Bridge key position disease prediction method and system based on graph convolutional neural network

Abstract

The invention relates to the technical field of bridge structure health monitoring, and discloses a bridge key position disease prediction method and system based on a graph convolutional neural network. According to a bridge design drawing and real-time sensor data, a dynamic topological graph reflecting an actual mechanical state of a bridge is constructed, and a real-time incremental graph convolutional network is designed to realize continuous optimization of a model. The mechanical relationship of the components is quantified into dynamically adjustable edge weights, so that the graph structure can reflect the real state of bridge load transmission in real time. The dynamic representation mode breaks through the limitation of a traditional fixed topological graph, so that the model can capture a disease propagation path more accurately. Meanwhile, an online gradient updating mechanism ensures that model parameters are continuously optimized along with new data, the inherent problem that a traditional offline training mode is difficult to adapt to data distribution drift and data lag is avoided, and prediction precision and timeliness are further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge structure health monitoring, and in particular to a method and system for predicting bridge key position diseases based on a graph convolutional neural network. Background Art

[0002] With the expansion of bridge construction and the increase in service life, the maintenance and monitoring of bridge defects have gradually become a key issue in bridge management. The prediction of bridge structure defects is an important part of ensuring the safe operation of bridges and extending their service life. In particular, the accurate identification and analysis of defects in key locations (such as cracks, corrosion, and stress anomalies) can provide a scientific basis for bridge maintenance and risk management.

[0003] At present, bridge disease prediction mainly relies on structural data collected by sensors and manual inspection results, which are analyzed to identify possible disease risk areas. However, traditional bridge health prediction methods often rely on manual experience and rule models, lacking the ability to deeply analyze complex bridge structure data, resulting in low accuracy in disease prediction at key locations; even if bridge disease prediction is performed through automated methods, it is mostly static analysis, which is difficult to meet the real-time requirements of bridge disease prediction.

[0004] Therefore, there is an urgent need for a bridge key position disease prediction method and system based on graph convolutional neural network, which can predict bridge diseases in real time and accurately. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a bridge key position disease prediction method based on graph convolutional neural network, comprising the following steps: Step S1: Acquire historical sensor data, historical environmental data and historical disease records of bridge components, wherein the historical sensor data includes stress, vibration frequency and displacement; Step S2: dynamically constructing a bridge topology map based on the bridge design drawings and real-time sensor data, wherein the bridge topology map uses bridge components as nodes and mechanical transmission relationships between components as edges; Step S3: fusing the historical sensor data, historical environmental data and historical disease records into node feature vectors, and inputting them into a real-time incremental graph convolutional network for training; Step S4: predicting the disease risk value of each node through the trained real-time incremental graph convolutional network according to the real-time sensor data and the real-time environmental data, and generating a risk list of key locations; Step S5: updating the edge weights of the bridge topology graph according to the real-time sensor data, triggering the online gradient update of the real-time incremental graph convolutional network; Step S6: Continue to predict the disease risk value of each node through the real-time incremental graph convolutional network after online gradient update, and incrementally generate a risk list of key locations.

[0006] Furthermore, in step S2, dynamically constructing a bridge topology map according to the bridge design drawings and real-time sensor data includes: Step S21: obtaining an initial bridge topology map according to the bridge design drawing; Step S22: Calculate the optimized weight of the edge according to the real-time sensor data; Step S23: dynamically adjusting the initial bridge topology map by optimizing the weights to construct a dynamic bridge topology map.

[0007] Furthermore, the calculation formula of the optimization weight is: ; In the formula, Indicates the component at the current time t i To component j The edge weights of represents the weight decay factor, Components representing real-time sensor acquisition i Transfer to component j The load value, Representation component j The design ultimate load value is Represents the edge weight at the previous moment.

[0008] Furthermore, in step S3, the historical sensor data, historical environmental data and historical disease records are fused into a node feature vector, specifically: Taking the material type and design life of the bridge component as static properties; Slicing the historical sensor data into time series features according to time windows; Performing word vector encoding on the temperature and humidity in the historical environmental data and the historical disease records; The static attributes, the temporal features and the word vector encoding are concatenated to generate multi-dimensional node features as the node feature vector.

[0009] Furthermore, in step S3, the real-time incremental graph convolutional network adopts a gated graph attention mechanism to calculate the node feature vector in such a way that: ; In the formula, Indicates l +1 layer node i The node feature vector of represents a nonlinear activation function, Representation Node i The set of neighbor nodes of Representation Node i With Node j The attention coefficient between Indicates l Layer trainable weight matrices, Indicates l Layer Node j The node feature vector of in, Calculated by the following formula: ; In the formula, MLP is a multi-layer perceptron. Representation Node i With Node j The node feature vector and the current moment component i To component j The edge weight The splicing result is represents the normalized exponential function.

[0010] Furthermore, in step S5, the edge weights of the bridge topology graph are updated according to the real-time sensor data to trigger the online gradient update of the real-time incremental graph convolutional network, specifically: Step S51: determining the update range of the edge weights of the bridge topology graph; Step S52: Compare the update range with the update threshold: If the update range is greater than or equal to the update threshold, triggering an online gradient update of the real-time incremental graph convolutional network; If the update range is smaller than the update threshold, the online gradient update of the real-time incremental graph convolutional network is not triggered.

[0011] Furthermore, the online gradient update in step S5 is: using an elastic weight consolidation algorithm to constrain the update amplitude of the model parameters of the real-time incremental graph convolutional network, so that the loss function of the real-time incremental graph convolutional network satisfies the following relationship: ; In the formula, represents the loss function of the real-time incremental graph convolutional network after online gradient update, represents the loss function of the current real-time incremental graph convolutional network, represents a hyperparameter, k represents the index of the trainable parameters of the real-time incremental graph convolutional network, F k represents the Fisher information matrix value of the kth trainable parameter, i krepresents the value of the kth trainable parameter of the real-time incremental graph convolutional network after online gradient update, i k，old Represents the value of the kth trainable parameter of the current real-time incremental graph convolutional network.

[0012] On the other hand, the present invention also provides a bridge key position disease prediction system based on graph convolutional neural network, which is used to execute the above-mentioned bridge key position disease prediction method based on graph convolutional neural network, and the prediction system includes: a historical data acquisition module, a bridge topology map construction module, a model training module, a prediction module, and a model update module; The historical data acquisition module is used to acquire historical sensor data, historical environmental data and historical disease records of bridge components, wherein the historical sensor data includes stress, vibration frequency and displacement; The bridge topology map construction module is used to dynamically construct a bridge topology map based on the bridge design drawings and real-time sensor data. The bridge topology map uses bridge components as nodes and mechanical transmission relationships between components as edges. The model training module is used to fuse the historical sensor data, historical environmental data and historical disease records into node feature vectors, and input them into a real-time incremental graph convolutional network for training; The prediction module is used to predict the disease risk value of each node based on the real-time sensor data and the real-time environmental data through the trained real-time incremental graph convolutional network to generate a risk list of key locations; The model updating module is used to update the edge weights of the bridge topology graph according to the real-time sensor data, triggering the online gradient update of the real-time incremental graph convolutional network.

[0013] The embodiments of the present invention have the following technical effects: When the present invention is based on the bridge key position disease prediction method based on the graph convolutional neural network, a dynamic topological graph reflecting the actual mechanical state of the bridge is constructed according to the bridge design drawings and real-time sensor data, and a real-time incremental graph convolutional network is designed to realize continuous optimization of the model. By quantifying the mechanical relationship of the components into dynamically adjustable edge weights, the graph structure can reflect the actual state of the bridge load transfer in real time. This dynamic characterization method breaks through the limitations of traditional fixed topological graphs, enabling the model to more accurately capture the disease propagation path. At the same time, the online gradient update mechanism ensures that the model parameters are continuously optimized with new data, avoiding the inherent problems of traditional offline training modes that are difficult to adapt to data distribution drift and data lag, and further improving prediction accuracy and timeliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 It is a flow chart of a method for predicting bridge key position defects based on a graph convolutional neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0017] The key positions of a bridge refer to those parts that are crucial to the safety of the overall structure, such as the main beam, piers, bearings, expansion joints, etc. They play a key role in supporting the bridge structure, and their health directly affects the load transfer, structural stability, seismic performance and durability of the bridge. They bear the main load transfer path (such as the connection node between the main beam and the pier), and their failure will cause chain structural damage; they are more significantly affected by environmental erosion, stress concentration and other factors, and are sensitive to diseases. If diseases occur in these positions, they may affect the stability and safety of the entire bridge. However, the field of bridge structure health monitoring has long faced the problem of insufficient disease prediction accuracy and poor real-time performance. Traditional methods rely on static models and manual experience, and cannot effectively capture the dynamic load distribution and structural relationship changes of bridges. The existing technology for identifying diseases mostly stays at the post-analysis level, lacking the foresight of potential risks. For example, fixed topology maps cannot reflect the dynamic adjustment of the mechanical relationship between components with load changes, which makes it difficult for the prediction model to accurately capture the evolution law of diseases at key locations. In addition, conventional graph convolutional networks use an offline training mode and cannot integrate new data in real time, causing damage prediction to lag behind actual structural state changes. The above defects make it difficult for existing systems to provide timely and effective decision support for bridge maintenance.

[0018] Based on this, the present invention provides a bridge key position disease prediction method based on graph convolutional neural network, such as Figure 1 As shown, the following steps are included: Step S1: Obtain historical sensor data, historical environmental data and historical disease records of bridge components, where the historical sensor data includes stress, vibration frequency and displacement; wherein stress is the force per unit area at a key position of the bridge, vibration frequency is the frequency of bridge vibration, and displacement is the offset of the deformation of the bridge structure.

[0019] After obtaining the data, the data may also be preprocessed. The preprocessing steps include: Missing value filling: use mean filling or linear interpolation method for missing data; Outlier removal: The outliers in the data were removed by triple standard deviation method.

[0020] Step S2: Based on the bridge design drawings and real-time sensor data, a bridge topology map is dynamically constructed. The bridge topology map uses bridge components as nodes and the mechanical transfer relationships between components as edges. Bridge components such as beams, columns, and bridge decks are represented as nodes in the bridge topology map, and the physical relationships between components (such as load transfer relationships) are represented as edges.

[0021] In some implementations, in step S2, dynamically constructing a bridge topology map based on the bridge design drawings and real-time sensor data includes: Step S21: obtaining an initial bridge topology map according to the bridge design drawing; Step S22: Calculate the optimized weight of the edge according to the real-time sensor data; Bridge design drawings only provide theoretical topology of component connection relationships, such as the theoretical connection between piers and main beams, but actual load transfer may shift due to construction errors or aging. The initial time t is dynamically calculated through real-time sensor data (such as stress distribution on the contact surface between piers and main beams). 0 The initial edge weight , modify the edge weights obtained for the static bridge design graph: ;(Formula 1) Initial moment component i Transfer to component j The load value, Design load values ​​for static bridge design drawings.

[0022] When a new component (such as a reinforced steel plate) is detected, the edge weights between it and the adjacent components are initialized according to the real-time sensor data (such as the vibration frequency of the reinforcement) to ensure that the topology is consistent with the actual structure.

[0023] In some implementations, when the sensor detects an abnormal load (such as a sudden increase in stress on the main beam caused by the passage of an overloaded vehicle), in response to the emergency, the weights of the relevant edges are dynamically updated to obtain the optimized weights of the edges, and the calculation formula is: ;(Formula 2) In the formula, Indicates the component at the current time t i To component j The edge weights of Represents the weight decay factor, which is used to control the balance between historical weight and current data. Components representing real-time sensor acquisition i Transfer to component j The load value, Representation component j The design ultimate load value is Represents the edge weight at the previous moment.

[0024] After the edge weights are updated, the gradient update of the real-time incremental graph convolutional network (RT-IGCN) is triggered (such as the EWC algorithm mentioned later), allowing the model to quickly adapt to the new mechanical state.

[0025] Step S23: dynamically adjust the initial bridge topology map by optimizing the weights to construct a dynamic bridge topology map.

[0026] The primary task of predicting defects at key locations of bridges is to accurately establish a mechanical correlation model between components. When constructing the initial bridge topology based on the bridge design drawings, each component is abstracted as a graph node, and the design load transfer path between components forms the initial edge connection. In the stage of dynamically constructing the bridge topology, real-time sensor data plays a key role. By calculating the dynamic ratio between the actual load transfer value and the design limit value between components, the edge weight is corrected online to obtain the optimized weight of the edge. In the specific implementation, the load sensor collects the force data of each connection point in real time, and normalizes the actual load value transferred from component i to component j with the design limit load of the connection point. By introducing a weight attenuation factor to balance the relationship between the historical state and the current load, an edge weight update formula with time memory characteristics is formed. This dynamic weight adjustment mechanism enables the bridge topology to accurately characterize the actual working state of the bridge, provide accurate neighborhood relationship input for subsequent graph convolution operations, and improve prediction accuracy.

[0027] Step S3: Fuse historical sensor data, historical environmental data, and historical disease records into node feature vectors and input them into the real-time incremental graph convolutional network for training. The task of this step is to fuse node features and input them into the real-time incremental graph convolutional network (RT-IGCN) for training. It is responsible for the initial training and feature processing of the model and belongs to the preparation stage before model deployment.

[0028] At this point, the input parameters of the real-time incremental graph convolutional network are constructed, including the bridge topology (representing the adjacency matrix) and the node feature vector. The bridge topology constructed in step S2 provides the graph structure, namely the adjacency matrix and edge weight information. The node feature vector in step S3 is the attribute data of each node.

[0029] The network structure of RT-IGCN (such as the gated graph attention layer and node feature calculation rules) is established for the first time in this step, completing the model architecture definition and completing the training based on the initial data.

[0030] In some implementations, in step S3, historical sensor data, historical environmental data, and historical disease records are fused into a node feature vector, specifically: The material type and design life of bridge components are used as static properties to reflect their inherent resistance to disease. For example, the material type is concrete and steel, and the design life is 50 years.

[0031] The historical sensor data is sliced ​​into time series features according to the time window, and the time series change rules of parameters such as stress and vibration frequency are extracted. For example, the sliding time window of the mean / variance of the stress value is 1 hour, and the time domain spectrum characteristics of the vibration frequency are extracted by FFT, and the coefficient is fitted according to the trend of the displacement change rate.

[0032] The temperature and humidity in the historical environmental data and the historical disease records are encoded by word vectors. For example, semantic encoding can be performed through natural language processing technology to convert them into quantifiable feature vectors. For example, the word vector encoding of temperature and humidity data is to map "high temperature and high humidity" to [0.7, -0.3]); "cracks" in the historical disease type are encoded as 128-dimensional vectors; the average annual crack expansion rate is used as a feature to count the frequency of disease occurrence.

[0033] Static attributes, temporal features and word vector encoding are concatenated to generate multi-dimensional node features as node feature vectors.

[0034] The physical sensor data, structural attributes, and environmental semantics are uniformly encoded, and multimodal features are fused to improve the model's ability to analyze the association of complex disaster-causing factors; through dynamic semantic coding, the words of historical disease records are vectorized, so that the model can learn the evolution patterns of diseases such as "cracks, rust, and peeling". The splicing of multidimensional features not only retains the independent information characteristics of various types of data, but also establishes cross-modal associations through feature space mapping to form a feature expression that comprehensively describes the health status of the component. This feature fusion method effectively solves the problem of missing feature information caused by traditional single-modal data analysis.

[0035] For example, in the scenario of bridge pier concrete carbonization prediction: Static properties: material type (concrete), design life (40 years), used to provide a benchmark for carbonation rate; Time series characteristics: Increased variance of stress fluctuations, indicating microcrack growth; Semantic coding: "Carbonization depth is 0.5 mm per year" in historical records is encoded as a trend vector; Fusion results: Static attributes, time series features and word vector encoding are spliced ​​to generate multi-dimensional node features as node feature vectors. The trained real-time incremental graph convolutional network is used to predict the carbonization area of ​​piers. Compared with the prediction results that only use sensor time series data (single modal input) without integrating cross-domain features such as material type and disease semantics, the error is reduced by 32%.

[0036] In some implementations, in step S3, the real-time incremental graph convolutional network uses a gated graph attention mechanism to calculate the node feature vector to satisfy: ; In the formula, Indicates l +1 layer node i The node feature vector of represents a nonlinear activation function, Representation Node i The set of neighbor nodes of Representation Node i With Node j The attention coefficient between Indicates l Layer trainable weight matrices, Indicates l Layer Node j The node feature vector of in, Calculated by the following formula: ; In the formula, MLP is a multi-layer perceptron. Representation Node i With Node j The node feature vector and the current moment component i To component j The edge weight The splicing result is represents the normalized exponential function.

[0037] In traditional graph convolution (GCN), the aggregation weight of neighbor nodes is only determined by the adjacency matrix (0 or 1), and the importance of different neighbors cannot be distinguished. This paper introduces a gated attention mechanism to Similarity with node features ( h i(l) and h j (l) ) Dynamically calculate the attention coefficient , dynamically assign neighbor weights to focus the model on critical paths, such as high stress transfer paths. For example, higher weights are given to neighbor nodes of high load transfer paths (such as pier-girder connections) to improve the model's prediction sensitivity to critical locations.

[0038] Introducing edge weights , combining the changes in mechanical state (such as sudden increase in bridge load) with feature propagation, enhancing the model's ability to respond to emergencies, making the model adapt to dynamic changes in bridge status (such as traffic flow fluctuations), and improving prediction accuracy.

[0039] In summary, the real-time incremental graph convolutional network uses a gated graph attention mechanism to achieve dynamic feature aggregation. In the process of network propagation at each layer, the attention weight of the node feature vector and the feature vector of its neighboring nodes is calculated. This weight not only considers the feature similarity between nodes, but also introduces the real-time edge weight as an important adjustment factor. The feature interaction strength reflecting the current mechanical state is obtained by jointly encoding the node features and edge weights through a multi-layer perceptron. This enables the real-time incremental graph convolutional network to dynamically adjust the importance of each connection when aggregating neighborhood information, and strengthen the feature learning of key load transfer paths. The gating mechanism further controls the scope of information propagation, avoids noise interference from irrelevant nodes, and improves the model's ability to capture local disease features.

[0040] Step S4: According to the real-time sensor data and real-time environmental data, the disease risk value of each node is predicted through the trained real-time incremental graph convolutional network to generate a risk list of key locations; Step S5: Update the edge weights of the bridge topology graph according to the real-time sensor data, triggering the online gradient update of the real-time incremental graph convolutional network; In some implementations, in step S5, the edge weights of the bridge topology graph are updated according to the real-time sensor data, triggering the online gradient update of the real-time incremental graph convolutional network, specifically: Step S51: determining the update range of the edge weights of the bridge topology graph; Step S52: Compare the update range with the update threshold: If the update range is greater than or equal to the update threshold, the online gradient update of the real-time incremental graph convolutional network is triggered; If the update range is smaller than the update threshold, the online gradient update of the real-time incremental graph convolutional network will not be triggered.

[0041] The purpose of online gradient update is to respond to changes in a timely manner. If each edge weight update triggers a parameter update, it may lead to a shortage of computing resources. In the present invention, an update threshold is set to balance the response speed and computing efficiency to meet the rapid response requirements of bridge prediction.

[0042] After the trained real-time incremental graph convolutional network is deployed, the edge weights are adjusted according to the real-time data during the model operation, and the model parameters are fine-tuned through online gradient updates (such as the EWC algorithm), but the model architecture is not changed.

[0043] When the edge weights of the bridge topology are updated, the real-time incremental graph convolutional network will use the new edge weights to affect the feature representation of the nodes, and then affect the output prediction results. The load of the bridge will fluctuate due to the traffic flow. Sensors are used to perceive the changes in the bridge in real time. By calculating the dynamic ratio of the actual load transfer value between components and the design limit value, the edge weights are corrected online to obtain the optimized weights of the edges. When the update range of the edge weights of the bridge topology is greater than or equal to the update threshold, the online gradient update of the real-time incremental graph convolutional network is triggered. The EWC algorithm is used to constrain the changes in the model parameters of the real-time incremental graph convolutional network to ensure that after the edge weights are updated, it will not forget the previously learned knowledge, further improving the real-time accuracy of the prediction results.

[0044] In some implementations, the online gradient update in step S5 is: using an elastic weight consolidation algorithm to constrain the update amplitude of the model parameters of the real-time incremental graph convolutional network, so that the loss function of the real-time incremental graph convolutional network satisfies the following relationship: ; In the formula, represents the loss function of the real-time incremental graph convolutional network after online gradient update, represents the loss function of the current real-time incremental graph convolutional network, represents a hyperparameter, k represents the index of the trainable parameters of the real-time incremental graph convolutional network, F k represents the Fisher information matrix value of the kth trainable parameter, i k represents the value of the kth trainable parameter of the real-time incremental graph convolutional network after online gradient update, i k，old Represents the value of the kth trainable parameter of the current real-time incremental graph convolutional network. The second term on the right side of the equal sign forces the model to retain historical knowledge (such as the learned crack prediction rules) when optimizing new tasks to avoid conflicts between new and old tasks, so that it can retain historical prediction results and be able to respond to large changes in time to optimize model parameters.

[0045] Edge Weight The update changes the adjacency matrix of the topological graph, causing the distribution of the model input data to change, which in turn affects the predicted value and loss function. , forming an implicit transmission path The edge weight update drives the input data change, which in turn triggers the model parameter update; the EWC algorithm ensures that historical knowledge is retained when the parameters are updated. The two form a closed loop to adapt to the dynamic environment caused by factors such as large stress changes caused by different vehicles on the bridge, and jointly improve the prediction robustness of the real-time incremental graph convolutional network in a dynamic environment.

[0046] In the process of incremental learning (online update), directly updating all parameters of the real-time incremental graph convolutional network will cause the model to forget historical tasks (catastrophic forgetting). Therefore, through the elastic weight consolidation (EWC) algorithm, the importance of each parameter in the real-time incremental graph convolutional network is quantified using the Fisher information matrix, the update amplitude of important parameters (such as the weight of the bridge pier node) is constrained, the historical parameters are protected, and the prediction failure caused by excessive parameter update is avoided.

[0047] After step S2, the dynamic construction of the bridge topology map is completed, and the real-time incremental graph convolution network is trained through step S3. In step S4, the disease risk value of each node is predicted through the trained real-time incremental graph convolution network according to the real-time sensor data and real-time environmental data, and a risk list of key locations is generated. In step S5, if the update range is greater than or equal to the update threshold, the online gradient update of the real-time incremental graph convolution network is triggered, but the real-time incremental graph convolution network triggered by the online gradient update is not used, and the data of step S4 is immediately re-predicted. Instead, it is necessary to wait for the next data trigger (such as the sensor data input of the new time window) to form a new prediction result, such as the prediction of step S4, to generate a new prediction result, and superimpose and update the risk list of key locations. A full-cycle closed loop from structural modeling to state tracking is formed, which overcomes the dual defects of rigid topological structure and lagging model update in the prior art. It effectively solves the catastrophic forgetting problem in traditional incremental learning, enables the model to continuously accumulate knowledge of bridge state evolution, and improves the stability of long-term prediction.

[0048] The key position risk list is a display of the predicted results of the key position defects of the bridge. This list not only contains the basic information of the traditional inspection report, but also integrates the real-time mechanical state analysis and defect evolution prediction. The specific composition is as follows: (1) Component topology identification Each bridge component is assigned a unique topological code and marked with spatial coordinates in the 3D BIM model. The coding system includes the component type (such as main beam T-01, bridge pier P-03), material code (C50 concrete / Mn steel), and hierarchical position in the design topology diagram (such as main load-bearing layer L1, auxiliary structure layer L2).

[0049] (2) Dynamic risk scoring matrix For example, a 0-1 standardized scoring system is used, which includes three dimensions: Instant risk value: the probability of the disease occurring at the current moment (e.g. crack risk 0.68); Trend change rate: the risk growth rate in the next N hours based on time series prediction (e.g. +12% / h); Environmental sensitivity: the weighted impact coefficient of environmental factors such as temperature, humidity, and wind speed on the current risk (e.g., humidity weight is 0.43); (3) Disease type atlas Specific disease forms are identified through multi-label classification technology, such as: Stress-type diseases: fatigue cracks, concrete spalling; Displacement type anomaly: bearing slip and beam end displacement exceed the standard; Environmental erosion: steel bar corrosion rate, chloride ion penetration depth; Each disease type is labeled with the confidence level (e.g., 92% confidence level for rust detection) and the associated sensor number (e.g., data from corrosion sensor S-205 is abnormal).

[0050] (4) Mechanical correlation network Visualize the role of the target component in the overall load transfer with a graphical structure: Upstream influence source: Indicate the load input path that causes the current risk (e.g., the vehicle load on the bridge deck is transmitted to the component through edge E-12) Downstream impact: predict the chain reaction components that may be triggered when the component fails (e.g., failure of P-03 will affect T-07 / T-09) Edge weight change curve: displays the real-time load transfer ratio of key connection edges and compares it with historical trends.

[0051] (5) Maintaining the decision tree That is, the disposal suggestions generated based on the knowledge graph include: Priority level: The degree of urgency calculated based on the risk value and environmental sensitivity (e.g. Level I must be handled within 2 hours).

[0052] Treatment method library: Recommended solutions matching historical maintenance cases (such as epoxy resin injection when crack width is greater than 2mm).

[0053] Resource scheduling tips: associated with the nearest maintenance team location and required material inventory status.

[0054] For example, the results of the critical location risk list can also be mapped onto the 3D bridge model, including: (1) Generate a 3D visualization scene using the bridge model and decompose the model into multiple sub-components; (2) Bridge components are color-coded according to their damage risk levels, with high-risk areas marked in red, medium-risk areas marked in orange, and low-risk areas marked in green. When the predicted risk value exceeds the preset threshold, an alarm is triggered and an emergency damage report is generated.

[0055] (3) Supports interactive operations between users and 3D models, including zooming in, zooming out, rotating, and clicking to query the location of the disease.

[0056] In addition, the three-dimensional display interface also provides a timeline function to trace back the changes in the disease and help users understand the evolution of the disease.

[0057] The 3D bridge visualization system generates a summary report on the health status of the bridge, which includes the overall health score of the bridge, a list of disease locations, a detailed description of key risk nodes, and screenshots of the 3D model. The health score is calculated based on the disease risk value and weight of each node, reflecting the overall operating status of the bridge. The list of disease locations is sorted by risk level, and details the number, disease type, risk probability and impact range of each key disease location. The screenshot of the 3D model can show the distribution of bridge diseases in an intuitive way. The generated summary report supports export in multiple formats, including PDF, Word and HTML, which is convenient for users to store, share and print. The report is automatically stored in the cloud server, and users can view or download it at any time through the system interface.

[0058] On the other hand, the present invention also provides a bridge key position disease prediction system based on graph convolutional neural network, which is used to execute the above-mentioned bridge key position disease prediction method based on graph convolutional neural network, and the prediction system includes: a historical data acquisition module, a bridge topology map construction module, a model training module, a prediction module, and a model update module; A historical data acquisition module is used to obtain historical sensor data, historical environmental data and historical disease records of bridge components. The historical sensor data includes stress, vibration frequency and displacement. The bridge topology construction module is used to dynamically construct a bridge topology based on the bridge design drawings and real-time sensor data. The bridge topology uses bridge components as nodes and the mechanical transmission relationships between components as edges. Model training module, which is used to fuse historical sensor data, historical environmental data and historical disease records into node feature vectors and input them into real-time incremental graph convolutional networks for training; The prediction module is used to predict the disease risk value of each node based on real-time sensor data and real-time environmental data through the trained real-time incremental graph convolutional network to generate a risk list of key locations; The model update module is used to update the edge weights of the bridge topology graph according to the real-time sensor data, triggering the online gradient update of the real-time incremental graph convolutional network.

[0059] Exemplarily, a user feedback module may also be included to support users to supplement or correct disease risk prediction results. The feedback information is used as new training data to optimize model parameters and improve the accuracy and robustness of the prediction.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A bridge key position disease prediction method based on graph convolutional neural network, characterized in that: The following steps are involved: Step S1: Acquire historical sensor data, historical environmental data and historical disease records of bridge components, wherein the historical sensor data includes stress, vibration frequency and displacement; Step S2: dynamically constructing a bridge topology map based on the bridge design drawings and real-time sensor data, wherein the bridge topology map uses bridge components as nodes and mechanical transmission relationships between components as edges; Step S3: fusing the historical sensor data, historical environmental data and historical disease records into node feature vectors, and inputting them into a real-time incremental graph convolutional network for training; Step S4: predicting the disease risk value of each node through the trained real-time incremental graph convolutional network according to the real-time sensor data and the real-time environmental data, and generating a risk list of key locations; Step S5: updating the edge weights of the bridge topology graph according to the real-time sensor data, triggering the online gradient update of the real-time incremental graph convolutional network; Step S6: Continue to predict the disease risk value of each node through the real-time incremental graph convolutional network after online gradient update, and incrementally generate a risk list of key locations.

2. The bridge key position disease prediction method based on graph convolutional neural network according to claim 1 is characterized in that: In step S2, dynamically constructing a bridge topology map based on the bridge design drawings and real-time sensor data includes: Step S21: obtaining an initial bridge topology map according to the bridge design drawing; Step S22: Calculate the optimized weight of the edge according to the real-time sensor data; Step S23: dynamically adjusting the initial bridge topology map by optimizing the weights to construct a dynamic bridge topology map.

3. The bridge key position disease prediction method based on graph convolutional neural network according to claim 2 is characterized in that: The calculation formula of the optimization weight is: ； In the formula, Represents the current component at time t i To component j The edge weights of represents the weight decay factor, Components representing real-time sensor acquisition i Transfer to component j The load value, Representation component j The design ultimate load value is Represents the edge weight at the previous moment.

4. The bridge key position disease prediction method based on graph convolutional neural network according to claim 1 is characterized in that: In step S3, the historical sensor data, historical environmental data and historical disease records are fused into a node feature vector, specifically: Taking the material type and design life of the bridge component as static properties; Slicing the historical sensor data into time series features according to time windows; Performing word vector encoding on the temperature and humidity in the historical environmental data and the historical disease records; The static attributes, the temporal features and the word vector encoding are concatenated to generate multi-dimensional node features as the node feature vector.

5. The bridge key position disease prediction method based on graph convolutional neural network according to claim 1 is characterized in that: In step S3, the real-time incremental graph convolutional network adopts a gated graph attention mechanism to calculate the node feature vector to meet the following conditions: ； In the formula, Indicates l +1 layer node i The node feature vector of represents a nonlinear activation function, Representation Node i The set of neighbor nodes of Representation Node i With Node j The attention coefficient between Indicates l Layer trainable weight matrices, Indicates l Layer Node j The node feature vector of in, Calculated by the following formula: ； In the formula, MLP is a multi-layer perceptron. Representation Node i With Node j The node feature vector and the current moment component i To component j The edge weight The splicing result is represents the normalized exponential function.

6. The bridge key position disease prediction method based on graph convolutional neural network according to claim 1 is characterized in that: In step S5, the edge weights of the bridge topology graph are updated according to the real-time sensor data to trigger the online gradient update of the real-time incremental graph convolutional network, specifically: Step S51: determining the update range of the edge weights of the bridge topology graph; Step S52: Compare the update range with the update threshold: If the update range is greater than or equal to the update threshold, triggering an online gradient update of the real-time incremental graph convolutional network; If the update range is smaller than the update threshold, the online gradient update of the real-time incremental graph convolutional network is not triggered.

7. The bridge key position disease prediction method based on graph convolutional neural network according to claim 1 is characterized in that: The online gradient update in step S5 is: using the elastic weight consolidation algorithm to constrain the update amplitude of the model parameters of the real-time incremental graph convolutional network, so that the loss function of the real-time incremental graph convolutional network satisfies the following relationship: ； In the formula, represents the loss function of the real-time incremental graph convolutional network after online gradient update, represents the loss function of the current real-time incremental graph convolutional network, represents a hyperparameter, k represents the index of the trainable parameters of the real-time incremental graph convolutional network, F k represents the Fisher information matrix value of the kth trainable parameter, θ k represents the value of the kth trainable parameter of the real-time incremental graph convolutional network after online gradient update, θ k，old Represents the value of the kth trainable parameter of the current real-time incremental graph convolutional network.

8. A bridge key position disease prediction system based on graph convolutional neural network, used to execute the bridge key position disease prediction method based on graph convolutional neural network according to any one of claims 1 to 7, characterized in that: The prediction system includes: a historical data acquisition module, a bridge topology map construction module, a model training module, a prediction module, and a model update module; The historical data acquisition module is used to acquire historical sensor data, historical environmental data and historical disease records of bridge components, wherein the historical sensor data includes stress, vibration frequency and displacement; The bridge topology map construction module is used to dynamically construct a bridge topology map based on the bridge design drawings and real-time sensor data. The bridge topology map uses bridge components as nodes and mechanical transmission relationships between components as edges. The model training module is used to fuse the historical sensor data, historical environmental data and historical disease records into node feature vectors, and input them into a real-time incremental graph convolutional network for training; The prediction module is used to predict the disease risk value of each node based on the real-time sensor data and the real-time environmental data through the trained real-time incremental graph convolutional network to generate a risk list of key locations; The model updating module is used to update the edge weights of the bridge topology graph according to the real-time sensor data, triggering the online gradient update of the real-time incremental graph convolutional network.

Images