Road engineering subgrade disease diagnosis method and system based on knowledge graph

Abstract

The present application relates to the technical field of data processing, in particular to a road engineering roadbed disease diagnosis method and system based on a knowledge graph. Geological exploration data, hydrological monitoring data and non-destructive testing data corresponding to a road stake number sequence are acquired, spatial topological edges are constructed using adjacent stake number nodes and disease entity nodes, time sequence evolution edges are constructed using the monitoring state of the same node at different time slices, and a space-time dynamic knowledge graph is generated. Disease characteristic entity sets of a roadbed region to be diagnosed at the current time and hydrogeological sequences within a preset time window are extracted, and in the space-time dynamic knowledge graph, the disease characteristic entity sets are used as a matching starting point, and a time sequence constraint subgraph isomorphism matching is performed in combination with the time labels of the hydrogeological sequences. The present application overcomes the defect that a static graph cannot reflect the dynamic spread path of a disease, and realizes time sequence tracing from the appearance to the historical geological cause of a roadbed disease.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method and system for diagnosing roadbed defects in road engineering based on knowledge graphs. Background Technology

[0002] Current roadbed defect diagnosis technologies typically employ a combination of multi-source physical detection data and knowledge graphs. In practice, ground-penetrating radar deployed along the road's chainage sequence acquires electromagnetic echo signals to determine the coordinates of the stratum reflection interface; pore water pressure sensors embedded within the roadbed depth acquire moisture content and water pressure data; and a falling-weighted deflectometer obtains the deflection basin time-history curve. The acquired geological exploration data, hydrological monitoring data, and non-destructive testing data are extracted as feature entities to construct a static roadbed knowledge graph. In this graph structure, nodes consist of chainage nodes, defect entity nodes, and geological attribute nodes; edges consist of undirected spatial topological edges reflecting the physical distance between adjacent chainages and static attribute edges reflecting the entity's category. When diagnosing a specific roadbed area, the non-destructive testing features of that area at the current moment are extracted. Matching defect entity nodes are searched in the static knowledge graph, and the surrounding geological attribute nodes are queried based on the spatial topological edges. Finally, based on pre-defined static association rules, the defect category and causal analysis results are output.

[0003] Based on the specific implementation of the above-mentioned existing technologies, there is a core technical problem: roadbed defects are affected by changes in groundwater level and long-term traffic loads, exhibiting a spatiotemporal dynamic characteristic of gradual spread along the road chainage. Existing static knowledge graphs only record the state and spatial adjacency relationships of a single time node, lacking the constraints of state evolution in the time dimension. This makes it impossible to track the dynamic spread path of defects over time during the diagnosis process, thus making it impossible to trace the current defect manifestations back to the original geological anomaly node in history, resulting in misjudgment of the cause of the defect. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for diagnosing roadbed defects in road engineering based on knowledge graphs, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A knowledge graph-based method for diagnosing roadbed defects in road engineering includes: acquiring geological exploration data, hydrological monitoring data, and non-destructive testing data corresponding to road station sequence; constructing spatial topological edges using adjacent station nodes and defect entity nodes in the road station sequence; constructing temporal evolution edges using the monitoring status of the same node in different time slices; and generating a spatiotemporal dynamic knowledge graph. Extract the disease feature entity set of the roadbed area to be diagnosed at the current moment and the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window; In the spatiotemporal dynamic knowledge graph, the disease feature entity set is used as the matching starting point, and combined with the time label of the hydrogeological sequence, subgraph isomorphic matching with temporal constraints is performed. Extract the successfully matched evolution path subgraphs and output the starting geological anomaly nodes in the evolution path subgraphs as the source diagnosis results of the roadbed area to be diagnosed.

[0006] Preferably, the acquisition of geological exploration data, hydrological monitoring data and non-destructive testing data corresponding to the road chainage sequence includes: acquiring electromagnetic echo signals by ground-penetrating radar deployed in the road chainage sequence, performing hyperbolic wave fitting processing on the electromagnetic echo signals to obtain the stratigraphic reflection interface coordinates, and using the stratigraphic reflection interface coordinates as the geological exploration data. The water content time series data and water pressure time series data are obtained by pore water pressure sensors buried in the subgrade depth corresponding to the road station sequence, and the water content time series data and the water pressure time series data are combined into the hydrological monitoring data. The deflection basin time history curves of the road station sequence are obtained by a falling weight deflectometer, and the deflection basin time history curves are used as the non-destructive testing data.

[0007] Preferably, the step of constructing a spatial topology edge using adjacent station nodes and disease entity nodes in the road station sequence includes: calculating the physical distance between the adjacent station nodes; when the physical distance is less than the lateral influence radius of the roadbed disease, establishing an undirected spatial topology edge between the adjacent station nodes, and using the physical distance as the initial weight attribute of the undirected spatial topology edge. The method of constructing a temporal evolution edge based on the monitoring states of the same node in different time slices includes: for the same station node, extracting two monitoring states with adjacent time tags, calculating the change in state features between the two monitoring states, and when the change in state features is greater than a preset state transition threshold, establishing a directed temporal evolution edge between the nodes corresponding to the two monitoring states, wherein the direction of the directed temporal evolution edge is from the earlier time tag to the later time tag.

[0008] Preferably, the generation of the spatiotemporal dynamic knowledge graph includes: constructing a graph data structure containing a set of nodes and a set of edges, wherein the set of nodes includes station nodes, disease entity nodes, geological anomaly nodes and time slice nodes, and the set of edges includes the spatial topology edges and the temporal evolution edges; The graph data structure is mapped to a graph database, the spatial topological edges are stored using an adjacency list, and the temporal evolution edges are stored using an event timeline linked list. When new time slice monitoring data is received, a new time slice node is added to the end of the event time axis linked list, and the change in state characteristics between the new time slice node and the previous adjacent time slice node is calculated. The time evolution edge is dynamically updated based on the change in state characteristics.

[0009] Preferably, the step of extracting the disease feature entity set of the roadbed area to be diagnosed at the current moment and the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window includes: performing time-frequency domain transformation on the non-destructive testing data of the roadbed area to be diagnosed at the current moment, extracting the frequency domain energy concentration band and local time domain distortion features, and combining the frequency domain energy concentration band and the local time domain distortion features into a disease feature vector; Search for the term with the highest similarity to the disease feature vector in the preset knowledge dictionary, and use the found term as the disease feature entity set; Hydrological monitoring data and geological exploration data of the roadbed area to be diagnosed within the preset time window are extracted and spliced ​​together according to the time tag sequence to obtain the hydrogeological sequence.

[0010] Preferably, in the spatiotemporal dynamic knowledge graph, taking the disease feature entity set as the matching starting point and combining the time label of the hydrogeological sequence to perform subgraph isomorphic matching with temporal constraints, the step includes: taking the nodes in the disease feature entity set as the root nodes of the candidate subgraphs and initializing the candidate subgraphs. Starting from the root node, reverse time backtracking is performed along the temporal evolution edge in the spatiotemporal dynamic knowledge graph. During the backtracking process, the monitoring status corresponding to the currently traversed node is compared with the data of the same time label in the hydrogeological sequence. If the deviation of the feature comparison is less than the preset matching tolerance, the currently traversed node is added to the candidate subgraph, and the backtracking continues along the temporal evolution edge until the backtracking time reaches the starting boundary of the preset time window. The candidate subgraph is then output as the successfully matched evolution path subgraph.

[0011] Preferably, the method further includes dynamically adjusting the initial weight attributes of the undirected spatial topological edges: obtaining the roadbed filler modulus and vehicle axle load distribution data between the adjacent station nodes; Based on the roadbed fill modulus and the vehicle axle load distribution data, calculate the area of ​​the stress diffusion angle superposition region between adjacent station nodes; The physical distance is corrected by using the area of ​​the stress diffusion angle superposition region to obtain the corrected distance; The initial weight attribute of the undirected spatial topology edge is updated from the physical distance to the corrected distance. When the corrected distance is greater than the lateral influence radius of the roadbed defect, the undirected spatial topology edge is deleted.

[0012] Preferably, the process of constructing the preset knowledge dictionary includes: obtaining historical roadbed defect excavation verification records, and extracting defect type labels and actual defect size labels from the historical roadbed defect excavation verification records; The non-destructive testing data corresponding to the historical roadbed defect excavation verification records are subjected to time-frequency domain transformation to extract the historical frequency domain energy concentration band and historical local time-domain distortion features; The historical frequency domain energy concentration band, the historical local time domain distortion features, the disease type label and the real disease size label are combined into the initial dictionary entries, and multiple initial dictionary entries are aggregated into the preset knowledge dictionary. When the number of newly acquired disease excavation verification records reaches the preset update number, new entries are generated using the newly acquired disease excavation verification records, and the new entries are incorporated into the preset knowledge dictionary.

[0013] Preferably, the step of comparing the monitoring status corresponding to the currently traversed node with the data of the same time label in the hydrogeological sequence includes: extracting the water content value and the stratigraphic reflection interface coordinate offset in the monitoring status corresponding to the currently traversed node. Obtain the target water content value and the target stratum reflection interface coordinate offset from the hydrogeological sequence that have the same time label as the currently traversed node; Calculate the first absolute difference between the stated moisture content value and the target moisture content value, and the second absolute difference between the formation reflection interface coordinate offset and the target formation reflection interface coordinate offset; The deviation of the feature comparison is calculated based on the weighted sum of the first absolute difference and the second absolute difference, and the preset matching tolerance is dynamically configured according to the geological complexity level of the roadbed area to be diagnosed.

[0014] A knowledge graph-based road engineering subgrade disease diagnosis system includes: a spatiotemporal dynamic knowledge graph generation device, used to acquire geological exploration data, hydrological monitoring data and non-destructive testing data corresponding to the road station sequence, construct spatial topological edges with adjacent station nodes and disease entity nodes in the road station sequence, construct temporal evolution edges with the monitoring status of the same node in different time slices, and generate a spatiotemporal dynamic knowledge graph. The feature extraction device is used to extract the set of disease feature entities of the roadbed area to be diagnosed at the current time and the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window; A temporal constraint matching device is used to perform subgraph isomorphic matching with temporal constraints in the spatiotemporal dynamic knowledge graph, taking the disease feature entity set as the matching starting point and combining the time label of the hydrogeological sequence. The dynamic source tracing output device is used to extract successfully matched evolution path subgraphs and output the starting geological anomaly nodes in the evolution path subgraphs as the source tracing diagnosis results of the roadbed area to be diagnosed.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs temporal evolution edges between the monitoring states of different time slices at the same node, associating discrete station geological data with detection data into a topological network with temporal constraints. During the diagnostic phase, using the disease feature entity set as a starting point and combining it with the time label of the hydrogeological sequence, it performs temporally constrained subgraph isomorphic matching, transforming the map query process from static attribute retrieval to dynamic evolution path extraction by tracing back along the time axis. This processing method can trace the current disease manifestation back to earlier time slices along the temporal evolution edges, directly extracting the initial geological anomaly node in the evolution path subgraph as the diagnostic result, eliminating the interference of accompanying phenomena in static association queries, and realizing temporal tracing of roadbed diseases from apparent characteristics to historical geological causes.

[0016] 2. This invention uses an adjacency list and an event timeline linked list to store spatial topological edges and temporal evolution edges respectively. When new time slice monitoring data is received, a node is appended to the end of the event timeline linked list and the temporal evolution edges are dynamically updated, enabling the knowledge graph structure to automatically expand with the monitoring cycle. By combining the subgrade filler modulus and vehicle axle load distribution data to calculate the area of ​​the stress diffusion angle superposition region, the weights of spatial topological edges between adjacent station nodes are corrected or deleted, integrating the physical and mechanical transmission laws of the road structure into the graph topological relationships. During the reverse backtracking process, feature comparison is performed by calculating the weighted sum and deviation of the water content value and the coordinate offset of the stratum reflection interface. Matching tolerance is configured according to the geological complexity level, so that the determination criteria for the backtracking path adapt to the data fluctuation characteristics under different geological environments. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-source monitoring data acquisition and preprocessing process of the present invention; Figure 2 This is a flowchart of the spatiotemporal dynamic knowledge graph construction and dynamic update process of the present invention; Figure 3 This is a flowchart of the extraction process for the feature entity set and hydrogeological sequence of roadbed defects according to the present invention. Figure 4 This is a flowchart illustrating the construction and dynamic updating of the preset knowledge dictionary in this invention. Figure 5This is a flowchart of the time-constrained subgraph isomorphic matching and disease tracing of the present invention; Figure 6 This is a flowchart illustrating the composition and data interaction of the roadbed disease diagnosis system of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please refer to Figure 1 This embodiment provides a method and system for diagnosing roadbed defects based on a knowledge graph. It acquires geological exploration data, hydrological monitoring data, and non-destructive testing data corresponding to road station sequence. The road station sequence is a continuous set of station numbers extending longitudinally along the road, with each station number corresponding to unique spatial coordinates and roadbed cross-sectional attributes. Spatial topological edges are constructed using adjacent station nodes and defect entity nodes in the road station sequence, and temporal evolution edges are constructed using the monitoring status of the same node at different time slices, generating a spatiotemporal dynamic knowledge graph. The defect feature entity set of the roadbed area to be diagnosed at the current moment and the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window are extracted. The roadbed area to be diagnosed is the continuous roadbed range defined by the station numbers exhibiting defect manifestations, and the preset time window is the range of continuous monitoring cycles traced back from the current moment. In the spatiotemporal dynamic knowledge graph, using the defect feature entity set as the matching starting point and combining the time labels of the hydrogeological sequence, subgraph isomorphic matching with temporal constraints is performed. Extract the successfully matched evolution path subgraphs and output the starting geological anomaly nodes in the evolution path subgraphs as the source diagnosis results of the roadbed area to be diagnosed.

[0020] Specifically, geological exploration data is obtained by performing hyperbolic wave fitting on the electromagnetic echo signals collected corresponding to the road station sequence. The fitting process is achieved using the following formula:

[0021] in, Horizontal position When the hyperbolic wave travels at a certain point, The horizontal center coordinates of the target object are: The longitudinal burial depth of the target body. The electromagnetic wave propagation speed of the roadbed medium. The travel time corresponds to the hyperbolic wave apex. The coordinate parameters of the stratum reflection interface are obtained by fitting the hyperbolic wave morphology in the electromagnetic echo signal using the least squares method, and these coordinate parameters are used as geological exploration data. Hydrological monitoring data consists of a combination of water content time-series data and water pressure time-series data collected by sensors buried within the roadbed depth. Both types of time-series data have corresponding time labels, and the time labels correspond to the monitoring period. Non-destructive testing data consists of deflection basin time-history curves corresponding to the road station sequence. These curves record the change in vertical deformation of the roadbed under impact loads over time.

[0022] Further, refer to Figure 2 A graph data structure is constructed, comprising a set of nodes and a set of edges. The node set includes station nodes, defect entity nodes, geological anomaly nodes, and time slice nodes. The edge set includes spatial topological edges and temporal evolution edges. For the construction of spatial topological edges, the physical distance between adjacent station nodes is calculated using the following formula:

[0023] in, and These are the three-dimensional spatial coordinates of two adjacent station nodes. This refers to the physical distance between adjacent station nodes. When the radius of influence of the roadbed defects is less than the lateral radius of influence, an undirected spatial topological edge is established between adjacent station nodes, and the physical distance is... As the initial weight attribute of the topological edge in undirected space, the lateral influence radius of roadbed defects is determined based on the type of roadbed fill material and the compaction characteristics.

[0024] For the construction of time-series evolution edges, for the same station number node, two monitoring states with adjacent time tags are extracted, and the two monitoring states correspond to... Time and The state feature vector at a given time point is composed of the feature parameters from the geological exploration data, hydrological monitoring data, and non-destructive testing data of the time slice corresponding to that node. The change in state features between two monitoring states is calculated using the following formula:

[0025] in, for The state feature vector at time t, for The state feature vector at time t, It is the L2 norm. This represents the change in state characteristics between two monitored states. When the change in state characteristics... When the threshold for state transition exceeds the preset threshold, a directed temporal evolution edge is established between the nodes corresponding to the two monitoring states. The direction of the directed temporal evolution edge is from the earlier time label to the later time label. The preset threshold for state transition is determined based on the stability characteristics of the roadbed fill material.

[0026] The above graph data structure is mapped to a graph database. An adjacency list stores spatial topology edges, and an event timeline linked list stores temporal evolution edges. The adjacency list uses each station node as its head, storing all adjacent nodes with spatial topological associations and the weight attributes of their corresponding edges. The event timeline linked list uses each station node as its body, linking all time slice nodes corresponding to that node in chronological order of their time tags. Each time slice node stores the monitoring state characteristics at the corresponding time. When monitoring data for a new time slice is received, a new time slice node is appended to the end of the event timeline linked list, and the change in state characteristics between the new time slice node and its previous adjacent time slice node is calculated. The temporal evolution edges are dynamically updated based on the change in state characteristics, completing the dynamic expansion of the spatiotemporal dynamic knowledge graph.

[0027] Further, refer to Figure 3 The area of ​​the roadbed to be diagnosed at the current moment is defined. This area is centered on the station number where the defects appear, extending longitudinally along the road to cover a continuous range of stations, and laterally covering the entire roadbed cross-section. The non-destructive testing data of the roadbed area at the current moment is then transformed in the time-frequency domain using a short-time Fourier transform. The calculation formula is as follows:

[0028] in, This represents the time-domain signal of the deflection basin time history curve. For window functions, For time variables, For frequency variables, This is a two-dimensional time-frequency matrix after time-frequency domain transformation. Energy statistics are performed on the time-frequency matrix to calculate the total energy corresponding to each frequency point. Continuous frequency bands with an energy proportion exceeding a preset energy threshold are selected as the concentrated energy frequency band in the frequency domain. Simultaneously, continuous signal segments with amplitudes exceeding a preset amplitude threshold are extracted from the time-domain signal as local time-domain distortion features. The concentrated energy frequency band and the local time-domain distortion features are combined to form a disease feature vector. The term with the highest similarity to the disease feature vector is searched in a preset knowledge dictionary. The similarity is calculated using cosine similarity, and the formula is:

[0029] in, The feature vector of the disease to be matched. For the first in the preset knowledge dictionary The standard feature vector corresponding to each term The cosine similarity between the two vectors is used as the disease feature entity set. The term with the highest similarity is selected as the disease feature entity set, which includes the corresponding disease type label and disease size label.

[0030] Hydrological monitoring data and geological exploration data of the roadbed area to be diagnosed within a preset time window are extracted and spliced ​​together according to the chronological order of the time tags. The spliced ​​sequence is then normalized to obtain the hydrogeological sequence. The calculation formula for the normalization process is as follows:

[0031] in, This is the spliced ​​vector of the original hydrogeological sequence. This represents the minimum value of each dimension of the feature in the sequence. This represents the maximum value of each dimension of the feature in the sequence. The normalized hydrogeological sequence has values ​​ranging from [0,1] for each dimension and is arranged in chronological order according to the time label.

[0032] Furthermore, in the spatiotemporal dynamic knowledge graph, nodes in the disease feature entity set are used as the root nodes of the candidate subgraphs. The candidate subgraphs are initialized with an initial node set containing only the root node and an empty initial edge set. Starting from the root node, reverse time backtracking is performed along the temporal evolution edges in the spatiotemporal dynamic knowledge graph. The direction of reverse time backtracking is from later time labels to earlier time labels, opposite to the direction of the temporal evolution edges. During the backtracking process, the monitoring status corresponding to the currently traversed node is compared with the data of the same time label in the hydrogeological sequence, and the deviation of the feature comparison is calculated. If the deviation of the feature comparison is less than the preset matching tolerance, the currently traversed node is added to the candidate subgraph, and the temporal evolution edge between this node and the previous node is added to the edge set of the candidate subgraph. Reverse backtracking continues along the temporal evolution edges until the backtracking time reaches the starting boundary of the preset time window. The candidate subgraph is then output as the successfully matched evolutionary path subgraph. The termination condition of the backtracking process is determined by the following formula:

[0033] in, This is the time label corresponding to the currently traced node. The starting boundary time of the preset time window is used. When the above formula is satisfied, the reverse backtracking process is terminated.

[0034] Extract the successfully matched evolutionary path subgraph, traverse all nodes in the subgraph, sort them according to time label from earliest to latest, and select the geological anomaly node with the earliest time label as the starting geological anomaly node. The criterion for determining a geological anomaly node is that the deviation between the monitoring state corresponding to the node and the normal roadbed state exceeds a preset anomaly threshold. Output the station location, time label, anomaly type, and anomaly parameters of the starting geological anomaly node as the source diagnosis result of the roadbed area to be diagnosed.

[0035] Table 1. Definition of Basic Attributes of Road Station Nodes Station node Station marker String type The unique code of the station number corresponds to the road mileage. Not empty, globally unique Station node Spatial coordinates Floating-point arrays Three-dimensional spatial coordinates of the roadbed section corresponding to the station number Includes three dimensions: vertical, horizontal, and elevation. Station node Roadbed cross-sectional properties Structured data The chainage corresponds to the basic attributes of the roadbed, such as the type of fill material, compaction degree, and fill height. Compliant with the specifications for road engineering geological investigation Time slice node Time tags timestamp type The collection time corresponding to the monitoring data Corresponding to the monitoring cycle, non-empty Time slice node State feature vector Floating-point arrays Feature set of geological, hydrological, and nondestructive testing data corresponding to time slices With fixed dimensions, normalized to the [0,1] interval Diseased entity node Disease type label Enumeration type The classification codes for roadbed diseases correspond to the standard classification of roadbed diseases. Not empty, conforms to the road engineering defect classification standard. Diseased entity node Disease size parameters Floating-point arrays The longitudinal length, transverse width, and vertical depth of the disease non-negative values Geological anomaly nodes Exception type label Enumeration type The classification and coding of geological anomalies correspond to stratigraphic, hydrological, and other anomaly types. Non-empty, conforms to the engineering geological anomaly classification standard. Geological anomaly nodes Abnormal feature parameters Floating-point arrays The characteristic values ​​and magnitude of change corresponding to the anomaly It includes two dimensions: absolute value and relative change.

[0036] The table above is used to standardize the attribute storage format and value constraints of various nodes in the spatiotemporal dynamic knowledge graph, ensuring that monitoring data from different sources can be uniformly mapped to the corresponding nodes, providing a standardized node attribute basis for the construction of spatial topological edges and temporal evolution edges, and ensuring the consistency and compatibility of graph data.

[0037] This embodiment fully implements the entire process of roadbed disease diagnosis based on spatiotemporal dynamic knowledge graph. It constructs spatial association between road station numbers and disease entities through spatial topological edges, and constructs temporal dimension association of node states through temporal evolution edges. The generated spatiotemporal dynamic knowledge graph can completely record the spatiotemporal evolution process of roadbed state. It achieves the tracing of disease manifestations to initial geological anomalies through subgraph isomorphic matching with temporal constraints, and solves the problem that static knowledge graphs cannot track the dynamic evolution path of diseases.

[0038] In a preferred embodiment, a standardized preprocessing procedure is performed on the multi-source monitoring data corresponding to the road station sequence. For the electromagnetic echo signal corresponding to the geological exploration data, background denoising, gain compensation, and time-depth conversion are performed sequentially to eliminate system noise and propagation attenuation in the signal. Then, hyperbolic wave fitting is used to extract the coordinates of the stratum reflection interfaces, which include the reflection interfaces corresponding to the top surface of the roadbed, the bottom surface of the subgrade, the base course of the roadbed, and the groundwater level. Each reflection interface corresponds to a continuous coordinate sequence. For the water content and water pressure time series data corresponding to the hydrological monitoring data, outlier removal and moving average filtering are performed sequentially to eliminate abrupt outliers and random noise in the data. The processed time series data are aligned according to time labels to ensure that the water content and water pressure data for the same time slice correspond. For the deflection basin time history curve corresponding to the non-destructive testing data, baseline correction and low-pass filtering are performed sequentially to eliminate baseline drift and high-frequency noise in the signal. The processed time history curve is used to extract the time-domain and frequency-domain characteristics of roadbed deformation.

[0039] Furthermore, dynamic weight adjustment is performed on the undirected spatial topological edges between adjacent chainage nodes. Data on the subgrade fill modulus and vehicle axle load distribution between adjacent chainage nodes are obtained. The subgrade fill modulus is calculated from geotechnical test data or deflection basin time history curves, while the vehicle axle load distribution data is obtained from a weighing monitoring system along the road, including statistical parameters such as axle load type and traffic frequency. Based on the subgrade fill modulus and vehicle axle load distribution data, the area of ​​the stress diffusion angle superposition region between adjacent chainage nodes is calculated. The formula for calculating the stress diffusion angle is:

[0040] in, The stress diffusion angle of the roadbed fill material. The resilient modulus of the roadbed fill material. Let be the Poisson's ratio of the subgrade fill material. The area of ​​the superimposed stress diffusion region along the longitudinal direction of the road is calculated based on the stress diffusion angle of adjacent station nodes. The calculation formula is:

[0041] in, The area of ​​the region where the stress diffusion angles are superimposed. This refers to the height of the roadbed fill. , These are the stress diffusion angles of the subgrade fill material corresponding to two adjacent station nodes. , These are the longitudinal coordinates of two adjacent station nodes. The point is the midpoint of the longitudinal coordinates of two adjacent station nodes, and the integration interval is the range of the longitudinal coordinates of the two station nodes.

[0042] The physical distance is corrected by using the area of ​​the region where the stress diffusion angle is superimposed, and the corrected distance is obtained by the following formula:

[0043] in, To correct the distance, This represents the initial physical distance between adjacent station nodes. This refers to the standard cross-sectional area of ​​the roadbed corresponding to adjacent station nodes. Let be the area of ​​the region where stress diffusion angles are superimposed. The initial weights of the undirected spatial topological edges are determined by physical distance. Updated to correct distance When correcting the distance When the radius of influence of the roadbed defects is greater than the lateral radius of influence, the undirected spatial topology edge is deleted to complete the dynamic adjustment of the spatial topology edge.

[0044] For the construction and dynamic updating of time-series evolution edges, for the event timeline linked list of the same station number node, when a new time slice node is added, the state feature vectors of the new node and the previous adjacent time slice node are extracted, and the change in state features between the two state feature vectors is calculated. When the change in state features is greater than a preset state transition threshold, a directed time-series evolution edge is established between the two nodes, with the edge pointing from the previous node to the new node. When the change in state features is less than or equal to the preset state transition threshold, no new time-series evolution edge is established; instead, the state feature vector of the new node is merged into the attributes of the previous node, compressing the storage size of the event timeline linked list. For established time-series evolution edges, when corrected historical monitoring data is received, the change in state features between the corresponding time slice nodes is recalculated. Based on the recalculated results, the existence and attribute parameters of the time-series evolution edge are adjusted to ensure consistency between the time-series evolution edge and the actual state changes of the roadbed.

[0045] Table 2. Spatial Topology Edge Weight Correction Parameters for Adjacent Station Number Nodes

[0046] The table above is used to standardize the parameter definitions, calculation methods and adjustment logic in the dynamic adjustment process of spatial topology edge weights, to ensure that the weight correction process can accurately reflect the physical and mechanical transmission laws of the roadbed structure, to provide standardized parameter basis for the dynamic update of spatial topology relationships, and to ensure that spatial topology edges can accurately characterize the transmission relationship of disease effects between adjacent station nodes.

[0047] This embodiment refines the preprocessing process of multi-source monitoring data, clarifies the construction rules and quantitative calculation methods of spatial topological edges and temporal evolution edges, and dynamically corrects the weights of spatial topological edges by using roadbed filler modulus and vehicle axle load distribution data. It integrates the physical and mechanical characteristics of the road structure into the topological relationship of the map, so that the spatial topological edges can accurately reflect the transmission relationship of disease influence between adjacent station nodes, and improve the fit between the map topology structure and the actual engineering characteristics of the roadbed.

[0048] In another preferred embodiment, time-domain and frequency-domain features are extracted from the non-destructive testing data of the subgrade area to be diagnosed to construct a disease feature vector. A short-time Fourier transform is performed on the preprocessed deflection basin time-history curve to obtain a two-dimensional time-frequency matrix. The energy in the column directions of the time-frequency matrix is ​​summed to obtain the total energy corresponding to each frequency point, generating a frequency-domain energy distribution curve. The frequency-domain energy distribution curve is smoothed, and peak points in the curve are identified. Using the peak points as centers, the curve is expanded to both sides until the energy decreases to a preset proportion of the peak energy, resulting in a continuous frequency band. This frequency band is designated as the concentrated frequency band of the frequency domain energy, and the start frequency, end frequency, peak frequency, and peak energy of the concentrated frequency band are extracted as frequency-domain feature parameters.

[0049] For the time-domain signal of the deflection basin time history curve, the mean and standard deviation of the time-domain signal of the normal roadbed section are calculated. Based on the mean and standard deviation, the time-domain amplitude threshold is determined. Continuous signal segments with amplitudes exceeding the amplitude threshold in the time-domain signal are extracted as local time-domain distortion features. The start time, end time, maximum amplitude, and duration of the local time-domain distortion features are extracted as time-domain feature parameters. The frequency-domain feature parameters are concatenated with the time-domain feature parameters to obtain an 8-dimensional defect feature vector. The defect feature vector is then normalized to unify the value range of each dimension parameter to the [0,1] interval.

[0050] The system searches a pre-defined knowledge dictionary for the term with the highest similarity to the disease feature vector. This dictionary uses a key-value pair structure, where each term corresponds to a unique disease instance. The key is the term's unique identifier, and the value includes the term's standard feature vector, disease type label, and actual disease size label. The system calculates the cosine similarity between the disease feature vector to be matched and the standard feature vector of each term in the dictionary. The term with the highest similarity is selected. If the maximum similarity is greater than a pre-defined similarity threshold, the term is included in the disease feature entity set. If the maximum similarity is less than or equal to the pre-defined similarity threshold, the disease feature vector to be matched is marked as an unknown disease type, and the corresponding feature parameters and anomaly message are output.

[0051] Further, refer to Figure 4 The process of constructing the pre-defined knowledge dictionary includes: acquiring historical roadbed defect excavation verification records, which contain the excavation location, excavation time, defect type, actual defect size, and description of the defect's cause, as well as the corresponding non-destructive testing data, hydrological monitoring data, and geological exploration data prior to excavation at the corresponding station. Defect type tags and actual defect size tags are extracted from these historical roadbed defect excavation verification records. Defect type tags include roadbed settlement, roadbed slippage, roadbed frost heave, roadbed voids, and insufficient roadbed compaction. Actual defect size tags include the longitudinal length, transverse width, and vertical depth of the defect.

[0052] For the non-destructive testing data corresponding to historical roadbed defect excavation verification records, the same time-frequency domain transformation method as for defect feature vector extraction is used to extract historical frequency domain energy concentration bands and historical local time-domain distortion features, which are combined into historical feature vectors. These historical feature vectors are then normalized to obtain standardized feature vectors. The standardized feature vectors, defect type labels, and actual defect size labels are combined into initial dictionary entries. Multiple initial dictionary entries are aggregated into a pre-defined knowledge dictionary. The initial pre-defined knowledge dictionary undergoes consistency verification by calculating the cosine similarity between the standard feature vectors of any two entries. When the similarity exceeds a pre-defined duplication threshold, the two entries are merged, retaining the entry with more complete defect type and size labels, and eliminating duplicate and redundant entries to ensure the uniqueness and representativeness of entries within the dictionary.

[0053] When the number of newly acquired excavation verification records for roadbed defects reaches the preset update quantity, new entries are generated using these records, following the same process as the initial entry construction. Consistency checks are performed on the new entries, removing duplicate entries whose similarity to existing entries exceeds a preset duplication threshold. The verified new entries are then incorporated into the preset knowledge dictionary. Simultaneously, historical entries in the dictionary that have exceeded their preset validity period undergo weight decay processing, reducing their priority in similarity matching. The weight decay coefficient is positively correlated with the entry's storage time, ensuring the dictionary can adapt to the evolving characteristics of roadbed defects as the engineering environment changes.

[0054] For the hydrogeological sequence of the roadbed area to be diagnosed, hydrological monitoring data and geological exploration data within a preset time window are extracted. The hydrological monitoring data includes water content and water pressure values ​​for each time slice, while the geological exploration data includes the stratigraphic reflection interface coordinate offset for each time slice. The stratigraphic reflection interface coordinate offset is the difference between the stratigraphic reflection interface coordinates of that time slice and the stratigraphic reflection interface coordinates of the initial exploration stage. Following the chronological order of the time tags, the water content, water pressure, and stratigraphic reflection interface coordinate offsets of the same time slice are concatenated to obtain the feature vector of a single time slice. The feature vectors of all time slices are arranged according to the time tag order to obtain the initial hydrogeological sequence. Linear interpolation is performed on the initial hydrogeological sequence to fill in the feature data of missing time slices, followed by normalization to obtain a standardized hydrogeological sequence for subsequent feature comparison.

[0055] Table 3. Preset Knowledge Dictionary Entry Structure and Attributes Entry identifier Unique ID String type Fixed-length encoding Globally unique, not empty Feature data Standard eigenvectors Floating-point arrays Fixed-length array, 8 dimensions Each dimension is normalized to the [0,1] interval and is non-empty. Tag data Disease type label Enumeration type Standard classification coding Conforms to the road engineering defect classification standard, non-empty Tag data Actual disease size label Floating-point arrays 3D array Includes vertical length, horizontal width, and vertical depth; non-negative values. Auxiliary data Excavation verification time timestamp type Standard Time Format Not empty, corresponding to the excavation time of the disease instance. Auxiliary data Geological environment label Enumeration type Geological type coding Conforms to engineering geological classification standards Weighted data Matching priority weight floating point Single precision numerical The value ranges from [0,1] and is positively correlated with the timeliness of the entry. Weighted data Weight decay coefficient floating point Single precision numerical The value ranges from [0,1] and is positively correlated with the entry storage time.

[0056] The table above is used to standardize the term structure, storage format and verification rules of the preset knowledge dictionary, to ensure the integrity, consistency and standardization of term data, to provide a reliable dictionary data foundation for matching disease feature vectors and extracting disease feature entity sets, and to ensure the accuracy and stability of disease feature matching.

[0057] This embodiment refines the method for extracting disease feature vectors and the quantification calculation process, clarifies the construction, verification, and dynamic update rules of the preset knowledge dictionary, extracts multi-dimensional features of non-destructive testing data through time-frequency domain transformation, and constructs a standardized knowledge dictionary by combining historical excavation verification data. This realizes the mapping from disease appearance features to standardized disease entities, provides a matching starting point for subsequent subgraph isomorphic matching, and ensures the consistency and reliability of disease feature extraction.

[0058] In yet another preferred embodiment, reference is made to Figure 5 In the spatiotemporal dynamic knowledge graph, disease entity nodes from the disease feature entity set are used as the root nodes of candidate subgraphs. The candidate subgraphs are initialized with a node set initially containing only the root node and an empty edge set. Starting from the root node, a reverse time backtracking is performed along the temporal evolution edges in the spatiotemporal dynamic knowledge graph. The direction of the reverse backtracking is opposite to the direction of the temporal evolution edges, i.e., traversing from late time labels to early time labels. During the backtracking process, for each node traversed, the corresponding time label is obtained, and the water content value and the stratigraphic reflection interface coordinate offset in the monitoring state corresponding to that node are extracted. The target water content value and the target stratigraphic reflection interface coordinate offset with the same time label as the currently traversed node are obtained from the hydrogeological sequence.

[0059] Calculate the first absolute difference between the numerical water content and the target water content, and the second absolute difference between the coordinate offset of the formation reflection interface and the coordinate offset of the target formation reflection interface. Based on the weighted sum of the first and second absolute differences, calculate the deviation of the feature comparison. The calculation formula is as follows:

[0060] in, This is due to the bias in feature matching. The first absolute difference, The second absolute difference, The weighting coefficient for the moisture content value. The weighting coefficients for the coordinate offsets of the formation reflection interface satisfy the following conditions: The weighting coefficients are determined based on the geological type of the roadbed area to be diagnosed.

[0061] The preset matching tolerance is dynamically configured based on the geological complexity level of the roadbed area to be diagnosed. The geological complexity level is divided into 1 to 5 levels according to the uniformity of stratum distribution, groundwater activity intensity, and the degree of geological hazard development. Level 1 corresponds to the simplest geological scenario, and level 5 corresponds to the most complex geological scenario. The dynamic configuration formula for the matching tolerance is:

[0062] in, For dynamically configured matching tolerance, Based on matching tolerance, The geological complexity level is represented by a value ranging from 1 to 5.

[0063] When the feature matching is biased Less than the preset matching tolerance When the node is traversed, it is added to the node set of the candidate subgraph, and the temporal evolution edge between that node and the previously traversed node is added to the edge set of the candidate subgraph. Backtracking continues along the temporal evolution edge. When the feature alignment deviation... Greater than or equal to the preset matching tolerance When the current branch is terminated, the traversal result of the current branch is discarded. During the backtracking process, if the currently traversed node has multiple reverse temporal evolution edges, each branch is backtracked independently to generate multiple corresponding candidate subgraphs.

[0064] After the backtracking process of all branches reaches the termination condition, the matching degree of all generated candidate subgraphs is calculated. The matching degree calculation formula is as follows:

[0065] in, The matching degree of the candidate subgraph. This represents the number of nodes in the candidate subgraph. The first candidate subgraph The feature alignment bias corresponding to each node, with the matching degree ranging from [0,1]. The candidate subgraph with the highest matching degree is selected as the successfully matched evolutionary path subgraph.

[0066] Extract the successfully matched evolutionary path subgraph, traverse all nodes in the subgraph, and sort them in ascending order according to their time labels from earliest to latest to obtain the node time series. Traverse the nodes in the time series and determine whether each node is a geological anomaly node. The criterion for a geological anomaly node is that the deviation between the monitoring state corresponding to the node and the normal roadbed state exceeds a preset anomaly threshold. The preset anomaly threshold is dynamically adjusted according to the geological complexity level. Select the earliest appearing geological anomaly node in the time series as the starting geological anomaly node in the evolutionary path subgraph. Combine the station location, time label, anomaly type, anomaly parameters, and evolutionary path correlation of the starting geological anomaly node to output the disease source diagnosis result for the roadbed area to be diagnosed.

[0067] Based on the above methods, refer to Figure 6 A corresponding roadbed defect diagnosis system is constructed, comprising a spatiotemporal dynamic knowledge graph generation device, a feature extraction device, a temporal constraint matching device, and a dynamic source tracing output device. The spatiotemporal dynamic knowledge graph generation device acquires geological exploration data, hydrological monitoring data, and non-destructive testing data corresponding to the road station sequence. It constructs spatial topological edges using adjacent station nodes and defect entity nodes in the road station sequence, and constructs temporal evolution edges using the monitoring status of the same node at different time slices, generating a spatiotemporal dynamic knowledge graph. The spatiotemporal dynamic knowledge graph generation device includes a data acquisition unit, a topological edge construction unit, and a graph storage and update unit. The data acquisition unit communicates with the front-end monitoring equipment, receives multi-source monitoring data, and performs preprocessing operations. The topological edge construction unit calculates the feature parameters of spatial topological edges and temporal evolution edges, constructing the corresponding edge structure. The graph storage and update unit is implemented based on a graph database, used to store the graph data structure and dynamically update the nodes and edges of the graph according to newly added monitoring data.

[0068] The feature extraction device, communicatively connected to the spatiotemporal dynamic knowledge graph generation device, is used to extract the set of disease feature entities of the roadbed area to be diagnosed at the current moment, as well as the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window. The feature extraction device includes a disease feature extraction unit, a knowledge dictionary matching unit, and a hydrogeological sequence generation unit. The disease feature extraction unit performs time-frequency domain transformation on the non-destructive testing data, extracts and normalizes it to obtain disease feature vectors. The knowledge dictionary matching unit accesses a preset knowledge dictionary, calculates the similarity of feature vectors, and matches them to obtain the set of disease feature entities. The hydrogeological sequence generation unit extracts multi-source data within the preset time window, splices and normalizes it to obtain the hydrogeological sequence.

[0069] The temporal constraint matching device is communicatively connected to the spatiotemporal dynamic knowledge graph generation device and the feature extraction device, respectively. It is used to perform temporally constrained subgraph isomorphic matching in the spatiotemporal dynamic knowledge graph, starting from the disease feature entity set and combining it with the time tags of the hydrogeological sequence. The temporal constraint matching device includes a backtracking traversal unit, a feature comparison unit, and a candidate subgraph generation unit. The backtracking traversal unit performs reverse time backtracking along the temporal evolution edge, traversing the time slice nodes in the graph. The feature comparison unit calculates the deviation between the monitoring status of the traversed nodes and the feature comparison of the hydrogeological sequence, determining whether the nodes meet the matching requirements. The candidate subgraph generation unit generates candidate subgraphs based on the traversal and comparison results, calculates the matching degree, and selects successfully matched evolutionary path subgraphs.

[0070] The dynamic source tracing output device, communicatively connected to the time-series constraint matching device, is used to extract successfully matched evolutionary path subgraphs and output the initial geological anomaly nodes in the evolutionary path subgraphs as the source tracing diagnosis results for the subgrade area to be diagnosed. The dynamic source tracing output device includes a path sorting unit, an anomaly node identification unit, and a diagnosis result output unit. The path sorting unit sorts the nodes in the evolutionary path subgraphs in ascending order according to time labels; the anomaly node identification unit identifies the initial geological anomaly nodes in the sorted node sequence; and the diagnosis result output unit generates standardized source tracing diagnosis results and outputs them to the corresponding display or storage terminal.

[0071] Table 4. Temporal Retrospective Feature Comparison and Matching Tolerance Configuration Table

[0072] The table above is used to standardize the configuration of feature comparison weight coefficients, matching tolerances and state transition thresholds under different geological complexity levels. This ensures that the feature comparison standards in the reverse backtracking process can adapt to the data fluctuation characteristics under different geological environments, balance the recall and precision of subgraph matching, and improve the reliability and adaptability of subgraph isomorphic matching.

[0073] This embodiment refines the complete process and quantitative calculation method of subgraph isomorphic matching with temporal constraints, clarifies the deviation calculation rules for feature comparison and the dynamic configuration mechanism of matching tolerance, and fully discloses the device composition and functional implementation of the roadbed disease diagnosis system based on the above method. By extracting the evolution path subgraph of the disease through reverse time backtracking, the system locates the starting geological anomaly node, realizing the temporal tracing of roadbed diseases from appearance to cause. The modular design of the system can adapt to the disease diagnosis needs of different road engineering scenarios.

Claims

1. A knowledge graph-based method for diagnosing roadbed defects in road engineering, characterized in that, include: Geological exploration data, hydrological monitoring data and non-destructive testing data corresponding to the road chainage sequence are obtained. Spatial topological edges are constructed using adjacent chainage nodes and disease entity nodes in the road chainage sequence. Temporal evolution edges are constructed using the monitoring status of the same node in different time slices, and a spatiotemporal dynamic knowledge graph is generated. Extract the disease feature entity set of the roadbed area to be diagnosed at the current moment and the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window; In the spatiotemporal dynamic knowledge graph, the disease feature entity set is used as the matching starting point, and combined with the time label of the hydrogeological sequence, subgraph isomorphic matching with temporal constraints is performed. Extract the successfully matched evolution path subgraphs and output the starting geological anomaly nodes in the evolution path subgraphs as the source diagnosis results of the roadbed area to be diagnosed.

2. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 1, characterized in that, The process of obtaining geological exploration data, hydrological monitoring data and non-destructive testing data corresponding to the road chain sequence includes: acquiring electromagnetic echo signals by ground penetrating radar deployed in the road chain sequence, performing hyperbolic wave fitting processing on the electromagnetic echo signals to obtain the stratigraphic reflection interface coordinates, and using the stratigraphic reflection interface coordinates as the geological exploration data. The water content time series data and water pressure time series data are obtained by pore water pressure sensors buried in the subgrade depth corresponding to the road station sequence, and the water content time series data and the water pressure time series data are combined into the hydrological monitoring data. The deflection basin time history curves of the road station sequence are obtained by a falling weight deflectometer, and the deflection basin time history curves are used as the non-destructive testing data.

3. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 1, characterized in that, The step of constructing a spatial topology edge using adjacent station nodes and disease entity nodes in the road station sequence includes: calculating the physical distance between the adjacent station nodes; when the physical distance is less than the lateral influence radius of the roadbed disease, establishing an undirected spatial topology edge between the adjacent station nodes, and using the physical distance as the initial weight attribute of the undirected spatial topology edge. The method of constructing a temporal evolution edge based on the monitoring states of the same node in different time slices includes: for the same station node, extracting two monitoring states with adjacent time tags, calculating the change in state features between the two monitoring states, and when the change in state features is greater than a preset state transition threshold, establishing a directed temporal evolution edge between the nodes corresponding to the two monitoring states, wherein the direction of the directed temporal evolution edge is from the earlier time tag to the later time tag.

4. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 1, characterized in that, The generation of the spatiotemporal dynamic knowledge graph includes: constructing a graph data structure containing a set of nodes and a set of edges. The set of nodes includes station nodes, disease entity nodes, geological anomaly nodes, and time slice nodes. The set of edges includes the spatial topology edges and the temporal evolution edges. The graph data structure is mapped to a graph database, the spatial topological edges are stored using an adjacency list, and the temporal evolution edges are stored using an event timeline linked list. When new time slice monitoring data is received, a new time slice node is added to the end of the event time axis linked list, and the change in state characteristics between the new time slice node and the previous adjacent time slice node is calculated. The time evolution edge is dynamically updated based on the change in state characteristics.

5. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 1, characterized in that, The step of extracting the disease feature entity set of the roadbed area to be diagnosed at the current time and the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window includes: performing time-frequency domain transformation on the non-destructive testing data of the roadbed area to be diagnosed at the current time, extracting the frequency domain energy concentration band and local time domain distortion features, and combining the frequency domain energy concentration band and the local time domain distortion features into a disease feature vector; Search for the term with the highest similarity to the disease feature vector in the preset knowledge dictionary, and use the found term as the disease feature entity set; Hydrological monitoring data and geological exploration data of the roadbed area to be diagnosed within the preset time window are extracted and spliced ​​together according to the time tag sequence to obtain the hydrogeological sequence.

6. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 1, characterized in that, In the spatiotemporal dynamic knowledge graph, taking the disease feature entity set as the matching starting point and combining the time label of the hydrogeological sequence, subgraph isomorphic matching with temporal constraints is performed, including: taking the nodes in the disease feature entity set as the root nodes of the candidate subgraphs and initializing the candidate subgraphs. Starting from the root node, reverse time backtracking is performed along the temporal evolution edge in the spatiotemporal dynamic knowledge graph. During the backtracking process, the monitoring status corresponding to the currently traversed node is compared with the data of the same time label in the hydrogeological sequence. If the deviation of the feature comparison is less than the preset matching tolerance, the currently traversed node is added to the candidate subgraph, and the backtracking continues along the temporal evolution edge until the backtracking time reaches the starting boundary of the preset time window. The candidate subgraph is then output as the successfully matched evolution path subgraph.

7. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 3, characterized in that, The method further includes dynamically adjusting the initial weight attributes of the undirected spatial topological edges: obtaining the roadbed filler modulus and vehicle axle load distribution data between adjacent station nodes; Based on the roadbed fill modulus and the vehicle axle load distribution data, calculate the area of ​​the stress diffusion angle superposition region between adjacent station nodes; The physical distance is corrected by using the area of ​​the stress diffusion angle superposition region to obtain the corrected distance; The initial weight attribute of the undirected spatial topology edge is updated from the physical distance to the corrected distance. When the corrected distance is greater than the lateral influence radius of the roadbed defect, the undirected spatial topology edge is deleted.

8. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 5, characterized in that, The process of constructing the preset knowledge dictionary includes: obtaining historical roadbed defect excavation verification records, and extracting defect type labels and actual defect size labels from the historical roadbed defect excavation verification records; The non-destructive testing data corresponding to the historical roadbed defect excavation verification records are subjected to time-frequency domain transformation to extract the historical frequency domain energy concentration band and historical local time-domain distortion features; The historical frequency domain energy concentration band, the historical local time domain distortion features, the disease type label and the real disease size label are combined into the initial dictionary entries, and multiple initial dictionary entries are aggregated into the preset knowledge dictionary. When the number of newly acquired disease excavation verification records reaches the preset update number, new entries are generated using the newly acquired disease excavation verification records, and the new entries are incorporated into the preset knowledge dictionary.

9. The method for diagnosing roadbed defects based on knowledge graphs in road engineering according to claim 6, characterized in that, The step of comparing the monitoring status corresponding to the currently traversed node with the data of the same time label in the hydrogeological sequence includes: extracting the water content value and the stratigraphic reflection interface coordinate offset in the monitoring status corresponding to the currently traversed node. Obtain the target water content value and the target stratum reflection interface coordinate offset from the hydrogeological sequence that have the same time label as the currently traversed node; Calculate the first absolute difference between the stated moisture content value and the target moisture content value, and the second absolute difference between the formation reflection interface coordinate offset and the target formation reflection interface coordinate offset; The deviation of the feature comparison is calculated based on the weighted sum of the first absolute difference and the second absolute difference, and the preset matching tolerance is dynamically configured according to the geological complexity level of the roadbed area to be diagnosed.

10. A road engineering subgrade disease diagnosis system based on knowledge graph, characterized in that, include: A spatiotemporal dynamic knowledge graph generation device is used to acquire geological exploration data, hydrological monitoring data and non-destructive testing data corresponding to road chainage sequences, construct spatial topological edges with adjacent chainage nodes and disease entity nodes in the road chainage sequence, construct temporal evolution edges with the monitoring status of the same node in different time slices, and generate a spatiotemporal dynamic knowledge graph. The feature extraction device is used to extract the set of disease feature entities of the roadbed area to be diagnosed at the current time and the hydrogeological sequence of the roadbed area to be diagnosed within a preset time window; A temporal constraint matching device is used to perform subgraph isomorphic matching with temporal constraints in the spatiotemporal dynamic knowledge graph, taking the disease feature entity set as the matching starting point and combining the time label of the hydrogeological sequence. The dynamic source tracing output device is used to extract successfully matched evolution path subgraphs and output the starting geological anomaly nodes in the evolution path subgraphs as the source tracing diagnosis results of the roadbed area to be diagnosed.

Images