating drastically between -5°C and 10°C, and temperature load spectrum; based on the dynamic mechanism of ballastless track, the system constructed a state transition matrix A(t); under the coupled effect of low temperature and high-frequency train load, the elements in matrix A(t) reflect the accelerated shrinkage and creep of CRTSIII type track slab concrete, as well as the physical laws of WJ-8 fastener elastic strips becoming stiffer and having reduced fatigue resistance at low temperatures.

[0120] For the WJ-8 fastener node Nodefastener_C at K100+020.350, the system uses its "fusion feature vector: containing a gap of 8.5mm and slight corrosion" at the beginning of the month's detection time t0 as the initial state h(t0). When the time step advances to the middle of the month t15, the system combines A(t15) with the cumulative tonnage passed in the current half month to calculate the predicted state attribute of Nodefastener_C at time t15. The value of "stiffness degradation dimension" in this attribute vector has decreased by 3% compared to time t0, and the "gap dimension" has increased slightly to 8.6mm due to temperature contraction. This node attribute is serialized and expanded into a set of time-series feature tensors containing a 30-day trend.

[0121] The system uses a spatiotemporal graph convolution operator to dynamically evolve the fastener node and its associated edges. In the spatial dimension, the operator detects that after the stiffness of Nodefastener_C (fastener) decreases, the support constraint of the node above it, Noderail_C (rail), weakens, resulting in an increase in local bending stress in the rail. Based on this, the system updates the mechanical transfer edge weight between Noderail_C and Nodefastener_C, dynamically adjusting the initial transfer weight of 0.43 in S121 to 0.39 through a nonlinear mapping function to simulate the weakening of support stiffness.

[0122] In the time dimension, the gated loop unit extracts the trend feature of the continuous decrease of the edge weight and encodes it into the attributes of the dynamically associated edge. After the above evolution, the initial static map of the track is completely updated into a "spatiotemporal dynamic knowledge map" that fluctuates with temperature and train load. This map not only records the micro-degradation of WJ-8 fasteners, but also characterizes the macro-trend of the relaxation of the mechanical coupling between the "rail-fastener-track slab" system over time.

[0123] In step S13, the system performs multidimensional feature marking on the WJ-8 fastener node (denoted as Nodefastener_C) at K100+020.350; the system analyzes the detected train data and marks its structural geometric deviation as "elastic bar gap 8.5mm, rail bottom slope deviation 0.2 degrees"; its vibration response as "high-frequency vibration acceleration RMS value at this location 12g, up 15% from the baseline"; its material degradation as "elastic bar surface rust area ratio 8%"; after normalization, the above three physical parameters are spliced ​​into the multidimensional initial feature vector of Nodefastener_C.

[0124] The system initiates a dynamic graph attention mechanism to perform multi-hop neighborhood aggregation. For the one-hop neighbor of Nodefastener_C, the rail node (Noderail_C) and the track slab node below it, Nodeslab_A, the system reads the dynamic mechanical transfer edge weight caused by cooling and fatigue in S122, which has decreased to 0.39 as before. Since the decrease in edge weight represents the weakening of support stiffness, the dynamic GAT will assign a higher abnormal stress allocation attention weight to Noderail_C when calculating attention. After two layers of GAT, i.e., two-hop aggregation, Nodefastener_C not only "sensed" its own gap reduction and corrosion, but also "sensed" the abnormal local bending stress transmitted from Noderail_C through the mechanical transfer edge, as well as the slight translational trend of Nodeslab_A (CRTSIII type track slab) caused by temperature contraction.

[0125] The system uses a state evolution update mechanism to process the above-mentioned aggregated feature sequence. By analyzing the changes in the aggregated features of the fastener node over the past 30 days, the state evolution operator found that the "stiffness degradation dimension" in its hidden state exhibits a nonlinear accelerating downward trend, which is highly consistent with the recent drastic cooling time series. The system outputs the spatiotemporal topological feature vector of Nodefastener_C. This vector is no longer isolated fastener detection data, but a comprehensive digital representation that integrates "fastener self-deterioration, local deformation + rail stress redistribution + track slab thermal shrinkage displacement, global mechanical transmission and coupling degradation trend". This high-order spatiotemporal topological feature provides a deep feature input with strong physical interpretation for the subsequent step S132 to combine previous fault content for logical deduction and accurately determine fault detection events such as "fastener fatigue fracture" or "track slab separation".

[0126] The system maps the "spatiotemporal topological feature vector of the WJ-8 fastener node (Nodefastener_C) at K100+020.350, which contains information such as gap widening, corrosion aggravation, and rail stress redistribution" to a high-dimensional manifold space. In this space, the feature points of healthy fasteners cluster to form a low-dimensional "healthy manifold cluster", while the feature points of Nodefastener_C deviate from this cluster and slide along a nonlinear manifold surface due to the degradation of multi-dimensional performance. The system extracts the local tangent space of this surface through a manifold learning algorithm and uses an autoencoder to reconstruct the pattern, filtering out high-frequency noise caused by the high-speed vibration of the detected train, and outputting a pure reconstructed feature combination. This combination accurately reflects the real physical principal components of the coupling between fastener stiffness degradation and rail deformation.

[0127] The system extracts past fault information of ballastless tracks from the database, such as the typical deterioration curves of "WJ-8 spring clip fatigue fracture" and "track slab environmental temperature shrinkage gap". The system initiates multi-level iterations: In the first iteration, it is found that the reconstruction feature of Nodefastener_C deviates abnormally by 1.8 from the healthy baseline, which is in the warning range. In the second iteration, the system performs dynamic time warping matching between the feature's time series and the deterioration curve of "WJ-8 spring clip fatigue fracture", and finds that its "evolution direction: stiffness drops sharply and vibration frequency shifts to the right" matches the fault mode with a high degree of 85%, and the evolution rate is accelerated by 20% due to the recent sharp temperature drop. The multi-level iterations finally determine that the fault evolution content of this node is "accelerated deterioration of spring clip fatigue under low temperature coupling".

[0128] The system outputs degradation stage events based on multi-level iteration results. For the WJ-8 fastener and its associated structure at K100+020.350, the system outputs two fault detection events at different stages: the first is a "moderate corrosion and stiffness degradation warning event" for the spring clip, which is output because the early degradation threshold has been exceeded at the current moment, with a deviation of 1.8, requiring enhanced monitoring; the second is a "critical warning event for fatigue fracture of the spring clip" based on accelerated evolution rate extrapolation, which is expected to exceed the critical safety threshold within the next 15 days. Thus, through manifold reconstruction and multi-level iteration, the system transforms the abstract topological features of the graph into staged fault detection events with clear engineering guidance significance, providing accurate semantic input for determining the fault area and formulating maintenance strategies in the subsequent step S14.

[0129] In step S14, the system dynamically identifies the "critical warning event of fatigue fracture of the elastic bar" output in S132. Starting from the "abnormal node: WJ-8 fastener node Nodefastener_C at mileage K100+020.350" anchored to this event, the system initiates reverse tracing in the spatiotemporal dynamic knowledge graph. The system traverses backward along the "mechanical transmission relationship" and finds that the rail node Noderail_C above Nodefastener_C has abnormal bending stress concentration, and the dynamic edge weight shows abnormal stress distribution. At the same time, tracing downward along the "functional dependency relationship", it finds that the CRTSIII type track slab node Nodeslab_A below it has a small vertical displacement and the stiffness support edge weight has significantly decreased. By calculating the contribution, the system confirms that the fault is not only a problem with the fastener itself, but its deformation force link can be traced back to the track slab node below.

[0130] The system detected a "moderate fastener degradation warning event" at the adjacent mileage K100+020.700. The system performed topological clustering analysis on the "related node sub-links: including Noderail_C, Nodefastener_C, Nodeslab_A and the node clusters of adjacent rows" obtained from tracing these two locations. Based on spatial adjacency and mechanical transmission density calculations, the clustering algorithm determined that these two abnormal areas have common stiffness attenuation characteristics at the bottom layer of the track slab and are spatially adjacent. The system then delineated the connected subgraph from K100+020.100 to K100+020.800 as a macroscopic "fault area," avoiding the limitation of only replacing a single fastener while ignoring the overall area settlement.

[0131] The system performs semantic labeling on the above-defined fault areas; the system extracts the abnormal deviation degree and edge weight attenuation rate of all nodes in the area, and calculates the overall stiffness of the area by weighted average to find that the stiffness of the area has decreased by 18%; based on this, the system labels the fault category of the fault area as "fatigue cascading failure of fasteners caused by uneven settlement of foundation support stiffness"; and labels its severity as "moderate degradation evolving into severe failure, requiring speed limit and planned maintenance".

[0132] For the fault area in section K100+020 delineated by S141, candidate nodes include "WJ-8 fastener node Nodefastener_C", "rail node Noderail_C", and "CRTSIII type track slab node Nodeslab_A". Observations show that fastener fracture and minor track slab settlement coexist, indicating a mixed correlation. The system performed counterfactual reasoning intervention: Do (normal) intervention was implemented on Nodefastener_C, and it was found that the overall abnormal deviation of the area decreased by only 15%, proving that fastener fracture was only a symptom; Do (no settlement) intervention was implemented on Nodeslab_A, and it was calculated that the overall deviation decreased by 65%. Based on this, the system decoupled the false correlation of "settlement and fracture occurring simultaneously" and confirmed Nodeslab_A (minor void / settlement under the track slab) as the actual root cause node.

[0133] Starting from Nodeslab_A, the system traces the node above it to Nodefastener_C based on the distribution of mechanical transfer edge weights, and finally points to the apparent node Noderail_C, indicating that the rail stress exceeds the limit. Combining the temporal degradation trajectory, the system generates a complete fault propagation path as follows: "Micro-settlement and voiding of the track slab base, root cause, lasting for 2 months → Decreased local support stiffness leading to excessive tensile stress on the WJ-8 fastener above, intermediate state, accelerated deterioration in the past week → Fatigue fracture of the fastener spring and geometric irregularity of the rail, apparent."

[0134] The system reads the current operating status: the high-speed rail line is in the peak of the Spring Festival travel rush, with an extremely high number of trains running daily, and only 44 hours of maintenance time at night. The system initiates multi-objective optimization and sets alternative strategies: the first strategy is "replacing the broken fastener overnight"; the second strategy is "replacing the fastener overnight and reinforcing the bottom of the track slab with polymer resin grouting"; the third strategy is "closing the line for 3 days and completely replacing the track slab". After Pareto optimization using the NSGA-II algorithm, the third strategy is eliminated due to the high cost of downtime. Although the first strategy has a low cost, the root cause is not eliminated, and the safety risk elimination rate does not meet the standard. The second strategy can be completed within the 44-hour maintenance window. Although the grouting cost is higher, it completely blocks the fault propagation path, and the safety risk elimination rate reaches 92%. The system finally recommends the second strategy as the optimal fault maintenance strategy output in the Pareto front, perfectly balancing the safety bottom line and maintenance resource constraints during the Spring Festival travel rush.

[0135] Please see Figure 2 The fault detection system based on knowledge graph-based track structure data is applied to the aforementioned fault detection method based on knowledge graph-based track structure data; the fault detection system based on knowledge graph-based track structure data includes:

[0136] Entity module 21 is used to determine track structure data based on track structure detection, and to perform spatial registration and cross-modal entity alignment of track structure data based on the absolute track mileage coordinate system, thereby determining the entity set of track structure and the set of topological relationships between entities;

[0137] The knowledge graph module 22 is used to obtain a three-dimensional model of the track structure, and combine the entity set of the track structure and the topological relationship set between the entities to determine multiple initial nodes and corresponding initial associated edges, further construct the corresponding initial knowledge graph, and combine the state transition matrix of the track structure to determine the corresponding spatiotemporal dynamic knowledge graph;

[0138] The fault detection module 23 is used to mark the node features in the spatiotemporal dynamic knowledge graph, perform neighborhood aggregation and state evolution update on the node features, extract the spatiotemporal topological features of each node, deduce the spatiotemporal topological features, and determine multiple fault detection events of the track structure in combination with the past fault content of the track structure.

[0139] The fault maintenance module 24 is used to identify multiple fault areas based on the identification of each fault detection event, mark the corresponding fault content, further combine the causal mechanism to determine the corresponding fault path, and determine the corresponding fault maintenance strategy based on the fault path and the current usage status of the track structure.

[0140] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fault detection method based on knowledge graph-based track structure data, characterized in that, include: Track structure data is determined based on track structure detection. Spatial registration and cross-modal entity alignment of the track structure data are performed based on the absolute track mileage coordinate system, thereby determining the entity set of the track structure and the set of topological relationships between entities. Obtain a three-dimensional model of the track structure, and determine multiple initial nodes and corresponding initial associated edges by combining the entity set of the track structure and the topological relationship set between entities. Further construct the corresponding initial knowledge graph, and determine the corresponding spatiotemporal dynamic knowledge graph by combining the state transition matrix of the track structure. The node features in the spatiotemporal dynamic knowledge graph are labeled, and the node features are aggregated in the neighborhood and updated in the state evolution. The spatiotemporal topological features of each node are extracted, the spatiotemporal topological features are deduced, and multiple fault detection events of the track structure are determined by combining the past fault content of the track structure. Multiple fault areas are identified based on the identification of each fault detection event, and the corresponding fault content is marked. Furthermore, the corresponding fault path is determined by combining the causal mechanism, and the corresponding fault maintenance strategy is determined based on the fault path and the current usage status of the track structure.

2. The fault detection method based on knowledge graph-based track structure data according to claim 1, characterized in that, The determination of track structure data based on track structure detection, and the spatial registration and cross-modal entity alignment of the track structure data based on the absolute track mileage coordinate system, thereby determining the entity set of the track structure and the set of topological relationships between entities, including: By using a comprehensive inspection train to dynamically inspect the track structure, multimodal inspection data of the track structure is obtained. This multimodal inspection data is then projected onto the absolute track mileage coordinate system and multi-scale spatial registration is performed. During the registration process, a cross-modal entity alignment mechanism is combined to determine the geometric morphological characteristics of the track components.

3. The fault detection method based on knowledge graph-based track structure data according to claim 2, characterized in that, The method for determining track structure data based on track structure detection, and performing spatial registration and cross-modal entity alignment of the track structure data based on the absolute track mileage coordinate system to determine the entity set of the track structure and the set of topological relationships between entities, also includes: The semantic features of track components are labeled and combined with the corresponding geometric features for feature-level fusion to eliminate spatial ambiguity under different detection perspectives and sampling frequencies, thereby outputting an optimized data combination. Based on this data combination, the track components of the track structure are mapped as graph entities. The spatial adjacency relationship, functional dependency relationship and mechanical transmission relationship between entities are extracted by combining track design specifications and line engineering rules, thereby determining the entity set of the track structure and the set of topological relationships between entities.

4. The fault detection method based on knowledge graph-based track structure data according to claim 1, characterized in that, The process involves obtaining a 3D model of the orbital structure, determining multiple initial nodes and corresponding initial associated edges by combining the entity set of the orbital structure and the set of topological relationships between entities, further constructing a corresponding initial knowledge graph, and determining the corresponding spatiotemporal dynamic knowledge graph by combining the state transition matrix of the orbital structure, including: The 3D model of the track structure is retrieved, and the entity set is mapped to the spatial grid of the 3D model of the track structure, serving as multiple initial nodes. Based on the set of topological relationships, corresponding weights are assigned between the corresponding initial nodes to generate corresponding initial associated edges. Finally, the corresponding initial knowledge graph is constructed by combining the working history of the track structure.

5. The fault detection method based on knowledge graph-based track structure data according to claim 4, characterized in that, The process of obtaining a three-dimensional model of the orbital structure, determining multiple initial nodes and corresponding initial associated edges by combining the entity set of the orbital structure and the set of topological relationships between entities, further constructing a corresponding initial knowledge graph, and determining the corresponding spatiotemporal dynamic knowledge graph by combining the state transition matrix of the orbital structure, also includes: The track structure state transition matrix is ​​obtained, which dynamically changes with train load and temperature load. The node attributes in the initial knowledge graph are serialized and expanded along the time dimension. The spatiotemporal graph convolution operator is used to update the weights of the initial associated edges in a time series and perform nonlinear feature mapping, so that the initial knowledge graph is updated with time series into a spatiotemporal dynamic knowledge graph that can characterize the time-varying mechanical properties and coupling degradation trend of the track structure.

6. The fault detection method based on knowledge graph-based track structure data according to claim 1, characterized in that, The node features in the labeled spatiotemporal dynamic knowledge graph are used to perform neighborhood aggregation and state evolution updates on these node features. The spatiotemporal topological features of each node are extracted, and these features are used to deduce multiple fault detection events of the track structure, including: In this spatiotemporal dynamic knowledge graph, the corresponding node features are marked based on the identification of the spatiotemporal dynamic knowledge graph. These node features include structural geometric deviations, vibration responses, and material degradation. A dynamic graph attention mechanism is used to perform multi-hop neighborhood aggregation on the node features, and a state evolution update mechanism is combined to determine the temporal degradation trajectory of the node features in order to extract the spatiotemporal topological features of each node. These spatiotemporal topological features integrate local deformation and global mechanical transmission.

7. The fault detection method based on knowledge graph-based track structure data according to claim 6, characterized in that, The node features in the labeled spatiotemporal dynamic knowledge graph are used to perform neighborhood aggregation and state evolution updates on the node features, and the spatiotemporal topological features of each node are extracted. The spatiotemporal topological features are deduced, and multiple fault detection events of the track structure are determined by combining the past fault content of the track structure. The method also includes: The spatiotemporal topological features are mapped to a high-dimensional manifold space, and manifold learning and pattern reconstruction are performed in this high-dimensional manifold space to output the reconstructed feature combination. At the same time, the past fault content of the track structure is obtained, and multi-level iteration is performed in combination with the reconstructed feature combination. During the iteration process, the abnormal deviation degree of the current spatiotemporal topological features and the fault evolution content are determined, and then multiple fault detection events of the track structure at different degradation stages are output.

8. The fault detection method based on knowledge graph-based track structure data according to claim 1, characterized in that, The process involves identifying multiple fault regions based on the recognition of various fault detection events, marking the corresponding fault content, further determining the corresponding fault path using a causal mechanism, and determining the corresponding fault maintenance strategy based on the fault path and the current usage status of the track structure, including: Each fault detection event is dynamically identified. During the identification process, a spatiotemporal dynamic knowledge graph is used to perform spatial reverse tracing and cluster analysis to determine multiple fault regions. In each fault region, the corresponding fault category and severity content are marked.

9. The fault detection method based on knowledge graph-based track structure data according to claim 8, characterized in that, The process of identifying multiple fault regions based on the identification of each fault detection event, marking the corresponding fault content, further determining the corresponding fault path using a causal mechanism, and determining the corresponding fault maintenance strategy based on the fault path and the current usage status of the track structure also includes: By intervening in multiple candidate nodes within the fault area through counterfactual reasoning, false associations and confounding factors between candidate nodes are decoupled, thereby determining the actual root cause node that caused the fault and the corresponding fault propagation path. Combined with the current usage status of the track structure, multi-objective optimization is performed to achieve Pareto optimization between track operation safety risks and maintenance resource costs, so as to output a fault maintenance strategy for the fault path.

10. A fault detection system based on knowledge graph-based track structure data, characterized in that, The fault detection system based on knowledge graph track structure data is applied to the fault detection method based on knowledge graph track structure data as described in any one of claims 1-9; The fault detection system based on knowledge graph-based track structure data includes: The entity module is used to determine track structure data based on track structure detection. It performs spatial registration and cross-modal entity alignment on the track structure data based on the absolute track mileage coordinate system, thereby determining the entity set of the track structure and the set of topological relationships between entities. The knowledge graph module is used to obtain a three-dimensional model of the track structure, and combine the entity set of the track structure and the topological relationship set between the entities to determine multiple initial nodes and corresponding initial associated edges, further construct the corresponding initial knowledge graph, and combine the state transition matrix of the track structure to determine the corresponding spatiotemporal dynamic knowledge graph; The fault detection module is used to label the node features in the spatiotemporal dynamic knowledge graph, perform neighborhood aggregation and state evolution update on the node features, extract the spatiotemporal topological features of each node, deduce the spatiotemporal topological features, and determine multiple fault detection events of the track structure in combination with the past fault content of the track structure. The fault maintenance module is used to identify multiple fault areas based on the identification of each fault detection event, mark the corresponding fault content, further combine the causal mechanism to determine the corresponding fault path, and determine the corresponding fault maintenance strategy based on the fault path and the current usage status of the track structure.

Images