Bridge construction quality real-time monitoring system and method based on artificial intelligence

Abstract

The application relates to the field of bridge construction quality management, and particularly discloses a bridge construction quality real-time monitoring system and method based on artificial intelligence. The system comprises: a multi-source data acquisition and fusion module for acquiring and fusing multi-modal construction data and outputting a high-order fusion feature vector; a quality analysis engine module for performing quality identification and prediction based on the fusion feature vector and outputting a structured quality analysis result; a knowledge graph and rule reasoning engine module for performing standard matching and reasoning on the analysis result and outputting problem causes and rectification suggestions; a real-time early warning and dynamic regulation module for performing global risk deduction and generating regulation instructions; and a digital twin and three-dimensional visualization module for realizing construction state visualization and decision assistance. Through multi-modal data deep fusion and dynamic rule evolution driven by a knowledge graph, the application realizes real-time monitoring, accurate early warning and closed-loop control of construction quality.

Description

Technical Field

[0001] This invention relates to the field of bridge construction quality management technology, and more specifically to a real-time monitoring system and method for bridge construction quality based on artificial intelligence. Background Technology

[0002] Currently, the quality of bridge construction directly affects the service life and operational safety of the structure. Traditional construction quality control mainly relies on manual inspection, offline sampling and testing, and experience-based judgment. This approach has several drawbacks, including long testing cycles, delayed feedback, limited testing points leading to incomplete coverage, inconsistent and subjective quality judgment standards, and a lack of real-time risk assessment and trend warning capabilities.

[0003] However, with the development of digital twin and intelligent monitoring technologies, existing research has proposed bridge construction monitoring systems based on digital twins. For example, by collecting construction data and constructing stress prediction models, construction status monitoring and scheme optimization can be achieved. However, existing technologies still generally have the following shortcomings: sensor data sources are singular, lacking effective integration of video information and equipment operation data; the technical solutions mainly focus on structural stress prediction, lacking intelligent identification capabilities for specific construction quality problems (such as rebar tying and prestressing tensioning); quality judgment relies on preset static thresholds, lacking intelligent reasoning and dynamic adjustment mechanisms based on construction specification clauses; and it is difficult to achieve automated root cause tracing and rectification decisions after quality problems are discovered.

[0004] Therefore, how to construct a bridge construction quality monitoring system with real-time perception, intelligent identification, automatic reasoning, dynamic early warning and closed-loop control capabilities is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide an artificial intelligence-based real-time monitoring system and method for bridge construction quality that overcomes or at least partially solves the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention relates to an artificial intelligence-based real-time monitoring system for bridge construction quality, comprising: The multi-source data acquisition and fusion module is used to collect multi-source sensing data during bridge construction and fuse the multi-source sensing data through a multi-modal data deep fusion model based on cross-attention mechanism to generate a high-order fusion feature vector.

[0008] The quality analysis engine module receives high-order fused feature vectors and performs construction quality identification and prediction tasks based on a deep learning model, outputting structured quality analysis results.

[0009] The knowledge graph and rule reasoning engine module receives structured quality analysis results, performs semantic matching and rule reasoning on the structured quality analysis results based on a pre-built construction specification knowledge graph, and generates reasoning results.

[0010] The real-time early warning and dynamic control module is used to receive structured quality analysis results and inference results, and generate risk early warning information.

[0011] The digital twin and 3D visualization module is used to receive high-order fusion feature vectors, structured quality analysis results, and risk warning information, and to map the received information onto the building information model for visualization.

[0012] Preferably, the multi-source data acquisition and fusion module includes: a distributed sensor network, a visual acquisition device, an industrial interface, and a multimodal data deep fusion model.

[0013] Distributed sensor networks are deployed at key structural parts of bridges to collect real-time data from these key structural parts, which is then used as sensor data.

[0014] Visual acquisition equipment is deployed in key construction areas to collect video footage and temperature distribution images of the construction process, serving as visual data.

[0015] The industrial interface is used to connect to the PLC of construction equipment to obtain equipment operation data and ratio data. It also obtains construction progress information and component information through the building information modeling system interface, and associates the component information with sensor data to generate construction equipment and business data.

[0016] The multimodal data deep fusion model uses two-dimensional spatial features extracted from visual data as query vectors and one-dimensional temporal dynamic features extracted from sensor data after encoding by a temporal convolutional network as key and value vectors. Cross-modal alignment and fusion are performed through a cross-attention mechanism to generate a high-order fused feature vector.

[0017] Preferably, the deep learning model loaded on the quality analysis engine module includes at least one of the following models: The concrete strength prediction model, based on mix proportion data, construction process data, environmental data, structural internal state data, and age data, uses a temporal convolutional network to predict the strength development of concrete at various ages, generate strength development curves and early warning information, and output structured quality analysis results.

[0018] The prestressed tensioning quality identification model receives high-order fusion feature vectors output by the multi-source data acquisition and fusion module, comprehensively determines the tensioning state category, and outputs structured quality analysis results.

[0019] The rebar tying quality inspection model is based on image data acquired by a multi-source data acquisition and fusion module. It uses a Mask R-CNN network for target detection and segmentation, and generates structured quality analysis results after comparing them with construction specification thresholds.

[0020] The temperature-visual linkage crack early identification model receives high-order fusion feature vectors output by the multi-source data acquisition and fusion module, comprehensively determines the coupling relationship between the internal hydration thermal gradient abrupt change and the surface infrared thermal field distribution, and outputs structured quality analysis results.

[0021] Preferably, the knowledge graph and rule reasoning engine module includes: The knowledge graph construction unit performs structured processing on the standard texts, process flow data, and historical engineering case data of bridge construction to construct a knowledge graph of construction standards.

[0022] The semantic matching and reasoning unit uses the BERT model to semantically encode the structured quality analysis results output by the quality analysis engine module and performs semantic matching with the standard text. The corresponding standard text is selected through similarity calculation, and the knowledge graph reasoning engine is combined to perform node path traversal and association calculation to generate reasoning results containing the causes of quality problems and rectification suggestions.

[0023] The dynamic compliance rule self-evolution unit presets the construction specification safety threshold. When the quality analysis engine module, combined with sensor data, finds that the construction specification safety threshold is met, but this results in a compliance critical deviation, the knowledge graph reasoning engine triggers rule reconstruction to dynamically shrink and correct the construction specification safety threshold.

[0024] Preferably, the real-time early warning and dynamic control module includes: The dynamic threshold calculation unit determines the construction stage, structural component type, and environmental conditions based on the building information model progress information. It obtains the construction stage coefficient, structural component coefficient, and environmental correction coefficient, and performs weighted adjustment on the construction specification safety threshold to generate a dynamic threshold.

[0025] The trend analysis unit uses an exponentially weighted moving average to smooth the sensor data and uses a cumulative control method to detect continuous deviations in the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is determined to be an abnormal trend.

[0026] The global risk simulation unit extracts the spatial topological relationships of the bridge structure based on the building information model, constructs a structural association graph model, and introduces a spatiotemporal graph convolutional network. It uses abnormal trends as the initial state vector and combines the physical force association rules and material properties in the construction specification knowledge graph to generate risk warning information.

[0027] Preferably, the digital twin and 3D visualization module includes: Based on the spatial coordinates and component numbers of the components in the building information model, they are uniformly bound with sensor-collected data, visual analysis results, and construction process information to form a component-time-data mapping relationship.

[0028] Transform sensor data and structured quality analysis results into visual representations.

[0029] Based on the spatial location of the components, the visual representation data is mapped to a digital twin and 3D visualization platform.

[0030] Secondly, the present invention also relates to a method for real-time monitoring of bridge construction quality based on artificial intelligence, comprising the following steps: S1. Collect multi-source sensing data during bridge construction; generate high-order fusion feature vectors through a multi-modal data deep fusion model based on cross-attention mechanism.

[0031] S2 receives high-order fusion feature vectors and performs construction quality identification and prediction tasks based on a deep learning model, outputting structured quality analysis results.

[0032] S3. Based on a pre-built knowledge graph of construction specifications, semantic matching and rule-based reasoning are performed on the structured quality analysis results to generate reasoning results.

[0033] S4. Receive the structured quality analysis results and inference results, and generate risk warning information.

[0034] S5. Map the high-order feature fusion vector, structured quality analysis results, and risk warning information onto the building information model for visualization.

[0035] Preferably, the multi-source sensing data acquisition and fusion steps in S1 include: Based on a distributed sensor network deployed at key structural parts of the bridge, real-time data from these key structural parts is collected as sensor data.

[0036] Based on visual acquisition equipment deployed in key construction areas, construction process videos and temperature distribution images are collected as visual data.

[0037] Based on the industrial interface of the PLC of the construction equipment, the system obtains equipment operation data and ratio data, acquires construction progress information and component information through the building information modeling system interface, and associates the component information with sensor data to generate construction equipment and business data.

[0038] Based on a multimodal data deep fusion model, two-dimensional spatial features extracted from visual data are used as query vectors, and one-dimensional temporal dynamic features extracted from sensor data after being encoded by a temporal convolutional network are used as key and value vectors. Cross-modal alignment and fusion are performed through a cross-attention mechanism to generate a high-order fusion feature vector.

[0039] Preferably, S3 also includes a dynamic compliance rule self-evolution step: A preset construction specification safety threshold is set. When the quality analysis engine module finds that the construction specification safety threshold is met by combining sensor data, but this results in a compliance threshold deviation, the knowledge graph reasoning engine triggers rule reconstruction to dynamically shrink and correct the construction specification safety threshold.

[0040] Preferably, the risk warning information generated in S4 includes: Based on the construction information model progress information, the construction stage, structural component type, and environmental conditions are determined. Construction stage coefficients, structural component coefficients, and environmental correction coefficients are obtained. The safety thresholds of the construction specifications are then weighted and adjusted to generate dynamic thresholds.

[0041] The sensor data is smoothed using an exponentially weighted moving average, and a cumulative control method is used to detect continuous deviations in the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is judged as an abnormal trend.

[0042] Based on Building Information Modeling (BIM), the spatial topological relationships of bridge structures are extracted, a structural association graph model is constructed, and a spatiotemporal graph convolutional network is introduced. Abnormal trends are used as initial state vectors, and combined with physical stress association rules and material properties in the construction specification knowledge graph, risk warning information is generated.

[0043] As can be seen from the above technical solution, compared with the prior art, the beneficial effects of the technical solution of the present invention include at least the following: 1. Achieve cross-modal physical law mapping and eliminate monitoring blind spots and false alarms: Through a multimodal data deep fusion model, spatial visual representations and internal physical quantities are dynamically aligned, and "surface anomalies" are automatically converged to "internal force mutation points", which significantly improves the accuracy of anomaly identification and source tracing capabilities in complex environments.

[0044] 2. Establish a global dynamic simulation mechanism for hidden dangers driven by both graph and twin: By introducing a spatiotemporal graph convolutional network, the impact path and weight of local quality defects on the overall structure are automatically simulated when they occur, realizing a technological leap from local alarm to global systemic risk evolution simulation.

[0045] 3. The machine has the ability to evolve dynamically in compliance with standards: The knowledge graph engine can capture the subtle hidden dangers of “complying with static specifications but causing nonlinear stress deterioration”, and adaptively shrink the safety threshold, so that the quality control system has the ability to learn deeply and adapt to the environment. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0047] Figure 1 This is a framework diagram of a construction quality monitoring system based on multi-source perception and knowledge graph reasoning provided in an embodiment of the present invention. Figure 2 This is a data processing flowchart of the artificial intelligence-based real-time monitoring system for bridge construction quality provided in this embodiment of the invention. Figure 3 This is a flowchart of a method for real-time monitoring of bridge construction quality based on artificial intelligence, provided in an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] like Figure 1 As shown in the figure, an embodiment of the present invention discloses a real-time monitoring system for bridge construction quality based on artificial intelligence, comprising: The multi-source data acquisition and fusion module is used to collect multi-source sensing data during bridge construction and fuse the multi-source sensing data through a multi-modal data deep fusion model based on cross-attention mechanism to generate a high-order fusion feature vector.

[0050] The multi-source data acquisition and fusion module includes: a distributed sensor network, visual acquisition equipment, industrial interfaces, and a multimodal data deep fusion model.

[0051] Distributed sensor networks are deployed at key structural parts of bridges to collect real-time data from these key structural parts, which is then used as sensor data.

[0052] Visual acquisition equipment is deployed in key construction areas to collect video footage and temperature distribution images of the construction process, serving as visual data.

[0053] The industrial interface is used to connect to the PLC of construction equipment to obtain equipment operation data and ratio data. It also obtains construction progress information and component information through the building information modeling system interface, and associates the component information with sensor data to generate construction equipment and business data.

[0054] The multimodal data deep fusion model uses two-dimensional spatial features extracted from visual data as query vectors and one-dimensional temporal dynamic features extracted from sensor data after encoding by a temporal convolutional network as key and value vectors. Cross-modal alignment and fusion are performed through a cross-attention mechanism to generate a high-order fused feature vector.

[0055] The quality analysis engine module communicates with the multi-source data acquisition and fusion module, receives high-order fusion feature vectors, performs construction quality identification and prediction tasks based on a deep learning model, and outputs structured quality analysis results.

[0056] The deep learning models loaded on the quality analysis engine module include at least one of the following models: The concrete strength prediction model, based on mix proportion data, construction process data, environmental data, structural internal state data, and age data, uses a temporal convolutional network to predict the strength development of concrete at various ages, generate strength development curves and early warning information, and output structured quality analysis results.

[0057] The prestressed tensioning quality identification model receives high-order fusion feature vectors output by the multi-source data acquisition and fusion module, comprehensively determines the tensioning state category, and outputs structured quality analysis results.

[0058] The rebar tying quality inspection model is based on image data acquired by a multi-source data acquisition and fusion module. It uses a Mask R-CNN network for target detection and segmentation, and generates structured quality analysis results after comparing them with construction specification thresholds.

[0059] The temperature-visual linkage crack early identification model receives high-order fusion feature vectors output by the multi-source data acquisition and fusion module, comprehensively determines the coupling relationship between the internal hydration thermal gradient abrupt change and the surface infrared thermal field distribution, and outputs structured quality analysis results.

[0060] The knowledge graph and rule reasoning engine module communicates with the quality analysis engine module, receives structured quality analysis results, and performs semantic matching and rule reasoning on the structured quality analysis results based on the pre-built construction specification knowledge graph to generate reasoning results.

[0061] The knowledge graph and rule-based reasoning engine module includes: The knowledge graph construction unit performs structured processing on the standard texts, process flow data, and historical engineering case data of bridge construction to construct a knowledge graph of construction standards.

[0062] The semantic matching and reasoning unit uses the BERT model to semantically encode the structured quality analysis results output by the quality analysis engine module and performs semantic matching with the standard text. The corresponding standard text is selected through similarity calculation, and the knowledge graph reasoning engine is combined to perform node path traversal and association calculation to generate reasoning results containing the causes of quality problems and rectification suggestions.

[0063] The dynamic compliance rule self-evolution unit presets the construction specification safety threshold. When the quality analysis engine module, combined with sensor data, finds that the construction specification safety threshold is met, but this results in a compliance critical deviation, the knowledge graph reasoning engine triggers rule reconstruction to dynamically shrink and correct the construction specification safety threshold.

[0064] The real-time early warning and dynamic control module communicates with the quality analysis engine module and the knowledge graph and rule reasoning engine module, respectively, to receive structured quality analysis results and reasoning results, and generate risk early warning information.

[0065] The real-time early warning and dynamic control module includes: The dynamic threshold calculation unit determines the construction stage, structural component type, and environmental conditions based on the building information model progress information. It obtains the construction stage coefficient, structural component coefficient, and environmental correction coefficient, and performs weighted adjustment on the construction specification safety threshold to generate a dynamic threshold.

[0066] The trend analysis unit uses an exponentially weighted moving average to smooth the sensor data and uses a cumulative control method to detect continuous deviations in the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is determined to be an abnormal trend.

[0067] The global risk simulation unit extracts the spatial topological relationships of the bridge structure based on the building information model, constructs a structural association graph model, and introduces a spatiotemporal graph convolutional network. It uses abnormal trends as the initial state vector and combines the physical force association rules and material properties in the construction specification knowledge graph to generate risk warning information.

[0068] The digital twin and 3D visualization module communicates with the multi-source data acquisition and fusion module, the quality analysis engine module, and the real-time early warning and dynamic control module, respectively. It receives high-order fusion feature vectors, structured quality analysis results, and risk warning information, and maps them onto the building information model for visualization.

[0069] The digital twin and 3D visualization module includes: Based on the spatial coordinates and component numbers of the components in the building information model, they are uniformly bound with sensor-collected data, visual analysis results, and construction process information to form a component-time-data mapping relationship.

[0070] Transform sensor data and structured quality analysis results into visual representations.

[0071] Based on the spatial location of the components, the visual representation data is mapped to a digital twin and 3D visualization platform.

[0072] This invention also relates to an artificial intelligence-based method for real-time monitoring of bridge construction quality, comprising the following steps: S1. Collect multi-source sensing data during bridge construction; generate high-order fusion feature vectors through a multi-modal data deep fusion model based on cross-attention mechanism.

[0073] The steps of multi-source sensing data acquisition and fusion include: Based on a distributed sensor network deployed at key structural parts of the bridge, real-time data from these key structural parts is collected as sensor data.

[0074] Based on visual acquisition equipment deployed in key construction areas, construction process videos and temperature distribution images are collected as visual data.

[0075] Based on the industrial interface of the PLC of the construction equipment, the system obtains equipment operation data and ratio data, acquires construction progress information and component information through the building information modeling system interface, and associates the component information with sensor data to generate construction equipment and business data.

[0076] Based on a multimodal data deep fusion model, two-dimensional spatial features extracted from visual data are used as query vectors, and one-dimensional temporal dynamic features extracted from sensor data after being encoded by a temporal convolutional network are used as key and value vectors. Cross-modal alignment and fusion are performed through a cross-attention mechanism to generate a high-order fusion feature vector.

[0077] S2 receives high-order fusion feature vectors and performs construction quality identification and prediction tasks based on a deep learning model, outputting structured quality analysis results.

[0078] S3. Based on a pre-built knowledge graph of construction specifications, semantic matching and rule-based reasoning are performed on the structured quality analysis results to generate reasoning results.

[0079] It also includes the self-evolution steps of dynamic compliance rules: A preset construction specification safety threshold is set. When the quality analysis engine module finds that the construction specification safety threshold is met by combining sensor data, but this results in a compliance threshold deviation, the knowledge graph reasoning engine triggers rule reconstruction to dynamically shrink and correct the construction specification safety threshold.

[0080] S4. Receive the structured quality analysis results and inference results, and generate risk warning information.

[0081] The generated risk warning information includes: Based on the construction information model progress information, the construction stage, structural component type, and environmental conditions are determined. Construction stage coefficients, structural component coefficients, and environmental correction coefficients are obtained. The safety thresholds of the construction specifications are then weighted and adjusted to generate dynamic thresholds.

[0082] The sensor data is smoothed using an exponentially weighted moving average, and a cumulative control method is used to detect continuous deviations in the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is judged as an abnormal trend.

[0083] Based on Building Information Modeling (BIM), the spatial topological relationships of bridge structures are extracted, a structural association graph model is constructed, and a spatiotemporal graph convolutional network is introduced. Abnormal trends are used as initial state vectors, and combined with physical stress association rules and material properties in the construction specification knowledge graph, risk warning information is generated.

[0084] S5. Map the high-order feature fusion vector, structured quality analysis results, and risk warning information onto the building information model for visualization.

[0085] The technical solution of the present invention will be described in detail through specific embodiments.

[0086] The system of this invention consists of six modules, which exchange data with each other via bus communication, database and message queue.

[0087] 1. Multi-source data acquisition and fusion module (1) Sensor data sources and deployment methods This system deploys distributed sensors at key structural locations of the bridge, as detailed below: Main beams, bridge towers, and cable sections: Vibrating wire strain gauges and fiber optic grating sensors are installed to collect structural stress and strain data; Bridge towers and formwork structure: Inclinometers are installed to monitor changes in structural attitude; In the beam body and key node areas: displacement gauges are installed to monitor settlement and deflection; Inside the concrete pouring area: Temperature and humidity sensors and concrete maturity sensors are embedded to collect data on temperature field changes and strength development. Environmental area (surrounding construction site): Deploy environmental temperature and humidity sensors to collect data on external climate conditions.

[0088] The aforementioned sensors are connected to the edge data acquisition unit via wired or wireless means to achieve real-time data acquisition.

[0089] (2) Visual data sources and deployment methods Visual acquisition equipment was deployed in key construction areas of the bridge. Prestressed tensioning end area: Industrial cameras are deployed to capture video of the tensioning process; Concrete pouring area: Industrial cameras are deployed to capture video of the pouring process; Rebar binding area: Industrial cameras are deployed to collect data on the rebar construction status; Infrared thermal imagers will be deployed on the bridge deck and structural surface areas to collect temperature distribution images and assist in crack identification.

[0090] Visual data undergoes initial processing via edge computing nodes.

[0091] (3) Sources of construction equipment and business data The system connects to the PLC of the construction equipment through an industrial interface to obtain equipment operation data: pump truck operating parameters, mixing plant mix ratio data, tensioning equipment tension force data, and grouting equipment operating status.

[0092] Specifically, the industrial interface is used to connect to the PLC of the construction equipment to obtain equipment operation data and allocation data; at the same time, it is used to parse and obtain construction progress information and structured data of corresponding components through the Building Information Modeling (BIM) system interface (such as a data interface based on the IFC standard format). The structured data is the component information (including the component's unique code, three-dimensional spatial coordinates, design dimensions and material properties).

[0093] In addition, the industrial interface synchronously receives real-time sensing data with precise timestamps from the distributed sensor network and vision acquisition devices in this system; this data is the sensor data.

[0094] Furthermore, the system utilizes a spatiotemporal registration algorithm, using the unique component code and timestamp as a joint index, to accurately map and associate the aforementioned static component information with dynamic sensor data in terms of spatial location and time dimension, ultimately generating construction equipment and business data that includes spatial topology and real-time status.

[0095] At the same time, construction progress information is obtained through the building information modeling system interface, and component information is associated with sensor data.

[0096] (4) Deep fusion of multimodal data based on cross-attention mechanism like Figure 2As shown, unlike traditional time-series monitoring or single-vision detection relying on a single sensor, this system constructs a spatiotemporal feature alignment and fusion model based on a cross-attention mechanism in the multimodal fusion stage. Specifically, two-dimensional spatial features extracted by machine vision cameras or infrared thermal imagers (e.g., thermal anomaly gradient distribution maps of suspected crack areas, fast and slow motion feature masks of tension jacks) are used as query vectors. Simultaneously, one-dimensional time-series dynamic data collected by vibrating wire strain gauges, displacement gauges, or pre-embedded temperature sensors associated with the corresponding time and spatial location are input into a temporal convolutional network for feature encoding to extract implicit stress / temperature abrupt change features, which are then used as key and value vectors.

[0097] Subsequently, cross-modal computation is performed through a cross-attention mechanism, enabling the system to automatically learn the mapping weights between spatial visual representations and internal physical quantities. This allows for high-precision convergence of "surface temperature anomalies / visual anomaly points" to "specific sensor locations where internal structural stress or parameters undergo drastic changes." This generates a high-order fusion feature vector containing multi-dimensional spatiotemporal attributes, which is then fed into the quality analysis engine module. This completely resolves the technical shortcomings of single-vision systems being susceptible to false alarms due to environmental interference, or the inability of a single sensor to deduce the source of spatial quality defects.

[0098] 2. Quality Analysis Engine Module The quality analysis engine module receives multimodal aligned datasets from the multi-source data acquisition and fusion module, and extracts corresponding features for different construction quality tasks to complete concrete strength prediction, prestressing tension anomaly identification, rebar binding quality inspection, and early crack identification. The output of the quality analysis engine module is a structured quality analysis result, which includes at least the predicted value, detection parameters, anomaly type, anomaly location, risk level, and corresponding component number.

[0099] (1) Concrete strength prediction model The concrete strength prediction model is used to predict the development of concrete strength during bridge construction and generate early warning information.

[0100] The model's input data includes: 1) Mix proportion data, including water-cement ratio, cement dosage, mineral admixture dosage, and additive dosage, obtained through the mixing plant system; 2) Construction process data, including pouring time, placement temperature, curing method, and curing time, obtained through the construction management system; 3) Environmental data, including ambient temperature and humidity, collected by on-site environmental sensors; 4) Internal structural condition data, including internal concrete temperature, humidity, and maturity, collected by sensors embedded in the beams, bridge towers, or abutments; 5) Age data, calculated from timestamps; 6) Strength label data, obtained from compressive strength tests of test blocks cured under the same conditions or standard test blocks.

[0101] First, the historical sample data mentioned above is input into the XGBoost model to assess the importance of each input factor and screen out the key influencing factors, which include parameters such as the peak internal temperature of concrete, the rate of temperature change, ambient temperature, ambient humidity, and age.

[0102] Then, the key influencing factors are constructed into multi-dimensional temporal input features according to the time series, and the input is used to train a temporal convolutional network to obtain a concrete strength prediction model.

[0103] During the online prediction phase, real-time temperature, humidity, and age data are input into the prediction model, which outputs predicted concrete strength values ​​for each future age and generates strength development curves based on the prediction results.

[0104] The intensity development curve is dynamically updated using age as the horizontal axis and predicted intensity as the vertical axis through a time series method.

[0105] The predicted strength is compared with the preset strength threshold. When the predicted strength does not meet the construction requirements within the target age period, an early warning message is generated. The early warning message includes the component number, the current predicted strength, the target strength and the expected time to meet the standard, and provides suggestions to extend the curing period or adjust the construction schedule.

[0106] The prediction results and early warning information are displayed through a digital twin visualization platform, in which the status of the corresponding components is marked in the three-dimensional bridge model, and the strength development curve and early warning information are displayed simultaneously.

[0107] (2) Prestressed tensioning quality identification model based on multimodal feature fusion The prestressing tensioning quality identification model is used to identify abnormal states during the tensioning construction process. Its input is no longer a single video, but directly receives a high-order fused feature vector of the tensioning process output by the multi-source data acquisition and fusion module. The specific workflow is as follows: First, the multi-source acquisition system aligns the slow / fast visual features of the tensioning end extracted by the industrial camera (extracted via a SlowFast network) with the one-dimensional time-series data of tension force, hydraulic pressure, and displacement synchronously transmitted via the PLC industrial bus, using the MMF-Net cross-attention mechanism, to form a "visual-mechanical" joint feature matrix. Subsequently, this joint feature matrix is ​​input into the anomaly classification fully connected layer (FC layer). This model no longer relies solely on visual "wire retraction" or solely on PLC "sudden drop in hydraulic pressure," but instead comprehensively determines the tensioning state category (including normal tensioning, minor wire slippage, wire breakage, slow retraction, and abnormal equipment vibration) through weight abrupt changes in the joint feature matrix. This approach utilizes the mutual verification of underlying data to completely eliminate false alarms caused by single visual occlusion or single sensor drift, and transmits the identified abnormal events (including precise timestamps and component building information model coordinates) to the knowledge graph reasoning module.

[0108] (3) Reinforcing bar binding quality inspection model The rebar binding quality inspection model is used to automatically detect and determine the quality of rebar binding during bridge construction.

[0109] The model's input data consists of image or video frame data acquired by industrial cameras deployed in the rebar tying operation area by the multi-source data acquisition and fusion module, combined with the component number, shooting location, timestamp, and building information model component information output by the system to form the detection input.

[0110] First, the acquired images of the rebar tying area are preprocessed, including image cropping, scale normalization, noise suppression, and brightness correction, to improve the recognizability of rebar edges and tying nodes. Then, the preprocessed images are input into the Mask R-CNN network, which uses its backbone feature extraction network to extract image features. Based on the region proposal network, candidate rebar target regions are generated, and the candidate regions are classified, bounding box regression is performed, and pixel-level mask prediction is performed to obtain the segmentation results of rebar strip regions, intersection node regions, and tying node regions.

[0111] In the post-processing stage, the segmentation results are morphologically corrected, contour extracted, and connected component analyzed to remove false targets and noisy regions. Then, combined with image calibration parameters or camera extrinsic parameters, the image pixel size is converted into the actual engineering size, and the center line, endpoint position, intersection position, and edge contour information of the rebar are extracted.

[0112] Based on the post-processing results, the rebar binding quality parameters are calculated, including: 1) The spacing between reinforcing bars is calculated by the actual distance between the center lines of adjacent reinforcing bars; 2) The lap length is calculated by the actual length of the overlapping area of ​​coaxial adjacent reinforcing bars; 3) The thickness of the protective layer is calculated by the minimum distance between the outer edge of the reinforcing bar and the boundary of the formwork or the design boundary of the component; 4) The integrity of the tying node is judged by the connectivity of the intersection node area, the number of nodes, and the shape characteristics of the nodes.

[0113] After completing the parameter calculations, the results of rebar spacing, lap length, protective layer thickness, and tying node integrity are compared with the construction specification thresholds pre-stored in the knowledge base to automatically determine whether the current rebar tying status meets the construction quality requirements. The determination results include at least the following categories: qualified, spacing out of tolerance, insufficient lap, insufficient protective layer, and missing tying.

[0114] The model's output includes: a rebar segmentation mask, rebar location frames, rebar centerlines, rebar spacing values, lap length values, protective layer thickness values, quality assessment results, and corresponding component numbers and shooting time information. The output is further transmitted to the knowledge graph inference module and the real-time early warning module to generate rectification suggestions and early warning information, which are then displayed on the digital twin visualization platform.

[0115] (4) Temperature-visual linkage early identification model for cracks like Figure 3 As shown, the input of this model is a high-order fused feature vector output from the multi-source data acquisition and fusion module. Traditional crack identification often only triggers an alarm when the crack is visible to the naked eye, while this model utilizes data fusion to achieve "pre-crack detection." In specific implementation: the model performs tensor splicing and attention weighting on the "suspected pixel-level thermal anomaly area mask" (two-dimensional spatial features) extracted from the infrared thermal imager by the U-Net network, and the temperature and humidity sensor data and displacement gauge micro-strain data (one-dimensional temporal features) pre-embedded in the corresponding concrete area at the same time. By comprehensively judging the coupling relationship between the internal hydration heat gradient abrupt change and the surface infrared thermal field distribution, the model can not only accurately locate the micron-level crack risk area, but also output the thermal anomaly intensity and cracking risk level. The judgment results (including crack coordinates, risk level, and associated internal temperature anomalies) are directly mapped to the digital twin visualization platform and simultaneously sent to the early warning module.

[0116] 3. Knowledge Graph and Rule Reasoning Engine Module (1) Knowledge Graph Construction Unit The construction specification knowledge graph is constructed by structuring bridge construction specification texts, construction process data, and historical project case data. First, the specification text is parsed to extract component types, process flows, and quality control indicators, constructing specification nodes. Then, based on Building Information Modeling (BIM) and construction processes, the relationships between components and processes are established. Further, the abnormal results output by the quality analysis engine module are structured to construct abnormal feature nodes. Finally, historical quality cases are associated with the specification text and abnormal features to form a complete knowledge graph structure.

[0117] (2) Semantic matching and reasoning unit Based on the construction specification knowledge graph, the BERT model is used to semantically encode the structured quality analysis results output by the quality analysis engine module and perform semantic matching with the specification text. The corresponding specification text is filtered through similarity calculation to realize automatic comparison between construction quality results and specification requirements. On this basis, the knowledge graph reasoning engine is used to traverse paths through the relationship between abnormal features and components, processes, specification texts and historical cases, and to perform matching calculations in combination with specification texts and historical cases, thereby deriving the causes of quality problems and rectification suggestions, and outputting the problem type, cause analysis, specification basis and handling measures.

[0118] The knowledge graph and rule-based reasoning engine module is used to interpret the test results output by the quality analysis engine module and generate decisions. The reasoning results include the problem type, possible causes, regulatory basis, and recommended measures. The regulatory basis explains the compliance relationship between the test results and construction specifications, while the recommended measures guide adjustments during construction. The reasoning results are further input into the real-time early warning and dynamic control module to generate specific construction control instructions. Simultaneously, they are input into the digital twin and 3D visualization platform to display the problem location, problem type, and handling suggestions in the 3D model. This achieves the transformation process from test results to regulatory judgment and then to construction decisions.

[0119] (3) Dynamic compliance rule self-evolution unit Existing digital twin monitoring systems typically rely on preset static thresholds for judgment, while the construction specification knowledge graph of this invention has the ability to dynamically evolve rules and learn itself for specific working conditions.

[0120] When the quality analysis engine module detects multiple similar "compliance critical offset" phenomena by combining sensor data, the knowledge graph inference engine will automatically trigger rule reconstruction. Specifically, the "compliance critical offset" discrimination mechanism is as follows: the quality analysis engine not only monitors the absolute amplitude of the quality indicator, but also uses a temporal convolutional network (TCN) to output the first-order rate of change (evolution speed) and the second-order rate of change (evolution acceleration) of the indicator. For example, under a specific environmental condition with drastic temperature differences, although the absolute amplitude of the fluctuation of the concrete cover thickness or tension of a local node has not exceeded the static tolerance threshold set by the national standard, its corresponding second-order rate of change shows an exponential increase over several consecutive time windows; or, combined with calculations using digital twins and finite element prediction models, it is found that this small geometric / mechanical offset has caused a step change in the stress concentration factor of the local node, showing a nonlinear deterioration trend.

[0121] At this point, the system automatically extracts the triplet of "special climate conditions - specific process parameters - digital twin anomaly evolution results" from the above-mentioned abnormal process and refines it into new derived control rules. Based on this, the knowledge graph, through adaptive weight updates, dynamically shrinks and corrects the original basic construction specification safety thresholds in a targeted manner (for example, automatically shrinking and updating the specification in the knowledge base under this extreme condition, which allows a 5mm protective layer error, to a more stringent 3mm red line warning threshold), thereby realizing the machine learning adaptation and self-evolution of bridge construction quality control threshold standards.

[0122] 4. Real-time early warning and dynamic control module The real-time early warning and dynamic control module is used to comprehensively evaluate the structured results output by the quality analysis engine module and the knowledge graph reasoning engine, and generate risk warnings and construction control suggestions.

[0123] First, based on the construction stage, structural component, and environmental conditions, the basic thresholds given in the specifications are modified to obtain dynamic thresholds. Specifically, the current construction stage is determined based on the building information model progress information, the structural component category is determined based on the component type, and construction stage coefficients, structural component coefficients, and environmental correction coefficients are set based on historical sensor data statistics or empirical parameters. These are then weighted to obtain dynamic thresholds, which are then adjusted to achieve personalized threshold settings for different working conditions.

[0124] Then, trend analysis is performed on the sensor data and prediction results. The data is smoothed by exponential weighted moving average, and the cumulative control method is used to detect the continuous deviation of the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is judged as an abnormal trend.

[0125] Global Evolutionary Inference Based on Spatiotemporal Graph Convolutional Network and Graph-Driven Model: To accurately assess the global stress impact of local quality defects (such as local micro-cracks, insufficient early strength of the closure segment, and slippage of a single cable) identified by the quality analysis engine module on the overall bridge structure, this module introduces a spatiotemporal graph convolutional network for dynamic risk inference. First, based on the Building Information Model (BIM) and digital twin 3D model, the spatial topological relationships of the bridge structure are extracted, a structural association graph model is constructed, and each structural unit (such as main beam segments, towers, and cables) is mapped as a graph node. The structural stress transmission relationship is then transformed into an adjacency matrix of the graph neural network. Specifically, the adjacency matrix The Middle The node and the first The adjacency weight elements between nodes The calculation formula is:

[0126] The meanings of the parameters in the formula are as follows: Represents nodes extracted from the building information model. With nodes The three-dimensional Euclidean distance between them; The distance Gaussian variance scaling parameter represents the control of the spatial decay rate; This represents nodes extracted from the construction specification knowledge graph and structural material properties. To the node The stiffness transfer coefficient (reflecting the ability to transmit mechanical forces).

[0127] When local nodes When a quality anomaly or a sudden change in multimodal fusion features occurs, the system extracts the initial state vector of the anomalous node. Subsequently, combining the physical stress correlation rules and material properties between the node and adjacent components in the construction specification knowledge graph, the system uses the graph convolution operation of the spatiotemporal graph convolutional network to dynamically propagate and aggregate the stress deterioration features of the anomalous node to the surrounding structural network along the adjacency matrix.

[0128] This module includes: The dynamic threshold calculation unit determines the current construction stage based on the building information model (BIM) progress information, identifies the structural component category and environmental conditions based on the component type, and obtains the construction stage coefficients accordingly. Structural component coefficient and environmental correction factor The system's preset construction specification basic safety thresholds. Perform dynamic weighting adjustments to generate dynamic thresholds. The specific formula for calculating the dynamic threshold is as follows:

[0129] in, and These are the stage weighting factor and the location weighting factor, respectively, set based on the statistical distribution of historical sensor data; For environment correction functions, specifically defined as follows: , As an environmentally sensitive factor, For real-time ambient temperature, To standardize the temperature, this calculation process enables the system to convert basic static thresholds into personalized dynamic thresholds for different complex operating conditions.

[0130] The trend analysis unit uses an exponentially weighted moving average to smooth the sensor data and uses a cumulative control method to detect continuous deviations in the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is determined to be an abnormal trend.

[0131] The global risk simulation unit extracts the spatial topological relationships of the bridge structure based on the building information model, constructs a structural association graph model, and introduces a spatiotemporal graph convolutional network. It uses abnormal trends as initial state vectors and combines the physical stress association rules and material properties in the construction specification knowledge graph to accurately deduce the abnormal diffusion path and generate risk warning information.

[0132] Among them, the physical force association rule characterizes the mechanical transmission mechanism and stress influence weight between adjacent structural components of the bridge. It is constructed by parsing the bridge structural design code text and extracting historical finite element simulation experience data. The material properties characterize the basic physical and mechanical parameters (such as elastic modulus and Poisson's ratio) required for the deduction. They are directly extracted by calling the component attribute database of Building Information Modeling (BIM) (such as material labels based on IFC data standards), or read in real time by connecting to the construction material inspection and mixing plant management database at the engineering site through a data interface.

[0133] This process calculates the stress redistribution state induced by local quality hazards, as well as the dynamic influence weights and attenuation laws of these hazards on connected and adjacent span components. The system thereby achieves global hazard evolution prediction driven by both "empirical map rules" and "real-time twin data," replacing the lagging approach of relying solely on offline finite element simulation for threshold alarms in traditional methods. It generates dynamic construction control commands with a global topological perspective (such as the linkage adjustment strategy of local reinforcement on far-end cable forces).

[0134] The output of this module is a structured risk assessment result, including risk level, abnormal location, trend status, scope of impact, and control recommendations. The result is then input into a digital twin and 3D visualization platform for display and can also be fed back to the construction execution system to guide on-site construction operations.

[0135] This module is located after the quality analysis engine module and the knowledge graph reasoning engine. Its inputs are the detection results and reasoning results, and its outputs are risk warnings and control decisions, thereby realizing a closed-loop process from quality detection to construction control.

[0136] 5. Digital Twin and 3D Visualization Platform The digital twin and 3D visualization module is used to spatially map and visualize the data output by the multi-source data acquisition and fusion module, the quality analysis engine module, and the real-time early warning module.

[0137] First, based on the spatial coordinates and component numbers of the components in the Building Information Model (BIM), sensor-collected data, visual analysis results, and construction sequence information are uniformly bound together. Specifically, the correspondence between sensor measurement points, video acquisition areas, and 3D components is established through component numbers, and the time sequence of multi-source data and construction sequences is aligned based on timestamps, thus forming a "component-time-data" mapping relationship.

[0138] Then, the sensor data and analysis results are converted into visual representations. Among these: (1) The temperature data is spatially interpolated using the inverse distance weighted interpolation method to generate a continuous temperature field distribution and map it as a temperature control cloud map; (2) Organize the tension force data collected by the tensioning equipment according to the time series to generate the tension force change curve; (3) Quantify the quality analysis results and calculate the quality score based on the degree of deviation between the test parameters and the specifications; (4) Spatial location of abnormal detection results to generate risk identification data of corresponding components.

[0139] Furthermore, based on the spatial location of the components, the aforementioned visualization data is mapped onto the surface of the 3D model. Specifically, temperature field data is mapped onto the component surface using color gradients to form a cloud map display; tensile force and strength data are associated with the corresponding components in the form of curves; and quality scores and abnormal states are displayed at the component locations using color indicators or icons.

[0140] When a user selects any component, the system retrieves the corresponding time-series data from the historical database based on the component number, including construction process records, sensor data, and analysis results, and displays them, thereby enabling component-level historical information traceability.

[0141] This module takes structured data and building information model information from various systems as input and outputs a 3D visualization. It also works in conjunction with the real-time early warning module to visualize the construction status and support decision-making.

[0142] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0143] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A real-time monitoring system for bridge construction quality based on artificial intelligence, characterized in that, include: The multi-source data acquisition and fusion module is used to collect multi-source sensing data during bridge construction and fuse the multi-source sensing data through a multi-modal data deep fusion model based on cross-attention mechanism to generate a high-order fusion feature vector. The quality analysis engine module is used to receive high-order fused feature vectors and perform construction quality identification and prediction tasks based on deep learning models, and output structured quality analysis results. The knowledge graph and rule reasoning engine module is used to receive structured quality analysis results, perform semantic matching and rule reasoning on the structured quality analysis results based on a pre-built construction specification knowledge graph, and generate reasoning results. The real-time early warning and dynamic control module is used to receive structured quality analysis results and inference results, and generate risk early warning information; The digital twin and 3D visualization module is used to receive high-order fusion feature vectors, structured quality analysis results, and risk warning information, and to map the received information onto the building information model for visualization.

2. The bridge construction quality real-time monitoring system based on artificial intelligence according to claim 1, characterized in that, The multi-source data acquisition and fusion module includes: a distributed sensor network, visual acquisition devices, industrial interfaces, and a multimodal data deep fusion model; Distributed sensor networks are deployed at key structural parts of bridges to collect real-time data from these key structural parts as sensor data. Visual acquisition equipment is deployed in key construction areas to collect video and temperature distribution images of the construction process as visual data. The industrial interface is used to connect to the PLC of construction equipment to obtain equipment operation data and ratio data. It also obtains construction progress information and component information through the building information modeling system interface, and associates the component information with sensor data to generate construction equipment and business data. The multimodal data deep fusion model uses two-dimensional spatial features extracted from visual data as query vectors and one-dimensional temporal dynamic features extracted from sensor data after encoding by a temporal convolutional network as key and value vectors. Cross-modal alignment and fusion are performed through a cross-attention mechanism to generate a high-order fused feature vector.

3. The bridge construction quality real-time monitoring system based on artificial intelligence according to claim 1, characterized in that, The deep learning models loaded on the quality analysis engine module include at least one of the following models: The concrete strength prediction model, based on mix proportion data, construction process data, environmental data, structural internal state data, and age data, uses a temporal convolutional network to predict the strength development of concrete at various ages, generate strength development curves and early warning information, and output structured quality analysis results. The prestressing tensioning quality identification model receives high-order fusion feature vectors output by the multi-source data acquisition and fusion module, comprehensively determines the tensioning state category, and outputs structured quality analysis results. The rebar binding quality inspection model is based on image data acquired by a multi-source data acquisition and fusion module. It performs target detection and segmentation through a MaskR-CNN network, and generates structured quality analysis results after comparing them with construction specification thresholds. The temperature-visual linkage crack early identification model receives high-order fusion feature vectors output by the multi-source data acquisition and fusion module, comprehensively determines the coupling relationship between the internal hydration thermal gradient abrupt change and the surface infrared thermal field distribution, and outputs structured quality analysis results.

4. The bridge construction quality real-time monitoring system based on artificial intelligence according to claim 2, characterized in that, The knowledge graph and rule-based reasoning engine module includes: The knowledge graph construction unit performs structured processing on the standard texts, process flow data, and historical engineering case data of bridge construction to construct a construction standard knowledge graph; The semantic matching and reasoning unit uses the BERT model to semantically encode the structured quality analysis results output by the quality analysis engine module and performs semantic matching with the standard text. The corresponding standard text is selected through similarity calculation, and the node path traversal and association calculation are combined with the knowledge graph reasoning engine to generate reasoning results containing the causes of quality problems and rectification suggestions. The dynamic compliance rule self-evolution unit presets the construction specification safety threshold. When the quality analysis engine module, combined with sensor data, finds that the construction specification safety threshold is met, but this results in a compliance critical deviation, the knowledge graph reasoning engine triggers rule reconstruction to dynamically shrink and correct the construction specification safety threshold.

5. The bridge construction quality real-time monitoring system based on artificial intelligence according to claim 4, characterized in that, The real-time early warning and dynamic control module includes: The dynamic threshold calculation unit determines the construction stage, structural part category and environmental conditions based on the progress information of the building information model, obtains the construction stage coefficient, structural part coefficient and environmental correction coefficient, and performs weighted adjustment on the safety threshold of the construction specification to generate a dynamic threshold. The trend analysis unit uses an exponentially weighted moving average to smooth the sensor data and uses a cumulative control method to detect continuous deviations in the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is judged as an abnormal trend. The global risk simulation unit extracts the spatial topological relationships of the bridge structure based on the building information model, constructs a structural association graph model, and introduces a spatiotemporal graph convolutional network. It uses abnormal trends as the initial state vector and combines the physical force association rules and material properties in the construction specification knowledge graph to generate risk warning information.

6. The bridge construction quality real-time monitoring system based on artificial intelligence according to claim 5, characterized in that, The digital twin and 3D visualization module includes: Based on the spatial coordinates and component numbers of the components in the building information model, they are uniformly bound with sensor-collected data, visual analysis results, and construction process information to form a mapping relationship between components, time, and data. Transform sensor data and structured quality analysis results into visual representations; Based on the spatial location of the components, the visual representation data is mapped to a digital twin and 3D visualization platform.

7. A method for real-time monitoring of bridge construction quality based on artificial intelligence, implemented based on the construction quality monitoring system based on multi-source perception and knowledge graph reasoning as described in any one of claims 1 to 6, characterized in that, Includes the following steps: S1. Collect multi-source sensing data during bridge construction; generate high-order fusion feature vectors through a multimodal data deep fusion model based on cross-attention mechanism; S2. Receive high-order fused feature vectors and perform construction quality identification and prediction tasks based on deep learning models, outputting structured quality analysis results; S3. Based on a pre-built construction specification knowledge graph, semantic matching and rule reasoning are performed on the structured quality analysis results to generate reasoning results; S4. Receive the structured quality analysis results and inference results, and generate risk warning information; S5. Map the high-order feature fusion vector, structured quality analysis results, and risk warning information onto the building information model for visualization.

8. The method for real-time monitoring of bridge construction quality based on artificial intelligence according to claim 7, characterized in that, The multi-source sensing data acquisition and fusion steps in S1 include: Based on a distributed sensor network deployed at key structural parts of the bridge, real-time data of the key structural parts of the bridge is collected as sensor data. Based on visual acquisition equipment deployed in key construction process areas, video of the construction process and temperature distribution images are collected as visual data. Based on the industrial interface of the PLC of the construction equipment, the system obtains equipment operation data and allocation data, obtains construction progress information and component information through the building information model system interface, and associates the component information with sensor data to generate construction equipment and business data. Based on a multimodal data deep fusion model, two-dimensional spatial features extracted from visual data are used as query vectors, and one-dimensional temporal dynamic features extracted from sensor data after being encoded by a temporal convolutional network are used as key and value vectors. Cross-modal alignment and fusion are performed through a cross-attention mechanism to generate a high-order fusion feature vector.

9. A method for real-time monitoring of bridge construction quality based on artificial intelligence according to claim 7, characterized in that, S3 also includes a dynamic compliance rule self-evolution process: A preset construction specification safety threshold is set. When the quality analysis engine module finds that the construction specification safety threshold is met by combining sensor data, but this results in a compliance threshold deviation, the knowledge graph reasoning engine triggers rule reconstruction to dynamically shrink and correct the construction specification safety threshold.

10. A method for real-time monitoring of bridge construction quality based on artificial intelligence according to claim 7, characterized in that, The risk warning information generated in S4 includes: Based on the progress information of the building information model, the construction stage, structural component type, and environmental conditions are determined. Construction stage coefficients, structural component coefficients, and environmental correction coefficients are obtained. The safety thresholds of the construction specifications are then weighted and adjusted to generate dynamic thresholds. The sensor data is smoothed using an exponentially weighted moving average, and a cumulative control method is used to detect continuous deviations in the data. When the smoothing result or cumulative deviation exceeds the dynamic threshold, it is judged as an abnormal trend. Based on Building Information Modeling (BIM), the spatial topological relationships of bridge structures are extracted, a structural association graph model is constructed, and a spatiotemporal graph convolutional network is introduced. Abnormal trends are used as initial state vectors, and combined with physical stress association rules and material properties in the construction specification knowledge graph, risk warning information is generated.

Images