Construction, management and maintenance integrated data management system for traffic infrastructure

Abstract

The invention discloses a construction, management and maintenance integrated data management system for traffic infrastructure, and relates to the technical field of traffic infrastructure full life cycle management, and the system comprises a dynamic space-time grid coding engine, a multi-modal feature vector fusion engine, a dual verification module and a self-evolution feedback module which are connected in sequence. According to the construction, management and maintenance integrated data management system for the traffic infrastructure, the problem of space-time reference dislocation of multi-source heterogeneous data is solved through a dynamic space-time grid coding engine, and accurate space association of a design model, construction records and monitoring data is achieved; the cross-modal fusion mechanism based on lightweight feature extraction and twin network hard constraint solves the problem of intelligent alignment of images, texts and time series data, and improves the real-time performance and accuracy of disease state perception. A physical rule engine preferential interception and dual verification collaborative architecture eliminates decision risks caused by data errors and solves the hidden danger of delay of structural safety early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of full life cycle management of transportation infrastructure, and specifically to an integrated data management system for the construction, management and maintenance of transportation infrastructure. Background Art

[0002] In the full lifecycle management of transportation infrastructure, achieving integrated data management for construction, management, and maintenance is a key goal for improving efficiency, ensuring safety, and optimizing decision-making. While existing data management systems integrate information from different stages to a certain extent, they face a significant and fundamental bottleneck when addressing actual business needs. The construction, management, and maintenance of transportation infrastructure involve an extremely complex and diverse array of data sources, including design models, construction records, manual inspection reports, sensor information collected from vehicles or aircraft, time-series data from fixed-point monitoring, and various image and video materials. These data vary significantly in format specifications, collection accuracy, update frequency, and, most critically, spatiotemporal references.

[0003] Existing systems generally lack efficient, automated core capabilities, and are unable to intelligently identify, compare, and correlate data from these diverse sources describing the same infrastructure entity or the same type of status within a unified spatiotemporal framework. For example, it is difficult to automatically match and correlate an abnormal location detected by automated inspection with defect descriptions in historical manual records and real-time changes captured by nearby sensors. This lack of cross-source, heterogeneous data fusion capabilities results in systems relying heavily on tedious manual data search, verification, and splicing, which is inefficient and prone to errors. The direct consequence is an inability to accurately capture subtle trends in infrastructure status over time, resulting in delayed early warning of potential risks and ultimately incomplete and in-time basis for maintenance decisions, making it difficult to support accurate judgments and actions based on full-lifecycle data. Therefore, the current technical challenge is how to efficiently and automatically achieve accurate comparison, correlation, and deep fusion of multi-source, heterogeneous infrastructure maintenance data within a unified spatiotemporal framework. Summary of the Invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: an integrated data management system for the construction, management and maintenance of transportation infrastructure, comprising a dynamic spatiotemporal grid coding engine, a multimodal feature vector fusion engine, a dual verification module and a self-evolution feedback module connected in sequence;

[0005] The dynamic spatiotemporal grid coding engine divides bridges and roads into grid cells with unique spatiotemporal codes, and integrates Beidou positioning data, BIM component coordinates, and sensor physical coordinates through a nonlinear coordinate mapping algorithm.

[0006] The multimodal feature vector fusion engine includes a structured data processing pipeline and an unstructured data processing pipeline, which respectively generate statistical feature vectors, visual feature vectors, and semantic feature vectors, and output a unified state description vector after alignment by a twin network;

[0007] The dual verification module is composed of a physical rule engine and a dynamic adversarial verification network in series. The physical rule engine presets a crack propagation model and a material fatigue equation, and the dynamic adversarial verification network includes a generator and a discriminator.

[0008] The self-evolution feedback module adjusts the grid size and fusion weight according to the verification result.

[0009] Preferably, the dynamic spatiotemporal grid coding engine performs dynamic grid adjustment operations, the crack detection area uses grid units with a preset side length of 5 to 10 cm, and the settlement monitoring area uses grid units with a preset side length of 1 to 2 m. The numerical range of the grid units preset to 5 to 10 cm and the grid units preset to 1 to 2 m is set according to the monitoring accuracy requirements.

[0010] Preferably, the nonlinear coordinate mapping algorithm includes: establishing a rigid transformation matrix between Beidou coordinates and the center point of the BIM component; when the deviation between the sensor coordinates and the grid center point exceeds a preset spatial deviation threshold, triggering grid redivision based on the sensor coordinates, wherein the preset spatial deviation threshold is set according to the sensor positioning accuracy and the grid resolution.

[0011] Preferably, the structured data processing pipeline uses a temporal feature extractor to calculate the mean, variance and linear regression slope;

[0012] In the unstructured data processing pipeline:

[0013] The image data is passed through a three-convolutional layer lightweight network to extract edge features;

[0014] The text data is used to extract crack and spalling disease entities through the named entity recognition model.

[0015] Preferably, the twin network alignment process includes:

[0016] The visual feature vector and the semantic feature vector are input into a two-branch network with shared weights;

[0017] Using spatiotemporal grid coding as a constraint for similarity calculation;

[0018] Outputs a state description vector with a 0-1 confidence score.

[0019] Preferably, the execution process of the physical rule engine includes: when the daily crack expansion in the state description vector exceeds a preset expansion rate threshold, determining that the concrete crack expansion equation is violated; triggering an alarm instruction and interrupting data output, wherein the preset expansion rate threshold is set based on the concrete material properties and engineering specifications.

[0020] Preferably, the dynamic adversarial verification network performs:

[0021] The generator simulates the association results based on historical maintenance records;

[0022] The discriminator compares the simulation results with the actual maintenance effect data;

[0023] The data output interface is activated only when the output confidence is not lower than the preset confidence threshold and passes the physical rule verification, where the preset confidence threshold is configured according to the maintenance simulation accuracy requirements.

[0024] Preferably, the connection relationship of the dual verification modules is:

[0025] The output end of the physical rule engine is connected to the input end of the dynamic adversarial verification network;

[0026] Data that fails validation is returned to the fusion engine via a feedback loop.

[0027] Preferably, the self-evolution feedback module performs:

[0028] According to the alarm frequency of the physical rule engine, the corresponding grid is subdivided into a preset scale ratio of the original size. The implementer can independently set the preset scale ratio according to the monitoring accuracy requirements;

[0029] Based on the decrease in the discriminator confidence, the fusion weights of the twin network are updated according to the gradient.

[0030] Preferably, the integrated data management system for construction, management and maintenance of transportation infrastructure further includes:

[0031] A visual decision terminal connected to the output of the double verification module;

[0032] A distributed database that stores linked data encoded in a spatiotemporal grid.

[0033] The present invention provides an integrated data management system for the construction, management and maintenance of transportation infrastructure. It has the following beneficial effects:

[0034] This integrated data management system for construction, management, and maintenance of transportation infrastructure utilizes a dynamic spatiotemporal grid coding engine to address the spatial and temporal datum misalignment of heterogeneous multi-source data, enabling precise spatial correlation of design models, construction records, and monitoring data. A cross-modal fusion mechanism based on lightweight feature extraction and hard constraints from twin networks solves the intelligent alignment of image, text, and time-series data, improving the real-time and accuracy of disease status perception. A physical rule engine prioritizes interception and dual verification architecture eliminates decision-making risks caused by data errors and addresses potential delays in structural safety warnings.

[0035] This integrated data management system for construction, management, and maintenance of transportation infrastructure features a self-evolving feedback module that empowers dynamic optimization. Its grid segmentation and weight adjustment mechanisms adapt to the long-term evolution of infrastructure, reducing reliance on manual parameter adjustments. Full-link lineage traceability, under a unified spatiotemporal coding index, makes data from the entire construction, management, and maintenance cycle traceable, verifiable, and reliable. This enables a shift in maintenance decision-making from passive response to proactive prediction, extending the service life of facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the overall framework of the present invention;

[0037] Figure 2 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] See also Figure 1 and Figure 2 , the present invention provides a technical solution: an integrated data management system for the construction, management and maintenance of transportation infrastructure, comprising a dynamic spatiotemporal grid coding engine, a multimodal feature vector fusion engine, a dual verification module and a self-evolution feedback module connected in sequence;

[0040] The dynamic spatiotemporal grid coding engine divides bridges and roads into grid cells with unique spatiotemporal codes, and integrates Beidou positioning data, BIM component coordinates, and sensor physical coordinates through a nonlinear coordinate mapping algorithm.

[0041] The multimodal feature vector fusion engine includes structured data processing pipelines and unstructured data processing pipelines, which generate statistical feature vectors, visual feature vectors, and semantic feature vectors respectively. After alignment through the twin network, the engine outputs a unified state description vector.

[0042] The dual verification module consists of a physical rule engine and a dynamic adversarial verification network in series. The physical rule engine presets the crack propagation model and material fatigue equation, and the dynamic adversarial verification network includes a generator and a discriminator.

[0043] The self-evolution feedback module adjusts the grid size and fusion weight according to the verification results.

[0044] It should be further explained that in the specific implementation process, the system first processes multi-source data through a dynamic spatiotemporal grid coding engine, including: After receiving Beidou positioning data, BIM component coordinates and sensor physical coordinates, the system divides entities such as bridge piers and road sections into dynamic grid units, and generates a unique spatiotemporal code for each unit. For crack detection scenarios, a fine grid with a side length of 5 to 10 cm is used, while for settlement monitoring scenarios, a coarse-grained grid with a side length of 1 to 2 meters is switched. When the deviation between the sensor coordinates and the grid center point exceeds 10 cm, the grid is automatically re-divided based on the high-precision sensor as the reference to ensure a unified spatial reference.

[0045] The multimodal feature vector fusion engine splits and translates the input data. This includes extracting the mean, variance, and trend slope of structured data, including vibration sensor time series, to form a statistical vector. For unstructured data, a three-layer convolutional lightweight network is used to extract edge feature vectors from crack images, and a named entity recognition model is used to extract semantic vectors such as "crack width" and "stripping location" from manual inspection text. All feature vectors are then fed into a twin network for cross-modal alignment. Using spatiotemporal grid encoding as a hard constraint, the network calculates the similarity weights between visual and semantic features, outputting a unified state description vector with a 0-1 confidence score.

[0046] The dual verification module strictly screens the fusion results, including: the physical rule engine presets a daily expansion threshold of 0.5 mm for concrete cracks. If the state vector shows that the widening of a crack in a single day exceeds the limit, the data output is immediately frozen and an alarm is issued; the data input through physical verification is dynamically adversarially verified in the network, the generator simulates the association results based on historical maintenance records, and the discriminator compares the simulation results with the actual maintenance effect data. The data output interface is activated only when the output confidence level is ≥0.85.

[0047] The self-evolutionary feedback module dynamically optimizes the system based on verification results. This includes: When the physical rule engine issues three consecutive alerts on the same grid area, it automatically subdivides that area into a quarter of its original size. Furthermore, if the discriminator in the dynamic adversarial verification network detects a continuous decrease in the confidence level of a particular data type, it gradient-updates the corresponding modal fusion weight coefficient in the twin network. Ultimately, the verified data is pushed to the visualization decision terminal and stored in a distributed database using a spatiotemporal grid encoding index.

[0048] The dynamic spatiotemporal grid coding engine dynamically adjusts the grid, using grid cells with a preset side length of 5 to 10 cm in the crack detection area and 1 to 2 m in the settlement monitoring area. It should be further explained that during implementation, the system dynamically configures the grid cell size based on the target type during initialization. When processing crack detection tasks, fine grid cells with a side length of 5 to 10 cm are automatically enabled. This size range can capture millimeter-level crack width variations and avoid data overload. In settlement monitoring scenarios, the system switches to a coarse-grained grid cell size of 1 to 2 meters, balancing the efficiency and accuracy requirements of large-scale deformation monitoring.

[0049] The gridding operation is performed in real time in the spatiotemporal coding engine. If crack and settlement monitoring data need to be processed simultaneously in the same area, the fine grid for crack detection is used first and the settlement sensor interpolation data is integrated.

[0050] The dynamic grid adjustment process follows preset rules, including: the minimum grid size is enabled by default for vulnerable areas such as bridge expansion joints and tunnel linings, while the maximum grid size is used for uniform structures such as roadbed fill areas.

[0051] When the physical coordinates of the tilt sensor deployed on the pier deviate from the current grid center by more than 10 cm, the surrounding 3×3 grid clusters are immediately re-divided based on the high-precision coordinates of the sensor until the deviation between the new grid center and the sensor coordinates is less than 2 cm.

[0052] The repartitioning process synchronously updates the spatial topology of BIM components, ensuring that different inspection data for the same component always belongs to the same grid code. Historical version data generated by grid size switching and repartitioning operations is stored by timestamp index to support retrospective analysis of state evolution.

[0053] The nonlinear coordinate mapping algorithm involves establishing a rigid transformation matrix between Beidou coordinates and the center point of a BIM component. When the deviation between the sensor coordinates and the grid center exceeds a preset spatial deviation threshold of 10 cm, grid re-division is triggered based on the sensor coordinates. It should be further explained that, in its implementation, when the system receives Beidou positioning data, it first solves its WGS-84 geodetic coordinates and converts them to a local engineering coordinate system. It then extracts the design coordinates of the component center point in the BIM model. A rigid transformation matrix is then established between the two, and the rotation and translation parameters are fitted using the least squares method to achieve initial alignment of the design coordinates with the measured coordinates.

[0054] If the physical coordinates of devices such as inclination sensors and strain gauges deployed on the surface of a component deviate from the current grid center point by more than 10 cm, the coordinate reference is determined to be inaccurate and the calibration procedure is immediately initiated, including: using the high-precision sensor coordinates as the reference point, freezing the data processing flow of the 3×3 grid area around the sensor, recalculating the grid boundaries and dividing new units according to the reference point coordinate values, until the deviation between the new grid center point and the sensor coordinates is less than 2 cm.

[0055] The repartitioning process automatically inherits the original grid's spatiotemporal encoding and adds a version identifier. The BIM component spatial topology database is also updated to ensure that multiple monitoring data for the same component can be traced back to the same coding system. In areas where multiple high-precision sensors exist, the device with the highest measurement accuracy is prioritized as the benchmark. If the accuracy is the same, the sensor closest to the grid's geometric center is selected to avoid global topology conflicts caused by local calibration. After calibration, the grid parameters and transformation matrix are incrementally updated to the spatial index library for subsequent automatic matching and retrieval of multi-source data.

[0056] The structured data processing pipeline uses a time series feature extractor to calculate the mean, variance, and linear regression slope;

[0057] In the unstructured data processing pipeline:

[0058] The image data is passed through a three-convolutional layer lightweight network to extract edge features;

[0059] The text data is used to extract crack and spalling disease entities through the named entity recognition model.

[0060] It should be further explained that, in the specific implementation process, after the structured data enters the processing pipeline, the time series data generated by vibration sensors, temperature and humidity monitors, etc. are calculated using a 15-minute sliding window to calculate the mean value to reflect the stability level, the variance to characterize the intensity of fluctuations, and the trend slope to identify the long-term evolution direction through least squares fitting; when the absolute value of the slope increases for three consecutive windows, the window is automatically shortened to 5 minutes to capture sudden abnormal fluctuations.

[0061] In unstructured data processing, crack images are input into a lightweight network with three convolutional layers. This network includes: the first layer uses 64-channel 3x3 convolution to extract coarse edges; the second layer uses 128-channel 1x1 convolution to enhance crack textures; and the third layer uses 64-channel 3x3 convolution to suppress noise. The network outputs a 512-dimensional edge feature vector. This structure has been verified by a bridge crack database and reduces the number of parameters to 18% of a standard ResNet while ensuring a 95% recognition rate.

[0062] The manual inspection text is processed by the named entity recognition model, and the "cracks", "peeling", and "water seepage" disease entities and their modifiers are preferentially extracted, such as "transverse crack width 3mm". If the same sentence contains a location description, such as "K25+300 pier top", the associated disease entities are generated to generate a spatial semantic vector; when the text description conflicts with the image detection result, such as the text reports "no cracks" but the image detects cracks, the image feature weight is automatically increased and the review process is triggered.

[0063] The twin network alignment process involves feeding the visual and semantic feature vectors into a two-branch network with shared weights; using the spatiotemporal grid code as a constraint for similarity calculation; and outputting a state description vector with a 0-1 confidence score. It should be further explained that, in the specific implementation, the visual and semantic feature vectors are fed into the two-branch network structure with shared weight parameters in parallel. The left branch processes the 512-dimensional edge feature vector extracted from the image, while the right branch processes the semantic vector of the defect generated from the text. When the input data carries the same spatiotemporal grid code, the similarity calculation module is activated, calculating the cosine similarity of the two feature vectors and multiplying it by the grid position weight coefficient. If the grid is located in a critical area such as a bridge expansion joint, the weight is increased to 1.5 times the baseline value.

[0064] The similarity calculation results are converted into confidence scores in the range of 0-1, including: when the score is higher than 0.7, the fused state description vector is directly generated; when the score is in the range of 0.4-0.7, the cross-modal enhancement module is triggered, and the attention mechanism is used to focus on key attributes such as crack width and position coordinates for secondary matching; when the score is lower than 0.4, it is judged as an invalid association, and the data set is automatically discarded and the grid coding anomaly is marked.

[0065] The output vector is accompanied by a confidence score and a list of key attribute matches. When the position coordinate deviation in the list exceeds 5 cm or the crack width difference is greater than 0.2 mm, the confidence score is forced to be downgraded to below 0.3 and returned to the feature extraction engine for reprocessing.

[0066] The execution process of the physical rule engine includes: when the daily crack expansion in the state description vector exceeds the preset expansion rate threshold >0.5mm, it is determined that the concrete crack expansion equation is violated; an alarm instruction is triggered and data output is interrupted, where the preset expansion rate threshold is set based on the concrete material properties and engineering specifications.

[0067] It should be further explained that, in the specific implementation process, the execution process of the physical rule engine includes: the system presets a daily concrete crack expansion threshold of 0.5 mm as the core judgment benchmark. This value is derived from the fracture mechanics model corresponding to the tensile strength limit of C40 concrete. When the fused state description vector shows that the daily width increment of a crack exceeds this limit, the data output stream of the grid unit is immediately frozen; the verification rules are automatically switched for different structural materials, namely: the fatigue stress amplitude threshold of 150 MPa is enabled for steel structures, and the daily change threshold of rutting depth of 3 mm is used for asphalt pavement.

[0068] When an alarm is triggered, the root cause analysis is initiated simultaneously, including: if the excessive crack is located in the pressure-bearing area of the pier, it is marked as a structural risk, and a red alert is generated and pushed to the maintenance terminal; if it is in a non-load-bearing area such as a crash barrier, it is marked as appearance damage and the drone review process is initiated.

[0069] For material fatigue equations, when the trend slope of vibration sensor data exceeds 0.15 for three consecutive hours after feature extraction, and the vehicle load spectrum during that period shows that heavy vehicles account for more than 40%, the steel structure's SN curve is automatically used to predict the remaining life. All frozen data generates a snapshot of abnormalities with spatiotemporal encoding, and historical records of similar incidents are linked for decision-making, until manual confirmation is made to unfreeze the data.

[0070] The dynamic adversarial verification network executes: The generator simulates association results based on historical maintenance records; the discriminator compares the simulation results with actual maintenance effect data; the data output interface is activated only when the output confidence is not less than a preset confidence threshold ≥0.85 and passes physical rule verification, where the preset confidence threshold is configured based on the required maintenance simulation accuracy. It should be further explained that in the specific implementation process, the dynamic adversarial verification network execution process includes: the generator simulates association results based on a typical disease pattern library in the maintenance records of the past three years, preferentially extracting historical cases with crack widths in the range of 0.2-0.5mm and located at the top of the pier to generate a virtual state vector; the discriminator receives the real-time fused data stream and compares the Mahalanobis distance between the current vibration spectrum of the component under test and the normal state after historical maintenance. When the distance value exceeds three times the standard deviation, the confidence score is automatically reduced.

[0071] The verification process implements a double interception mechanism, namely: if the physical rule engine has marked a structural risk, a basic confidence level of 0.3 is directly output; only when the physical verification passes and the discriminator confidence level is ≥0.85, the data output interface is activated and a status report with a green trust mark is generated.

[0072] For data with a confidence level between 0.75 and 0.85, the system automatically links the three most recent high-definition drone inspection images of that grid cell for manual review and decision-making assistance. After three consecutive verifications of a data stream with decreasing confidence, the feature fusion engine's weight coefficient reset process is triggered, freezing the data output for that grid cell until the reset is complete. All intercepted data generates a spatiotemporal-encoded verification log, annotating root causes such as "physical rule conflict," "inadequate judgment confidence," or "continuous decline anomaly." This log is summarized weekly to drive optimization of the self-evolutionary module.

[0073] The dual verification module is connected as follows: the output of the physical rule engine is connected to the input of the dynamic adversarial verification network; data that fails verification is returned to the fusion engine via a feedback loop. It should be further explained that, in its implementation, the dual verification module executes the following steps: the physical rule engine prioritizes receiving the state description vector output by the feature fusion engine. If it detects a crack's daily expansion exceeding 0.5 mm or a steel structure stress amplitude exceeding 150 MPa, it immediately freezes the grid data stream and generates a red alarm code. Only vectors that pass physical verification are transmitted to the dynamic adversarial verification network.

[0074] The adversarial network discriminator calculates the Mahalanobis distance between real-time data and historical normal states. If the confidence score is ≥0.85, the green pass instruction is activated and the data is pushed to the decision-making terminal. If the confidence score is in the range of 0.75-0.85, the three most recent drone images are automatically associated to generate a yellow review report. If it is lower than 0.75, the data return mechanism is triggered, that is, vectors that fail any verification carry an error code and are returned to the feature fusion engine.

[0075] Reflowing data triggers a priority reset, including: vectors that fail physical rules are forced to re-extract raw sensor data, and vectors that fail adversarial verification are adjusted for twin network fusion weights and then re-aligned. If the same grid data is reflowed three times in a row, a comprehensive verification process for the grid's spatiotemporal encoding is initiated, and associated business processing is suspended.

[0076] The self-evolution feedback module performs:

[0077] Based on the alarm frequency of the physical rule engine, the corresponding grid is subdivided into a preset scale ratio of 1 / 4 of the original size. The implementer can independently set the preset scale ratio according to the monitoring accuracy requirements;

[0078] Based on the decrease in the discriminator's confidence, the twin network fusion weights are updated according to the gradient. It should be further explained that in the specific implementation process, the self-evolution feedback execution process includes: when the physical rule engine triggers an alarm for a specific grid unit, the system automatically records the alarm frequency of that grid; if the same grid triggers three consecutive alarms due to excessive crack expansion, the grid in that area is immediately subdivided into a quarter of its original size, and edge computing nodes are deployed in the new grid to increase the sampling frequency.

[0079] For the confidence score output by the dynamic adversarial verification network, the average weekly confidence value of each grid unit is monitored; when the confidence of a specific grid decreases by more than 0.15 for two consecutive weeks, the weight update program is automatically started, that is: if the main reason for the decrease is the conflict between image and text features, the weight coefficient of the text branch in the twin network will be reduced by 30%, and the weight of the image branch will be increased to 1.3 times the original value.

[0080] After the weights are adjusted, the last 72 hours of data are reprocessed. If the average confidence level rises above the baseline, the optimization is confirmed to be effective. If it continues to decline, expert intervention mode is triggered, freezing automatic updates and issuing a structural health assessment request. All optimization records are archived by grid code. When new infrastructure with similar structures is connected to the system, the corresponding optimization parameters are automatically loaded as the initial configuration.

[0081] The integrated data management system for the construction, management, and maintenance of transportation infrastructure also includes: a visual decision terminal connected to the output of the dual verification module; and a distributed database that stores related data in space-time grid coding. It should be further explained that in the specific implementation process, the data output and storage execution process includes:

[0082] The visual decision-making terminal receives the green pass data stream activated by the dual verification module in real time. Structural risk areas marked by physical rules, such as crack-exceeding-limit grids, are marked with flashing red dots in the three-dimensional BIM model, and a comparison of the three most recent high-definition drone images is displayed. The defect data with a verification confidence level ≥0.85 is generated into a status card with a trusted mark, and similar historical maintenance plans and material consumption lists are automatically associated.

[0083] The distributed database uses spatiotemporal grid encoding as the primary index key, and data lineage information is synchronously recorded during storage. Specifically, structured data retains the sensor serial number and acquisition timestamp, while unstructured data is appended with a lightweight network feature extraction parameter version. When a user queries the evolution of cracks in a pier from 2023 to 2024, the system prioritizes accessing thermal data from edge nodes. If a complete analysis requires access to the central database, a bandwidth estimate is automatically generated. Grid cells that continuously trigger the circuit breaker mechanism are marked with a yellow warning in the database. Subsequent queries force the loading of the physical rule engine's verification log and coordinate repartitioning records to ensure complete and traceable decision information.

[0084] It should be further explained that during the specific implementation process, when the system starts, the dynamic grid coding engine loads the infrastructure BIM model and decomposes entities such as bridge piers and road sections into grid cells based on spatial location and time dimensions. Each cell generates a unique spatiotemporal code as a data association benchmark. The coding rules integrate Beidou satellite positioning coordinates, the coordinates of the center points of design model components, and the physical coordinates of sensors. When processing crack detection tasks, small-sized grid cells are automatically used to capture millimeter-level changes, while settlement monitoring switches to large-sized grid cells to improve processing efficiency. If an area has both types of detection requirements, the small-sized grid is retained first and the large-sized monitoring data is interpolated and integrated. The grid division process monitors coordinate deviations in real time. When the measured coordinates of the fixed sensor deviate from the grid center point by more than the normal positioning error, the grid boundary is immediately recalculated based on the high-precision sensor. The new grid inherits the original spatiotemporal code and adds a version identifier to ensure the traceability of historical data.

[0085] The multimodal feature vector fusion engine classifies and processes the input data, including: calculating the mean of time series data such as vibration sensors according to a fixed window to reflect the stable state, the variance to characterize the intensity of fluctuations, and the trend slope to indicate the direction of evolution. When the trend slope continues to increase, the analysis window is automatically shortened to capture sudden anomalies. The crack image is input into a lightweight convolutional network, which extracts edge features through three specially designed convolutional layers. The first layer captures the general outline, the second layer enhances the crack texture, and the third layer suppresses environmental noise, outputting a feature vector representing the crack morphology. The manual inspection text is processed by a dedicated entity recognition model, and key information such as crack width and spalling location is extracted in a targeted manner to generate a semantic vector. When the text description conflicts with the image recognition result, the system automatically increases the image feature weight and triggers the on-site review process.

[0086] Feature vectors are fed into a two-branch shared weight network for cross-modal alignment, and similarity is calculated using a spatiotemporal grid encoding as a hard constraint. Similarity results are converted to a unified confidence score. This includes: High-confidence data is directly fused; medium-confidence data initiates a secondary matching mechanism, focusing on core attributes such as crack location and width, and comparing them with historical records; low-confidence data is automatically discarded and marked as grid anomalies. The output vectors are accompanied by an attribute checklist. If the position coordinate deviation exceeds the typical measurement error or the crack width difference exceeds the accuracy of the detection instrument, the confidence score is forcibly downgraded and the system returns to the feature extraction stage.

[0087] The physical rule engine presets a material behavior model, including key parameters such as the daily extension threshold of concrete cracks and the stress amplitude limit of steel structures. When the fused data exceeds the threshold, the output stream is immediately frozen, and a red or yellow warning is generated according to the risk level of the structural part. The data input through physical verification is fed into the dynamic adversarial verification network. The generator simulates the correlation results based on the historical maintenance case library, and the discriminator compares the difference between the real-time data and the historical normal state. The green output channel is activated only when the credibility evaluation reaches a high confidence standard and the physical verification passes. Data that fails the verification is returned to the processing engine with an error code. Data that fails the physical rule re-extracts the original sensor readings, and data with insufficient credibility adjusts the feature fusion weights and is aligned again.

[0088] The self-evolving feedback module continuously monitors verification results. For example, when a specific grid cell repeatedly triggers physical alarms, it automatically subdivides the grid in that area to one-quarter its original size and increases the data collection frequency. If the credibility of a particular data fusion continues to decline, the system gradually reduces the weight coefficients of the conflicting modes and reprocesses recent data streams to verify the results after optimization. All optimized parameters are archived and stored by structure type, and when new infrastructure of the same type is built, the optimal configuration parameters are automatically loaded to initialize the system.

[0089] The visualization terminal receives green channel data, highlights risk areas in the 3D model, and links historical treatment plans. The distributed database stores data using spatiotemporal grid codes as the primary index key, synchronously recording the data source and processed version. User queries prioritize access to hot data from edge nodes, and automatically indicate bandwidth requirements when access to the central database is required for complete analysis. For grid cells that repeatedly fail verification, coordinate calibration records and physical rule alarm logs are forcibly linked to ensure complete and traceable decision information.

[0090] A method for integrated data management of construction, management and maintenance of transportation infrastructure, comprising the following steps:

[0091] Step S1: The dynamic spatiotemporal gridding engine loads the BIM model, initializes the grid cell size according to the structure type, receives Beidou coordinates, sensor coordinates, and design coordinates, and generates a unique spatiotemporal code through nonlinear mapping; when the deviation between the high-precision sensor coordinates and the grid center exceeds the conventional positioning error, it automatically triggers local grid re-division and appends a version identifier.

[0092] Step S2: The multimodal feature processing engine diverts input data, including: extracting mean, variance, and trend features from the sliding window of time series data; extracting edge feature vectors from crack images through a lightweight convolutional network; extracting disease entities from manual inspection text to generate semantic vectors; and giving priority to image data and initiating review when image and text features conflict.

[0093] Step S3: The shared weighted twin network performs cross-modal alignment with spatiotemporal coding as a constraint, calculates feature similarity and outputs a fusion vector with credibility evaluation; when the position or size attribute deviation exceeds the accuracy of the detection instrument, the credibility is forcibly downgraded and reflux processing is performed.

[0094] Step S4: The physical rule engine prioritizes verifying the fusion vector, including: freezing the data flow when the metamaterial behavior threshold is reached and issuing an alarm based on the risk level of the structural part; and inputting the verified vector into the dynamic adversarial verification network.

[0095] Step S5: The adversarial network generator simulates the association results based on the historical case library, and the discriminator compares the real-time data with the historical normal state; the green output channel is activated only when the high confidence standard and physical verification are passed at the same time; the failed data is diverted and recirculated according to the error type.

[0096] Step S6: The self-evolution module monitors and verifies the results, including: automatically subdividing the size of grid cells that continuously trigger physical alarms and increasing the sampling rate; when the credibility of specific modal fusion continues to decrease, gradient adjusting the feature weights and verifying the optimization effect.

[0097] Step S7: The visualization terminal receives green channel data, highlights the risk area in the three-dimensional model and associates it with historical treatment plans; the distributed database stores the full-link lineage information using spatiotemporal coding indexes.

[0098] The dynamic spatiotemporal grid coding engine solves the problem of spatiotemporal benchmark misalignment in multi-source heterogeneous data, enabling precise spatial correlation of design models, construction records, and monitoring data. A cross-modal fusion mechanism based on lightweight feature extraction and hard constraints from twin networks solves the problem of intelligent alignment of image, text, and time-series data, improving the real-time and accuracy of disease status perception. A physical rule engine prioritizes interception and dual verification, eliminating decision-making risks caused by data errors and addressing potential delays in structural safety warnings.

[0099] The self-evolving feedback module empowers the system with dynamic optimization capabilities. The grid subdivision and weight adjustment mechanism adapts to the long-term evolution of infrastructure, reducing reliance on manual parameter adjustments. Full-link lineage traceability, under a unified spatiotemporal coding index, makes data from the entire construction, management, and maintenance cycle traceable, verifiable, and reliable. This enables a shift in maintenance decision-making from passive response to proactive prediction, extending facility lifespan.

[0100] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0101] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A data management system for the construction, management and maintenance of transportation infrastructure, characterized by: It includes a dynamic spatiotemporal grid coding engine, a multimodal feature vector fusion engine, a double verification module and a self-evolution feedback module connected in sequence; The dynamic spatiotemporal grid coding engine divides bridges and roads into grid cells with unique spatiotemporal codes, and integrates Beidou positioning data, BIM component coordinates, and sensor physical coordinates through a nonlinear coordinate mapping algorithm. The multimodal feature vector fusion engine includes a structured data processing pipeline and an unstructured data processing pipeline, which respectively generate statistical feature vectors, visual feature vectors, and semantic feature vectors, and output a unified state description vector after alignment by a twin network; The dual verification module is composed of a physical rule engine and a dynamic adversarial verification network in series. The physical rule engine presets a crack propagation model and a material fatigue equation, and the dynamic adversarial verification network includes a generator and a discriminator. The self-evolution feedback module adjusts the grid size and fusion weight according to the verification result.

2. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 1, characterized in that: The dynamic spatiotemporal grid coding engine performs a dynamic grid adjustment operation. The crack detection area uses grid units with a preset side length of 5 to 10 cm, and the settlement monitoring area uses grid units with a preset side length of 1 to 2 m.

3. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 2, characterized in that: The nonlinear coordinate mapping algorithm includes: establishing a rigid transformation matrix between Beidou coordinates and the center point of the BIM component; when the deviation between the sensor coordinates and the grid center point exceeds a preset spatial deviation threshold, triggering grid re-division based on the sensor coordinates.

4. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 1, characterized in that: The structured data processing pipeline uses a temporal feature extractor to calculate the mean, variance, and linear regression slope; In the unstructured data processing pipeline: The image data is passed through a three-convolutional layer lightweight network to extract edge features; The text data is used to extract crack and spalling disease entities through the named entity recognition model.

5. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 4, characterized in that: The twin network alignment process includes: The visual feature vector and the semantic feature vector are input into a two-branch network with shared weights; Using spatiotemporal grid coding as a constraint for similarity calculation; Output a state description vector with a 0-1 confidence score.

6. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 1, characterized in that: The execution process of the physical rule engine includes: when the daily crack expansion amount in the state description vector exceeds a preset expansion rate threshold, determining that the concrete crack expansion equation is violated; triggering an alarm instruction and interrupting data output.

7. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 6, characterized in that: The dynamic adversarial verification network performs: The generator simulates the association results based on historical maintenance records; The discriminator compares the simulation results with the actual maintenance effect data; The data output interface is activated only when the output confidence is not lower than the preset confidence threshold and passes the physical rule verification.

8. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 1, characterized in that: The connection relationship of the dual verification modules is: The output end of the physical rule engine is connected to the input end of the dynamic adversarial verification network; Data that fails validation is returned to the fusion engine via a feedback loop.

9. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 7, characterized in that: The self-evolution feedback module performs: Based on the alarm frequency of the physical rule engine, the corresponding grid is subdivided into a preset scale ratio of the original size; Based on the decrease in the discriminator confidence, the fusion weights of the twin network are updated according to the gradient.

10. The integrated data management system for construction, management and maintenance of transportation infrastructure according to claim 1, characterized in that: The integrated data management system for construction, management and maintenance of transportation infrastructure also includes: A visual decision terminal connected to the output of the double verification module; A distributed database that stores linked data encoded in a spatiotemporal grid.

Images