Bridge maintenance evaluation method based on digital twinning

Abstract

The application discloses a bridge maintenance evaluation method based on digital twinning, and belongs to the field of bridge digital twinning and bridge maintenance evaluation. In order to solve the problem that multi-source detection data before bridge maintenance construction is difficult to be uniformly mapped to the same bridge state, resulting in the lack of continuity of disease degree, maintenance range and construction risk judgment, the application constructs a bridge digital twinning component base, a multi-source evidence set, a bridge component state diagram, an evidence confidence gate and a physical constraint space-time diagram Transformer, and combines a maintenance knowledge graph to output a component-level evaluation result, so that the technical effects of continuous bridge state before construction, component-level and traceable evaluation are realized.

Description

Technical Field

[0001] This invention relates to the field of bridge digital twins and bridge maintenance assessment, and particularly to a bridge maintenance assessment method based on digital twins. Background Technology

[0002] As existing bridges age, bridge maintenance work typically requires a comprehensive review of inspection records, sensor monitoring data, defect images, and historical maintenance information for components such as beams, pavement, bearings, and expansion joints to determine the extent of damage, the scope of repair, and construction risks. In current engineering practice, BIM models, inspection forms, monitoring systems, and maintenance records are often maintained by different systems, resulting in inconsistent data formats, time granularity, and component location methods. This makes it difficult to establish a continuous understanding of component-level conditions before construction.

[0003] Existing bridge maintenance assessment methods often rely on single inspection results or manually compiled data for judgment, which can easily lead to inconsistencies in the location of the same defect across different data sources, inconsistent descriptions of severity, and a lack of continuous tracking of post-repair condition changes. For components with load transfer and adjacent influences, if structural connections, spatial adjacency, historical defect associations, and construction risk propagation relationships are not considered simultaneously, the determination of the scope of maintenance and risk level may lack evidence of inter-component relationships.

[0004] Furthermore, the credibility, timeliness, and degree of conflict of multi-source evidence are often simply averaged or handled by human experience in existing assessments, lacking a mechanism to transform the confidence level of evidence into node feature fusion weights and model attention constraints. When there is a large amount of conflicting evidence or significant deviations in similar historical cases, the assessment results also lack confidence intervals, verification markers, and priority inspection components, making it difficult to meet the traceable and verifiable decision-making needs before bridge maintenance and construction.

[0005] Therefore, a bridge maintenance assessment method is needed to address the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a bridge maintenance assessment method based on digital twins. Addressing the problems in existing technologies such as scattered sources of multi-source detection data before bridge maintenance construction, inconsistent component state mapping, and lack of continuity in assessing the degree of damage and construction risks, this invention proposes a technical solution that establishes a digital twin component base for bridges, unifies component evidence mapping, constructs a bridge component state diagram that updates over time, sets evidence confidence gating, and uses a physical constraint spatiotemporal graph Transformer combined with a maintenance knowledge graph for assessment. This invention achieves the technical effects of continuous component-level assessment, outputting credible maintenance ranges, and forming a chain of evidence.

[0007] This invention provides a bridge maintenance assessment method based on digital twins, comprising: S1, reading the building information model or industrial base model of the target bridge, extracting component numbers, spatial locations, connection relationships, and material properties, and generating a bridge digital twin component base including beams, bridge deck pavement, bearings, and expansion joints; S2, converting on-site inspection records, sensor monitoring data, defect image recognition results, and historical maintenance data into inspection events, monitoring features, image defect features, and maintenance events, respectively, and mapping them to component nodes according to component numbers, spatial locations, and timestamps to form a component evidence set; S3, based on the component nodes and the component evidence set, establishing a system including structural connection edges, spatial phases, and other parameters. S4. Calculate the source credibility, time decay coefficient, component matching confidence, and evidence conflict degree for the component evidence set, generate evidence gating coefficients, and use the evidence gating coefficients to fuse node features and adjust the attention weights of the bridge component state diagram; S5. Input the fused bridge component state diagram into the trained physical constraint spatiotemporal graph Transformer, combine it with the historical defect treatment records of the maintenance knowledge graph, and output the component defect level, state confidence, recommended maintenance scope, construction risk level, confidence interval, verification mark, and evidence chain.

[0008] Optionally, S1 includes:

[0009] Analyze the global identifiers, component categories, axis coordinates, elevations, spans, material design parameters, and support relationships of components in Building Information Modeling (BIM) or Industrial Basic Modeling (IMM).

[0010] Generate temporary component numbers for model components with missing component numbers according to component category, spatial bounding box, and adjacent connectors;

[0011] A one-to-one correspondence index table is established between the global component identifier, the component number, and the temporary component number, and the index table is used as the basis for component location in subsequent evidence mapping and state diagram updates.

[0012] Optionally, S2 includes:

[0013] Extract inspection time, inspection location, disease type, disease size, and treatment suggestions from on-site inspection records to generate inspection events;

[0014] Statistical characteristics of strain, deflection, vibration frequency, temperature, and displacement are calculated from sensor monitoring data according to a preset monitoring window to generate monitoring features;

[0015] Crack length, crack width, spalling area, and seepage area are extracted from the disease image recognition results to generate image disease features;

[0016] Extract maintenance time, maintenance location, maintenance method, and post-maintenance inspection results from historical maintenance data to generate maintenance events;

[0017] The inspection events, monitoring features, image defect features, and maintenance events are written into the component evidence set according to the component number matching result, the spatial nearest component matching result, and the timestamp.

[0018] Optionally, S3 includes:

[0019] Using component nodes as graph nodes, structural connection edges are generated based on model connection relationships. Spatial adjacent edges are generated based on component bounding box distances not exceeding a preset spatial adjacency threshold. Load transfer edges are generated based on support reaction force paths, main beam force transmission paths, and bridge deck pavement force transmission paths. Historical defect association edges are generated based on records of the same or adjacent components exhibiting the same defect type within a historical time window.

[0020] The preset spatial adjacency threshold is determined by the minimum bounding box size of the bridge model components and the manually set maintenance inspection radius, and the historical time window is determined by the maintenance cycle of similar bridges in the maintenance database.

[0021] Each time a new set of component evidence is received, the state version number, state timestamp, and node feature vector of the corresponding component node are updated to the bridge component state diagram.

[0022] Optionally, S4 includes:

[0023] The reliability of the data source is calculated based on the metrological calibration records, the status of the data acquisition equipment, the qualifications of the personnel recording the data, and the completeness of the maintenance data archives.

[0024] The time decay coefficient is calculated using an exponential decay function based on the interval between the evidence collection time and the assessment time.

[0025] Calculate the component matching confidence based on the component number matching result, spatial distance matching result, and image location matching result;

[0026] The degree of conflict of evidence is calculated based on the differences in disease level, size, and deviation of monitoring characteristics of the same component node under the same disease type.

[0027] The source credibility, time decay coefficient, component matching confidence and evidence conflict degree are input into the gating mapping function to obtain the evidence gating coefficient of the corresponding evidence. The evidence gating coefficient is used to determine the fusion weight of the feature vector of the corresponding evidence entering the node, and to correct the spatial attention weight and temporal attention weight.

[0028] Optionally, the physical constraint spatiotemporal graph Transformer in step S5 includes a spatial encoding layer, a temporal encoding layer, and an evaluation output layer;

[0029] The spatial coding layer generates a spatial attention mask based on structural connection edges, spatial adjacent edges, load transfer edges, component stiffness, distance between components, and load path direction, and uses the evidence gating coefficient as a multiplicative adjustment amount for the spatial attention weight.

[0030] The time coding layer encodes the node state sequence according to the detection time, maintenance time, and monitoring window, and uses the degree of evidence conflict as the uncertainty input feature of the time coding layer.

[0031] The evaluation output layer generates component defect level, status confidence, recommended repair scope, and construction risk level based on the outputs of the spatial coding layer and the temporal coding layer.

[0032] Furthermore, the physical constraint spatiotemporal graph Transformer is trained in the following way: a training sample containing historical bridge inspection records, historical sensor monitoring records, historical maintenance data, and post-maintenance re-inspection results is constructed.

[0033] The actual level of component defects, the scope of repair, and the results of construction risk verification are used as the supervision labels;

[0034] Add the following to the model loss function: disease level classification loss, maintenance scope regression loss, construction risk classification loss, and physical constraint loss;

[0035] The physical constraint loss is calculated based on the direction of disease propagation, load transfer direction and consistency of post-repair state changes of adjacent components, and is used to constrain the spatial attention distribution and temporal attention distribution after model training.

[0036] Furthermore, the historical records of disease treatment in the maintenance knowledge graph include: establishing entities and relationships in the maintenance knowledge graph based on bridge type, component category, material properties, disease type, disease size range, monitoring characteristic range, repair method, and treatment results;

[0037] Generate a retrieval vector based on the component attributes, node feature vectors, and disease types of the assessment object;

[0038] The retrieval score of the historical disease treatment record is determined based on the cosine similarity between the retrieval vector and the historical disease treatment record vector, as well as the consistency between the component category and the disease type.

[0039] Historical disease treatment records with retrieval scores not less than the preset case retrieval threshold are used as case evidence in the evaluation output layer, and the deviation of their treatment results is input into the evaluation output layer.

[0040] Furthermore, the confidence interval, verification mark, and priority inspection component are generated in the following manner: the width of the defect level interval and the width of the construction risk interval are calculated based on the node-level evidence conflict degree, the risk propagation amount of adjacent components, and the deviation of the treatment result.

[0041] Among them, the risk propagation of adjacent components is determined by the severity of the defects of adjacent components, the direction of the load transfer edge, and the spatial distance between adjacent edges;

[0042] When the width of the disease level range is greater than the preset disease review threshold, or the width of the construction risk range is greater than the preset construction risk review threshold, a review mark is generated.

[0043] The component nodes that rank first in terms of their contribution value to the width of the disease level range or the width of the construction risk range are identified as priority components for re-inspection.

[0044] Furthermore, the evidence chain includes component node identifiers, inspection events involved in the evaluation, monitoring features, image defect features, maintenance events, evidence gating coefficients, spatial attention weights, temporal attention weights, historical defect treatment records, and the basis for generating review markers;

[0045] When outputting the recommended maintenance scope, component nodes whose component defects reach the preset maintenance trigger level are used as basic maintenance components, and component nodes that have load transfer edges with the basic maintenance components and whose risk propagation amount of adjacent components is not less than the preset propagation threshold are used as extended maintenance components. The basic maintenance components and extended maintenance components are then combined to form the recommended maintenance scope.

[0046] The beneficial effects of this invention are:

[0047] 1. By reading the building information model or industrial base model and establishing a component index table, inspection events, monitoring features, image defect features and maintenance events are uniformly mapped to component nodes such as beams, bridge deck pavement, bearings and expansion joints, so that scattered data can be merged into the same bridge digital twin component base, improving the continuity of pre-construction status identification and the consistency of component positioning.

[0048] 2. By constructing a bridge component state diagram that includes structural connection edges, spatial adjacent edges, load transfer edges, and historical defect association edges, and by subjecting the spatial attention of the physical constraint spatiotemporal diagram Transformer to connection, stiffness, distance, and load path constraints, and by modeling the temporal attention according to inspection time, maintenance time, and monitoring window, the defect level, suggested maintenance scope, and construction risk level can be output based on considering the risk propagation between components and the changes in state after maintenance.

[0049] 3. The evidence confidence gating module calculates the source confidence, time decay coefficient, component matching confidence, and evidence conflict degree. The evidence gating coefficient is used for node feature fusion and attention weight adjustment. At the same time, confidence intervals, verification marks, and priority supplementary inspection components are generated based on evidence conflict degree, risk propagation of adjacent components, and historical case deviation, thereby improving the credibility, verifiability, and traceability of multi-source bridge maintenance assessment results. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 A flowchart of a bridge maintenance assessment method based on digital twins;

[0052] Figure 2 This is a flowchart of the evidence gating process in step S4 of the present invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figures 1-2 The bridge maintenance assessment method based on digital twins includes: S1. Reading the building information model or industrial base model of the target bridge, extracting component numbers, spatial locations, connection relationships, and material properties, and generating a bridge digital twin component base containing beams, bridge deck pavement, bearings, and expansion joints; S2. Converting on-site inspection records, sensor monitoring data, defect image recognition results, and historical maintenance data into inspection events, monitoring features, image defect features, and maintenance events respectively, and mapping them to component nodes according to component numbers, spatial locations, and timestamps to form a component evidence set; S3. Based on the component nodes and the component evidence set, establishing a system including structural connection edges, spatial adjacent edges, and other data. S4. Calculate the source credibility, time decay coefficient, component matching confidence, and evidence conflict degree for the component evidence set, generate evidence gating coefficients, and use the evidence gating coefficients to fuse node features and adjust the attention weights of the bridge component state diagram; S5. Input the fused bridge component state diagram into the trained physical constraint spatiotemporal graph Transformer, combine it with the historical defect treatment records of the maintenance knowledge graph, and output the component defect level, state confidence, recommended maintenance scope, construction risk level, confidence interval, review mark, and evidence chain.

[0055] In this specific embodiment, S1 includes:

[0056] The model parsing module of the bridge maintenance assessment platform reads the architectural information model or industrial foundation model of the target bridge and writes the project identifier, coordinate reference, floor or bridge span partition, component type dictionary and material library version of the model file into the bridge model parsing cache. In this specific embodiment, the model file is stored in a unified coordinate system. The component record includes at least the global component identifier, component category, axis coordinates, elevation, span, spatial bounding box, connection port, material design parameters and support relationship. The material design parameters include elastic modulus, design strength grade, initial value of component stiffness reduction and durability grade. The model parsing module first removes duplicates according to the global component identifier, and then filters out candidate nodes of beam, bridge deck pavement, bearing and expansion joint according to component category.

[0057] For model components with component numbers, the platform uses the component number as the primary locator key for the component node. Simultaneously, it writes the component's global identifier, component category, center coordinates, component length, width, height, elevation range, span range, and set of adjacent connectors into the component node record. The component's center coordinates are calculated from the minimum and maximum corner points of the component's bounding box, using the following formula: In the formula, For the first The center coordinates of each component node and These are the minimum and maximum corner points of the bounding box of the component, respectively. Used for subsequent spatial matching, construction of spatial adjacent edges, and verification of image localization results;

[0058] For model components with missing component numbers, the model parsing module generates temporary component numbers based on component category, spatial bounding box, and adjacent connectors. The generation rule is to first read the component category code, the span number of the bridge it belongs to, the center coordinate bin along the bridge direction, and the hash value of the connection port, and then calculate the temporary number key. In the formula, For component category codes, Bridge span number The binning results are centered on the coordinates. For the set of adjacent connectors, The platform's pre-defined deterministic hash function, when two candidate components... When the bounding boxes are identical and the overlap rate is greater than the calibrated overlap threshold, the record with the more recent model update time is retained, and the other record is marked as a duplicate candidate.

[0059] The model parsing module establishes a component index table. The component index table uses the global component identifier, component number, and temporary component number as three entry keys, and stores the one-to-one correspondence between the entry keys, bridge span partition, component category, material property pointer, support relationship pointer, spatial bounding box, and node version number. In this specific embodiment, a record in the index table includes the global component identifier as the original global component identifier of the model, the temporary component number when the component number is empty, the component category as support, the bridge span partition as the second span, the material property pointer pointing to the support material parameter record, and the support relationship pointer pointing to the adjacent beam and support anchor point record. Among them, the pier and abutment are only stored as support anchor point fields and are not used as component nodes in the final evaluation output. This record is used by S2 to locate the inspection position and monitoring point to the same component node.

[0060] During the material property organization process, the platform generates component material property vectors based on material design parameters. These vectors include concrete strength grade codes, steel reinforcement or steel component strength codes, normalized elastic modulus values, durability grade codes, and normalized support stiffness values. Normalization is based on the upper and lower limits and design benchmarks in the material parameter calibration table for similar bridges. It employs a method of dividing the parameter difference by the range of parameters within the same unit, truncating the result to a value range of 0 to 1. The calculation formula is as follows: In the formula, These are the measured or designed values ​​of the material parameters. and For the same unit of reference range, For zero protection quantity of the same unit, the obtained Write the initial values ​​for the component node features;

[0061] When a model component is missing any key field in component category, spatial bounding box, or support relationship, the platform does not delete the component record. Instead, it writes it into the list of components to be verified and uses the component's global identifier as a temporary location key. The component to be verified only participates in evidence reception and manual completion prompts, and does not participate in load transfer edge calculation. After the fields are completed, it is restored as a valid component node according to the same index table.

[0062] The model parsing module also establishes a transformation record between the bridge's local coordinate system and the original model coordinate system based on the model coordinate datum. The transformation record includes the coordinate origin, unit vector along the bridge direction, unit vector across the bridge direction, elevation datum, and coordinate version number. When the model contains both building information model and industrial base model, the platform uses the component global identifier and spatial bounding box overlap rate as the primary key to merge components with the same name. The merged component nodes retain the source model identifier, geometric source priority, and attribute source priority, so that subsequent inspection, monitoring, and image positioning all fall on the same digital twin component base.

[0063] After the component base is generated, the platform performs an integrity check. The check items include whether the beam node has at least one structural connection edge candidate, whether the support node has an upper beam relationship and a lower support anchor point relationship, whether the bridge deck pavement node has a corresponding beam coverage relationship, and whether the expansion joint node has an adjacent beam end or abutment anchor point relationship. Nodes that fail the check are written into the base quality record. The base quality record saves the missing fields, the source of the candidate completion, the virtual support anchor point identifier, and the flag indicating whether entry into the S2 evidence mapping is allowed.

[0064] The platform ultimately generates a digital twin component base for the bridge. This digital twin component base includes a component node table, a component index table, a material property table, a support relationship table, a support anchor point table, and a model version record. The component node table covers the component nodes of the beam, bridge deck pavement, bearings, and expansion joints. The support anchor point table stores the spatial anchor points and support boundaries of non-evaluation objects such as piers and abutments. The component index table serves as the component positioning basis for subsequent evidence mapping and state diagram updates. The material property table, support relationship table, and support anchor point table serve as inputs for S3 to establish load transfer edges and S5 to generate physical constraint spatial attention masks.

[0065] In this specific embodiment, S2 includes:

[0066] The evidence conversion module reads the bridge digital twin component base and component index table generated by S1, and receives on-site inspection records, sensor monitoring data, defect image recognition results and historical maintenance data respectively. All input records are first converted into unified evidence records. The unified evidence record includes evidence type, original source, original record identifier, candidate component number, spatial location, collection timestamp, defect type, numerical features, text summary, source file version and parsing status. The parsing status is used to identify whether the field is complete, the space is to be matched, the time is to be aligned or the source is to be verified.

[0067] For on-site inspection records, the evidence conversion module extracts inspection time, inspection location, disease type, disease size, and treatment suggestions from inspection forms, mobile terminal records, or maintenance management ledgers. It also breaks down disease size into length, width, area, and depth fields, with units uniformly set to meters, millimeters, square meters, and millimeters, respectively. When the inspection location contains component numbers, it directly queries the component index table. When the inspection location only contains bridge spans, station numbers, or text, it generates a set of candidate components based on the location semantic dictionary and component spatial bounding boxes. Subsequently, it forms an inspection event and writes it into the inspection event sub-table.

[0068] For sensor monitoring data, the evidence conversion module calculates statistical characteristics from the time series of strain gauges, deflectometers, vibration acquisition devices, temperature sensors, and displacement gauges according to a preset monitoring window. In this specific embodiment, the preset monitoring window is given by the monitoring configuration table and is calculated in parallel using the most recent 24-hour window and the most recent 7-day window. For any sensor quantity... In the window The mean, peak value, and standard deviation within the range are calculated as follows: , and We obtain the following formula: For window The sensor readings, aligned by timestamp, are used to obtain statistical characteristics of strain, deflection, vibration frequency, temperature, and displacement, which are then written into the monitoring characteristics sub-table.

[0069] For the disease image recognition results, the evidence conversion module reads the disease category, pixel mask, localization box, recognition confidence, shooting posture and image acquisition time output by the image recognition model, and converts the pixel scale to the engineering scale in combination with the camera calibration parameters. The crack length is obtained by multiplying the skeleton line pixel length by the scale coefficient, the crack width is obtained by statistically analyzing the width in the local normal direction of the mask, and the peeling area and seepage area are obtained by multiplying the mask area by the square of the scale. When the recognition confidence is lower than the image feature reception threshold, the image record is retained but its parsing status is written as low confidence pending review.

[0070] For historical maintenance data, the evidence conversion module extracts maintenance time, maintenance location, maintenance method, and post-maintenance inspection results from maintenance work orders, completion records, and post-maintenance inspection records. It also maps maintenance methods to standard method labels such as crack sealing, pavement replacement, support adjustment, expansion joint replacement, or structural reinforcement. The post-maintenance inspection results include inspection time, changes in disease level, changes in size, and recovery of monitoring features. The maintenance event also carries the component number before maintenance and the location result of post-maintenance inspection, which are used for S3 historical disease association edges and S5 maintenance knowledge graph retrieval.

[0071] The evidence mapping module writes evidence into the component evidence set based on component number matching results, spatial nearest component matching results, and timestamps. Component number matching results are preferentially obtained from the component index table. Spatial nearest component matching results are calculated based on the minimum distance from the evidence's spatial location to the component's bounding box. The matching score is determined by... We obtain the following formula: Indicates the hit value of component number. This represents the minimum distance from the location of the evidence to the bounding box of the candidate component. For spatial matching radius, Indicates the image location hit value. , and From the evidence mapping weight table and sum to one, The larger the value, the more credible the match between the evidence and the component node;

[0072] When inspection events, monitoring features, image defect features, and maintenance events simultaneously hit the same component node and their timestamps fall within the same evaluation period, the evidence mapping module merges them into different evidence items within the same component evidence set. For each evidence item, it saves the evidence type, component node identifier, original record identifier, acquisition timestamp, defect type, numerical feature vector, and matching score. In the parsing state, when the evidence only has spatial matching but no component number matching, the platform writes the evidence into the candidate evidence branch and reduces the component matching confidence in S4;

[0073] The evidence conversion module performs deduplication on duplicate evidence with the same original record identifier and the same component node identifier. The deduplication rule is to retain the record with the latest collection timestamp and the highest completeness of the parsed fields, and at the same time write the source file version of the merged record into the evidence source list. If a certain type of evidence is empty in the current evaluation period, only missing type metadata and insufficient evidence markers are generated, and no real evidence items are generated to participate in the gating mapping. The gating coefficient of the missing metadata is forced to be zero and is not included in the evidence quantity, average conflict degree and time series statistics.

[0074] Before the component evidence set is written, the evidence conversion module performs unit unification and boundary verification on all numerical features. Length fields are unified to meters or millimeters, area fields are unified to square meters, temperature fields are unified to degrees Celsius, and frequency fields are unified to Hertz. For numerical features that need to be used as model input, the platform reads the upper and lower limits of the same unit in the feature benchmark table, generates 0 to 1 interval features by dividing the difference by the same unit range, and truncates to the boundary value when the range is exceeded. At the same time, the untruncated original value and truncation mark are retained in the evidence record.

[0075] The evidence conversion module also performs time alignment processing. On-site inspection records and disease image recognition results are entered into the event sequence according to the collection timestamp, sensor monitoring data are entered into the feature sequence according to the monitoring window end time, and historical maintenance data are entered into the maintenance sequence according to the maintenance completion time and the post-maintenance re-inspection time. When the time difference of multiple types of evidence for the same component node exceeds the allowable range of the evaluation cycle, the platform retains the original timestamp and writes the time to be aligned status. After completing S2, the component evidence set is output for use in S3 state diagram construction and S4 evidence gating calculation.

[0076] The evidence mapping weight table, spatial matching radius, and image feature reception threshold are obtained through historical manually labeled samples. The labeled samples include inspection records with confirmed component numbers, sensor records with measurement point coordinates, and manually reviewed defect image recognition results. The platform searches for weight combinations with the goal of minimizing the component matching error rate and sets the spatial matching radius to the high quantile value of the error of manually labeled landing points for the same bridge type and component category. Parameter records are saved separately for different bridge types, component categories, and sensor layout versions, so that the S2 evidence mapping parameters can be configured according to the engineering object.

[0077] The platform ultimately outputs a component evidence set, which is grouped according to the component node identifier. Each group includes four types of sub-records: inspection events, monitoring features, image defect features, and maintenance events, as well as their component number matching results, spatial nearest component matching results, and timestamps. The component evidence set is written into the evidence partition of the bridge assessment database and serves as the input for S3 to establish the bridge component status diagram and S4 to calculate the evidence gating coefficient.

[0078] In this specific embodiment, S3 includes:

[0079] The state diagram construction module reads the component node table generated by S1 and the component evidence set generated by S2, and uses each valid component node as a graph node in the bridge component state diagram. The graph node record includes the component node identifier, component category, bridge span zone, and center coordinates. The system includes a bounding box, material property vector, support relationship, node feature vector, status version number, status timestamp, and the most recent evidence reception time. The initial value of the node feature vector is obtained by splicing the material property vector, geometric scale normalized value, defect statistics vector, and maintenance history vector. The geometric scale normalization is formed by dividing the original scale by the reference scale and truncating it to the range of 0 to 1, with the same bridge component scale as the denominator.

[0080] The state diagram construction module generates structural connection edges based on the model connection relationships. When two components have direct connection records in the model support relationship table or connection port table, a structural connection edge is generated and the edge type, start node, end node, connection port, initial value of connection strength, and edge version number are written. When the connection port is missing but the support relationship table has the same support anchor point or the same beam end associated, the edge is marked as a structural connection pending verification state. The support anchor point is only saved as a virtual endpoint in the edge attributes and is not included in the component defect level output set.

[0081] The state graph construction module generates spatially adjacent edges based on the bounding box distance of the components. With components The bounding box distance between them is denoted as ,when Spatial adjacent edges are generated in real time, with a preset spatial adjacency threshold. The minimum bounding box size of the bridge model components and the manually set maintenance inspection radius are determined by the following formula: In the formula, and These are the feature dimensions of the two component bounding boxes, respectively. For maintenance and inspection radius, The spatial adjacency coefficient. The unit is meters and should be written into the graph parameter record;

[0082] The state diagram construction module generates load transfer edges based on the support reaction force path, the main beam force transmission path, and the bridge deck pavement force transmission path. The platform first reads the load path directions from the support to the beam, the beam to the adjacent beam segment, and the bridge deck pavement to the beam from the support relationship table, and then generates directed load transfer edges according to the edge direction. The load transfer edge record includes the start node, end node, load path direction, path type, component stiffness, normalized value of support reaction force, and path confidence. The path confidence is determined by the existence of structural connection edges and the integrity of support relationships.

[0083] The state graph construction module generates historical disease association edges based on historical disease records. When the same component or adjacent components are within a historical time window... When records of the same type of defect appear within a given set, historical defect association edges are generated. Adjacent components are determined by the union of already generated edges, prioritizing structural connection edges, followed by load transfer edges, and then spatial adjacency edges. Any matching relationship qualifies a component to enter the candidate adjacent set. (Historical time window) The maintenance cycle is determined by the maintenance database of similar bridges, and the edge weights are determined according to... calculate, The larger the value, the stronger the correlation with historical diseases;

[0084] For the component evidence set written by S2, the state diagram construction module reads the corresponding inspection events, monitoring features, image defect features and maintenance events according to the component node identifier, and encodes the defect type, defect size, monitoring statistical features, post-maintenance re-inspection results and evidence timestamp as node feature increments. The node feature increments are aggregated according to the same defect type, and different defect types occupy different slots in the node feature vector to avoid defects such as cracks, spalling, water seepage and support abnormalities being covered as a single state;

[0085] Each time a new set of component evidence is received, the state graph construction module updates the state version number, state timestamp, and node feature vector of the corresponding component node. The version update rule is as follows: The node feature vector update rule is as follows In the formula, To update the feature vector of the previous node, This is an increment of node features in the formation of this evidence. The state retention factor is given by the evaluation cycle configuration table;

[0086] When the same component node receives multiple pieces of evidence with different detection times, maintenance times, and monitoring windows within the same evaluation cycle, the state graph construction module does not directly cover the node state. Instead, it writes the node state sequence according to the detection time, maintenance time, and monitoring window. Each frame in the node state sequence includes a state timestamp, evidence type mask, node feature vector, number of evidence, and source summary. The node state sequence is read by the time coding layer of S5.

[0087] When a new set of component evidence points to the component to be verified or when there are multiple candidate components in the spatial matching result, the state graph construction module attaches the evidence to the candidate node set and generates candidate edges. Candidate edges are not included in the physical constraint calculation. Candidate edges retain the candidate component node, spatial distance, matching score and reason for verification. If the gating coefficient calculated by S4 is not lower than the candidate conversion threshold and the manual or rule verification is passed, it will be converted to the corresponding edge type in the next graph version. Otherwise, it will be written into the list of failed edges and removed from the input of S5.

[0088] The defect statistics slots in the graph node feature vector are filled by the inspection events, monitoring features, and image defect features of S2. The platform establishes slots such as cracks, spalling, water seepage, support anomalies, and expansion joint anomalies according to defect types. Each slot stores the defect level code, size normalization value, monitoring deviation normalization value, most recent evidence time, and evidence quantity. Size normalization uses the defect size benchmark of the same type of component as the denominator, and monitoring deviation normalization uses the alarm benchmark of the same unit in the sensor monitoring configuration table as the denominator. Both are truncated to the 0 to 1 interval and then written into the node feature vector.

[0089] The state diagram construction module saves the edge source and update conditions for each type of edge. Structural connection edges are updated when the model version changes, spatial adjacent edges are updated when the component bounding box or spatial adjacency threshold changes, load transfer edges are updated when the support relationship or component stiffness parameters change, and historical defect-related edges are updated when new maintenance events or re-inspection results are written. When an edge update causes a change in the downstream attention mask, the platform records the edge version number and the trigger evidence batch for S5 to locate the source of the change in the assessment results.

[0090] The state graph construction module executes after completing node and edge updates. Figure 1 Consistency verification includes checking whether structural connection edges are connected to valid component nodes or support anchor points, and whether the bounding box distance between adjacent spatial edges is still no greater than [value missing]. 1. Whether the load transfer direction is consistent with the support relationship table; 2. Whether the historical defect association edge falls within the historical time window. Edges that fail the verification are written into the list of failed edges. The list of failed edges saves the reason for failure, the original edge version number, the batch of triggering evidence, and the recovery conditions.

[0091] For the node state sequence in the bridge component state diagram, the state diagram construction module saves the original state source according to the detection time, maintenance time and monitoring window respectively, and the state frames from different sources do not overlap with each other; each state frame exposes the frame identifier, evidence item identifier, original node features, candidate edge identifier and insufficient evidence mark to S4, and exposes the feature pointer and time position field after S4 gating to S5, so that the interface of state diagram construction, gating fusion and model inference remains consistent.

[0092] The state graph construction module ultimately generates a bridge component state graph, which includes graph nodes, structural connection edges, spatial adjacency edges, load transfer edges, historical defect association edges, node state sequences, and graph version records. The graph version records store the graph construction time, evidence batches, and spatial adjacency thresholds. Historical time window The bridge component state diagram, along with the set of node state version numbers, serves as the input for S4 evidence gating fusion and S5 physical constraint spatiotemporal graph Transformer inference.

[0093] In this specific embodiment, S4 includes:

[0094] The evidence gating module reads the component evidence set formed by S2 and the bridge component status diagram formed by S3, and calculates the source credibility, time decay coefficient, component matching confidence and evidence conflict degree for each piece of evidence. Before the calculation, the measurement calibration record, acquisition equipment status, recorder qualification, maintenance data archive completeness, acquisition time, candidate component matching information and historical status of the same type of defect in the evidence record are loaded into the evidence gating cache. The gating cache uses the component node identifier and evidence batch as the joint key.

[0095] The reliability of the data source is jointly determined by the metrological calibration records of the data source, the status of the acquisition equipment, the qualifications of the personnel recording the data, and the completeness of the maintenance data archive. The evidence gating module normalizes the four components into... After the interval, press Calculate, where, As evidence The credibility of the source Indicates the validity of metrological calibration. Indicates the status of the data acquisition device. Indicates the qualifications of the recorder. Indicates the completeness of maintenance data archiving, weighted. to The values ​​are derived from the source credibility weight table and sum to one.

[0096] The time decay coefficient is calculated based on the interval between the evidence collection time and the evaluation time. The evidence gating module reads the evaluation time. and evidence collection time First, estimate the time deviation based on the clock calibration records from the data source; if the calibrated... Still later If the evidence is written into the timestamp anomaly verification branch and a timestamp anomaly verification mark is forcibly generated, the time decay coefficient is truncated to one instead of being set to zero, so that the key evidence retains a traceable entry point. If the time deviation cannot be calibrated, the evidence will only have its source credibility reduced and the verification mark will be retained.

[0097] The component matching confidence score is calculated based on the component number matching result, spatial distance matching result, and image location matching result. The evidence gating module reuses the matching score of S2. It also supplements the image localization residuals and component number hit status; the platform first uses the minimum distance from the evidence location to the bounding box of the candidate component. Divide by the matching radius of the same unit space Obtain the distance normalized ratio Press again Calculate the component matching confidence score, where, The hit value is the component number. Locate matching values ​​for the image. , and From the component matching weight table and the sum is one. The larger the value, the more reliable the component matching;

[0098] The degree of evidence conflict is calculated based on the differences in disease level, size, and deviation of monitoring characteristics of the same component node under the same disease type. The evidence gating module first reads the baseline state of the same disease type in the node's historical state sequence. If it is the first assessment or the historical baseline is missing, the statistical prior of the disease of the same bridge type and component category is used as the baseline. If the statistical prior is also missing, the median state of multi-source evidence in the current assessment period is used as the temporary baseline, and the source of the baseline is recorded in the conflict evidence list. Then the calculation is performed. ;

[0099] Before calculating the degree of evidence conflict, the evidence gating module performs dimensionless normalization on the differences in disease level, size, and deviation of monitoring features. The difference in disease level is divided by the maximum span of the disease level, the size difference is divided by the size benchmark of the same type of disease, and the deviation of monitoring features is divided by the corresponding sensor alarm benchmark. The three components are then truncated to the range of 0 to 1. This processing makes the source confidence, time decay coefficient, component matching confidence, and degree of evidence conflict all dimensionless gating inputs, which can be combined in the same gating mapping function.

[0100] The evidence gating module will assess the credibility of the source. Time decay coefficient Component matching confidence And the degree of conflict of evidence Input the gating mapping function to obtain the evidence gating coefficients. In the formula, The larger the value, the higher the fusion weight of the evidence entering the node's feature vector. , or If any value is below the corresponding receiving threshold, then The calculated value is retained, but this evidence is marked as low-confidence evidence and written into the review source field;

[0101] When fusing node features, the evidence gating module performs weighted fusion on the evidence feature vectors of the same component node, and the fusion rule is as follows: In the formula, For component nodes Features of evidence after fusion As evidence eigenvectors, When the protection level is zero, When the node fusion threshold is lower than the threshold, the platform retains the feature vector of the previous version node in S3 and marks the node as insufficient evidence.

[0102] The evidence gating module uses evidence gating coefficients to correct the spatial attention weights and temporal attention weights of the bridge component state diagram, and the spatial edges arrive Correction weights according to Obtain the time frame Correction weights according to We obtain the following formula: As the initial weights of the space, As the initial weight for time, and These represent the mean values ​​of the evidence gating coefficients at both ends of the spatial edge and within the temporal frame, respectively. The mean of the degree of conflict of evidence within the time frame. and Provided by the attention regulation parameter table;

[0103] When the number of highly conflicting pieces of evidence for the same component node exceeds the conflict review threshold, the evidence gating module does not delete the evidence, but instead adds the conflicting evidence and its degree of conflict to the list. The disease type, source record, and candidate component nodes are written into the conflict evidence list, and the identifier of the conflict evidence list is written into the verification input slot of the node feature vector, so that S5 can trace the source of conflict when outputting confidence interval and verification mark;

[0104] The evidence gating module also maintains a gating parameter table, which includes source credibility weight, time decay constant, spatial matching radius, conflict weight, node fusion threshold, spatial attention adjustment coefficient, and temporal attention adjustment coefficient. Each parameter record stores the bridge type, component category, evidence type, parameter value, valid range, calibration source, and version number. When no parameter record matches the evidence type, the platform uses the default parameters for the same bridge type and component category and writes the reason for the mismatch into the evidence chain candidate field.

[0105] The evidence gating module ultimately outputs gating features, an evidence gating coefficient table, corrected initial values ​​of spatial attention weights, corrected initial values ​​of temporal attention weights, and a list of conflicting evidence. Among these, the original node state sequence output by S3 is not overwritten, S4 performs gating fusion on the evidence items only within each state frame to obtain gating features and writes them to the gating version record as additional fields, and S5 reads both the original node state sequence and the gating features to avoid the same evidence being counted repeatedly.

[0106] In this specific embodiment, S5 includes:

[0107] The evaluation reasoning module reads the fused bridge component state diagram output by S4 in the digital twin-based bridge maintenance evaluation method. The fused bridge component state diagram includes graph nodes, four types of edges, and fused node features. Evidence gating coefficient Conflict of evidence Initial values ​​for spatial attention weights and temporal attention weights are set. The evaluation inference module simultaneously loads the trained physical constraint spatiotemporal graph Transformer and maintenance knowledge graph. The model version record includes training sample batch, parameter version, feature dimension, label dictionary, physical constraint weights, and case retrieval threshold.

[0108] The physical constraint spatiotemporal graph Transformer includes a spatial encoding layer, a temporal encoding layer, and an evaluation output layer. The spatial encoding layer takes the node feature matrix and edge set of the bridge component state graph as input, and the temporal encoding layer takes the state sequence of each component node as input. The evaluation output layer receives the output of the spatial encoding layer, the output of the temporal encoding layer, and the case evidence vector obtained from the maintenance knowledge graph retrieval, and generates the component defect level, state confidence, component-level maintenance scope label, and construction risk level. Confidence intervals, review marks, priority re-inspection components, and evidence chains are generated by the uncertainty generation module and evidence chain module after the evaluation output layer and summarized into the final output of S5.

[0109] In this specific implementation, the spatial coding layer includes two spatial Transformer blocks, and the temporal coding layer includes two temporal Transformer blocks. The input dimension of each block is determined by concatenating the node feature dimension, the gating feature dimension, and the temporal position coding dimension. The spatial attention mask is applied to the attention score before the softmax function. Invalid mask positions are set to negative infinity, making their softmax weight zero. The temporal state sequence is truncated according to the most recent monitoring window or padded to a fixed length according to the previous valid state. The padded frame is marked as insufficient evidence and does not participate in the evidence count.

[0110] The spatial coding layer generates a spatial attention mask based on structural connection edges, spatially adjacent edges, load transfer edges, component stiffness, inter-component distances, and load path directions. The evaluation and inference module first performs a spatial attention mask on any node pair. Read edge type set and component stiffness normalization value and Distance between components and consistency of load path direction Press again:

[0111] Computational space values ​​of interest;

[0112] In the formula, , and These represent the existence values ​​of structural connection edges, spatially adjacent edges, and load transfer edges, respectively.

[0113] The time-coding layer performs positional encoding on the node state sequence according to the detection time, maintenance time, and monitoring window. The evaluation and inference module sorts the state frames of each component node by state timestamp and encodes the detection time interval, maintenance interval, and monitoring window number into time-position vectors. Generate, where, For nodes In the status frame Time location encoding, , and The time encoding function obtained from model training. , and These are the detection time interval, the post-maintenance interval, and the monitoring window number, respectively.

[0114] The temporal coding layer uses the degree of evidence conflict as an uncertainty input feature, and the evaluation inference module uses the average degree of evidence conflict within the same state frame. The number of highly conflicting pieces of evidence and the markers for insufficient evidence are concatenated into the time state vector, and then... The time-coded output is obtained, where, For state frames The node feature vectors, For time coding layer, It is written to the node time state cache when When the conflict input threshold is exceeded, the state frame increases the interval width contribution in the evaluation output layer;

[0115] The physical constraint spatiotemporal graph Transformer was trained using historical samples before deployment. These samples were constructed from historical bridge inspection records, historical sensor monitoring records, historical maintenance data, and post-maintenance re-inspection results. The samples used the actual defect level of the components, the scope of maintenance, and the results of construction risk verification as supervisory labels. The model loss function was... In the formula, Losses are classified according to disease severity. To cover the scope of repair losses, Classify losses according to construction risks. For physical constraint loss, , and From the training configuration table;

[0116] Physical constraint loss is calculated based on the consistency of the direction of disease propagation, load transfer, and post-repair state changes of adjacent components. The training module reads the direction of disease level change, load path direction, and post-repair state change for nodes with load transfer edges or historical disease-related edges, and then... Calculate, where, This indicates a penalty when the direction of spatial attention is inconsistent with the direction of disease propagation or load transfer. This represents the penalty term when the post-repair status change is inconsistent with the historical re-inspection results. After training, this loss constrains the spatial attention distribution and the temporal attention distribution.

[0117] The maintenance knowledge graph establishes entities and relationships based on bridge type, component category, material properties, defect type, defect size range, monitoring characteristic range, repair method, and treatment results. The knowledge graph also records historical defect treatment record vectors, treatment result deviations, re-inspection results, case time, and source project identifiers. Vector generation uses a combination of structured field binning embedding and relational graph embedding. Bridge type, component category, material properties, and defect type are embedded using field embedding, while entity relationships are trained using TransE to obtain graph embeddings. Training samples are derived from historical defect treatment records and post-repair re-inspection results, and the index is updated according to new case batches.

[0118] The retrieval score for historical disease treatment records is determined based on the cosine similarity between the retrieval vector and the historical disease treatment record vector, the consistency of component categories, and the consistency of disease types. The calculation formula is as follows: In the formula, For the first A vector of historical disease treatment records, For component category consistency values, This value represents the consistency of disease type. to The weights are derived from the case retrieval weight table and sum to one. The larger the value, the more suitable the historical records are as case evidence for evaluating the output layer;

[0119] When the retrieval score of the historical disease treatment record is not less than the preset case retrieval threshold, the evaluation reasoning module uses the record as case evidence in the evaluation output layer and inputs the treatment result deviation into the evaluation output layer. The treatment result deviation includes the difference between the actual repair scope and the model suggested scope, the difference between the construction risk review result and the prediction result, and the change of disease level after the repair re-inspection. Historical records that do not reach the case retrieval threshold are not included in the case evidence vector, but their retrieval score and the reason for not using them are written into the evidence chain candidate area.

[0120] The evaluation output layer receives spatial coding layer output, temporal coding layer output, and case evidence vector, all normalized to the same unit reference and truncated to the 0-1 interval. This dimensionless input generates the probability of the defect level, state confidence, component-level maintenance range label probability, and construction risk probability. The length, area, temperature, frequency, displacement, and defect size fields from the engineering quantity are all divided by the range to the 0-1 interval, and the component-level maintenance range label indicates whether the component is included in the subsequent recommended maintenance range set.

[0121] The uncertainty generation module generates confidence intervals, review markers, and priority re-inspection components based on node-level evidence conflict degree, risk propagation amount of adjacent components, and deviation of handling results. The risk propagation amount of adjacent components is calculated according to... Calculate, where, Distance between components Divide by the same unit spatial adjacency threshold The obtained distance normalized ratio is truncated to the interval between 0 and 1. The larger the value, the stronger the risk propagation from adjacent components to the current component;

[0122] The uncertainty generation module calculates the width of the disease level interval and the width of the construction risk interval, using the following formula: and In the formula, This refers to the width of the disease severity range. The width of the construction risk zone, For node-level evidence conflict degree, The normalized value for the deviation of the processing result is used, and all inputs are dimensionless values ​​in the range of 0 to 1;

[0123] When the range of disease levels is wide The value exceeds the preset disease review threshold, or the width of the construction risk range. When the risk exceeds the preset construction risk review threshold, the uncertainty generation module generates a review marker and will then review the relevant information. or The component node with the highest contribution value is determined as the priority component to be inspected. The contribution value is calculated by the risk propagation of adjacent components, the degree of conflict of relevant evidence for that component, and the deviation of the handling result. For nodes with the same ranking, the upstream node of the load transfer edge is selected first.

[0124] When outputting the suggested maintenance scope, the evaluation layer first outputs the component-level maintenance scope label probability for each component, and takes the component nodes whose component defect level reaches the preset maintenance trigger level or whose component-level maintenance scope label probability is not lower than the maintenance inclusion threshold as basic maintenance components; then the scope merging rule reads the load transfer edge and the risk propagation amount of adjacent components, and takes the component nodes that have a load transfer edge with the basic maintenance component and whose risk propagation amount of adjacent components is not less than the preset propagation threshold as extended maintenance components. The basic maintenance components and extended maintenance components are merged to form the set of suggested maintenance scopes for the entire bridge.

[0125] The evidence chain module generates an evidence chain for each output node. The evidence chain includes component node identifier, inspection events involved in the evaluation, monitoring features, image disease features, maintenance events, evidence gating coefficient, spatial attention weight, temporal attention weight, historical disease treatment records, treatment result deviation, basis for generating review markers, basis for generating priority re-inspection components and suggested maintenance scope. The evidence chain is stored using the output batch number and component node identifier as a joint index.

[0126] The evaluation reasoning module ultimately writes the component defect level, status confidence, component-level maintenance scope label, recommended maintenance scope set for the entire bridge, construction risk level, confidence interval, review mark, priority components for supplementary inspection, and evidence chain into the bridge maintenance evaluation result table. It also links the result table to the map version record of the bridge component status diagram after this fusion, enabling maintenance personnel to trace each digital twin bridge maintenance evaluation output along the component index table, component evidence set, gated fusion record, spatiotemporal attention weight, and historical defect treatment record.

[0127] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0128] This invention uses a bridge digital twin component base to constrain model components, on-site inspections, sensor monitoring, defect image recognition, and historical maintenance data to the same component node. It also saves structural connections, spatial adjacency, load transfer, and historical defect associations through a bridge component state diagram. This allows pre-construction condition assessment to no longer be limited to isolated inspection results, but to continuously reflect defect evolution, maintenance impact, and risk propagation along component relationships and time series, thus more directly solving the problem of difficulty in unifying the mapping of multi-source data.

[0129] The evidence confidence gating module does not change the basic object of bridge condition assessment, but improves the node feature fusion and attention weight adjustment methods to address the issues of confidence differences and conflicts among multi-source evidence. The physical constraint spatiotemporal graph Transformer and the maintenance knowledge graph are used together to form component defect level, maintenance scope, construction risk level, confidence interval and evidence chain, so that when the assessment results are not confident enough, they can output review marks and priority components for inspection, which is more suitable for traceable decision-making before bridge maintenance construction.

Claims

1. A bridge maintenance assessment method based on digital twins, characterized in that, include: S1. Read the building information model or industrial base model of the target bridge, extract the component number, spatial location, connection relationship and material properties, and generate a bridge digital twin component base containing beam body, bridge deck pavement, bearing and expansion joint component nodes. S2. Convert on-site inspection records, sensor monitoring data, defect image recognition results, and historical maintenance data into inspection events, monitoring features, image defect features, and maintenance events, respectively. Map these to component nodes according to component number, spatial location, and timestamp to form a component evidence set. S3. Based on the component nodes and component evidence set, establish a bridge component status diagram that includes structural connection edges, spatial adjacent edges, load transfer edges, and historical defect association edges. Update the node status according to inspection time, maintenance time, and monitoring window. S4. Calculate the source credibility, time decay coefficient, component matching confidence and evidence conflict degree for the component evidence set, generate evidence gating coefficient, and use the evidence gating coefficient to fuse node features and adjust the attention weight of the bridge component state diagram. S5. Input the fused bridge component state diagram into the trained physical constraint spatiotemporal graph Transformer, combine it with the historical disease treatment records of the maintenance knowledge graph, and output the component disease level, state confidence, recommended maintenance scope, construction risk level, confidence interval, verification mark and evidence chain.

2. The bridge maintenance assessment method based on digital twins according to claim 1, characterized in that, S1 includes: Analyze the global identifiers, component categories, axis coordinates, elevations, spans, material design parameters, and support relationships of components in Building Information Modeling (BIM) or Industrial Basic Modeling (IMM). Generate temporary component numbers for model components with missing component numbers according to component category, spatial bounding box, and adjacent connectors; A one-to-one correspondence index table is established between the global component identifier, the component number, and the temporary component number, and the index table is used as the basis for component location in subsequent evidence mapping and state diagram updates.

3. The bridge maintenance assessment method based on digital twins according to claim 1, characterized in that, S2 includes: extracting inspection time, inspection location, disease type, disease size, and treatment suggestions from on-site inspection records to generate inspection events; calculating statistical characteristics of strain, deflection, vibration frequency, temperature, and displacement from sensor monitoring data according to a preset monitoring window to generate monitoring features; extracting crack length, crack width, spalling area, and seepage area from disease image recognition results to generate image disease features; extracting maintenance time, maintenance location, maintenance method, and post-maintenance re-inspection results from historical maintenance data to generate maintenance events; and writing the inspection events, monitoring features, image disease features, and maintenance events into a component evidence set according to component number matching results, spatial nearest component matching results, and timestamps.

4. The bridge maintenance assessment method based on digital twins according to claim 1, characterized in that, S3 includes: using component nodes as graph nodes, generating structural connection edges based on model connection relationships, generating spatially adjacent edges based on component bounding box distances not exceeding a preset spatial adjacency threshold, generating load transfer edges based on support reaction paths, main beam force transmission paths, and bridge deck pavement force transmission paths, and generating historical defect association edges based on records of the same component or adjacent components exhibiting the same defect type within a historical time window; wherein, the preset spatial adjacency threshold is determined by the minimum bounding box size of the bridge model components and the manually set maintenance inspection radius, and the historical time window is determined by the maintenance cycle of similar bridges in the maintenance database; each time a new set of component evidence is received, the state version number, state timestamp, and node feature vector of the corresponding component node are updated to the bridge component state graph.

5. The bridge maintenance assessment method based on digital twins according to claim 1, characterized in that, S4 includes: calculating the source credibility based on the measurement and calibration records of the data source, the status of the acquisition equipment, the qualifications of the recording personnel, and the completeness of the maintenance data archive; calculating the time decay coefficient using an exponential decay function based on the interval between the evidence acquisition time and the evaluation time; calculating the component matching confidence based on the component number matching result, spatial distance matching result, and image positioning matching result; calculating the evidence conflict degree based on the difference in disease level, size difference, and monitoring feature deviation of the same component node under the same disease type; inputting the source credibility, time decay coefficient, component matching confidence, and evidence conflict degree into a gating mapping function to obtain the evidence gating coefficient of the corresponding evidence. The evidence gating coefficient is used to determine the fusion weight of the feature vector of the corresponding evidence entering the node and to correct the spatial attention weight and temporal attention weight.

6. The bridge maintenance assessment method based on digital twins according to claim 1, characterized in that, The physical constraint spatiotemporal graph Transformer in step S5 includes a spatial encoding layer, a temporal encoding layer, and an evaluation output layer. The spatial encoding layer generates a spatial attention mask based on structural connection edges, spatially adjacent edges, load transfer edges, component stiffness, distance between components, and load path direction, and uses the evidence gating coefficient as a multiplicative adjustment amount for the spatial attention weight. The temporal encoding layer encodes the node state sequence according to the detection time, maintenance time, and monitoring window, and uses the evidence conflict degree as the uncertainty input feature of the temporal encoding layer. The evaluation output layer generates component defect level, status confidence, recommended repair scope, and construction risk level based on the outputs of the spatial coding layer and the temporal coding layer.

7. The bridge maintenance assessment method based on digital twins according to claim 6, characterized in that, The physical constraint spatiotemporal graph Transformer is trained as follows: a training sample is constructed containing historical bridge inspection records, historical sensor monitoring records, historical maintenance data, and post-maintenance re-inspection results; the actual defect level of the components, the maintenance scope, and the construction risk verification results are used as supervision labels; defect level classification loss, maintenance scope regression loss, construction risk classification loss, and physical constraint loss are added to the model loss function; wherein, the physical constraint loss is calculated based on the consistency of defect propagation direction, load transfer direction, and post-maintenance state change of adjacent components, and is used to constrain the spatial attention distribution and temporal attention distribution after model training.

8. The bridge maintenance assessment method based on digital twins according to claim 6, characterized in that, The historical disease treatment records combined with the maintenance knowledge graph include: establishing entities and relationships in the maintenance knowledge graph based on bridge type, component category, material properties, disease type, disease size range, monitoring feature range, repair method, and treatment results; generating retrieval vectors based on the component properties, node feature vectors, and disease types of the assessment object; determining the retrieval score of the historical disease treatment records based on the cosine similarity between the retrieval vector and the historical disease treatment record vector, as well as the consistency of component category and disease type; and using historical disease treatment records with retrieval scores not less than the preset case retrieval threshold as case evidence for the assessment output layer, and inputting the deviation of their treatment results into the assessment output layer.

9. The bridge maintenance assessment method based on digital twins according to claim 8, characterized in that, The confidence interval, verification mark, and priority inspection component are generated as follows: the width of the defect level interval and the width of the construction risk interval are calculated based on the node-level evidence conflict degree, the risk propagation amount of adjacent components, and the deviation of the handling result; wherein, the risk propagation amount of adjacent components is jointly determined by the defect level of adjacent components, the load transfer edge direction, and the spatial adjacent edge distance; when the defect level interval width is greater than the preset defect verification threshold, or the construction risk interval width is greater than the preset construction risk verification threshold, a verification mark is generated; the component node whose contribution value to the defect level interval width or construction risk interval width is ranked first is determined as the priority inspection component.

10. The bridge maintenance assessment method based on digital twins according to claim 9, characterized in that, The evidence chain includes component node identifiers, inspection events involved in the assessment, monitoring features, image defect features, maintenance events, evidence gating coefficients, spatial attention weights, temporal attention weights, historical defect treatment records, and the basis for generating review markers. When outputting the recommended maintenance scope, component nodes whose defect level reaches the preset maintenance trigger level are used as basic maintenance components, and component nodes that have a load transfer edge with the basic maintenance component and whose risk propagation amount of adjacent components is not less than the preset propagation threshold are used as extended maintenance components. The basic maintenance components and extended maintenance components are merged to form the recommended maintenance scope.

Images