A highway structure multi-source data acquisition and monitoring system based on digital twinning

Abstract

The application discloses a kind of based on digital twinning highway structure multi-source data acquisition and monitoring system, comprising: data acquisition and pre-processing module, acquisition data, pre-processing generation node input vector set;Graph structure construction module, constructs graph structure;Improved EvolveGCN calculation module, executes state residual gate mechanism and adjacent matrix dynamic reconstruction mechanism, outputs evolution characteristic sequence;Time series attention aggregation module, executes time series attention aggregation, generates fusion structure state vector;Digital twinning modeling module generates virtual node graph and is associated with node historical state trajectory;State comparison and health determination module, execute real-time state comparison, generate health determination label;Self-adapting correction and closed loop evolution module, execute self-adapting correction, complete closed loop evolution.The application improves the real-time, accuracy and self-adapting ability of highway structure health monitoring through multi-source data fusion and improved EvolveGCN structure, and is suitable for complex bridge and road monitoring.

Description

Technical Field

[0001] This invention relates to the field of civil engineering monitoring technology, and in particular to a multi-source data acquisition and monitoring system for highway structures based on digital twins. Background Technology

[0002] Existing highway structural safety monitoring primarily relies on single-type sensors and centralized data recording. Monitoring parameters mainly include vibration, tilt, and temperature / humidity. Some systems transmit data wirelessly, but this suffers from signal attenuation, high energy consumption, and insufficient real-time performance in complex terrain or long-distance road sections. Traditional monitoring devices often operate independently as distributed nodes, lacking unified data structure management and dynamic topology sensing capabilities. This results in weak data correlation between different monitoring points, making it difficult to reflect the overall evolution characteristics of the highway structure. At the data processing level, conventional systems generally use fixed models or static algorithms to fit trends in the monitoring data, lacking dynamic modeling of time-series characteristics and the ability to fuse features from multiple data sources, making it difficult to identify potential trends in structural changes.

[0003] Existing research has introduced multi-parameter wireless acquisition devices and achieved high transmission reliability through LoRa and 4G communication. However, it remains limited to data acquisition and one-way transmission, lacking multi-node association modeling methods based on graph structures, and thus unable to perform real-time topology updates and state evolution analysis in the cloud. Monitoring results typically rely on manual threshold judgments, making it difficult for the system to form automated decision-making and intelligent feedback mechanisms, resulting in delayed monitoring responses. The lack of a unified digital twin mapping mechanism means that the state correspondence between the physical road structure and the virtual monitoring model cannot be kept synchronized, resulting in low model update frequency and an inability to support long-term adaptive evolution requirements.

[0004] Therefore, existing technologies have significant shortcomings in multi-source data fusion, dynamic modeling of graph structures, and digital twin feedback correction, and cannot meet the comprehensive requirements of real-time performance, accuracy, and intelligent evolution in highway structure monitoring. How to provide a multi-source data acquisition and monitoring system for highway structures based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multi-source data acquisition and monitoring system for highway structures based on digital twins. This invention utilizes multi-sensor fusion, graph neural network evolution calculation, and adaptive correction mechanisms to achieve dynamic monitoring and health assessment of highway structures, possessing advantages such as strong real-time performance, high recognition accuracy, and strong closed-loop evolution capability.

[0006] According to an embodiment of the present invention, a multi-source data acquisition and monitoring system for highway structures based on digital twins includes: The data acquisition and preprocessing module is used to acquire multi-source sensor data of the highway structure, perform time alignment and feature stitching operations, and generate a set of node input vectors. The graph structure construction module is used to build the graph structure of the highway structure based on the set of node input vectors, and generate the initial node feature matrix and adjacency matrix; The improved EvolveGCN computation module is used to receive the initial node feature matrix and adjacency matrix, execute the state residual gating mechanism and the adjacency matrix dynamic reconstruction mechanism, and output the evolution feature sequence. The temporal attention aggregation module is used to perform temporal attention aggregation operations on evolutionary feature sequences, extract spatial structure dependencies and temporal trend variability, and generate fused structural state vectors. The digital twin modeling module is used to generate a virtual node diagram of the highway structure based on the fused structural state vector and associate the historical state trajectory of each node. The status comparison and health assessment module is used to perform real-time status comparison in the virtual node graph, identify structural trend shifts and abnormal node activation areas, and generate health assessment labels. The adaptive correction and closed-loop evolution module is used to receive health determination labels, perform adaptive correction on state residual gating parameters and adjacency reconstruction functions, and realize closed-loop evolution.

[0007] Optionally, modules can be integrated using the following methods: Collect highway structure data, perform preprocessing to construct a set of node input vectors; A graph structure is constructed based on the set of node input vectors, generating an initial node feature matrix and adjacency matrix; The initial node feature matrix and adjacency matrix are input into the improved EvolveGCN structure. The improved EvolveGCN structure enhances the ability to retain the evolution state of historical nodes by introducing a state residual gating mechanism and introduces a dynamic reconstruction mechanism for the adjacency matrix. The edge weights and connection structure are adjusted according to the convolution output of the previous time step, and the evolution feature sequence is output. Temporal attention aggregation is performed on the evolutionary feature sequence to extract spatial structure dependence and temporal trend variability, forming a fused structural state vector; A virtual node graph of a highway structure is generated based on a digital twin model driven by a fused structural state vector, and the historical state trajectory of each node is associated with it. Perform real-time status comparison in the virtual node graph to identify structural trend shifts and abnormal node activation areas, and output health judgment labels. The health assessment label is fed back to the improved EvolveGCN structure, and adaptive correction is performed on the state residual gating parameters and the adjacency reconstruction function to complete the closed-loop evolution.

[0008] Optionally, the data of the highway structure specifically includes vibration, tilt, temperature and humidity, and image data of the highway structure; the preprocessing step includes time alignment and feature stitching of the highway structure data.

[0009] Optionally, the operation steps of the improved EvolveGCN structure specifically include: The initial node feature matrix and adjacency matrix are input into the convolutional trajectory stack of the evolution graph. At each time step, the following operations are performed: Extract the residual vector between the current node feature matrix and the convolutional output of the previous time step, and construct the state residual input by combining it with the hidden state of the previous time step; Calculate the gating weight coefficients of the residual input through a state residual gating mechanism to control the injection ratio, and output the node evolution state; Execute a dynamic adjacency matrix reconstruction mechanism based on the node evolution state, and adaptively update the edge connection relationships and weights of the current adjacency matrix using feature similarity and historical structural information; Input the node evolution state and the reconstructed adjacency matrix into the graph convolutional kernel, perform feature propagation and aggregation, and output the convolutional feature vector for that time step; Store the convolutional feature vector into the evolution graph convolutional trajectory stack for use as the state residual input and adjacency matrix update in the next time step; After all time steps are completed, the convolutional feature vectors of each time step are aggregated sequentially to generate an evolutionary feature sequence, which serves as the temporal attention aggregation input.

[0010] Optionally, the operation steps of the state residual gating mechanism specifically include: At each time step, the difference vector between the current node feature and the convolutional feature of the previous time step is extracted, and combined with the hidden state matrix of the previous time step to form the state residual input. Input the state residual into the gating function and calculate the weight matrix of the update gate and the reset gate; The hidden state matrix is ​​feature-adjusted based on the reset gate weight matrix to obtain the state adjustment matrix; the state residual input is linearly mapped and activated with the state adjustment matrix to generate the candidate state matrix; the candidate state matrix is ​​weighted and fused with the hidden state of the previous time step based on the update gate weight matrix to output the evolution state of the node at the current time step.

[0011] Optionally, the operation steps of the adjacency matrix dynamic reconstruction mechanism include: At each time step, the node evolution state output from the previous time step and the node input vector set at the current time step are received to form the structural reconstruction input data. Based on the structural reconstruction input data, the structural feature difference degree between all node pairs is calculated. The structural feature difference degree is mapped and compared with the initial adjacency matrix to identify edge pairs with weights below a set threshold. The corresponding edge connections are deleted from the initial adjacency matrix to form a preliminary screening and reduction adjacency matrix. According to the feature similarity between nodes in the current evolution state and the frequency of historical connections, an edge connection supplementary candidate set is generated. For each edge pair in the edge connection supplementary candidate set, a structural consistency verification operation is performed. Edge pairs that satisfy the similarity condition and the continuity of the evolution trajectory are added to the preliminary screening and reduction adjacency matrix to form an updated adjacency matrix. The updated adjacency matrix is ​​input to the graph convolution unit at the current time step as the graph propagation structure input. It is used together with the current node evolution state for feature propagation and state update. The updated adjacency matrix and the updated node evolution state are written into the evolution graph convolution trajectory stack.

[0012] Optionally, the process of generating the fused structure state vector includes: The system receives the evolutionary feature sequence output from the improved EvolveGCN structure as the input for temporal fusion. It then performs vector normalization on the evolutionary feature sequence in time-step order to unify the feature scale of each time step, forming a standard temporal feature sequence. A feature difference map between time steps is constructed to quantify the magnitude of state changes between adjacent time nodes and extract evolutionary trend factors. The standard temporal feature sequence and evolutionary trend factors are jointly input into the attention guidance module, and temporal attention weights are calculated based on the sensitivity of node evolution changes. The node features at each time step are weighted and fused according to the temporal attention weights, outputting the first-stage temporal fusion representation. A structural consistency mapping operation is performed on the first-stage temporal fusion representation, and a topology-aligned feature set is constructed by combining the adjacency matrix of the current time step with the node evolution state. Finally, the topology-aligned feature set is compressed and reconstructed to output the fused structural state vector.

[0013] Optionally, the process of generating the virtual node diagram of the highway structure includes: The system receives the fused structural state vector and extracts the evolutionary feature representation of each node. It then constructs a structural mapping rule set, which is preset based on the node functions, physical connections, and sensor deployment information of the actual highway structure. Based on the structural mapping rule set, it maps the nodes in the fused structural state vector to the structural skeleton of the virtual node graph, initially constructing the virtual node graph. It calls historical state trajectory records and injects the historical evolutionary features corresponding to each mapped node into the node attributes of the virtual node graph, constructing a multi-temporal state stack structure. Combining the adjacency structure information in the convolutional trajectory stack of the evolution graph, it dynamically infers the connection relationships between virtual nodes, constructs a virtual topology structure that conforms to the evolutionary trend, and outputs a virtual node graph containing node evolutionary attributes and inference edge structures.

[0014] Optionally, the generation process of the health determination label includes: The system receives node evolution attributes and adjacent structure information from a virtual node graph to construct a state comparison input set. It extracts real node observation data for corresponding time steps from a digital twin model and synchronizes the correspondence between virtual and real nodes based on node indices. For each pair of virtual and real nodes, it performs state difference calculations to extract structural displacement change rate, vibration amplitude deviation, and temperature and humidity gradient difference, forming a set of difference vectors. This set of difference vectors is input into a trend offset detection unit, and the system analyzes the changing trends of node feature sequences based on a time sliding window to extract abnormal offset factors. Node activation response values ​​are calculated based on the abnormal offset factors and compared with the system's preset evolution threshold to generate abnormal node activation identifiers. Spatial clustering is performed on all abnormal node activation identifiers to extract high-risk areas with connectivity or similar feature distributions in the virtual node graph. Based on the structural type, abnormal activation density, and trend duration of the high-risk areas, a health assessment label is generated, including the structural health level, abnormal location index, and trend evolution direction.

[0015] Optionally, the operation steps of the adaptive correction process specifically include: The system receives the structural health level, abnormal location index, and trend evolution direction recorded in the health assessment label, and constructs a feedback correction input set. Based on the abnormal location index, it locates the corresponding node in the evolution graph convolutional trajectory stack, extracts the node's historical state trajectory and the adjacent structure information of the current time step, forming the node's historical state sequence. It performs abnormal intensity quantization on the historical state sequence, calculates the node's abnormal intensity value according to the health level and trend evolution direction, and normalizes the node's abnormal intensity value into a gating adjustment factor and an adjacency reconstruction factor. It maps the gating adjustment factor to the node sub-vectors of the update gate weight matrix and reset gate weight matrix of the state residual gating mechanism, performs proportional scaling up or down operations on the node sub-vectors according to the node index, with the scaling up or down operations constrained by preset upper and lower limits, and writes it back to the gating parameter storage. It uses the adjacency reconstruction factor as an edge weight adjustment coefficient to apply to the adjacent edge weights of the located node in the adjacent structure information, performs multiplicative updates on each adjacent edge according to the edge weight adjustment coefficient, and performs normalization processing on the updated edge weights. Consistency verification is performed on the updated gating parameters and adjacency structure information, including checking whether changes in the gating parameters of the adjusted nodes cause the adjacency matrix to become unconnectable or the weights to be abnormal. If the verification fails, the parameters are restored to the previous available parameters according to the rollback strategy and the rollback flag is recorded. The verified gating parameters and adjacency structure information are written into the evolution graph convolutional trajectory stack, triggering a single correction iteration of the improved EvolveGCN, generating the corrected node feature sequence and adjacency structure information. The corrected node feature sequence is input into the digital twin model to update the evolutionary attributes of the corresponding nodes in the virtual node graph and recalculate the local health judgment index. The correction effect is evaluated based on the recalculated local health judgment index. If the abnormal intensity decreases by more than a threshold and the health level improves to a preset threshold, the gating parameters and adjacency reconstruction function parameters of this round are marked as valid. The correction and update operations are repeated, iterating until the preset stability condition is reached or the maximum number of iterations is reached. When the preset stability condition is met, the current gating parameters and adjacency reconstruction function parameters are confirmed as long-term correction parameters, written into the online parameter storage, and the correction log and correction results are output for auditing and manual confirmation.

[0016] The beneficial effects of this invention are: (1) Realize the fusion acquisition and topological modeling of multi-source heterogeneous data. Vibration, tilt, temperature and humidity and image monitoring data are acquired through multi-parameter wireless acquisition devices. The node input vector set is constructed by combining time alignment and feature splicing operations. The overall correlation expression of the highway structure status is realized in the graph structure, which improves the integrity and timeliness of the monitoring data.

[0017] (2) Realize dynamic evolution monitoring based on the improved EvolveGCN structure. The ability to retain the state of historical nodes is enhanced by the state residual gating mechanism, and the adaptive adjustment of the connection relationship between nodes is realized by combining the adjacency matrix dynamic reconstruction mechanism. This enables continuous tracking and high-precision identification of structural change trends, thereby improving the evolution stability and prediction accuracy of the monitoring model.

[0018] (3) Construct a digital twin feedback closed-loop correction system. A virtual node graph is generated by using the fused structural state vector and a virtual-real mapping relationship is established. Real-time state comparison and health judgment are performed in the virtual space. The health judgment label is fed back to the improved EvolveGCN structure to perform adaptive correction, realizing bidirectional dynamic coupling between the physical structure and the virtual model, and significantly improving the self-learning and self-repair capabilities of the highway structure monitoring system. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a module structure connection diagram of a multi-source data acquisition and monitoring system for highway structures based on digital twins proposed in this invention. Fig. 2 This is a flowchart illustrating the data acquisition and preprocessing module and the graph structure construction module of a multi-source data acquisition and monitoring system for highway structures based on digital twins, as proposed in this invention. Fig. 3 This is a schematic diagram illustrating the digital twin modeling and virtual node graph generation process of a multi-source data acquisition and monitoring system for highway structures based on digital twins, as proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figs. 1-3 A multi-source data acquisition and monitoring system for highway structures based on digital twins, comprising: The data acquisition and preprocessing module is used to acquire multi-source sensor data of the highway structure, perform time alignment and feature stitching operations, and generate a set of node input vectors. The graph structure construction module is used to build the graph structure of the highway structure based on the set of node input vectors, and generate the initial node feature matrix and adjacency matrix; The improved EvolveGCN computation module is used to receive the initial node feature matrix and adjacency matrix, execute the state residual gating mechanism and the adjacency matrix dynamic reconstruction mechanism, and output the evolution feature sequence. The temporal attention aggregation module is used to perform temporal attention aggregation operations on evolutionary feature sequences, extract spatial structure dependencies and temporal trend variability, and generate fused structural state vectors. The digital twin modeling module is used to generate a virtual node diagram of the highway structure based on the fused structural state vector and associate the historical state trajectory of each node. The status comparison and health assessment module is used to perform real-time status comparison in the virtual node graph, identify structural trend shifts and abnormal node activation areas, and generate health assessment labels. The adaptive correction and closed-loop evolution module is used to receive health determination labels, perform adaptive correction on state residual gating parameters and adjacency reconstruction functions, and realize closed-loop evolution.

[0022] In this embodiment, the modules are interconnected using the following method: Collect highway structure data, perform preprocessing to construct a set of node input vectors; A graph structure is constructed based on the set of node input vectors, generating an initial node feature matrix and adjacency matrix; The initial node feature matrix and adjacency matrix are input into the improved EvolveGCN structure. The improved EvolveGCN structure enhances the ability to retain the evolution state of historical nodes by introducing a state residual gating mechanism and introduces a dynamic reconstruction mechanism for the adjacency matrix. The edge weights and connection structure are adjusted according to the convolution output of the previous time step, and the evolution feature sequence is output. Temporal attention aggregation is performed on the evolutionary feature sequence to extract spatial structure dependence and temporal trend variability, forming a fused structural state vector; A virtual node graph of a highway structure is generated based on a digital twin model driven by a fused structural state vector, and the historical state trajectory of each node is associated with it. Perform real-time status comparison in the virtual node graph to identify structural trend shifts and abnormal node activation areas, and output health judgment labels. The health assessment label is fed back to the improved EvolveGCN structure, and adaptive correction is performed on the state residual gating parameters and the adjacency reconstruction function to complete the closed-loop evolution.

[0023] In this embodiment, the data of the highway structure specifically includes vibration, tilt, temperature and humidity, and image data of the highway structure; the preprocessing step includes time alignment and feature stitching of the highway structure data.

[0024] In this embodiment, the operation steps of the improved EvolveGCN structure specifically include: The initial node feature matrix and adjacency matrix are input into the convolutional trajectory stack of the evolution graph. At each time step, the following operations are performed: Extract the residual vector between the current node feature matrix and the convolutional output of the previous time step, and construct the state residual input by combining it with the hidden state of the previous time step; Calculate the gating weight coefficients of the residual input through a state residual gating mechanism to control the injection ratio, and output the node evolution state; Execute a dynamic adjacency matrix reconstruction mechanism based on the node evolution state, and adaptively update the edge connection relationships and weights of the current adjacency matrix using feature similarity and historical structural information; Input the node evolution state and the reconstructed adjacency matrix into the graph convolutional kernel, perform feature propagation and aggregation, and output the convolutional feature vector for that time step; Store the convolutional feature vector into the evolution graph convolutional trajectory stack for use as the state residual input and adjacency matrix update in the next time step; After all time steps are completed, the convolutional feature vectors of each time step are aggregated sequentially to generate an evolutionary feature sequence, which serves as the temporal attention aggregation input.

[0025] In this embodiment, the operation steps of the state residual gating mechanism specifically include: At each time step, the difference vector between the current node feature and the convolutional feature of the previous time step is extracted, and combined with the hidden state matrix of the previous time step to form the state residual input. In the improved EvolveGCN structure, a state residual gating mechanism is introduced to enhance the graph neural network's ability to model the evolution trend of node states. This mechanism requires the node state output of the previous time step to be referenced at each time step, and the aforementioned node state output is the hidden state matrix.

[0026] The hidden state matrix is ​​generated during the graph convolution propagation stage and is retained in the evolution graph convolution trajectory stack during time step progression, serving as an important context input for subsequent time steps. The difference between the node features of the current time step and the node features of the previous time step is concatenated with this hidden state matrix to form the state residual input. The state residual input is then fed into the gating function to calculate the weight matrices of the update gate and the reset gate. Specifically, the gating function dynamically calculates the update and adjustment ratio of node states based on the state residual input, ensuring that the evolution of each node state can adaptively adjust according to historical information and current input characteristics. The gating function consists of two parts: an update gate and a reset gate.

[0027] Update gate: The update gate controls the update ratio of the current node state. Based on the state residual input, the update gate calculates a weight value, which indicates how the node state at the current time step should migrate to the historical state. The weight matrix of the update gate is calculated by weighted combination of the state residual input and the features of the historical nodes. Through the update gate, the model decides how to use historical information to update the state of the current node.

[0028] Reset Gate: The function of the reset gate is to control the degree of influence of the historical node state on the current node state. The reset gate generates a weight value by calculating the relationship between the node features of the current time step and the state residual input. This weight value represents the proportion of historical node information that should be retained in the current time step. The weight matrix of the reset gate is calculated using the state residual input and determines how to "reset" the contribution of historical information in the next state calculation.

[0029] Through the two gating mechanisms mentioned above, the model can dynamically adjust the relationship between the state residual input and the historical node state, deciding which information should be retained and which information should be updated or reset. Finally, the node state processed by the gating function will be used as the output of the current time step for subsequent graph convolution operations and stored in the evolution graph convolution trajectory stack to provide input for the next time step.

[0030] The hidden state matrix is ​​feature-adjusted based on the reset gate weight matrix to obtain the state adjustment matrix; the state residual input is linearly mapped and activated with the state adjustment matrix to generate the candidate state matrix; the candidate state matrix is ​​weighted and fused with the hidden state of the previous time step based on the update gate weight matrix to output the evolution state of the node at the current time step.

[0031] In this embodiment, the operation steps of the adjacency matrix dynamic reconstruction mechanism include: At each time step, the node evolution state output from the previous time step and the node input vector set at the current time step are received to form the structural reconstruction input data. Based on the structural reconstruction input data, the structural feature difference degree between all node pairs is calculated. The structural feature difference degree is used to measure the connectivity adaptability of nodes in the current evolution stage. The structural feature difference degree is mapped and compared with the initial adjacency matrix to identify edge pairs with weights below a set threshold. The corresponding edge connections are deleted from the initial adjacency matrix to form a preliminary screening and reduction adjacency matrix. According to the feature similarity between nodes in the current evolution state and the frequency of historical connections, an edge connection supplementary candidate set is generated. For each edge pair in the edge connection supplementary candidate set, a structural consistency verification operation is performed. Edge pairs that satisfy the similarity condition and the continuity of the evolution trajectory are added to the preliminary screening and reduction adjacency matrix to form an updated adjacency matrix. The updated adjacency matrix is ​​input to the graph convolution unit at the current time step as the graph propagation structure input. It is used together with the current node evolution state for feature propagation and state update. The updated adjacency matrix and the updated node evolution state are written into the evolution graph convolution trajectory stack.

[0032] In this embodiment, the process of generating the fusion structure state vector includes: The system receives the evolutionary feature sequence output from the improved EvolveGCN structure as the input for temporal fusion. It then performs vector normalization on the evolutionary feature sequence in time-step order to unify the feature scale of each time step, forming a standard temporal feature sequence. A feature difference map between time steps is constructed to quantify the magnitude of state changes between adjacent time nodes and extract evolutionary trend factors. The standard temporal feature sequence and evolutionary trend factors are jointly input into the attention guidance module, and temporal attention weights are calculated based on the sensitivity of node evolution changes. The node features at each time step are weighted and fused according to the temporal attention weights, outputting the first-stage temporal fusion representation. A structural consistency mapping operation is performed on the first-stage temporal fusion representation, and a topology-aligned feature set is constructed by combining the adjacency matrix of the current time step with the node evolution state. Finally, the topology-aligned feature set is compressed and reconstructed to output the fused structural state vector.

[0033] In this embodiment, the process of generating the virtual node diagram of the highway structure includes: The system receives a fused structural state vector and extracts the evolutionary feature representation of each node, which is used as input for virtual modeling. It then constructs a structural mapping rule set, which is preset based on the node functions, physical connections, and sensor deployment information of the actual highway structure to guide the structural mapping of real nodes to virtual nodes. Based on the structural mapping rule set, the nodes in the fused structural state vector are mapped to the structural skeleton of the virtual node graph, initially constructing the virtual node graph. Historical state trajectory records are retrieved, and the historical evolutionary features corresponding to each mapped node are injected into the node attributes of the virtual node graph, constructing a multi-temporal state stacked structure. Combining the adjacency structure information in the convolutional trajectory stack of the evolution graph, the connection relationships between virtual nodes are dynamically inferred, constructing a virtual topology structure that conforms to the evolutionary trend, and outputting a virtual node graph containing node evolutionary attributes and inference edge structures.

[0034] In this invention, adjacency structure information refers to the graph topology connection data used and updated by the improved EvolveGCN structure during graph convolution propagation, mainly describing whether there are connections between nodes and the strength of those connections. This information exists in the form of an adjacency matrix and is dynamically adjusted at each time step according to the node evolution state.

[0035] The adjacency structure information includes not only the set of connecting edges between nodes in the graph structure, but also the connection weight of each edge, which represents the strength of the association between two nodes. During graph evolution, the adjacency structure information is continuously updated through a dynamic reconstruction mechanism of the adjacency matrix, reflecting the real-time evolution of node relationships.

[0036] Furthermore, adjacency structure information is one of the fundamental data sources for generating the virtual node graph topology. When the digital twin model constructs the virtual node graph, the system will call the adjacency structure information stored in the convolutional trajectory stack of the current evolution graph to infer the connectivity and structural similarity between virtual nodes, thereby constructing a virtual topology graph that corresponds to the logical structure of the real highway.

[0037] In this embodiment, the process of generating the health assessment label includes: The system receives node evolution attributes and adjacent structure information from a virtual node graph to construct a state comparison input set. It extracts real node observation data for corresponding time steps from a digital twin model and synchronizes the correspondence between virtual and real nodes based on node indices. For each pair of virtual and real nodes, it performs state difference calculations to extract structural displacement change rate, vibration amplitude deviation, and temperature and humidity gradient difference, forming a set of difference vectors. This set of difference vectors is input into a trend offset detection unit, and the system analyzes the changing trends of node feature sequences based on a time sliding window to extract abnormal offset factors. Node activation response values ​​are calculated based on the abnormal offset factors and compared with the system's preset evolution threshold to generate abnormal node activation identifiers. Spatial clustering is performed on all abnormal node activation identifiers to extract high-risk areas with connectivity or similar feature distributions in the virtual node graph. Based on the structural type, abnormal activation density, and trend duration of the high-risk areas, a health assessment label is generated, including the structural health level, abnormal location index, and trend evolution direction.

[0038] In this embodiment, the operation steps of the adaptive correction process specifically include: The system receives the structural health level, abnormal location index, and trend evolution direction recorded in the health assessment label, and constructs a feedback correction input set. Based on the abnormal location index, it locates the corresponding node in the evolution graph convolutional trajectory stack, extracts the node's historical state trajectory and the adjacent structure information of the current time step, forming the node's historical state sequence. It performs abnormal intensity quantization on the historical state sequence, calculates the node's abnormal intensity value according to the health level and trend evolution direction, and normalizes the node's abnormal intensity value into a gating adjustment factor and an adjacency reconstruction factor. It maps the gating adjustment factor to the node sub-vectors of the update gate weight matrix and reset gate weight matrix of the state residual gating mechanism, performs proportional scaling up or down operations on the node sub-vectors according to the node index, with the scaling up or down operations constrained by preset upper and lower limits, and writes it back to the gating parameter storage. It uses the adjacency reconstruction factor as an edge weight adjustment coefficient to apply to the adjacent edge weights of the located node in the adjacent structure information, performs multiplicative updates on each adjacent edge according to the edge weight adjustment coefficient, and performs normalization processing on the updated edge weights. Consistency verification is performed on the updated gating parameters and adjacency structure information. This includes checking whether changes in the gating parameters of the adjusted nodes cause the adjacency matrix to become disconnected or the weights to be abnormal. If the verification fails, the parameters are restored to the previous available parameters according to the rollback strategy, and the rollback flag is recorded. The verified gating parameters and adjacency structure information are written into the evolution graph convolutional trajectory stack, triggering a single correction iteration of the improved EvolveGCN, generating the corrected node feature sequence and adjacency structure information. The corrected node feature sequence is input into the digital twin model to update the evolutionary attributes of the corresponding nodes in the virtual node graph and recalculate the local health assessment index. The correction effect is evaluated based on the recalculated local health assessment index. If the abnormality intensity decreases by more than a threshold and the health level improves to a preset threshold, the gating parameters and adjacency reconstruction function parameters of this round are marked as valid. The correction and update operations are repeated, iterating until the preset stability condition is reached or the maximum number of iterations is reached. When the preset stability condition is met, the current gating parameters and adjacency reconstruction function parameters are confirmed as long-term correction parameters, written into the online parameter storage, and the correction log and correction results are output for auditing and manual confirmation. Example 1: To verify the feasibility of this invention, it was applied to the structural health monitoring of highway bridges. Traditional monitoring methods typically rely on a single type of sensor, such as a vibration sensor, which can only monitor one aspect of the bridge. Furthermore, the data cannot be processed in real time and lacks intelligent analysis and feedback mechanisms. To address these issues, this invention combines data acquisition from multiple sensors with digital twin technology to provide a more efficient monitoring solution.

[0039] First, the system deploys multiple wireless sensor nodes, including vibration sensors, temperature and humidity sensors, tilt sensors, and image sensors. These sensors transmit data through wireless communication modules, LoRa and 4G modules. By strategically placing the sensors, real-time monitoring of different parts of the bridge is achieved, collecting multi-dimensional data such as bridge surface cracks, temperature and humidity changes, and structural tilt. Unlike traditional single data acquisition methods, this invention enhances the comprehensiveness and accuracy of the data through the fusion of multiple sensors.

[0040] All collected data underwent preprocessing, including time alignment and feature concatenation, to generate a set of node input vectors. In the graph structure construction module, based on these input vector sets, we constructed a graph structure for each node of the highway bridge, generating an initial node feature matrix and adjacency matrix, and establishing a preliminary graph model of the highway structure. Through this graph model, we can better capture the correlation between structures, thereby providing support for subsequent analysis.

[0041] The system uses an improved EvolveGCN computation module, employing a state residual gating mechanism and a dynamic adjacency matrix reconstruction mechanism, to evolve the initial node feature matrix and adjacency matrix, thus obtaining the evolutionary feature sequence of the bridge structure. During this process, the dynamic updates of node states and adjacency structures enable the system to continuously track changes in the bridge state and detect potential structural problems in a timely manner.

[0042] The temporal attention aggregation module extracts the spatial structural dependencies and temporal trend changes of the bridge structure to generate a fused structural state vector. This vector provides a key input for digital twin modeling. Through the digital twin modeling module, a virtual node map of the highway bridge is generated and compared with the state of the actual nodes in real time. The trend deviation of the structure and the abnormal activation areas of the nodes are identified. This information is used to generate health judgment labels and is fed back to the improved EvolveGCN structure to perform adaptive correction and complete closed-loop evolution.

[0043] To verify the effectiveness of this invention, we compared the monitoring results of traditional monitoring methods with those of this invention. Table 1 shows a comparison of the data collected during the monitoring process, including vibration amplitude, temperature and humidity changes, structural tilt, number of image analyses, and number of abnormal node activations.

[0044] Table 1: Comparison of Structural Health Monitoring Data for Highway Bridges

[0045] Based on the data in the table above, we can compare the performance of traditional monitoring methods and the monitoring method of this invention in monitoring the structural health of highway bridges, and analyze the results as follows: In terms of vibration amplitude detection, traditional monitoring methods can only detect vibrations as small as 0.01 mm, while the monitoring method of this invention can detect larger vibration amplitudes (0.03 mm); the difference in vibration amplitude reaches 200%, indicating that the system of this invention can more sensitively capture minute changes in the structure and detect potential structural problems in advance.

[0046] Regarding temperature and humidity changes, traditional monitoring methods record temperature and humidity changes of 0.2℃ and 3%, respectively, while the monitoring method of this invention can detect higher changes (0.5℃ and 6%). This indicates that the system of this invention has a significant improvement in sensitivity to environmental changes and can provide more comprehensive monitoring data, especially under complex environmental conditions, making the assessment of the impact on highway bridges more accurate.

[0047] Regarding the monitoring of structural tilt, the traditional method detects a tilt of 0.02°, while the monitoring method of this invention can detect a larger tilt (0.05°), a difference of 150%. This means that the monitoring system of this invention can provide earlier warnings when structural deformation occurs, helping to carry out maintenance and repair in advance.

[0048] In terms of image analysis and detection, traditional monitoring methods only perform detection three times per month, while this invention, through its image sensor, can perform image analysis frequently, up to ten times per month, a difference of 233%. This increased frequency means that this invention can more comprehensively and in real-time detect problems such as surface cracks and settlement on structures, greatly improving the health monitoring capabilities of bridges.

[0049] In addition, the number of abnormal node activations is an indicator that traditional monitoring methods have failed to detect, while the monitoring system of this invention detected 5 abnormal node activations; this shows that the invention can identify and report potential problems in the bridge structure and provide early warnings.

[0050] Finally, regarding the number of health assessment tags issued, traditional monitoring methods did not generate health assessment tags, while this invention, through the combination of digital twins and graph neural network structures, successfully generated three health assessment tags, further demonstrating the accuracy and intelligent feedback capabilities of this invention in highway bridge health monitoring.

[0051] In summary, this invention, by combining multi-source data acquisition with deep learning algorithms, not only improves the real-time performance and accuracy of the monitoring system, but also provides a more comprehensive health assessment and early warning capability, which is advantageous compared to traditional monitoring methods.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-source data acquisition and monitoring system for highway structures based on digital twins, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source sensor data of the highway structure, perform time alignment and feature stitching operations, and generate a set of node input vectors. The graph structure construction module is used to build the graph structure of the highway structure based on the set of node input vectors, and generate the initial node feature matrix and adjacency matrix; The improved EvolveGCN computation module is used to receive the initial node feature matrix and adjacency matrix, execute the state residual gating mechanism and the adjacency matrix dynamic reconstruction mechanism, and output the evolution feature sequence. The temporal attention aggregation module is used to perform temporal attention aggregation operations on evolutionary feature sequences, extract spatial structure dependencies and temporal trend variability, and generate fused structural state vectors. The digital twin modeling module is used to generate a virtual node diagram of the highway structure based on the fused structural state vector and associate the historical state trajectory of each node. The status comparison and health assessment module is used to perform real-time status comparison in the virtual node graph, identify structural trend shifts and abnormal node activation areas, and generate health assessment labels. The adaptive correction and closed-loop evolution module is used to receive health determination labels, perform adaptive correction on state residual gating parameters and adjacency reconstruction functions, and realize closed-loop evolution.

2. The multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 1, characterized in that, The modules are connected in the following way: Collect highway structure data, perform preprocessing to construct a set of node input vectors; A graph structure is constructed based on the set of node input vectors, generating an initial node feature matrix and adjacency matrix; The initial node feature matrix and adjacency matrix are input into the improved EvolveGCN structure. The improved EvolveGCN structure enhances the ability to retain the evolution state of historical nodes by introducing a state residual gating mechanism and introduces a dynamic reconstruction mechanism for the adjacency matrix. The edge weights and connection structure are adjusted according to the convolution output of the previous time step, and the evolution feature sequence is output. Temporal attention aggregation is performed on the evolutionary feature sequence to extract spatial structure dependence and temporal trend variability, forming a fused structural state vector; A virtual node graph of a highway structure is generated based on a digital twin model driven by a fused structural state vector, and the historical state trajectory of each node is associated with it. Perform real-time status comparison in the virtual node graph to identify structural trend shifts and abnormal node activation areas, and output health judgment labels. The health assessment label is fed back to the improved EvolveGCN structure, and adaptive correction is performed on the state residual gating parameters and the adjacency reconstruction function to complete the closed-loop evolution.

3. The multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 2, characterized in that, The data of the highway structure specifically includes vibration, tilt, temperature and humidity, and image data of the highway structure; the preprocessing steps include time alignment and feature stitching of the highway structure data.

4. The multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 3, characterized in that, The specific operational steps of the improved EvolveGCN structure include: The initial node feature matrix and adjacency matrix are input into the convolutional trajectory stack of the evolution graph. At each time step, the following operations are performed: Extract the residual vector between the current node feature matrix and the convolutional output of the previous time step, and construct the state residual input by combining it with the hidden state of the previous time step; Calculate the gating weight coefficients of the residual input through a state residual gating mechanism to control the injection ratio, and output the node evolution state; Execute a dynamic adjacency matrix reconstruction mechanism based on the node evolution state, and adaptively update the edge connection relationships and weights of the current adjacency matrix using feature similarity and historical structural information; Input the node evolution state and the reconstructed adjacency matrix into the graph convolutional kernel, perform feature propagation and aggregation, and output the convolutional feature vector for that time step; Store the convolutional feature vector into the evolution graph convolutional trajectory stack for use as the state residual input and adjacency matrix update of the next time step; After all time steps are completed, sequentially aggregate the convolutional feature vectors of each time step to generate an evolutionary feature sequence, which serves as the temporal attention aggregation input.

5. A multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 4, characterized in that, The specific operational steps of the state residual gating mechanism include: At each time step, the difference vector between the current node feature and the convolutional feature of the previous time step is extracted, and combined with the hidden state matrix of the previous time step to form the state residual input. Input the state residual into the gating function and calculate the weight matrix of the update gate and the reset gate; The hidden state matrix is ​​feature-adjusted based on the reset gate weight matrix to obtain the state adjustment matrix; the state residual input is linearly mapped and activated with the state adjustment matrix to generate the candidate state matrix; the candidate state matrix is ​​weighted and fused with the hidden state of the previous time step based on the update gate weight matrix to output the evolution state of the node at the current time step.

6. The multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 5, characterized in that, The operation steps of the adjacency matrix dynamic reconstruction mechanism include: At each time step, the node evolution state output from the previous time step and the node input vector set at the current time step are received to form the structural reconstruction input data. Based on the structural reconstruction input data, the structural feature difference degree between all node pairs is calculated. The structural feature difference degree is mapped and compared with the initial adjacency matrix to identify edge pairs with weights below a set threshold. The corresponding edge connections are deleted from the initial adjacency matrix to form a preliminary screening and reduction adjacency matrix. According to the feature similarity between nodes in the current evolution state and the frequency of historical connections, an edge connection supplementary candidate set is generated. For each edge pair in the edge connection supplementary candidate set, a structural consistency verification operation is performed. Edge pairs that satisfy the similarity condition and the continuity of the evolution trajectory are added to the preliminary screening and reduction adjacency matrix to form an updated adjacency matrix. The updated adjacency matrix is ​​input to the graph convolution unit at the current time step as the graph propagation structure input. It is used together with the current node evolution state for feature propagation and state update. The updated adjacency matrix and the updated node evolution state are written into the evolution graph convolution trajectory stack.

7. A multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 6, characterized in that, The process of generating the fusion structure state vector includes: The system receives the evolutionary feature sequence output from the improved EvolveGCN structure as input for temporal fusion. It then performs vector normalization on the evolutionary feature sequence in time-step order to unify the feature scale of each time step, forming a standard temporal feature sequence. A feature difference map between time steps is constructed to quantify the state change amplitude between adjacent time nodes and extract evolutionary trend factors. The standard temporal feature sequence and evolutionary trend factors are jointly input into the attention guidance module, and temporal attention weights are calculated based on the node evolution sensitivity. The node features at each time step are weighted and fused according to the temporal attention weights, outputting the first-stage temporal fusion representation. A structural consistency mapping operation is performed on the first-stage temporal fusion representation, and a topology-aligned feature set is constructed by combining the adjacency matrix of the current time step with the node evolution state. Finally, the topology-aligned feature set is compressed and reconstructed to output a fused structural state vector.

8. A multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 7, characterized in that, The process of generating the virtual node diagram of the highway structure includes: The system receives the fused structural state vector and extracts the evolutionary feature representation of each node. It then constructs a structural mapping rule set, which is preset based on the node functions, physical connections, and sensor deployment information of the actual highway structure. Following the structural mapping rule set, it maps the nodes in the fused structural state vector to the structural skeleton of the virtual node graph, initially constructing the virtual node graph. It calls historical state trajectory records and injects the historical evolutionary features corresponding to each mapped node into the node attributes of the virtual node graph, constructing a multi-temporal state stacked structure. Combining the adjacency structure information in the convolutional trajectory stack of the evolution graph, it dynamically infers the connection relationships between virtual nodes, constructs a virtual topology structure that conforms to the evolutionary trend, and outputs a virtual node graph containing node evolutionary attributes and inference edge structures.

9. A multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 8, characterized in that, The process of generating the health assessment label includes: The system receives node evolution attributes and adjacent structure information from a virtual node graph to construct a state comparison input set. It extracts real node observation data for corresponding time steps from a digital twin model and synchronizes the correspondence between virtual and real nodes based on node indices. For each pair of virtual and real nodes, it performs state difference calculations to extract structural displacement change rate, vibration amplitude deviation, and temperature and humidity gradient difference, forming a set of difference vectors. This set of difference vectors is input into a trend offset detection unit, and the system analyzes the changing trends of node feature sequences based on a time sliding window to extract anomaly offset factors. Based on the anomaly offset factors, it calculates node activation response values ​​and compares them with the system's preset evolution threshold to generate node anomaly activation identifiers. It performs spatial clustering operations on all node anomaly activation identifiers to extract high-risk areas with connectivity or similar feature distributions in the virtual node graph. Based on the structural type, anomaly activation density, and trend duration of the high-risk areas, it generates health assessment labels, including structural health level, anomaly location index, and trend evolution direction.

10. A multi-source data acquisition and monitoring system for highway structures based on digital twins according to claim 9, characterized in that, The specific steps of the adaptive correction process include: The system receives the structural health level, abnormal location index, and trend evolution direction recorded in the health assessment label, and constructs a feedback correction input set. Based on the abnormal location index, it locates the corresponding node in the evolution graph convolutional trajectory stack, extracts the node's historical state trajectory and the adjacent structure information of the current time step, forming the node's historical state sequence. It performs abnormal intensity quantization on the historical state sequence, calculates the node's abnormal intensity value according to the health level and trend evolution direction, and normalizes the node's abnormal intensity value into a gating adjustment factor and an adjacency reconstruction factor. It maps the gating adjustment factor to the node sub-vectors of the update gate weight matrix and reset gate weight matrix of the state residual gating mechanism, performs proportional scaling up or down operations on the node sub-vectors according to the node index, with the scaling up or down operations constrained by preset upper and lower limits, and writes it back to the gating parameter storage. It uses the adjacency reconstruction factor as an edge weight adjustment coefficient to apply to the adjacent edge weights of the located node in the adjacent structure information, performs multiplicative updates on each adjacent edge according to the edge weight adjustment coefficient, and performs normalization processing on the updated edge weights. Consistency verification is performed on the updated gating parameters and adjacency structure information, including checking whether changes in the gating parameters of the adjusted nodes cause the adjacency matrix to become unconnectable or the weights to be abnormal. If the verification fails, the parameters are restored to the previous available parameters according to the rollback strategy and the rollback flag is recorded. The verified gating parameters and adjacency structure information are written into the evolution graph convolutional trajectory stack, triggering a single correction iteration of the improved EvolveGCN, generating the corrected node feature sequence and adjacency structure information. The corrected node feature sequence is input into the digital twin model to update the evolutionary attributes of the corresponding nodes in the virtual node graph and recalculate the local health judgment index. The correction effect is evaluated based on the recalculated local health judgment index. If the abnormal intensity decreases by more than a threshold and the health level improves to a preset threshold, the gating parameters and adjacency reconstruction function parameters of this round are marked as valid. The correction and update operations are repeated, iterating until the preset stability condition is reached or the maximum number of iterations is reached. When the preset stability condition is met, the current gating parameters and adjacency reconstruction function parameters are confirmed as long-term correction parameters, written into the online parameter storage, and the correction log and correction results are output for auditing and manual confirmation.

Images