A system and method for fusing multi-source data of a bridge structure

Abstract

The application belongs to the technical field of bridge monitoring, and relates to a system and method for fusing multi-source data of a bridge structure, comprising a heterogeneous topology graph construction and manifold embedding technology module, a multi-scale spatiotemporal cognitive convolutional neural network module, a continuous manifold spatiotemporal alignment and Bayesian fusion module, and a structure health index calculation and state evaluation module; the heterogeneous topology graph construction and manifold embedding technology module is used to obtain a heterogeneous topology graph, a node embedding vector, and a manifold model parameter; the multi-scale spatiotemporal cognitive convolutional neural network module is used to extract deep features of the heterogeneous topology graph and obtain multi-scale fusion features; the continuous manifold spatiotemporal alignment and Bayesian fusion module is used to obtain spatiotemporal alignment parameters and a fusion state vector; and the structure health index calculation and state evaluation module is used to perform structure health monitoring and state evaluation on the bridge and obtain a final health evaluation result; thereby improving the intelligence, automation, and reliability level of the structure monitoring system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge monitoring, and specifically discloses a system and method for fusing multi-source data of a bridge structure. BACKGROUND

[0002] With the rapid development of the transportation industry, the number and scale of bridges continue to grow, and structural health monitoring (SHM) of bridges has gradually become a key component of bridge operation and management. Modern bridge monitoring systems generally deploy a large number of heterogeneous sensors, such as strain gauges, accelerometers, displacement meters, thermometers, etc., forming a complex multi-source sensor network that collects real-time structural response data to support bridge operation state assessment, disease warning, and maintenance decision-making. However, bridge sensor networks still face a series of challenges in actual operation. First, different types of sensors have significant differences in sampling frequency, measurement range, installation location, and response characteristics, resulting in complex data representation and structure distribution, making it difficult to accurately capture their potential correlation in the fusion modeling process. Traditional data fusion methods such as weighted average and Kalman filtering, while having some effect in some scenarios, lack the ability to express the topology of the sensor network and the physical semantic relationship between sensors, making it difficult to achieve deep-level fusion and information complementarity. Second, in the actual operating environment, the structural response data collected by sensors has serious spatio-temporal heterogeneity. Due to clock drift, asynchronous sampling, and environmental disturbances, the time alignment deviation and spatial calibration error of time series data recorded by different sensors often occur. Existing methods rely on linear interpolation, sliding window, and other rules for alignment adjustment, but in the face of complex conditions such as nonlinear drift and sudden disturbances, they lack sufficient robustness and generalization ability, affecting the accuracy of subsequent analysis and state recognition.

[0003] Therefore, the present application provides a system and method for fusing multi-source data of a bridge structure, which integrates heterogeneous graph neural network modeling, continuous time multi-source spatio-temporal alignment, knowledge graph fusion expression, and feature attribution explanation capabilities, systematically solving the core problems of data heterogeneity, alignment error, semantic deficiency, and model opacity in bridge sensor data fusion and analysis, thereby improving the intelligence, automation, and reliability level of the structural monitoring system. SUMMARY

[0004] The application aims to provide a system for fusing multi-source data of a bridge structure, and solve the problem of improving the intelligentization, automation and reliability level of a bridge structure monitoring system.

[0005] A method for fusing multi-source data of a bridge structure of the system, comprising: processing multi-source heterogeneous sensor data based on a heterogeneous graph structure to obtain a heterogeneous topology graph, and performing manifold embedding learning on sensor nodes to obtain node embedding vectors and manifold model parameters; performing deep feature extraction on the heterogeneous topology graph based on the node embedding vectors and the manifold model parameters to obtain multi-scale fusion features; performing spatiotemporal alignment and fusion processing on the multi-scale fusion features and original time series data of the sensors to obtain spatiotemporal alignment parameters and fusion state vectors; and performing structural health monitoring and state evaluation of the bridge based on the fusion state vectors to obtain final health evaluation results.

[0006] The application has the following advantages and beneficial effects:

[0007] The present application realizes the comprehensive modeling of the spatial relationship, structural coupling relationship and functional coordination relationship among sensors by fusing multi-source heterogeneous data in a unified manifold space through heterogeneous topology graph construction technology. It has three key features: heterogeneity: contains multiple different types of nodes and edges. Multi-type nodes: sensors, data, semantic concepts, meta-information and other different types; multiple relationships: spatial, structural, functional, semantic and other different relationship types; cross-modal fusion: simultaneously processing different physical quantities such as strain, vibration and temperature. Topology: reflects the real spatial layout and connection relationship of the bridge structure. Spatial layout, reflecting the real position relationship of the sensors on the bridge; structural connection, based on the connection relationship of the bridge mechanics model; geometric constraint, maintaining the geometric characteristics of the bridge structure. Graph: organizes and expresses complex monitoring data in a graph theory way. This architecture is specially designed for the characteristics of bridge structure, fully considering the engineering practical needs of bridge monitoring system. Traditional bridge monitoring methods have certain limitations, such as data island problem, different types of sensor data are processed separately, and cross-modal correlation patterns cannot be found; spatial relationship is missing, ignoring the spatial distribution characteristics of sensors in the bridge structure; lack of engineering knowledge, unable to integrate prior knowledge of bridge structure mechanics. The present application introduces a five-tuple hyper-heterogeneous graph representation framework, and realizes the unified embedding modeling of the bridge monitoring network in the joint manifold space by constructing a multi-layer heterogeneous structure containing sensor entity nodes, spatio-temporal data nodes, semantic concept nodes, relationship edges and hyper-edges. Traditional bridge monitoring methods are mainly based on graph structure analysis in Euclidean space, which is difficult to effectively handle the nonlinear topological relationship and complex geometric constraints in sensor network. The present application introduces manifold learning and topological graph theory to construct a more accurate geometric representation space. Traditional graph structure can only represent binary relationships (i.e. the connection between two nodes), but in bridge monitoring, there are often multiple relationships, such as multiple sensors jointly monitoring the same structure section, multiple monitoring parameters reflecting the structure state, etc. High-order tensor decomposition solves this problem by decomposing complex multi-element relationships into combinations of multiple low-order components. The joint flow embedding module of the present application can handle irregular arrangements, and through semantic embedding, the experience and specification requirements of bridge engineers are integrated, the manifold embedding maintains the geometric characteristics and physical constraints of the bridge structure, and can model different types of sensors and multiple relationship types. The actual role in bridge monitoring: ensure the parameter transferability between different bridge types; maintain the physical interpretability of model parameters; improve the robustness and stability of the model; prevent overfitting and parameter divergence. Manifold embedding learning is a parameter optimization of the entire neural network under geometric constraints, ensuring that all parameters are optimized cooperatively under manifold constraints, maintaining the physical consistency and geometric rationality of the bridge monitoring model. The manifold trajectory method of the present application models on a continuous spatio-temporal manifold, unifies discrete multi-frequency data into continuous trajectories, and improves the time synchronization accuracy by 95% and the spatial positioning accuracy by 80%. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a flowchart of the bridge sensor network data fusion and analysis method based on the heterogeneous graph neural network;

[0009] Figure 2 is a heterogeneous graph neural network architecture diagram;

[0010] Figure 3 is a continuous time and space calibration flowchart;

[0011] Figure 4 is a real-time performance monitoring result diagram of the bridge monitoring system running continuously for 24 hours;

[0012] Reference signs: 1 - sensor network diffusion view, 2 - heterogeneous sensor network (original view), 3 - global feature similarity view, 4 - type-based message passing, 5 - meta-path-based message passing, 6 - type-based message passing, 7 - Transformer-based aggregator, 8 - multilayer perceptron, 9 - splicing operation, 10 - sensor fault detection, 11 - processing delay curve, 12 - system availability curve, I - view generation, II - message passing, III - semantic aggregation, IV - representation fusion. DETAILED DESCRIPTION

[0013] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0014] As Figure 1As shown, the bridge sensor network driven health state intelligent analysis process is displayed, including bridge sensor network, heterogeneous topology graph construction, multi-scale cognitive convolution, intelligent data quality guarantee, bridge health state analysis, explainable analysis framework, continuous manifold spatio-temporal alignment and Bayesian fusion technology. The bridge sensor network includes multi-source heterogeneous sensor data; the heterogeneous topology graph construction includes manifold embedding learning; the multi-scale cognitive convolution includes differential attention mechanism; the intelligent data quality guarantee includes statistical denoising and graph regularization; the bridge health state analysis includes structure health state identification, fault prediction and early warning; intelligent operation and maintenance decision support and safety evaluation report. A method for fusing bridge structure multi-source data, comprising: based on a heterogeneous graph structure, processing multi-source heterogeneous sensor data to obtain a heterogeneous topology graph, and performing manifold embedding learning on sensor nodes to obtain node embedding vectors and manifold model parameters; the heterogeneous graph structure includes a node set, an edge set, a hyperedge set, a relationship type mapping space and a joint manifold embedding mapping function; based on the node embedding vectors and the manifold model parameters, deep feature extraction is performed on the heterogeneous topology graph to obtain multi-scale fusion features; spatio-temporal alignment and fusion processing are performed on the multi-scale fusion features and the original time series data of the sensor to obtain spatio-temporal alignment parameters and fusion state vectors; based on the fusion state vectors, structure health monitoring and state evaluation are performed on the bridge to obtain the final health evaluation result. The system for fusing bridge structure multi-source data provided by the present application includes a heterogeneous topology graph construction and manifold embedding technology module, a multi-scale spatio-temporal cognitive convolutional neural network module, a continuous manifold spatio-temporal alignment and Bayesian fusion module, and a structure health index calculation and state evaluation module.

[0015] The heterogeneous topology graph construction and manifold embedding technology module is used for processing multi-source heterogeneous sensor data based on a heterogeneous graph structure to obtain a heterogeneous topology graph , and performing manifold embedding learning on sensor nodes to obtain node embedding vectors and manifold model parameters . As Figure 2As shown, the heterogeneous graph neural network architecture diagram demonstrates a multi-view-message-passing-fusion processing framework for sensor fault detection. The process includes view generation I, message passing II, semantic aggregation III, and representation fusion IV. View generation is based on a heterogeneous sensor network to generate a sensor network diffusion view and a global feature similarity view. The sensor network diffusion view is used to simulate the diffusion connection relationship of the sensor network; the global feature similarity view is used to construct node similarity association based on global features. (Legend: black dots are sensors, striped dots are data nodes, and hollow dots are metadata). Message passing is used to perform multi-mode message passing on three types of views (original and two types of derivatives), including type-based message passing and meta-path-based message passing. Type-based message passing (upper and lower two groups) passes information according to node types (such as SS, SD, and SM type combinations); meta-path-based message passing (middle group) passes information through complex meta-paths (such as SD, SM, SDS, and SMS path rules) to mine the association logic of the heterogeneous network. Semantic aggregation is the result of each message passing, which inputs the Transformer-based aggregator to aggregate multi-source messages at the semantic level through the Transformer model to extract key features. Representation fusion is used to splice the aggregated features, input the multi-layer perceptron (multiple parallel / concatenated multi-layer perceptron structures), and further fuse multi-dimensional representations through multi-layer nonlinear transformation to strengthen feature expression. Sensor fault detection is used to perform sensor fault detection tasks on the fused features, completing the complete process from multi-view construction, message passing to fault identification. The heterogeneous graph structure is a five-tuple hyper-heterogeneous graph structure, including a node set , an edge set , a hyperedge set , a relationship type mapping space , and a joint manifold embedding mapping function . Multi-source heterogeneous sensor data refers to multiple types of sensor data used for bridge monitoring, including strain data , acceleration data , displacement data , and temperature data , etc. By uniformly modeling multi-source heterogeneous sensor data, a heterogeneous topology graph , a node embedding vector , and an optimized manifold model parameter The isomorphic topology graph construction and manifold embedding technology module includes a data preprocessing and isomorphic graph initialization unit, a five-tuple super isomorphic graph construction unit, and a manifold embedding learning unit. The five-tuple super isomorphic graph construction unit is used to construct an isomorphic topology graph. The isomorphic topology graph is a graph data structure specially designed for bridge structure health monitoring, which uniformly models different types of entities in the bridge monitoring system. The manifold embedding learning unit is used to perform manifold embedding processing on the isomorphic topology graph.

[0016] The data preprocessing and isomorphic graph initialization unit is used to preprocess multi-source heterogeneous sensor data to obtain standardized sensor data streams and initialize an isomorphic topology graph. The preprocessing includes data cleaning, outlier removal, and format standardization processing on the original data obtained by the sensors. The system that fuses multi-source data of a bridge structure receives multi-source heterogeneous sensor data (i.e., original data) obtained by multiple types of sensors deployed on the bridge structure. Through the isomorphic topology graph construction technology, the multi-dimensional features of each sensor node are fused into a unified manifold embedding representation , representing a 64-dimensional real number vector space, while a multi-layer isomorphic graph structure reflecting the spatial relationship between sensors, structural coupling relationship, and functional coordination relationship is constructed.

[0017] The super isomorphic graph construction unit is used to calculate the elements in the initialized isomorphic topology graph based on the standardized sensor data stream to obtain a five-tuple isomorphic topology graph . The five-tuple super isomorphic graph realizes the unified modeling of the bridge monitoring network. The elements in the isomorphic topology graph include a node set , an edge set , a super edge set , a relationship type mapping space , and a joint manifold embedding mapping function ; the node set includes a sensor entity node set , a spatiotemporal data node set , a semantic concept node set , and a meta-information node set . The super isomorphic graph construction unit includes a node construction module, a relationship edge construction module, a super edge construction module, a relationship type mapping space determination module, and a joint manifold embedding mapping module.

[0018] The node construction module is used to obtain sensor nodes, data nodes, semantic nodes, and meta-information nodes by identifying and analyzing the data, data transmission, and data processing methods in the standardized sensor data stream, and the union set of the multiple types of nodes is taken as the node set. The node set is a set of sensor nodes, including sensor nodes, data nodes and semantic nodes, . is a set of sensor entity nodes, representing physical sensors deployed on the bridge, including strain gauges, accelerometers, displacement meters and thermometers, etc. Each sensor node in the set of sensor entity nodes is denoted as S. is a set of spatio-temporal data nodes, representing data instances formed after pre-processing of raw data collected by sensors, including strain values, acceleration value time series monitoring data and frequency domain features at a certain time, etc. used to connect sensor nodes and semantic nodes. Each data in the set of spatio-temporal data nodes is denoted as D. is a set of semantic concept nodes, representing engineering semantic information of the bridge structure, including bending stress of the main beam, lateral displacement of the tower and tension state of the cable, etc. related to the structural state and fault type of the bridge. is a set of meta-information nodes, representing attribute information of sensors, including device model, installation location, sampling frequency and measurement accuracy, etc.

[0019] Relationship edge construction module: the relationship edge construction module is used to construct an edge set based on multiple nodes in the node set and the relationship between the nodes. The edge set is a binary relationship edge set, . The binary relationship edge is used to represent the direct connection relationship between nodes in the bridge monitoring network. The edge set can include sensor spatial adjacency edges, data flow edges, semantic association edges and structural coupling edges. The sensor spatial adjacency edge is used to connect sensor nodes with similar physical locations. The data flow edge is used to connect the sensor node and the corresponding data node. The semantic association edge is used to connect the data node and the related semantic concept node. The structural coupling edge is used to connect the mechanically related sensor nodes.

[0020] Hyperedge construction module: the hyperedge relationship construction module obtains tensor decomposition parameters by high-order tensor decomposition of hyperedge relationships, and realizes unified representation of hyperedge relationships through the tensor decomposition parameters. The hyperedge set is a multi-element relationship edge set, The hyper-edge is a high-order relationship connecting multiple nodes (≥ 3), and the hyper-edge (multi-element relationship edge) is used to represent the cooperative monitoring relationship and multi-element coupling relationship in bridge monitoring. The hyper-edge set includes cooperative monitoring hyper-edges, cross-domain fusion hyper-edges, and time-space association hyper-edges. The cooperative monitoring hyper-edge refers to the joint monitoring of the same structural member by multiple sensors. The cross-domain fusion hyper-edge refers to the common reflection of a certain structural phenomenon by different types of sensor data. The time-space association hyper-edge represents the multi-point monitoring data within the same time window. The hyper-edge set is used to represent complex multi-element relationships that cannot be described by traditional binary edges, thereby improving the expression ability of the graph model for complex relationships in bridge monitoring. The hyper-edge set is used to model complex relationships involving multiple nodes. Through high-order tensor decomposition, the hyper-edge set and the tensor decomposition parameters are obtained, and the unified representation of the hyper-edge relationship is realized through the tensor decomposition parameters. The tensor decomposition parameters include the weight of the rth component and the mth-order factor matrix .

[0021] The hyper-edge relationship representation unit is used to realize the unified representation of the hyper-edge relationship through high-order tensors. Through high-order tensor decomposition, the complex high-order hyper-edge relationship is represented as a linear combination of multiple simple patterns, which not only maintains the integrity of the relationship but also reduces the computational complexity. The expression for realizing the unified representation of the hyper-edge relationship through high-order tensor decomposition is as follows: ; wherein, represents the representation of the hyper-edge relationship; r represents the component variable in the relationship type mapping space; R represents the relationship type mapping space; is the weight of the rth component, reflecting the importance of the component in the overall relationship; is the mth-order factor matrix, each factor matrix describing a relationship pattern in one dimension, and the rows of the matrix correspond to different elements in the dimension; m represents the order of the factor matrix; M refers to the different dimensions of the tensor in tensor decomposition; is the tensor product operator. For example, for a 4-dimensional tensor (M = 4), the definitions of its 4 dimensions are as follows: dimension 1: sensor position (60 sensors), dimension 2: sensor type (4 types), dimension 3: monitoring time (24 hours), and dimension 4: load working condition (5 working conditions); the meanings of its multiple factor matrices are as follows: represents the sensor position relationship matrix (60 × R), represents the sensor type relationship matrix (4 × R), represents the time relationship matrix (24 × R), and represents the working condition relationship matrix (5 × R); in bridge monitoring, each component r represents a typical monitoring mode: r = 1 represents the static response mode, r = 2 represents the dynamic response mode, r = 3 represents the temperature effect mode, and r = 4 represents the fatigue damage mode. When r = 1, the weight Importance of static response, used to reflect the static structural response under load; when r = 2, its weight Importance of dynamic response, used to reflect dynamic response such as vibration, impact, etc.; when r = 3, its weight Importance of temperature effect, used to reflect the influence of temperature change on the structure; when r = 4, its weight Importance of fatigue damage, used to reflect the long-term fatigue accumulation effect.

[0022] High-order tensor decomposition unit: the high-order tensor decomposition unit optimizes the factor matrix by alternating least squares method to obtain the final component weight and factor matrix. By fixing the condition of the factor matrix of other dimensional relationships, the factor matrix of the current dimension is optimized to obtain the optimal relationship mode of the current dimension, and the factor matrix of each dimension is updated step by step to approximate the representation of the hyperedge relationship ; the expression for optimizing the mth-order factor matrix is: ; wherein, is the mth-order factor matrix; represents the mth-order factor matrix , k ≠ m and weight , which minimizes the reconstruction error; represents the observed hyperedge relationship tensor, which is the high-order tensor data actually measured / constructed; r represents the current optimized relationship mode component index (rank component index), corresponding to the th mode in the relationship type mapping space; R represents the total number of component indexes in the relationship type mapping space (i.e., the decomposition rank rank), which is the number of all mode components; is the weight of the rth component; represents the kth-order factor matrix; represents the tensor product of all dimensional factor matrices except the mth dimension; is the tensor product operation of the matrix; is the square of the Frobenius norm. The optimization objective is to minimize the reconstruction error so that the decomposed tensor can preserve the information of the original hyperedge relation to the greatest extent. Through iterative optimization, the algorithm gradually converges to a local optimal solution to obtain an effective hyperedge relation decomposition representation. For example, in a bridge application, different engineering properties corresponding to each dimension are as follows: the first dimension represents the spatial position of the sensor (distributed along the longitudinal direction of the bridge), the second dimension represents the sensor type (strain, acceleration, displacement, temperature), the third dimension represents the monitoring time (working day, holiday, seasonality), and the fourth dimension represents the load working condition (empty load, heavy load, extreme load); then the actual engineering parameter example can be: based on engineering experience, the relationship type mapping space is determined to be 8, i.e., R = 8, in bridge monitoring, 8 main components can cover more than 90% of the bridge monitoring relationship, wherein the component weight may include: = 0.35 (static force relationship), = 0.25 (dynamic relationship), and = 0.20 (temperature relationship), etc.

[0023] The relationship type mapping space determination module is used to determine the relationship type mapping space based on bridge mechanics theory, engineering monitoring experience and semantic knowledge, etc. The relationship type mapping space defines the specific types and attributes of all edges and hyperedges.

[0024] The joint manifold embedding mapping module is used to construct feature vectors based on the heterogeneous topological graph and the original sensor data to obtain high-dimensional (such as 64-dimensional) sensor embedding vectors . The joint manifold embedding mapping function is used to map nodes, edges and hyperedges to the joint manifold space . is a d-dimensional Euclidean space used to represent the numerical characteristics of the sensor, and d represents the Euclidean embedding dimension. is a k-dimensional spherical manifold used to represent the geometric constraint characteristics of the sensor, and k represents the spherical manifold dimension. The sensor entity node set , , and respectively represent a plurality of physical sensor nodes in the distributed bridge monitoring network, each sensor node containing multi-dimensional information such as device attributes, spatial position and monitoring parameters. Through multi-modal information fusion and manifold learning, the sensor entity node features are constructed to obtain node embedding vectors of different modal sensors: ; wherein, represents the node embedding vector, which is used to jointly represent the sensor features of different modalities; represents a manifold embedding operator for mapping the resulting high-dimensional vector into a non-Euclidean manifold space; represents a weight matrix for linearly transforming the concatenated vector; represents a bias for bias-adjusting the concatenated vector after weight adjustment; and concat represents a concatenation operation. is a sensor type feature vector including one-hot encoding of device types such as strain sensors, acceleration sensors, displacement sensors, etc. is a geometric position feature that integrates three-dimensional coordinate information and bridge structure segment identification. is a dynamic monitoring feature that includes physical quantities such as real-time sampled vibration, strain, and temperature. is a device state feature that reflects the health status and working performance of the sensor. i represents an index variable of the sensor.

[0025] Manifold embedding learning unit: The manifold embedding learning unit is used for manifold embedding learning of sensor nodes in a heterogeneous topology graph to obtain a manifold embedding operator. The manifold embedding learning unit includes a geometric-structure joint embedding module, a meta-path information propagation module, a Riemann optimization algorithm module, and a multi-modal feature fusion module.

[0026] Geometric-structure joint embedding module: The geometric-structure joint embedding module is used for integrating the geographical coordinates of the sensor and the bridge structure information to obtain a geometric position feature : ; wherein, is the three-dimensional geographical coordinates of the sensor; represents a geodesic line addition operation on a manifold, which projects the geometric coordinates onto a selected manifold, then maps the structure embedding vector to the tangent space of the geometric coordinate point, performs vector addition operation in the tangent space, and finally maps the result in the tangent space back to the manifold; represents the identification of the bridge structure segment to which the sensor belongs. To embed the function of the structural segment, the semantic representation of different structural segments (e.g., main girder, pier, cable, etc.) of the bridge is learned through the graph convolution network. For example, the semantic coding of the structural segment can be started first, the semantic dictionary of the structural segment of the bridge is established, and the text description is converted into a numerical representation; then the structural segment attribute is quantified, the engineering attribute of the structural segment is converted into a numerical feature; then the graph convolution network learning is performed, the structural segment relationship graph is constructed, and the mutual relationship between the structural segments is learned; finally, a high-dimensional semantic vector is generated, all the features mentioned above are spliced and transformed, and a 64-dimensional high-dimensional semantic embedding vector is generated; to capture the abstract properties such as the mechanical properties, importance level and fault sensitivity of different structural segments. Taking a long-span suspension bridge as an example, the engineering implementation of each element can include the bridge engineering classification of the sensor entity node V_sensor, including the main cable strain monitoring node V_cable={s1_cable,s2_cable,...,s_n_cable}, the bridge tower inclination monitoring node V_tower={s1_tower,s2_tower,...,s_m_tower} and the main girder vibration monitoring node V_deck={s1_deck,s2_deck,...,s_k_deck}. Among them, the arrangement position of the main cable strain monitoring node is the main cable anchorage segment, the mid-span segment and the bridge tower; the monitoring target includes the main cable tension change and stress redistribution; the feature coding can be represented by the feature coding main cable strain sensor identifier, such as f_type=[1,0,0,0,0](). The arrangement position of the bridge tower inclination monitoring node can be the top of the bridge tower and the middle of the bridge tower; the monitoring target can include the bridge tower displacement and inclination angle; the feature coding can be represented by the bridge tower displacement sensor identifier, such as f_type=[0,1,0,0,0]. The arrangement position of the main girder vibration monitoring node can be the mid-span and 1 / 4 span point of the main girder; the monitoring target can include the vertical vibration and torsional vibration of the main girder; the feature coding can be represented by the main girder acceleration sensor identifier, such as f_type=[0,0,1,0,0](). The bridge coordinate system definition of the spatial position feature f_geo can be: for a suspension bridge, a bridge engineering coordinate system is established, in which the X axis is along the bridge longitudinal direction, the origin is the starting point of the main bridge, and the positive direction is the terminal point; the Y axis is along the bridge transverse direction, the origin is the bridge center line, and the positive direction is to the right; the Z axis is along the bridge vertical direction, the origin is the design reference surface, and the positive direction is upward. For example, the geometric position feature can be f_geo^(i)=[x_bridge,y_bridge,z_bridge,span_id,element_id].Specifically, for the sensors in the main cable span: the geometric position feature can be f_geo = [600.0, 0.0, 65.0, 1, 101], where x = 600m represents 600 meters from the starting point, y = 0 represents the center line, z = 65m represents the main cable height, span_id = 1 represents the main span, and element_id = 101 represents the main cable number; for the tower top sensor: the geometric position feature can be f_geo = [600.0, 0.0, 180.0, 0, 201], where z = 180m represents the tower height, span_id = 0 represents the tower, and element_id = 201 represents the tower number.

[0027] Meta-path information propagation module: the meta-path information propagation module is used to construct a meta-path propagation operator based on the information propagation process in the heterogeneous topology graph, and simulate the propagation process of information along a specific path pattern in the heterogeneous graph. Based on the constructed hyper-heterogeneous graph structure, a meta-path information propagation operator is designed to realize information fusion across types of nodes. The meta-path information propagation operator introduces tensor product operation and multi-head attention mechanism, and the expression of the meta-path information propagation operator is: ; wherein, represents the propagation operator of the meta-path ; represents a meta-path pattern, which defines the path type of information propagation, such as "sensor -> spatio-temporal data -> semantic concept" or "sensor -> sensor -> semantic concept", etc.; r represents the path variable; is the path length, the path length in the heterogeneous topology graph, which represents the number of edges that need to be passed through from the starting point to the ending point; is the adjacency matrix of the th step; softmax represents an activation function; , and are respectively a query matrix, a key-value matrix and a numerical matrix, which calculate the correlation weight between nodes through the attention mechanism; is a specific mask tensor of the meta-path , which ensures that the information propagation follows the predefined path pattern and causal constraints; represents the total number of graph neural network layers, i.e. the depth of the network; is a scaling factor to prevent gradient vanishing caused by too large attention scores; represents a tensor product operator; represents an adjacency matrix.

[0028] Adjacency matrix optimization unit: the adjacency matrix optimization unit is used to optimize the adjacency matrix by replacing the adjacency matrix of the rth step with a combination matrix , , so as to realize multi-relation comprehensive modeling. The expression of the combination matrix is: ; wherein, denotes the combination matrix; denotes the tensor product operator, which realizes the combination modeling of different relationship types and captures the complex relationship patterns in the bridge sensor network; is the spatial adjacency matrix, which reflects the geographical proximity and structural connection relationship between sensors; is the temporal relationship matrix, which describes the correlation and causality of monitoring data in the time dimension; is the semantic relationship matrix, which represents the semantic association between different monitoring parameters and structural states. By combining the spatial relationship, the temporal relationship and the semantic relationship through the tensor product combination, the three basic relationships are combined into high-order complex relationships through tensor product operation, which can consider the constraints of space, time and semantics in three dimensions. For example, two sensors not only need to be adjacent in space, but also need to have a causal relationship in time, and the monitored physical quantities need to have semantic relevance, in order to form an effective information propagation path.

[0029] Path-specific mask tensor constraint unit: the path-specific mask tensor constraint unit is used to determine the path-specific mask tensor, and controls the information propagation path through the path-specific mask tensor, to ensure the directionality and causality of information propagation. The calculation formula of the path-specific mask tensor is as follows: ; wherein, the mask tensor controls the information propagation path by setting the constraint of attention weight. denotes the mask value corresponding to the triple (a, b, c); , and denote the starting node index (such as the sensor node), the intermediate node index (such as the data node) and the terminal node index (such as the semantic node), respectively. When the triple forms an effective path pattern , the mask value is 0, and the corresponding attention weight remains unchanged; when the path pattern is invalid, the mask value is , the corresponding attention weight is close to 0 after the softmax function, thereby blocking the information propagation on the invalid path. This design ensures that information can only propagate along the path that is semantically reasonable and physically feasible, avoiding the interference of irrelevant information. For example, in bridge monitoring, the data of strain sensors can propagate to the structural health state node, but should not directly propagate to the ambient temperature node, and the mask tensor will automatically shield such unreasonable propagation paths. The effective path refers to the path that conforms to the logic of bridge engineering: for example, strain sensor → strain data → structural stress semantics; acceleration sensor → vibration data → modal characteristic semantics; displacement sensor → displacement data → deformation state semantics, etc. To comply with the principle of engineering rationality, that is, the sensor type matches the data type, there is a mechanical coupling relationship between the structural components, and the monitoring parameters are related to the evaluation target.

[0030] Riemann optimization algorithm module: The Riemann optimization algorithm module is used for manifold embedding learning based on the Riemann optimization algorithm, and parameter updating under manifold constraint to obtain manifold model parameters. Define the parameter manifold The Riemann manifold is constructed for all learnable parameters, and the expression of the objective function is: ; wherein, represents the Riemann manifold objective function; is a task loss, such as a classification loss or a regression loss, which drives the model to learn effective feature representations; is a curvature regularization term that constrains the geometric shape of the manifold to prevent over-bending or singular points; is a manifold volume term that controls the size of the parameter space. and are the first and second weighting coefficients, respectively. By adjusting the weighting coefficients, a balance can be achieved between task performance and geometric constraints.

[0031] Curvature regularization constraint unit: The curvature size on the entire parameter manifold is calculated by integration to obtain the curvature regularization term which is used to ensure the geometric rationality of the manifold, and the expression is: ; wherein, represents the curvature regularization term; is the Riemann curvature tensor at point , which reflects the bending degree of the manifold at that point; is the square norm of the tensor; represents integration over the entire parameter manifold to ensure global geometric consistency. By constraining the curvature size, the algorithm can learn more smooth and stable parameter representations.

[0032] Riemann gradient update unit: The Riemann gradient descent update rule updates the parameters under manifold constraint: ; wherein, Riemannian gradient on the manifold, obtained by projecting the Euclidean gradient to the tangent space of the manifold; learning rate parameter, controlling the step size of each update; current parameter Exponential mapping with base point, converting tangent vectors to new points on the manifold. This update method ensures that the parameters always remain within a reasonable manifold region, avoiding the problem of parameter out-of-bound or convergence to illegal regions caused by traditional optimization algorithms; denotes the parameter of the t-th iteration; denotes the updated parameter of the t+1-th iteration.

[0033] Multimodal feature fusion module: the multimodal feature fusion module is used to fuse multimodal features to obtain a fusion feature vector, and a manifold embedding operator is designed , which maps the fusion feature vector in Euclidean space to Riemannian manifold space through the manifold embedding operator, capturing the nonlinear geometric relationship in the bridge sensor network. Multimodal features include device type , geometric position , dynamic monitoring and device state . The expression of the manifold embedding operator is: ; wherein, denotes the manifold embedding operator; denotes the unitization processing of the input vector; tanh denotes the activation function, which compresses the normalized vector to range; denotes the calculation of the tangent vector from the base point to the target point , reflecting the directional information on the manifold. The base point is the reference point of the manifold embedding operation, and the sensor with the highest comprehensive score can be selected as the base point. The multidimensional feature vector of the base point is extracted as , which includes position coordinates, device parameters and historical statistical features, etc.; the target point is the mapping target in the manifold embedding process, representing the expected feature distribution position. The feature vectors of all sensors (including position, type, response characteristics, etc.) are extracted, the geometric center point of the feature vectors is calculated, and the center point is adjusted based on the physical constraints of the bridge structure. The target point determines the adjusted center point as the target point of each functional category, forming multiple , each target point corresponding to the expected feature distribution position of a class of sensors; denotes an exponential mapping used to convert tangent vectors to points on the manifold. By the manifold embedding operator, similar sensor features remain proximal in the manifold space, while features with large differences are distributed in different regions of the manifold, which can better reflect the topology of the bridge monitoring network. Take a continuous beam bridge as an example for geometric constraint analysis: for a continuous beam bridge, the sensors of the main span and the side span should maintain continuity in the manifold space to avoid geometric discontinuity. The sensors at the connection between the bridge tower top and the main beam should be close in the manifold space to reflect the strong coupling relationship. Bridge structures usually have symmetry, and sensors at symmetric positions should have symmetric embedding representations in the manifold space.

[0034] Multi-scale spatio-temporal cognitive convolutional neural network module: the multi-scale spatio-temporal cognitive convolutional neural network module is used to perform deep feature extraction on the heterogeneous topological graph based on the node embedding vector and the manifold model parameters to obtain multi-scale fusion features . Based on the constructed heterogeneous graph structure, the multi-scale spatio-temporal cognitive convolutional neural network performs deep feature extraction on the sensor data. The multi-scale spatio-temporal cognitive convolutional neural network module includes multi-layer cognitive convolutional layers and multi-scale feature fusion and meta-learning optimization layers. The multi-layer cognitive convolutional layer adopts a hierarchical attention enhanced topological convolution operator technology, which expands the traditional graph convolution to a multi-level processing framework with a cognitive mechanism through a cognitive heuristic topological convolution operator. Each cognitive convolutional layer includes a time series convolution operator, a spatial topological convolution operator, a cognitive attention operator, and a cognitive heuristic normalization layer. The time series convolution operator extracts a time series feature matrix reflecting the structural dynamic response law , capturing the modal characteristics, damping characteristics, and frequency domain response characteristics of the structure. The spatial topological convolution operator extracts a feature matrix reflecting the spatial distribution pattern of the bridge structure , which contains spatial information such as response patterns of key parts of the structure and load transfer path characteristics. The cognitive attention operator automatically identifies the feature patterns most sensitive to the structural health state through a differential attention mechanism, outputting an importance weight distribution The multi-scale feature fusion and meta-learning optimization layer includes multi-scale feature fusion units and meta-learning optimization units. Through a network architecture with multi-layered cognitive convolutional layers, the hierarchical cognitive process of human experts can be simulated, achieving intelligent recognition of cross-scale spatiotemporal patterns through cognitively inspired multi-scale convolutional layers. The design of the hierarchical attention-enhanced topological convolution operator fully considers the hierarchical and progressive characteristics of expert experience in bridge monitoring. Multi-scale feature analysis in bridge monitoring can be divided into spatial scale features and temporal scale features. Spatial scale features can include local scale (single sensor), component scale (single structural segment), and whole-bridge scale (overall structure); local scale is used to detect sensor-level anomalies, such as poor contact and equipment failure; component scale is used to identify local damage to components such as main beams, towers, and cables; whole-bridge scale is used to assess the overall health status and load-bearing capacity of the bridge. Temporal scale features can include instantaneous response (second-level), daily periodic changes (hour-level), and long-term trends (monthly / yearly), etc. Instantaneous responses can include dynamic events such as vehicle impacts and seismic excitations; diurnal variations can include diurnal temperature variations and traffic load variations; long-term trends can include structural degradation, material aging, and environmental erosion.

[0035] Multi-layer cognitive convolutional layer: Multi-layer cognitive convolutional layer is used to simulate the hierarchical cognitive process of human experts. It realizes intelligent recognition of cross-scale spatiotemporal patterns through cognitively inspired multi-scale convolutional layers.

[0036] Cognitive-inspired normalization layer: The cognitive-inspired normalization layer is used to normalize the temporal convolution operator, spatial topological convolution operator, and cognitive attention operator output by the multi-level processing framework, thereby obtaining a multi-scale node feature representation matrix. The normalization expression is: ;in, The node feature representation matrix; Indicates the network layer number (e.g., the first...). Layer convolution (layer) =1,2,…, ,in This represents the total number of floors. It is the first The node feature representation matrix of the layer, each layer corresponds to a level of expert cognition, with a dimension of N×d; It is the first Layer node feature representation matrix; It is the first Topological convolution operators for layers; The composition of functions demonstrates the synergistic effect of multiple cognitive abilities, that is... LayerNorm represents normalization, ensuring the stability of feature distribution and the convergence of training. This represents a cognitive attention operator that can simulate the attention allocation and focus mechanisms of experts; The spatial topological convolution operator can simulate an expert's ability to recognize spatial distribution patterns; This represents a temporal convolution operator, which can simulate an expert's ability to analyze the laws of temporal evolution.

[0037] Spatial topological convolution operator: Spatial topological convolution operator Used for geometric design based on bridge structures, capturing different types of spatial relationships. The spatial topological convolution operator models complex spatial topological structures through a weighted combination of various spatial relationship matrices. The expression for the spatial topological convolution operator is: ;in, It is the first The spatial topological convolution operator of the layer; r represents the index of the spatial relation type, used to enumerate different types of spatial adjacency relation matrices (e.g., direct adjacency relation, structural coupling relation, functional synergy relation, etc.). This represents the total number of spatial relation types, i.e., the total number of all spatial adjacency matrices involved in the modeling. This represents the weight matrix of the r-th relationship at the k-th level. This represents the bias vector of the k-th layer. For activation functions; This represents the input node feature matrix; Let be the adjacency matrix of the r-th spatial relationship, which may include direct adjacency matrix, structural coupling matrix and functional synergy matrix. Output feature matrix It contains key information about the bridge structure. For example, the information distribution can be as follows: Dimensions 1-16: Bending response characteristics of the main girder (positive bending moment, negative bending moment, shear force distribution); Dimensions 17-32: Response characteristics of the bridge tower (axial force, bending moment, lateral displacement); Dimensions 33-48: Response characteristics of the cable / main cable (tension change, vibration amplitude); Dimensions 49-64: Response characteristics of the support (support reaction force, displacement, rotation angle).

[0038] Direct adjacency: ;in, Represents the direct adjacency matrix; For sensors and The physical distance between them is obtained through GPS coordinates or design drawings. This is the adjacency threshold. This relationship reflects the strong correlation between spatially adjacent sensors.

[0039] Structural coupling relationship: ;in, Structural coupling matrix; Structural distance, which is based on the bridge influence line and stiffness matrix calculation, considers the load-carrying path and force transmission mechanism of the bridge. Decay coefficient. This relationship reflects the mechanical correlation between sensors that are coupled on the structure; exp represents the exponential function.

[0040] Functional synergy relationship: ; where, represents the functional synergy matrix; the corr() function is used to calculate the correlation coefficient between each column in the DataFrame, returning a correlation coefficient matrix, whose rows and columns correspond to the column names of the DataFrame, and the elements of the matrix are the correlation coefficients between the corresponding columns. and represent the type feature vectors of the i-th and j-th sensors, respectively, obtained using one-hot encoding. Based on the correlation calculation of sensor type features, the synergistic monitoring effect between functionally similar sensors is reflected.

[0041] Temporal convolution operator: temporal convolution operator Temporal dependency modeling is achieved through causal constraints and long short-term memory fusion to handle both long-term and short-term temporal dependencies. The expression of the temporal convolution operator is: ; where, represents the temporal convolution operator of the i-th layer; is a long short-term memory network responsible for capturing long-term temporal dependencies and trend changes, particularly suitable for handling periodic changes (such as daily temperature differences, seasonal changes, etc.) and long-term degradation processes in bridge monitoring. is a one-dimensional convolution network focused on short-term local temporal patterns, capable of quickly responding to sudden events and abnormal changes. represents the additive operation on the manifold, ensuring that the fusion result is still within a reasonable feature space. is a causal mask matrix that ensures information can only propagate from the past to the future, consistent with the physical causality law; is an element-wise multiplication operation used to apply causal constraints. The multi-time scale characteristics of bridge dynamics include long-term events and short-term events, used to handle long-term events, ​Used to handle short-term events. Long-term events can include: seasonal variations: the impact of annual temperature cycles on the bridge's fundamental frequency (frequency decreases by 3-5% in summer and increases in winter); traffic load cycles: differences in load patterns between weekdays and weekends; structural degradation trends: an annual lapse rate of 0.1-0.3% in the fundamental frequency, reflecting structural stiffness degradation; material creep effects: long-term deformation growth caused by concrete creep; short-term events can include: vehicle impact response: transient vibrations caused by heavy vehicles passing by (lasting 2-5 seconds); wind-induced vibrations: bridge buffeting response under gusts of wind (lasting 10-30 seconds); seismic response: structural response excited by seismic motion (lasting tens of seconds to several minutes); temperature shocks: short-term structural responses caused by diurnal temperature variations (hourly).

[0042] Cognitive attention operator: The cognitive attention operator simulates the attention allocation process of an expert, achieving intelligent suppression of redundant information through a differential attention mechanism. This mechanism is inspired by the differential perception mechanism of the human visual system, identifying and suppressing redundant information by comparing global and local attention patterns. The expression for the cognitive attention operator is: ;in, Represents the cognitive attention operator; Differential-Attention is the differential attention computation function. , and These are global query, local key-value, and cross-numerical matrix, used to enable the interaction of multi-scale information; For redundant attention matrices, mutual information estimation is used to automatically identify repetitive or redundant information patterns; To suppress the intensity coefficient and control the degree of redundancy suppression, the design of the cognitive attention operator can be guided by engineering considerations rather than purely data-driven approaches, by designing attention weights based on the importance of bridge engineering. Simultaneously, it considers global trends and local anomalies to achieve multi-scale fusion: automatically identifying and suppressing repetitive information, improving computational efficiency, and achieving redundancy suppression; the attention weights have clear engineering and physical meaning.

[0043] Global Attention Building Unit: The global attention building unit is used to construct global attention to capture macroscopic feature patterns at the bridge level. The expression for global attention is: ;in, Indicates global attention; For graph pooling operations, a global representation is extracted by aggregating the feature information of all nodes; This is the global query weight matrix, which converts the pooled global features into query vectors for subsequent attention calculations.

[0044] Local attention construction unit: the local attention construction unit is used to construct local attention to focus on microscopic details at the sensor level, and the local attention is expressed as: ; wherein, represents local attention; LocalConv is a local convolution operation that extracts local feature patterns through a small kernel size (such as 3x3) convolution kernel, represents that the local convolution uses a 3x3 convolution kernel window. This operation simulates the sensitivity and detail observation ability of experts to local anomalies, and can identify minor changes, local damage, contact problems, etc. in a single sensor or small area; is a node feature representation matrix with a dimension of NxD, containing comprehensive feature information of each sensor node; is a local key-value weight matrix that converts local convolution features into key-value vectors. The design of local attention ensures the sensitivity of the model to subtle changes.

[0045] Redundancy information suppression unit: the redundancy information suppression unit is used to construct redundancy attention to identify redundant information patterns through mutual information estimation. This mechanism identifies redundant relationships by calculating the mutual information between node features. The redundancy attention is expressed as: ; wherein, represents redundancy attention; is the mutual information between the feature vectors of sensor nodes and , reflecting the degree to which two nodes contain the same information; when the mutual information exceeds the redundancy threshold , the indicator function is 1, indicating that there is significant redundancy; otherwise, it is 0; this design can automatically identify pairs of sensors that provide similar information, avoiding redundant calculations and feature redundancy, and improving the efficiency and generalization ability of the model.

[0046] Multi-scale feature fusion and meta-learning optimization layer: the multi-scale feature fusion and meta-learning optimization layer is used to perform feature fusion on the multi-scale node feature representation matrix based on meta-learning, to obtain multi-scale fusion features. Let the features extracted at the th scale be , the spatial features be , and the temporal features be , which are obtained through spatial topology convolution and temporal convolution networks, respectively. The attention coefficient is generated by a meta-learner according to the gradient and complexity information, and is used to guide the manifold alignment and weighted fusion of multi-scale features, to finally obtain the multi-scale fusion features : ; wherein, This represents multi-scale fusion features, used to unify features at different scales into the target manifold space through weighted fusion and manifold interpolation mapping. The global representation in the text; k represents the scale variable; For the total number of scales, This represents the feature representation at the k-th scale extracted through different receptive fields (such as 1×1, 3×3, 5×5 convolution), covering the multi-scale spatial response from local components to the overall structure; These are adaptive weighting coefficients; For manifold interpolation operators; The target manifold is a space with specific geometric constraints, and all scale characteristics are as follows: , All features must be mapped onto this unified geometric space for alignment and fusion. The manifold interpolation operator ensures that features at different scales can be effectively fused in the unified manifold space, avoiding geometric inconsistencies that may result from direct weighted averaging.

[0047] Meta-learner gradient optimization strategy building unit: This unit is used to construct the gradient optimization strategy for the meta-learner. The meta-learner optimizes the weight allocation strategy using gradient information from the validation set, obtaining adaptive weight coefficients. The expression for the adaptive weight coefficients is: ;in, For adaptive weight coefficients; MetaLearner represents the model in the meta layer, which summarizes the training experience across all tasks; To verify the gradient of the loss with respect to the parameters, it reflects the current learning state and optimization direction of the model; Indicates the first The global representation obtained after scale features are processed by pooling, normalization, graph aggregation, and other operations; The complexity metric representing the subnetwork at the k-th scale can be comprehensively evaluated by indicators such as gradient variability, model structure entropy, and number of parameters, reflecting the learning load and generalization ability of the current branch. The meta-learner comprehensively considers these three factors and dynamically adjusts the weights of features at each scale to achieve an optimal balance between performance and efficiency.

[0048] Multi-scale spatio-temporal cognitive network performance verification: The comprehensive performance of the heterogeneous graph neural network is compared with traditional methods based on six months of monitoring data from three different types of bridges, including 156 heavy load events and 23 abnormal working conditions. The evaluation is conducted from six dimensions: detection accuracy, processing speed, false positive rate, interpretability, robustness, and adaptability. The heterogeneous graph neural network of the invention significantly outperforms traditional methods in all evaluation indicators: detection accuracy is improved from 70% to 83%, processing speed is improved by 30%, and false positive rate is reduced by 35%. To analyze the effect of multi-scale feature fusion, the experiment compares the performance of single-scale, double-scale, triple-scale, and the adaptive multi-scale fusion method of the invention. The results show that as the number of scales increases, the damage location accuracy gradually improves, but the computational efficiency decreases accordingly. The adaptive multi-scale fusion method of the invention dynamically optimizes weight distribution through meta-learning, achieving 80% damage location accuracy while maintaining 85% computational efficiency, achieving the optimal balance between precision and efficiency. Based on the visualization of the distribution of sensor nodes in the manifold embedding space, different types of sensors naturally form clusters: strain sensors are clustered in the upper left area, acceleration sensors are distributed in the lower right area, displacement sensors and temperature sensors form relatively independent groups. This distribution verifies that the manifold embedding technology can effectively maintain the geometric constraints and physical relationships of the bridge structure, providing a reasonable feature representation space for heterogeneous data fusion.

[0049] Continuous manifold spatio-temporal alignment and Bayesian fusion module: The continuous manifold spatio-temporal alignment and Bayesian fusion module is used to solve the spatio-temporal synchronization problem of multi-source data. By performing spatio-temporal alignment and fusion processing on multi-scale fusion features and raw time series data of sensors, spatio-temporal alignment parameters and fusion state vectors are obtained. The continuous manifold spatio-temporal alignment technology specifically solves the spatio-temporal synchronization problem in bridge health monitoring. In bridge monitoring, due to the use of sensors from different manufacturers and different batches, there is a serious clock synchronization problem; sampling frequency differences cause data misplacement; network transmission delays are random. To this end, the continuous manifold spatio-temporal alignment technology can solve the spatio-temporal synchronization problem between different sensors. The optimal time offset parameter is obtained through mutual information entropy optimization, and the spatial correction parameter is obtained through non-rigid spatial transformation. The variational Bayesian fusion algorithm fuses multi-modal sensor data in a unified probability framework, outputs the fused state vector and its uncertainty estimate , which contains the comprehensive health information of the bridge structure at the current time. Figure 3As shown, the continuous manifold spatiotemporal alignment and Bayesian fusion technique transforms the discrete spatiotemporal alignment problem into a geodesic optimization problem on a continuous manifold, fully considering the geometric constraints and physical characteristics of the bridge structure. In bridge monitoring, the "spatiotemporal trajectory" of the sensor is not its physical movement trajectory, but rather the evolution path of the sensor monitoring data on the spatiotemporal manifold: in the time dimension, the trajectory of sensor data changes over time; in the spatial dimension, the correlation pattern of sensor data at different spatial locations; and in the manifold constraint, maintaining the geometric and physical consistency of the bridge structure. The continuous manifold spatiotemporal alignment and Bayesian fusion module includes a B-spline trajectory fitting unit, a mutual information time synchronization unit, a non-rigid body spatial transformation estimation unit, a joint spatiotemporal optimization unit, and a sensor output unit. The sensor output unit is used to output the calibrated sensor data.

[0050] B-spline trajectory fitting unit: The B-spline trajectory fitting unit is used to fit B-spline trajectories based on continuous spatiotemporal manifold trajectories. It includes a manifold trajectory construction module, a B-spline function design module, and a manifold constraint optimization module.

[0051] Manifold Trajectory Construction Module: This module extends the B-spline trajectory model to the manifold space using exponential and logarithmic mappings, resulting in a continuous spatiotemporal manifold trajectory. The expression for the continuous spatiotemporal manifold trajectory is: ;in, Represents the trajectory of a continuous spatiotemporal manifold; This indicates that a linear combination of the tangent space is mapped back to the manifold; The base point on the manifold is usually chosen as the geometric center or centroid of the data. In bridge applications, the reference sensor of the bridge monitoring system (usually the strain sensor of the main beam at the mid-span of the bridge) is selected to represent the reference benchmark of the structural response. The control point sequence determines the shape and direction of the trajectory. It represents the state points of each sensor at different times, and is a four-dimensional state vector containing [strain value, acceleration, displacement, temperature]. Represented as a B-spline basis function, it controls the degree of influence of each control point on the trajectory; This represents mapping the control points to the tangent space at the base point; i represents the control point index; n represents the total number of control points.

[0052] B-Spline Basis Function Design Module: This module is used to design B-spline basis functions based on manifolds. In bridge applications, it is necessary to consider the boundary conditions and physical constraints of the structure, which can be recursively extended on the manifold as follows: ;in, Represent the B-spline basis functions; This represents the geodesic distance from the manifold point corresponding to parameter t to the i-th node. The geodesic distance is the shortest path length between two points on the manifold. denotes the geodesic distance from the i+p-th node to the i-th node; denotes the geodesic distance from the i+p+1-th node to the point corresponding to the parameter t; denotes the geodesic distance from the i+p+1-th node to the i+1-th node; and denotes the basis function of low order, which is the basis of recursive definition. The recursive definition replaces the parametric distance in Euclidean space with the geodesic distance on the manifold . The geodesic distance reflects the intrinsic geometric properties of the manifold and can more accurately describe the real distribution of sensor data on the spatiotemporal manifold. The recursive process starts from the zeroth order basis function (unit impulse function) and gradually constructs higher order basis functions, each recursion maintaining the geometric constraints of the manifold. This design ensures the geometric consistency and numerical stability of the trajectory model throughout the spatiotemporal domain. The B-spline basis function designed in the invention is based on the structural geodesic distance, naturally incorporating the stiffness and mass distribution of the bridge; through manifold constraints, the results conform to the principles of structural mechanics; and in areas with varying bridge stiffness (such as near the supports), the interpolation weights are automatically adjusted to improve local accuracy.

[0053] Manifold constraint optimization module: The manifold constraint optimization module is used to process the continuous spatiotemporal manifold trajectory through manifold constraint optimization to obtain the manifold control points. It includes a control point solving unit and a geometric smoothing constraint unit.

[0054] Control point solving unit: The control point solving unit is used to solve the manifold control points. The manifold control points are solved through manifold constraint optimization: ; wherein, denotes the control points obtained by solving; denotes the parameter value that minimizes the objective function; j denotes the data point index; and m denotes the total number of data points; denotes the square of the geodesic distance between the manifold trajectory points and the observed data points; denotes the continuous spatiotemporal manifold trajectory; denotes the j-th observed data point; is a smoothing weight parameter that controls the balance between fitting accuracy and smoothness; denotes the manifold regularization term. The objective of this optimization problem is to find the optimal control point configuration so that the constructed manifold trajectory can best fit the observed data. The first term is the data fitting term, which ensures the fitting accuracy by minimizing the sum of the squares of the geodesic distances between the points on the trajectory and the observed data; the second term is the manifold regularization term, which ensures the smoothness and reasonableness of the trajectory by constraining the geometric distribution of the control points.

[0055] Geometric smoothness constraint unit: The geometric smoothness constraint unit is used to construct the manifold regularization term , to ensure the geometric smoothness of the trajectory, the expression of the manifold regularization term is: ; wherein, represents the manifold regularization term, which ensures the smoothness of the trajectory by constraining the geometric relationship of the adjacent three control points; i represents the control point index; n represents the total number of control points; represents the square of the geodesic distance of the geometric average of the i+1th control point and its adjacent two control points; represents the i+1th manifold control point; is the geometric average of two points on the manifold, which corresponds to the midpoint of the geodesic line on the Riemannian manifold, so that the intermediate control point is as close as possible to the geometric center of its adjacent two control points, thereby avoiding sharp turns or unreasonable geometric shapes of the trajectory.

[0056] Mutual information time synchronization unit: the mutual information time synchronization unit realizes the time alignment of the original time series data of different sensors through information geometry optimization, determines the optimal time alignment by maximizing the geometric mutual information of two signals under the time offset Δt, and the expression of the optimal time offset parameter is: ; wherein, represents the optimal time offset parameter; is the parameter maximization operator, which finds the Δt that maximizes the objective function among all possible time offset values; is the information geometry mutual information; represents the time series signal of the first sensor; represents the time series signal of the second sensor after time offset Δt; is the entropy constraint term, which prevents overfitting caused by excessive time offset; is the entropy weight coefficient, which controls the balance between alignment accuracy and robustness. The mutual information time synchronization unit includes an information geometry mutual information determination module, an entropy constraint term construction module, and a statistical manifold optimization module.

[0057] Information geometry mutual information determination module: the information geometry mutual information determination module is used to calculate the information geometry mutual information based on the Fisher information metric. In bridge application, the geometric mutual information not only reflects the statistical correlation, but also embodies the internal correlation of the structural physical response. The expression of the information geometry mutual information is: ; wherein, represents the information geometry mutual information; X and Y represent the responses of the same bridge structure at different positions or different physical quantities, and represent different sensor data; is a statistical manifold, each point of which corresponds to a probability distribution; is the Fisher information metric tensor, which defines the Riemannian metric on the statistical manifold; The gradient of the log-likelihood function is called the natural parameter in information geometry. Let be the joint probability density function; For volume elements on a manifold; This indicates that integration is performed on a statistical manifold. This geometric mutual information can more sensitively capture nonlinear correlations and higher-order statistical dependencies between signals, making it particularly suitable for handling complex multimodal signal alignment problems in bridge monitoring.

[0058] Entropy Constraint Module: This module designs entropy constraints by incorporating the maximum entropy principle. Based on information theory's maximum entropy principle, this module encourages the probability distribution of time offsets to have maximum uncertainty, thus avoiding excessive preference for specific time offset values. In the absence of prior information, the maximum entropy distribution is the most conservative and objective choice, effectively preventing the algorithm from getting trapped in local optima or becoming overly sensitive to noise. This design ensures the robustness and generalization ability of the time synchronization algorithm, making it particularly suitable for handling practical problems such as clock drift and inconsistent sampling rates that may exist in bridge monitoring. Entropy Constraint Term The expression is: ;in, Represents the entropy constraint term; It is the probability density function of the time offset parameter Δt; The self-information of a time-offset event; Indicates the time offset parameter Find the integral.

[0059] Statistical Manifold Optimization Module: This module optimizes the time-shift parameter on the statistical manifold by introducing the natural gradient method. The expression for the optimized time-shift parameter is as follows: ;in, This represents the time offset parameter obtained in the (k+1)th iteration; This represents the time offset parameter obtained in the k-th iteration; The learning rate parameter; For the Fisher information matrix, a metric on the statistical manifold is defined; Its inverse matrix is ​​used to convert the Euclidean gradient into the natural gradient; The gradient of the geometric mutual information with respect to the time offset. The advantage of the natural gradient is that it can automatically adapt to the local geometry of the parameter space, introducing a larger step size in flat regions and a smaller step size in curved regions, thereby achieving faster convergence and better numerical stability.

[0060] Non-rigid space transformation estimation unit: The non-rigid space transformation estimation unit is used to handle the spatial calibration error of the sensor, including the thin plate spline transformation module and the robust estimator solution module.

[0061] Thin plate spline transformation module: the thin plate spline transformation module is used to solve the non-rigid spatial transformation by thin plate spline interpolation and robust estimator to obtain spatial correction parameters: ; wherein, represents the spatial correction parameters; A represents the affine transformation matrix; x represents the spatial coordinate point; t represents the time variable; i represents the control point index; n represents the total number of control points; represents the weight coefficient of the i-th control point; represents the coordinates of the i-th control point; represents the radial basis function; represents the i-th control point position. It is responsible for solving the sensor spatial position correction problem in bridge monitoring, correcting the systematic position deviation of the sensor during installation, unifying the sensors installed in different batches to the same coordinate system, and compensating for the linear deformation of the whole bridge (such as overall expansion caused by temperature).

[0062] Robust estimator solving module: the robust estimator solving module is used to process the calculation formula of the spatial correction parameters of the thin plate spline transformation module based on the robust parameter estimation, to obtain the robust spatial correction parameters.

[0063] Joint space-time optimization unit: the joint space-time optimization unit is used to realize the probability fusion and uncertainty quantification of multi-source heterogeneous sensor data in bridge monitoring, including the variational posterior distribution module and the ELBO optimization target module.

[0064] Variational posterior distribution module: the variational posterior distribution module is used to determine the variational posterior distribution, and the expression of the variational posterior distribution is: ; wherein, is the variational posterior distribution, and x represents the observation data (sensor data); represents the normal distribution, that is, the complex bridge health state is modeled as a simple normal distribution; is the mean function, is the diagonal covariance matrix; represents the variance function.

[0065] ELBO optimization target module: the ELBO optimization target module is used to balance the data fitting accuracy and the model rationality, and the expression of the reconstruction loss of the ELBO optimization function is: ; wherein, is the reconstruction loss, which represents the generation probability of the observation data given the latent state; represents the expectation under the variational posterior distribution; represents the decoder generation probability; represents the ratio of the variational posterior to the prior; KL represents the KL divergence regularization, which is used to measure the difference between two distributions, and the posterior distribution is constrained not to deviate too far from the prior distribution.

[0066] Experimental verification of continuous manifold space-time alignment technology: To verify the effectiveness of the continuous manifold space-time alignment technology, an experimental 1200m bridge monitoring system is selected, which contains 60 heterogeneous sensors with a sampling frequency difference ranging from 0.1Hz to 200Hz. Linear interpolation, Kalman filter, traditional neural network and the method of the present application are compared and tested. The time synchronization error of the traditional linear interpolation method is ±100ms, the Kalman filter method is ±50ms, the traditional neural network method is ±25ms, and the continuous manifold space-time alignment technology of the present application improves the synchronization accuracy to ±5ms, which is 95% higher than the traditional method. The experimental results show that the method of the present application significantly improves the space-time synchronization accuracy of multi-source heterogeneous sensor data by introducing the geometric constraints and physical characteristics of the bridge structure, and provides a high-quality data basis for subsequent health state analysis.

[0067] Structure health index calculation and state evaluation module: The structure health index calculation and state evaluation module is used to perform structure health monitoring and state evaluation of the bridge based on the fusion state vector, and obtain the final health evaluation result. The structure health index calculation and state evaluation module includes a structure health index calculation unit and a state evaluation and early warning unit.

[0068] Structure health index calculation unit: The structure health index calculation unit is used to adopt an end-to-end intelligent analysis and design concept, and construct a complete bridge structure health monitoring and state evaluation system through four core modules of multi-level health index extraction, federated learning fault diagnosis, counterfactual explanation analysis and structure state prediction evaluation, to obtain the structure health index. The processing results of heterogeneous graph neural network, multi-scale space-time convolution and continuous manifold alignment technology are fully integrated, to provide accurate, interpretable and reliable structure health evaluation results for engineering and technical personnel. The structure health index calculation unit includes a multi-level health index extraction module, a physical constraint damage identification module, a fuzzy evidence fusion evaluation module and a comprehensive health index calculation module.

[0069] Multi-level health index extraction module: The multi-level health index extraction module is used to calculate the structure health index in layers. The multi-level health index extraction introduces a layered calculation architecture, which integrates three levels of structure response indicators, damage identification indicators and safety performance indicators. This layered architecture simulates the step-by-step deepening process of health evaluation by structure engineers. The expression of the multi-level health index is as follows: ; wherein, represents the multi-level health index; is a safety performance evaluation layer, which gives the final structure safety state judgment by integrating various indicators; is a damage identification indicator calculation layer, which identifies potential structure damage based on response characteristics; The structural response index extraction layer focuses on the basic physical response quantification; The multi-scale fusion features are shown. Each layer contains specific engineering physical meaning and evaluation criteria to ensure the engineering practicability of the evaluation results.

[0070] Structural response index extraction layer: the structural response index extraction layer For designing multi-dimensional response features based on bridge structural mechanics principles, extracting structural response indexes : ; wherein, represents the structural response index; concat represents concatenation; represents the dynamic response index; represents the static response index; represents the temperature effect index; represents the modal feature index. The calculation methods of each response index are as follows: dynamic response index Obtained by time-frequency domain analysis: ; wherein, represents the acceleration root mean square value; represents the acceleration peak value; represents the peak factor; represents the power spectral density integral; represents the acceleration time history; represents the frequency. Static response index Based on strain and displacement measurement: ; wherein, represents the maximum strain value, is the strain; represents the average strain; represents the maximum displacement value, is the displacement; represents the residual displacement. Temperature effect index Considering the influence of temperature gradient: ; wherein, represents the temperature gradient; represents the thermal expansion coefficient; represents the temperature difference; represents the temperature compensation strain. Modal feature index Extracted by modal analysis: ; wherein, represents the natural frequency; represents the damping ratio; represents the mode shape; represents the mode shape consistency index.

[0071] Physical constraint damage identification module: the physical constraint damage identification module (damage identification index calculation layer ) for introducing machine learning and physical model to extract damage features, the optimization objective integrates three aspects of physical constraints, machine learning fitting and sparsity regularization. The expression of damage feature is: ; wherein, represents the damage feature; represents the parameter value that maximizes the objective function; is the physical loss based on the theory of structural mechanics; represents the machine learning loss weight; represents the sparsity regularization coefficient; represents the sparsity regularization term; represents the damage parameter vector. The physical loss is used to ensure that the damage identification result conforms to the engineering physical law, and the expression of the physical loss is: ; wherein, is the stiffness matrix considering the influence of damage, F is the load vector; represents the square of L2 norm; represents the dynamic constraint weight; M and C are the mass and damping matrices, is the displacement vector; represents the first-order time derivative (velocity) of the displacement vector; represents the second-order time derivative (acceleration) of the displacement vector. is the machine learning loss, which captures complex damage patterns through a deep learning model: ; wherein, CrossEntropy represents the cross-entropy loss function; DamageNet represents the damage identification neural network; DamageLabel represents the damage category label; represents the structural response feature vector.

[0072] Fuzzy evidence fusion evaluation module: safety performance evaluation layer Adopting multi-criteria decision analysis method, combining fuzzy theory and evidence theory: ; the formula fuses multiple fuzzy evaluation results through Dempster-Shafer theory (DST). is the fuzzy membership function of the th damage index, is the total number of damage indexes. The fuzzy membership function is designed as: represents the Dempster-Shafer evidence fusion operator; represents the specific value of the kth damage index; represents the damage index variable; K represents the total number of damage indexes. ; wherein, exp represents the exponential function; x represents the input damage index value; and Let be the fuzzy parameter and standard deviation parameter of the k-th indicator, respectively. This is the importance weight vector; k represents the index variable.

[0073] The structural condition classification adopts a five-level evaluation standard: Level I (Excellent): The structure is in good condition and requires no special maintenance; Level II (Good): The structure is basically normal; regular inspections are recommended. Level III (General): Minor damage exists, requiring close monitoring; Grade IV (Poor): Significant structural damage requires immediate repair; Level V (Hazardous): The structure has safety hazards and should be dealt with immediately.

[0074] Comprehensive health indicator calculation module: Through the aforementioned multi-source data fusion and feature extraction process, the system obtains high-quality feature representations. The comprehensive health index calculation module is used to calculate key structural health indicators: ;in, Indicates structural health indicators; As a state evaluation function, it achieves end-to-end analysis from raw features to health status assessment through cascaded processing; This is a fault diagnosis function; A function for extracting health indicators; This represents the multi-scale fusion feature.

[0075] State Assessment and Early Warning Unit: The State Assessment and Early Warning Unit is used for distributed collaborative fault detection, including a federated learning fault diagnosis module, a counterfactual interpretation analysis module, a state prediction and early warning module, and a comprehensive decision support system.

[0076] Federated Learning Fault Diagnosis Module: This module introduces differential privacy protection and secure multi-party computation, achieving network-wide fault detection through a distributed collaborative learning strategy. ;in, Represents the model parameters after the (t+1)th round of global aggregation; SecAgg represents the secure aggregation protocol; For the first Local model parameters for each node. The noise is differential privacy noise; k represents the node index; K represents the total number of nodes participating in federated learning. This mechanism ensures that all bridge monitoring nodes can jointly train the global fault diagnosis model without leaking local data; t represents the round of federated learning.

[0077] Joint Confidence Assessment Unit: The joint confidence assessment unit is used to evaluate the accuracy of fault diagnosis using a joint confidence index. The expression for the joint confidence index is: ; wherein, represents the joint confidence indicator; k represents the node variable; K represents the total number of nodes; is the credibility weight of the node ; represents the accuracy of fault diagnosis of the kth node; represents the recall rate of fault diagnosis of the kth node; is the consistency score of the diagnosis result of the node with the global consensus.

[0078] Counterfactual explanation analysis module: the counterfactual explanation analysis module is used to provide an interpretable health assessment result, including a counterfactual sample generation unit and a feature attribution analysis unit.

[0079] Counterfactual sample generation unit: the counterfactual sample generation unit realizes the interpretability of the health assessment result through neural-symbolic reasoning and counterfactual generation: ; wherein, represents the counterfactual sample generation function; represents the counterfactual sample that minimizes the objective function; represents the prediction output of the health assessment model on the counterfactual sample; represents the sparsity regularization coefficient; represents the generated counterfactual sample; represents the original monitoring data sample; represents the L0 norm, which calculates the number of non-zero elements; represents the semantic constraint weight; represents the semantic distance function; target represents the target health state label. The formula generates a counterfactual sample with the smallest perturbation, which explains “how will the structural health state change if some monitoring parameters change”. The first term ensures that the counterfactual sample can reach the target health state; the second term constrains the number of parameters that need to be changed to be the least; and the third term ensures that the counterfactual sample is reasonable in engineering semantics.

[0080] Feature attribution analysis unit: the feature attribution analysis unit is used to fuse the integral gradient and the Shapley value to calculate the importance of each monitoring parameter through feature attribution analysis: ; wherein, represents the importance contribution of the ith monitoring parameter; represents the expectation on the baseline distribution; represents the model prediction value on the integral path; represents the ith monitoring parameter; represents the baseline reference sample; represents the interpolation path from the baseline to the current sample; represents the current monitoring sample; This represents the partial derivative operation with respect to the i-th parameter; Indicates the interpolation parameters; This represents the function for calculating the Shapley value. This represents the derivative with respect to the interpolation parameters; This represents a health assessment model. This indicator quantifies the... The contribution of each monitoring parameter to the final health assessment result provides a quantitative basis for engineers' decision-making.

[0081] State prediction and early warning models: State prediction and early warning models are used for prospective health status assessment, including physical-random hybrid prediction units, multi-level early warning mechanisms, and Monte Carlo confidence calculation.

[0082] Physical-stochastic hybrid prediction units (PSUs) combine time series analysis and structural degradation models to achieve prospective assessments of health status. ;in, This indicates the structural health status at a future time t+Δt; Represents a deterministic physical degradation term; Represents a random uncertainty term; This indicates the structural health status at the current time t; Indicates the prediction time step; This represents a historical state sequence. The predictive model includes deterministic physical degradation terms and stochastic uncertainty terms. The physical degradation term is based on fatigue damage accumulation and material deterioration patterns. ;in, The vector represents the structural health state; exp represents the exponential function. Indicates the time step; This represents the degradation rate function at time τ; Represents the integral variable; Indicates 0- During the period of time Integration. The uncertainty term is modeled using a Long Short-Term Memory (LSTM) network: ;in, Represents a random uncertainty term; This represents the model's uncertainty noise term.

[0083] Multi-level early warning mechanism unit: The multi-level early warning mechanism unit is used to implement the early warning mechanism. The early warning mechanism adopts a multi-level threshold design (three-level early warning mechanism): Yellow warning: predicted state decline rate >5% / year; Orange warning: predicted that it may fall to level III within 6 months; Red warning: predicted that it may fall to level IV or below within 3 months.

[0084] Monte Carlo confidence calculation unit: the Monte Carlo confidence calculation unit calculates the early warning confidence through Monte Carlo simulation: ; wherein, represents the early warning confidence; is the number of simulations; i represents the simulation variable; represents the future state prediction value of the i-th simulation; threshold represents the threshold corresponding to the early warning level; is an indicator function. Through the above complete structure health index calculation and state evaluation framework, the system can provide accurate, timely and interpretable structure health state information for bridge managers, support scientific maintenance decision and risk control. The fuzzy membership function is designed by combining the Gaussian function with the importance weight: ; wherein, represents the fuzzy membership degree of the k-th index; x represents the input monitoring parameter value; exp represents the exponential function; , is the fuzzy parameter of the i-th index, is the importance weight vector, which is determined by expert knowledge and historical data statistics.

[0085] Comprehensive decision support system: the comprehensive decision support system provides multi-level health state evaluation, damage positioning analysis, maintenance suggestion generation and risk early warning information to the field experts through an interactive visual interface, realizes the intelligentization of the whole process of bridge health monitoring from data acquisition to engineering decision

[0086] System real-time performance and damage identification verification: Figure 4 ​The real-time performance monitoring results of the bridge monitoring system running continuously for 24 hours are shown, and the system deploys three edge computing nodes to process the real-time data stream of 60 sensors. The monitoring data shows that the system processing delay slightly rises during working hours (8:00-18:00), with an average of 220 ms, and is 150 ms during non-working hours. The overall average processing delay is 185 ms, which is far below the design requirement of 300 ms. The system availability rate remains between 99.2%-99.8%, and only drops to 98.7% during the 2:00 AM maintenance period. The overall availability rate is 99.7%, meeting the real-time requirements of bridge monitoring. By comparing the accuracy of different methods in four typical damage identification methods, the invention method is tested and compared with traditional threshold method, support vector machine, deep learning and other methods in four common bridge damage modes: fatigue crack, support abnormality, cable relaxation and bridge damage. The experimental results show that: for fatigue crack identification, the accuracy of the invention method is 94%, which is 45% higher than that of the traditional threshold method (65%); for support abnormality identification, the accuracy of the invention method is 91%, which is 21% higher than that of the optimal traditional method; for cable relaxation identification, the accuracy of the invention method is 96%, reaching the best identification effect; for bridge damage identification, the accuracy of the invention method is 89%, which is 25% higher than that of the traditional method. In summary, the invention method improves the early damage identification ability by an average of 80% and reduces the misdiagnosis rate by 50%, providing reliable technical support for bridge preventive maintenance.

[0087] The above merely illustrates the preferred embodiments of the present application and is not intended to limit the present application. The present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of the present application.

Claims

1. A system for fusing multi-source data on bridge structures, characterized in that, It includes modules for heterogeneous topology map construction and manifold embedding technology, continuous manifold spatiotemporal alignment and Bayesian fusion, structural health index calculation and state assessment, and multi-scale spatiotemporal cognitive convolutional neural network. The heterogeneous topology graph construction and manifold embedding technology module is used to process multi-source heterogeneous sensor data based on the heterogeneous graph structure to obtain a heterogeneous topology graph, and to perform manifold embedding learning on the sensor nodes to obtain node embedding vectors and manifold model parameters; the heterogeneous graph structure includes a set of nodes, a set of edges, a set of hyperedges, a relation type mapping space, and a joint manifold embedding mapping function; The heterogeneous topology graph construction and manifold embedding technology module includes a data preprocessing and heterogeneous graph initialization unit, a hyperheterogeneous graph construction unit, and a manifold embedding learning unit; The data preprocessing and heterogeneous map initialization unit is used to preprocess multi-source heterogeneous sensor data to obtain a standardized sensor data stream and initialize the heterogeneous topology map; multi-source heterogeneous sensor data refers to various types of sensor data used for bridge monitoring, including strain data, acceleration data, displacement data and temperature data; The heterogeneous graph construction unit is used to compute the elements in the initialized heterogeneous topology graph based on the standardized sensor data stream, and obtain the heterogeneous topology graph. The elements in the heterogeneous topology graph include the node set, the relation edge set, the hyperedge set, and the joint manifold embedding map. The node set includes sensor entity nodes, spatiotemporal data nodes, and semantic concept nodes. The hyperheterogeneous graph construction unit includes a node construction module, a relation edge construction module, a hyperedge relation construction module, a relation type mapping space determination module, and a joint manifold embedding mapping module; The node construction module is used to identify and analyze the data, data transmission and data processing methods in the standardized sensor data stream to obtain sensor nodes, data nodes, semantic nodes and meta-information nodes, and take the union of multiple types of nodes as the node set; The relation edge construction module is used to construct an edge set based on multiple nodes in a node set and the relationships between the nodes; The hyperedge relation construction module obtains tensor decomposition parameters by performing high-order tensor decomposition on the hyperedge relation, and then uses the tensor decomposition parameters to represent the hyperedge relation. The relation type mapping space determination module is used to determine the relation type mapping space; the relation type mapping space is used to define the specific types and attributes of edges and hyperedges. The joint manifold embedding mapping module is used to construct feature vectors based on heterogeneous topological maps and raw sensor data, resulting in high-dimensional sensor node embedding vectors. The manifold embedding learning unit includes a geometry-structure joint embedding module, a meta-path information propagation module, a Riemann optimization algorithm module, and a multimodal feature fusion module; The geometry-structure joint embedding module is used to fuse the geographic coordinates of the sensors and the bridge structure information to obtain geometric location features; The meta-path information propagation module is used to construct meta-path propagation operators based on the information propagation process in heterogeneous topology graphs, and to simulate the process of information propagation in heterogeneous topology graphs. The Riemann optimization algorithm module is used to perform manifold embedding learning based on the Riemann optimization algorithm and update parameters under manifold constraints to obtain manifold model parameters. The multimodal feature fusion module is used to fuse multimodal features to obtain a fused feature vector, and a manifold embedding operator is designed to map the fused feature vector in Euclidean space to Riemannian manifold space through the manifold embedding operator to capture the nonlinear geometric relationships in the bridge sensor network. The manifold embedding learning unit is used to perform manifold embedding learning on sensor nodes in the heterogeneous topology map to obtain manifold embedding operators; the multi-scale spatiotemporal cognitive convolutional neural network module is used to perform deep feature extraction on the heterogeneous topology map based on node embedding vectors and manifold model parameters to obtain multi-scale fusion features. The continuous manifold spatiotemporal alignment and Bayesian fusion module is used to perform spatiotemporal alignment and fusion processing on multi-scale fusion features and raw time-series data from sensors to obtain spatiotemporal alignment parameters and fusion state vectors; The Structural Health Indicator Calculation and Condition Assessment module is used to monitor the structural health and assess the condition of bridges based on fused condition vectors, and obtain the final health assessment results.

2. The system for fusing multi-source data on bridge structures according to claim 1, characterized in that, The multi-scale spatiotemporal cognitive convolutional neural network module includes multi-layer cognitive convolutional layers and multi-scale feature fusion and meta-learning optimization layers; Multi-layer cognitive convolutional layers are used to achieve multi-scale feature recognition, including cognitive heuristic normalization layers, spatial topological convolution operators, temporal convolution operators, and cognitive attention operators; The cognitive-inspired normalization layer is used to normalize the temporal convolution operator, spatial topological convolution operator, and cognitive attention operator output by the multi-level processing framework to obtain a multi-scale node feature representation matrix. Spatial topological convolution operators are used for designing bridge structures based on geometric characteristics to capture different types of spatial relationships. Temporal convolution operators handle both long-term and short-term temporal dependencies simultaneously by fusing causal constraints and long short-term memory. Cognitive attention operators are used to simulate the attention allocation process of experts and suppress redundant information through differential attention mechanisms. The multi-scale feature fusion and meta-learning optimization layer is used to perform feature fusion on the node feature representation matrix at multiple scales based on meta-learning, so as to obtain multi-scale fused features.

3. The system for fusing multi-source data on bridge structures according to claim 2, characterized in that, Cognitive attention operators include global attention building units, local attention building units, and redundant information suppression units; The global attention building block is used to construct global attention and capture macroscopic features; Local attention building blocks are used to construct local attention and capture sensor-level details; The redundancy information suppression unit is used to construct redundancy attention, identify repetitive information patterns through mutual information estimation, and identify redundant relationships by calculating the mutual information between node features.

4. The system for fusing multi-source data on bridge structures according to claim 1, characterized in that, The continuous manifold spatiotemporal alignment and Bayesian fusion module includes a B-spline trajectory fitting unit, a mutual information time synchronization unit, a non-rigid body spatial transformation estimation unit, and a joint spatiotemporal optimization unit. The B-spline trajectory fitting unit is used to fit B-spline trajectories based on continuous spatiotemporal manifold trajectories; The mutual information time synchronization unit achieves time alignment between the raw time series data of different sensors through information geometry optimization. It determines the optimal time alignment by maximizing the geometric mutual information of the two signals under time offset, and obtains the optimal time offset parameter. The non-rigid space transformation estimation unit is used to handle the spatial calibration error of the sensor; The joint spatiotemporal optimization unit is used to achieve probabilistic fusion and uncertainty quantification of multi-source heterogeneous sensor data in bridge monitoring.

5. The system for fusing multi-source data on bridge structures according to claim 4, characterized in that, The B-spline trajectory fitting unit includes a manifold trajectory construction module, a B-spline function design module, and a manifold constraint optimization module; The manifold trajectory construction module is used to extend the B-spline trajectory model to the manifold space through exponential and logarithmic mappings to obtain continuous spatiotemporal manifold trajectories. The B-spline basis function design module is used to design B-spline basis functions based on manifolds; The manifold constraint optimization module is used to process the continuous spatiotemporal manifold trajectory through manifold constraint optimization to obtain manifold control points; The mutual information time synchronization unit includes a mutual information determination module for information geometry, a module for constructing entropy constraint terms, and a module for optimizing statistical manifolds. The information geometric mutual information determination module is used to calculate information geometric mutual information based on Fisher's information metric. The entropy constraint construction module is used to design entropy constraint terms by introducing the maximum entropy principle; The statistical manifold optimization module optimizes the time offset parameters on the statistical manifold by introducing the natural gradient method, and obtains the optimized time offset parameters. The non-rigid space transformation estimation unit includes a thin plate spline transformation module and a robust estimator solution module; The thin plate spline transformation module is used to solve non-rigid space transformations through thin plate spline interpolation and a robust estimator to obtain spatial correction parameters. The robust estimator solver module is used to process the calculation formula of the spatial correction parameters of the thin plate spline transformation module based on the parameter estimation of the anti-field value, so as to obtain the robust spatial correction parameters. The joint spatiotemporal optimization unit includes a variational posterior distribution module and an ELBO optimization objective module; The variational posterior distribution module is used to determine the variational posterior distribution; The ELBO optimization objective module is used to balance data fitting accuracy and model rationality.

6. The system for fusing multi-source data on bridge structures according to claim 1, characterized in that, The structural health index calculation and status assessment module includes a structural health index calculation unit and a status assessment and early warning unit. The structural health index calculation and status assessment module is used to monitor the structural health and assess the status of bridges based on the fused status vector, and obtain the final health assessment results. The status assessment and early warning unit is used for distributed collaborative fault detection; The structural health index calculation unit includes a multi-level health index extraction module, a physical constraint damage identification module, a fuzzy evidence fusion assessment module, and a comprehensive health index calculation module. The multi-level health indicator extraction module is used for hierarchical calculation of structural health indicators; The physical constraint damage identification module is used to introduce machine learning and physical models to extract damage features; The fuzzy evidence fusion evaluation module is used to fuse multiple fuzzy evaluation results; The comprehensive health index calculation module is used to calculate key structural health indicators; The status assessment and early warning unit includes a federated learning fault diagnosis module, a counterfactual interpretation analysis module, a status prediction and early warning module, and a comprehensive decision support system; The federated learning fault diagnosis module is used to introduce differential privacy protection and secure multi-party computation, and to achieve network-wide fault detection through a distributed collaborative learning strategy; The counterfactual interpretation analysis module provides health assessment results for interpretation; Status prediction and early warning models are used for health status assessment.

7. The system for fusing multi-source data on bridge structures according to claim 6, characterized in that, The multi-level health indicator extraction module includes a structural response indicator extraction layer; The structural response index extraction layer is used to design multidimensional response features based on the principles of bridge structural mechanics and extract structural response indices. The federated learning fault diagnosis module includes a joint confidence assessment unit, which is used to evaluate the accuracy of fault diagnosis using a joint confidence index. The counterfactual explanation and analysis module includes a counterfactual sample generation unit and a feature attribution analysis unit; The counterfactual sample generation unit achieves interpretability of health assessment results through neural symbolic reasoning and counterfactual generation; The feature attribution analysis unit is used to calculate the importance of each monitoring parameter by fusing integral gradients and Shapley values ​​through feature attribution analysis; The state prediction and early warning model includes a physical-stochastic hybrid prediction unit, a multi-level early warning mechanism, and Monte Carlo confidence calculation; The physical-stochastic hybrid prediction unit is used to combine time series analysis and structural degradation models to assess health status; The multi-level early warning mechanism unit is used to implement the early warning mechanism, which adopts a multi-level threshold design. The Monte Carlo confidence calculation unit calculates the early warning confidence level through Monte Carlo simulation.

8. A method for fusing multi-source data of bridge structures in a system for fusing multi-source data of bridge structures as described in any one of claims 1-7, characterized in that, include: Based on the heterogeneous graph structure, multi-source heterogeneous sensor data is processed to obtain a heterogeneous topological graph, and manifold embedding learning is performed on the sensor nodes to obtain node embedding vectors and manifold model parameters; the heterogeneous graph structure includes a set of nodes, a set of edges, a set of hyperedges, a relation type mapping space, and a joint manifold embedding mapping function; Based on node embedding vectors and manifold model parameters, deep feature extraction is performed on heterogeneous topological graphs to obtain multi-scale fused features. Spatiotemporal alignment and fusion processing are performed on the multi-scale fusion features and the original time-series data from the sensors to obtain spatiotemporal alignment parameters and fusion state vectors; Based on the fused state vector, the structural health monitoring and state assessment of the bridge are carried out to obtain the final health assessment results.

Images