Bridge structure service state space correlation diagnosis method based on graph network and state space model

Abstract

The application proposes a bridge structure service state space-time correlation diagnosis method based on a graph network and a state space model. The method adopts a space-time graph structure, constructs a space-time graph fusion learning module, combines a graph neural network and a time series data deep learning, simultaneously considers data features of bridge structure component service states in space and time two dimensions, more accurately models and predicts dynamic space-time evolution of a single bridge complex component system. The application adopts a state space model, constructs a structured state space sequence, and introduces a selective mechanism to dynamically adjust model behavior and realize optimized utilization of computing resources. Through space-time feature decomposition and fusion of different complexity bridge multi-type component service state monitoring data, the application better adapts to processing requirements of multi-type, multi-scale and multi-granularity bridge structure component service state space-time data, and improves method universality and prediction efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical fields of bridge engineering, structural health monitoring, spatio-temporal data analysis and deep learning, in particular to a bridge structure service state spatio-temporal correlation diagnosis method based on a graph network and a state space model. The method can be directly applied to the fields of intelligent bridge, bridge intelligent operation and maintenance, smart infrastructure, intelligent disaster prevention and mitigation, etc. BACKGROUND

[0002] As a key component of urban transportation system, the stability of the service state of a bridge is directly related to the smoothness and safety of urban transportation. As a complex system composed of multiple types of structural components, a bridge is inevitably affected by the coupling of various factors such as natural environment, traffic load, material aging, etc. during long-term service, and the structural performance will continuously degrade, and even the extreme situation of overall bridge collapse safety accident due to local component failure may occur. Therefore, it is of great significance to accurately and efficiently predict the service state of a bridge structure to ensure the safety of urban transportation and improve the efficiency of bridge operation and maintenance management.

[0003] Traditional bridge service state diagnosis methods mainly rely on the prediction of the service state of local components of a single bridge, and the research objects are often limited to local components such as bridge towers, main beams and cables, without utilizing the shared characteristics of multiple types of responses of bridge structures, ignoring the spatio-temporal correlation of similarity and damage evolution between different components of a bridge, resulting in that the correlation and synergy effect between components of a single bridge are not fully explored.

[0004] At present, deep learning technology has developed rapidly, and domestic and foreign research has carried out research on predicting the service state of a bridge using deep learning methods. However, the current research mainly focuses on:

[0005] (1) Bridge single component state diagnosis based on deep learning: for example, damage location and severity assessment of a cable-stayed bridge based on a message passing neural network, using cable force data of multiple cables as input, and outputting the position and cross-sectional area of the damaged cable, only considering the spatial correlation of different cables on a single bridge;

[0006] (2) For key component structure state diagnosis, due to the mechanical correlation and deformation coordination characteristics between components of a single bridge, the spatio-temporal correlation between multiple types of responses of a single bridge is crucial to state diagnosis, but the current research has not involved;

[0007] (3) In addition, existing research often only targets a single type of data, such as acceleration, cable force, vertical displacement, etc., and has not established a general deep learning method for spatio-temporal correlation prediction of multiple types, multiple scales and multiple granularities of bridge service state spatio-temporal data (including structural dynamic response, static response and quasi-static response, etc.).

[0008] In recent years, spatio-temporal graph structure and state space model provide a new idea and advanced algorithm for studying spatio-temporal correlation diagnosis, and the spatio-temporal graph neural network shows a broad application prospect in intelligent diagnosis of the service state of bridge structure. However, the current research still faces great challenges. The problems to be solved urgently include: how to construct a suitable spatio-temporal graph structure to accurately reflect the complex spatio-temporal correlation relationship between different components of a single bridge, how to design an effective spatio-temporal graph neural network model to capture the spatio-temporal dependence relationship of the service state evolution of the bridge structure component, and how to fully utilize the spatio-temporal data of the bridge multi-type component state monitoring to improve the prediction accuracy and efficiency. SUMMARY

[0009] The purpose of the present application is to solve the problems in the prior art, and a bridge structure service state spatio-temporal correlation diagnosis method based on graph network and state space model is proposed.

[0010] The present application is realized by the following technical solutions, and the bridge structure service state spatio-temporal correlation diagnosis method based on graph network and state space model is proposed, which comprises the following steps:

[0011] Step one: determine the bridge structure and multi-type component object to be studied, and establish a single bridge graph structure model considering the spatio-temporal correlation of different component responses;

[0012] Step two: design a spatio-temporal graph prediction task based on the multi-type sensing data of the bridge structure component service state, and construct the corresponding input and output data sets;

[0013] Step three: establish a bridge structure service state spatio-temporal correlation diagnosis model based on graph network and state space model, and the graph selective state space module of the bridge structure service state spatio-temporal correlation diagnosis model comprises a normalization layer, a dynamic filtering graph convolution layer, a splicing layer, a linear layer, a segmentation layer, a one-dimensional convolution layer, a SiLU activation layer, a graph state space selection mechanism, an element-wise multiplication layer and a residual connection layer, wherein the dynamic filtering graph convolution layer learns the complex spatial dependence relationship of the bridge structure multi-type component service state sequence data, and the graph state space selection mechanism learns the evolution and spatio-temporal correlation of the bridge multi-type component service state;

[0014] Step four: training and prediction of the bridge structure service state spatio-temporal correlation diagnosis model; the bridge structure service state spatio-temporal correlation diagnosis model is trained based on gradient descent error back propagation, and the trained model is used for spatio-temporal correlation prediction and diagnosis of the service state of the key components including quasi-static response and dynamic response.

[0015] Further, the step one specifically comprises:

[0016] Step one: Bridge multi-type component service state multi-source sensing data collection and preprocessing;

[0017] (1) Bridge multi-type component service state multi-source sensing data collection: collect the key component service state historical data of quasi-static response and dynamic response;

[0018] (2) Bridge multi-type component service state multi-source sensing data preprocessing: clean the collected data, remove outliers and missing values, normalize multi-type data, and ensure data consistency and accuracy;

[0019] Step two: Establish a single bridge graph structure model considering the spatio-temporal correlation of different components;

[0020] (1) Single bridge graph structure node definition: each key component structure on the single bridge is regarded as a node in the graph, and the node attribute is quasi-static response and dynamic response;

[0021] (2) Single bridge graph structure edge definition: according to the spatial connection relationship between the multi-type components of the bridge structure, the edges between the nodes are defined, and the edges are undirected edges;

[0022] (3) Single bridge graph structure generation: the single bridge graph structure is represented by G=(V;E), where V is the set of component nodes, and E is the set of edges; according to the definition of nodes and edges, the single bridge graph structure is generated; according to the definition of nodes and edges, the adjacency matrix A of the single bridge graph structure is determined.

[0023] Further, the step two specifically comprises:

[0024] Step two one: for bridge structure service state diagnosis task, taking the historical monitoring data of quasi-static response and dynamic response as input, predicting the future bridge multi-type component quasi-static response and dynamic response time course as output;

[0025] Step two two: the input data set is embedded as Where N is the number of key components of the single bridge, T is the time step of the input data, and D is the feature dimension, including the quasi-static response and dynamic response of the structure; the corresponding output data set is embedded as Where T p is the predicted output data time step;

[0026] Step two three: for multi-type, multi-scale and multi-granularity bridge structure multi-type component service state spatio-temporal data, including quasi-static response and dynamic response of the structure, three different time granularities are used as input data, including recent data X r , periodic data X c and trend data X qThe recent data is a historical time series directly adjacent to the prediction period, the periodic data is composed of segments with an integer multiple of a of the same time period as the prediction period, and the trend data is composed of segments with an integer multiple of β of the same time period as the prediction period.

[0027] Further, the establishment of the bridge structure service state space correlation diagnosis model specifically includes:

[0028] Step three one: the three time granularity data X r , X c and X q of the bridge structure multi-type component service state are weighted and fused after being output by N space-time graph convolution blocks;

[0029] Step three two: the output of the M graph selective state space module passed to the input X r is weighted and fused again;

[0030] Step three three: the final output result Y is obtained after the full connection layer;

[0031] Further, the dynamic filtering graph convolution layer specifically refers to learning the multi-hop connection relationship in the graph and the dynamic change of the graph structure information; the calculation method is that the recent data X r of the bridge structure multi-type component service state is transformed by layer normalization and then passed to the dynamic filtering graph convolution layer:

[0032]

[0033] In the formula, A is the original adjacency matrix of the single bridge graph structure, D is the degree matrix, A norm is the normalized adjacency matrix, represents the power series of the adjacency matrix, B is the basic filter for dynamically adjusting the adjacency matrix of the single bridge graph structure, T(B) represents the linear transformation of the basic filter B matrix, λ is the weight coefficient between the original adjacency matrix and the transformed filter artificially assigned, is the dynamically adjusted adjacency matrix, are the weight and bias respectively, and B are to be optimized parameters, is the result of the i-th dynamic filtering graph convolution layer, H cat is the splicing result; Each element in and B is initialized by randomly taking values from a uniform distribution with a mean of 0 and a range of m is a hyperparameter, and N is the number of components included in the single bridge structure.

[0034] Further, the graph state space selection mechanism is specifically: after the dynamic filtering graph convolution layer, for learning the service state evolution and spatio-temporal correlation of the bridge multi-type component, the specific method is: the result H cat Output after linear layer core_sec Split into core sequence H core And secondary sequence H sec The two sequences have the same dimension, and the core sequence is transmitted to the graph state space selection mechanism GSSSM after one-dimensional convolution and SiLU activation, to obtain the output H gsssm The calculation formula is as follows:

[0035]

[0036] The graph state space selection mechanism GSSSM includes a state space selection mechanism parameter calculation update module and a graph selective scanning module. The output H gsssm And the secondary sequence after SiLU activation are multiplied element by element, and then H out :

[0037] H out = W out (H gsssm ⊙SiLU(H sec ))+b out (3)

[0038] In the formula, W out And b out Respectively represent the weight and bias parameters to be optimized; H out And the layer normalized X r Residual connection is obtained, and the output of the graph selective state space module of the bridge structure service state spatio-temporal correlation diagnosis model is obtained.

[0039] Further, the state space selection mechanism parameter calculation update module of the graph state space selection mechanism GSSSM has the following specific steps:

[0040] (1) Calculate the input-independent parameters A and D, A log Each row of the matrix is initialized as the logarithm of the continuous integer sequence from 1 to the column number, and the elements in D are initialized as 1. A is a state transition matrix, and D is a direct gain;

[0041] (2) Process the input H c ′ ore Using a linear layer to obtain H c ″ ore ;

[0042] (3) Split H c ″ oreThe parameters Δ, B, and C related to the service condition monitoring data input of multi-type components of bridge structures are obtained. The dimensions of B and C are the same, B and C are the input matrix and output matrix respectively, and Δ represents the conversion parameter from the continuous state space model to the discrete state space model.

[0043] (4) Use the linear layer and SoftPlus activation function to adjust Δ to obtain Δ′;

[0044] (5) Call the graph selective scanning module GraphSelectiveScan and pass in H c ' ore ,Δ′,A,B,C,D perform update calculation and get the output H gsssm ;

[0045] The parameter calculation and update formula of the state space selection mechanism GSSSM is:

[0046]

[0047] Furthermore, the graph selective scanning module GraphSelectiveScan of the graph state space selection mechanism GSSSM has the following specific steps:

[0048] (1) Obtaining the dynamically adjusted monomer bridge graph structure adjacency matrix from the dynamic filter graph convolution layer Initialize the filling matrix and convert the adjacency matrix Copy to the upper left corner of the filled matrix to get P pad_adj Filling matrix, P pad_adj Perform matrix multiplication matmul on the filling matrix and Δ′ to obtain the updated parameter Δ″;

[0049] (2) Use Δ″ to discretize and update A and B to obtain ΔA and ΔB, where einsum represents the Einstein summation convention;

[0050] (3) Initialize the state Z in the state space selection mechanism to 0, and perform state Z from 1 to T for the time step i in the service status monitoring data of multiple types of components of the bridge structure. i Update and output H i Calculation of

[0051] (4) Output H i Use stack operation to get H stack , and finally H c ' ore The weight coefficient is superimposed on the direct gain D to obtain H gsssm ;

[0052] The graph selective scanning calculation formula of the graph state space selection mechanism GSSSM is:

[0053]

[0054] The application further provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the bridge structure service state space correlation diagnosis method based on a graph network and a state space model when executing the computer program.

[0055] The application further provides a computer readable storage medium for storing computer instructions, wherein the computer instructions implement the steps of the bridge structure service state space correlation diagnosis method based on a graph network and a state space model when executed by a processor.

[0056] Compared with the prior art, the application has the beneficial effects that:

[0057] (1) By embedding different time granularity perception data (recent data, periodic data and trend data), the model can capture the dynamic change rule of multi-type, multi-scale and multi-granularity bridge structure multi-type component service state data on different time scales, thereby improving the universality, accuracy and robustness of the model.

[0058] (2) The dynamic filtering graph convolution network module is adopted to improve the processing capacity of the bridge structure multi-type component service state perception space-time data including noise and uncertainty.

[0059] (3) By designing the segmentation operation and the residual connection, the features are more effectively separated and fused, and more rich space-time information levels in the bridge structure multi-type component service state monitoring data are captured, thereby improving the representation ability and prediction performance of the space-time correlation prediction model.

[0060] (4) The dynamic parameter adjustment mechanism and the efficient selective mechanism are designed in the space-time graph selective state space module, and by dynamically adjusting the state space model parameters (including the state transition matrix, the control input matrix and the output matrix), important time series information can be automatically selected according to the bridge structure multi-type component service state monitoring data input features, the dynamic feature selection depending on the input is realized, and the flexibility and adaptability of the model are improved.

[0061] (5) The model has a linear time complexity, significantly reduces the calculation cost, and improves the processing performance. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on the provided drawings also belong to the protection scope of the present application.

[0063] Figure 1 is a bridge structure service state spatiotemporal correlation diagnosis method flowchart based on a graph network and a state space model;

[0064] Figure 2 is a single bridge graph structure model schematic diagram considering the spatiotemporal correlation of different component responses; wherein (a) is a cable-stayed bridge, and (b) is a suspension bridge;

[0065] Figure 3 is a bridge structure service state spatiotemporal correlation diagnosis model architecture diagram based on a graph network and a state space model;

[0066] Figure 4 is a graph selective state space module architecture schematic diagram of a bridge structure service state spatiotemporal correlation diagnosis model;

[0067] Figure 5 is a dynamic filtering graph convolution layer architecture schematic diagram of a bridge structure service state spatiotemporal correlation diagnosis model;

[0068] Figure 6 is a state space selection mechanism parameter calculation and update method schematic diagram of a bridge structure service state spatiotemporal correlation diagnosis model;

[0069] Figure 7 is a graph selective scanning method schematic diagram of a bridge structure service state spatiotemporal correlation diagnosis model. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0071] In combination with Figures 1-7 , the present application proposes a bridge structure service state spatiotemporal correlation diagnosis method based on a graph network and a state space model, which comprises the following steps:

[0072] Step one: determining the bridge structure and multiple types of component objects to be studied, and establishing a single bridge graph structure model considering the spatiotemporal correlation of different component responses;

[0073] The step one specifically comprises:

[0074] Step one one: bridge multi-type component service state multi-source sensing data collection and preprocessing;

[0075] (1) Bridge multi-type component service state multi-source sensing data collection: Collect the key component service state historical data of quasi-static response (main beam vertical displacement, main beam key section strain, support reaction force, cable force, etc.) and dynamic response (main beam vertical, transverse, longitudinal vibration acceleration, etc.);

[0076] (2) Bridge multi-type component service state multi-source sensing data preprocessing: Clean the collected data, remove outliers and missing values, normalize the multi-type data, and ensure the consistency and accuracy of the data;

[0077] Step one two: Establish a single bridge graph structure model considering the spatio-temporal correlation of different component responses;

[0078] (1) Single bridge graph structure node definition: Each key component structure on the single bridge is regarded as a node in the graph, and the node attribute is quasi-static response (main beam vertical displacement, main beam key section strain, support reaction force, cable force, etc.) and dynamic response (main beam vertical, transverse, longitudinal vibration acceleration, etc.);

[0079] (2) Single bridge graph structure edge definition: According to the spatial connection relationship between the multi-type components of the bridge structure, the edges between the nodes are defined, and the edges are undirected edges;

[0080] (3) Single bridge graph structure generation: The single bridge graph structure is represented by G=(V;E), where V is the set of component nodes, and E is the set of edges; According to the definition of nodes and edges, the single bridge graph structure is generated; According to the definition of nodes and edges, the adjacency matrix A of the single bridge graph structure is determined.

[0081] Taking a large bridge as an example, the single bridge graph structure model considering the spatio-temporal correlation of different component responses is as shown in Figure 2 .

[0082] The graph structure of a certain cable-stayed bridge is as shown in Figure 2 (a), which has 2 bridge towers, 4x18 pairs of parallel steel wire cables are arranged on the whole bridge, and the main beam section has 36 sections. Therefore, according to the above method, the whole bridge is modeled as 2 bridge tower nodes, 72 cable nodes, 36 main beam nodes and 2 foundation nodes.

[0083] A certain suspension bridge is as shown in Figure 2(b) As shown, the symmetrically arranged half-bridge has 1 bridge tower, 2 pairs of main cables in total, 2x14 pairs of parallel steel wire suspender cables on the main cables, and 14 main beam segments. Therefore, the symmetrically arranged half-bridge is modeled as 1 bridge tower node, 2 main cable nodes, 28 suspender cable nodes, 14 main beam nodes, and 1 foundation node according to the above method.

[0084] Step two: design a spatiotemporal graph prediction task based on the multi-type sensing data of the bridge structure component in service state, and construct a corresponding input and output data set;

[0085] The step two specifically includes:

[0086] Step two one: for the bridge structure service state diagnosis task, the historical monitoring data of the pseudo-static response and the dynamic response are taken as the input, and the future bridge multi-type component pseudo-static response and dynamic response time history are taken as the output; the change degree of the predicted value and the true value is taken as the basis for service state diagnosis. If the predicted value and the true value have a small deviation, it means that the service state has not changed and is in a healthy state; otherwise, it means that the corresponding response bridge component has deteriorated;

[0087] Step two two: the input data set is embedded as where N is the number of key components of the single bridge, T is the input data time step, and D is the feature dimension, including the structural pseudo-static response and the dynamic response; the corresponding output data set is embedded as where T p is the predicted output data time step;

[0088] Step two three: for the multi-type, multi-scale, and multi-granularity bridge structure multi-type component service state spatiotemporal data, including the structural pseudo-static response and the dynamic response, three different time granularities are taken as the input data, including the recent data X r , the periodic data X c , and the trend data X q . The recent data is a historical time series directly adjacent to the prediction period, the periodic data is composed of segments with an integer multiple of α of the past days in the same time period as the prediction period, and the trend data is composed of segments with an integer multiple of β of the past days in the same time period as the prediction period. Among them, the period window is expressed by variables α and β, which can take values: 1 represents 1 day, 7 represents 1 week, 30 represents 1 month, 365 represents 1 year, etc. The length of the recent data, the periodic data, and the trend data (i.e., the number of time steps T contained in each segment of data) is set according to actual prediction needs to adapt to different prediction targets and historical data characteristics.

[0089] Step three: a bridge structure service state spatiotemporal correlation diagnosis model based on a graph network and a state space model is established, a graph selective state space module of the bridge structure service state spatiotemporal correlation diagnosis model comprises a normalization layer, a dynamic filtering graph convolution layer, a splicing layer, a linear layer, a segmentation layer, a one-dimensional convolution layer, a SiLU activation layer, a graph state space selection mechanism, an element-wise multiplication layer and a residual connection layer, wherein the dynamic filtering graph convolution layer learns a complex spatial dependence relationship of bridge structure multi-type component service state sequence data, and the graph state space selection mechanism learns bridge multi-type component service state evolution and spatiotemporal correlation;

[0090] The bridge structure service state spatiotemporal correlation diagnosis model is established in particular as follows:

[0091] Step three one: three time granularity data X r , X c and X q of the bridge structure multi-type component service state are weighted and fused through results output by N spatiotemporal graph convolution blocks respectively;

[0092] Step three two: outputs of the M graph selective state space modules to which the input X r is transmitted are weighted and fused again;

[0093] Step three three: an ultimate output result Y is obtained after a fully connected layer;

[0094] The bridge structure service state spatiotemporal correlation diagnosis model based on the graph network and the state space model has the architecture as shown in Figure 3 , wherein N and M are the number of the spatiotemporal graph convolution modules and the graph selective state space modules, and is a preset hyperparameter, the output of the previous module is transmitted to the next module as the input, and the transmission is sequentially performed forward.

[0095] The graph selective state space module of the bridge structure service state spatiotemporal correlation diagnosis model in step three two has the architecture as shown in Figure 4 , and mainly comprises a layer normalization layer, a dynamic filtering graph convolution layer learning a complex spatial dependence relationship of bridge structure multi-type component service state sequence data, a splicing layer, a linear layer, a segmentation layer, a one-dimensional convolution layer, a SiLU activation layer, a graph state space selection mechanism learning bridge multi-type component service state evolution and spatiotemporal correlation, an element-wise multiplication layer and a residual connection layer.

[0096] The dynamic filtering graph convolution layer specifically refers to learning a dynamic change of multi-hop connection relationship and graph structure information in a graph; and a calculation method thereof is that recent data X r of the bridge structure multi-type component service state is transformed through the layer normalization and then transmitted to the dynamic filtering graph convolution layer:

[0097]

[0098] Where A is the original adjacency matrix of the monomer bridge graph structure, D is the degree matrix, and A norm is the normalized adjacency matrix, represents the power series of the adjacency matrix, B is the basic filter used to dynamically adjust the adjacency matrix of the monomer bridge graph structure, T(B) represents the linear transformation of the basic filter B matrix, and λ is the weight coefficient between the manually assigned original adjacency matrix and the transformed filter. is the dynamically adjusted adjacency matrix, are weights and biases respectively, and B are parameters to be optimized, is the result of the i-th dynamic filter graph convolution layer, H cat is the splicing result; Each element in B is initialized from a mean of 0 and a range of Random values ​​are taken from the uniform distribution; m is a hyperparameter, and N is the number of components contained in a single bridge structure.

[0099] The graph state space selection mechanism is specifically: after the dynamic filtering graph convolution layer is performed, it is used to learn the service state evolution and spatiotemporal correlation of multiple types of bridge components. The specific method is: the result H of the dynamic filtering graph convolution layer is spliced cat The output H after the linear layer core_sec Split into core sequence H core and the minor sequence H sec , the two sequences have the same dimension, and the core sequence is passed to the graph state space selection mechanism GSSSM after one-dimensional convolution and SiLU activation, and the output H is obtained. gsssm , the calculation formula is as follows:

[0100]

[0101] The graph state space selection mechanism GSSSM includes two parts: the state space selection mechanism parameter calculation and update module and the graph selective scanning module. gsssm Multiply element-wise with the secondary sequence activated by SiLU, and then pass through the linear layer to obtain H out :

[0102] H out =W out (H gsssm ⊙SiLU(H sec ))+b out (3)

[0103] Where W out and b out Respectively represent the weight and bias parameters to be optimized; H outX r The residual connection is performed to obtain the output of the graph selective state space module of the bridge structure service state space correlation diagnosis model.

[0104] The state space selection mechanism parameter calculation updating module of the graph state space selection mechanism GSSSM has the following specific steps:

[0105] (1) Calculate the input-independent parameters A and D, A log Each row of the matrix is initialized as a sequence of consecutive integers from 1 to the column number, the elements in D are initialized as 1, A is a state transition matrix, and D is a direct gain;

[0106] (2) Process the input H c ′ ore using a linear layer to obtain H c ″ ore ;

[0107] (3) Split H c ″ ore to obtain parameters Δ, B, and C related to the bridge structure multi-type component service state monitoring data input, B and C have the same dimension, B and C are input and output matrices respectively, and Δ represents the conversion parameter from the continuous state space model to the discrete state space model;

[0108] (4) Adjust Δ using a linear layer and a SoftPlus activation function to obtain Δ′;

[0109] (5) Call the graph selective scanning module GraphSelectiveScan and input H c ′ ore , Δ′, A, B, C, and D to perform updating calculation to obtain the output H gsssm ;

[0110] The state space selection mechanism parameter calculation updating formula of the graph state space selection mechanism GSSSM is as follows:

[0111]

[0112] The graph selective scanning module GraphSelectiveScan of the graph state space selection mechanism GSSSM has the following specific steps:

[0113] (1) Obtain the dynamically adjusted single bridge graph structure adjacency matrix from the dynamic filtering graph convolution layer Initialize the padding matrix (filled with 1), copy the adjacency matrix to the upper left corner of the padding matrix to obtain P pad_adj padding matrix, P pad_adjPerform matrix multiplication matmul on the filling matrix and Δ′ to obtain the updated parameter Δ″;

[0114] (2) Use Δ″ to discretize and update A and B to obtain ΔA and ΔB, where einsum represents the Einstein summation convention;

[0115] (3) Initialize the state Z in the state space selection mechanism to 0, and perform state Z from 1 to T for the time step i in the service status monitoring data of multiple types of components of the bridge structure. i Update and output H i Calculation of

[0116] (4) Output H i Use stack operation to get H stack , and finally H c ' ore Add the weight coefficient and direct gain D to get H gsssm ;

[0117] The graph selective scanning calculation formula of the graph state space selection mechanism GSSSM is:

[0118]

[0119] Step 4: Training and Predicting a Spatiotemporal Correlation Diagnosis Model for Bridge Structure Service Condition: This model is trained based on gradient descent error backpropagation. The trained model is then used to perform spatiotemporal correlation prediction and diagnosis of the service conditions of key components, including pseudo-static and dynamic responses. This establishes a spatiotemporal correlation diagnosis system for bridge structure service conditions.

[0120] The step 4 specifically includes:

[0121] Step 41: Based on the gradient descent error back propagation training graph selective state space model, the optimization algorithm includes AddmW, SGD and other optimizers, and the training objective function adopts the relative mean square error (RMSE) with regularization constraints, so that the generated pseudo-static response (main beam vertical displacement, main beam key section strain, support reaction, cable force, etc.) and dynamic response (main beam vertical, lateral, longitudinal vibration acceleration, etc.) and other key components of the service state prediction value are as close as possible to the actual value of the actual monitoring data:

[0122]

[0123] Where, is the time history data of the service state (including pseudo-static response and dynamic response) of the key components of the bridge predicted by the model in the future k steps, X (t+1):(t+k)is the real future k-step bridge key component service state (including quasi-static response and dynamic response) time series data; N is the number of components contained in the bridge structure, k is the time step, D is the feature dimension, including quasi-static response (main beam vertical displacement, main beam key section strain, support reaction, cable force, etc.) and dynamic response (main beam vertical, transverse and longitudinal vibration acceleration, etc.); B represents the batch size, and lambda is the regularization coefficient, and R(Theta) is the regularization term.

[0124] Step four two: for the service state diagnosis task of single bridge multiple types of components, the trained model is used to predict the key component service state including quasi-static response (main beam vertical displacement, main beam key section strain, support reaction, cable force, etc.) and dynamic response (main beam vertical, transverse and longitudinal vibration acceleration, etc.).

[0125] Step four three: the change degree of the predicted value and the real value is taken as the service state diagnosis basis, if the deviation between the predicted value and the real value is small, it indicates that the service state does not change and is in a healthy state, otherwise, it indicates that the corresponding bridge component has state deterioration, so as to establish a bridge structure service state space-time correlation diagnosis system.

[0126] The application proposes a bridge structure service state space-time correlation diagnosis method based on a graph network and a state space model, which captures the complex space-time correlation and space-time evolution law in the service state monitoring data of different components of a single bridge, aims to improve the service state diagnosis ability, accuracy, efficiency and method universality of the bridge structure, and the innovation points are:

[0127] (1) the space-time graph structure is adopted, the space-time graph fusion learning module is constructed, the graph neural network and the time series data deep learning are combined, the data characteristics of the bridge structure component service state in the space and time dimensions are considered, and the dynamic space-time evolution of the complex component system of the single bridge is more accurately modeled and predicted;

[0128] (2) the state space model is adopted, the structured state space sequence is constructed, and the selective mechanism is introduced to dynamically adjust the model behavior and realize the optimized use of computing resources;

[0129] (3) the space-time characteristics of different complexity bridge multiple type component service state monitoring data are decomposed and fused, the processing requirements of the space-time data of the multiple type, multiple scale and multiple granularity bridge structure components are better adapted, and the method universality and prediction performance are improved.

[0130] The application also proposes an electronic device, which comprises a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the bridge structure service state space-time correlation diagnosis method based on the graph network and the state space model when executing the computer program.

[0131] The application further provides a computer readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the bridge structure service state space correlation diagnosis method based on a graph network and a state space model.

[0132] The memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can be a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, and not limitation, many forms of RAM can be used, such as a static random access memory (SRAM), a dynamic random access memory (DRAM), a synchronous dynamic random access memory (SDRAM), a double data rate SDRAM (DDR SDRAM), an enhanced SDRAM (ESDRAM), a synchlink DRAM (SLDRAM), and a direct rambus RAM (DRRAM). It is noted that the memory described in the methods of the present application is intended to include, but not be limited to, these and any other suitable types of memory.

[0133] In the above embodiments, all or part of the method can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the method can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), optical media (such as high-density digital video disc (digital video disc, DVD)), or semiconductor media (such as solid state disc (solid state disc, SSD)) and the like.

[0134] In the implementation process, each step of the above method can be completed by integrated logic circuit of hardware in the processor or instruction in the form of software. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware processor execution or executed by combination of hardware and software modules in the processor. The software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0135] It should be noted that the processor in the embodiments of the present application can be an integrated circuit chip with processing capability of signals. In the implementation process, each step of the method embodiments can be completed by integrated logic circuits or instructions in the form of software of the hardware in the processor. The processor mentioned above can be a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as hardware code processing for execution, or executed by a combination of hardware and software modules in the code processing. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, and other mature storage media in the art. The storage medium is located in the storage, and the processor reads the information in the storage, and combines the hardware to complete the steps of the above method.

[0136] The bridge structure service state space correlation diagnosis method based on graph network and state space model is described in detail above. The principles and implementation manners of the present application are described by applying specific examples. The above embodiment is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed, and the above description should not be understood as the limitation of the present application.

Claims

1. A spatiotemporal correlation diagnosis method for bridge structure service status based on graph network and state space model, characterized by: The method comprises the following steps: Step 1: Determine the bridge structure and multi-type component objects to be studied, and establish a single bridge graph structure model that considers the temporal and spatial correlation of the responses of different components; Step 2: Design a spatiotemporal graph prediction task based on multi-type perception data of the service status of bridge structural components and construct the corresponding input and output datasets; Step 3: Establish a spatiotemporal correlation diagnosis model for the service status of bridge structures based on graph networks and state space models. The graph selective state space module of the spatiotemporal correlation diagnosis model for the service status of bridge structures includes a normalization layer, a dynamic filtering graph convolution layer, a splicing layer, a linear layer, a segmentation layer, a one-dimensional convolution layer, a SiLU activation layer, a graph state space selection mechanism, an element-by-element multiplication layer, and a residual connection layer. The dynamic filtering graph convolution layer learns the complex spatial dependencies of the service status sequence data of multiple types of components of bridge structures, and the graph state space selection mechanism learns the evolution and spatiotemporal correlation of the service status of multiple types of components of bridge structures. Step 4: Training and prediction of the spatiotemporal correlation diagnosis model for the service status of bridge structures. The spatiotemporal correlation diagnosis model for the service status of bridge structures is trained based on gradient descent error back propagation. The trained model is used to perform spatiotemporal correlation prediction and diagnosis of the service status of key components, including pseudo-static response and dynamic response. The dynamic filtering graph convolution layer specifically refers to the dynamic changes of the multi-hop connection relationship and graph structure information in the learning graph; its calculation method is to convert the recent data X of the service status of multiple types of components of the bridge structure into r After being transformed by layer normalization, it is passed to the dynamic filter graph convolution layer; recent data is a historical time series directly adjacent to the forecast period; Where A is the original adjacency matrix of the monomer bridge graph structure, D is the degree matrix, and A norm is the normalized adjacency matrix, represents the power series of the adjacency matrix, B is the basic filter used to dynamically adjust the adjacency matrix of the monomer bridge graph structure, T(B) represents the linear transformation of the basic filter B matrix, and λ is the weight coefficient between the manually assigned original adjacency matrix and the transformed filter. is the dynamically adjusted adjacency matrix, are weights and biases respectively, and B are parameters to be optimized, is the result of the i-th dynamic filter graph convolution layer, H cat is the splicing result; Each element in B is initialized from a mean of 0 and a range of Random values ​​are taken from the uniform distribution; m is a hyperparameter, and N is the number of components contained in a single bridge structure; The graph state space selection mechanism is specifically: after the dynamic filtering graph convolution layer is performed, it is used to learn the service state evolution and spatiotemporal correlation of multiple types of bridge components. The specific method is: the result H of the dynamic filtering graph convolution layer is spliced cat The output H after the linear layer core_sec Split into core sequence H core and the minor sequence H sec , the two sequences have the same dimension, and the core sequence is passed to the graph state space selection mechanism GSSSM after one-dimensional convolution and SiLU activation, and the output H is obtained. gsssm , the calculation formula is as follows: The graph state space selection mechanism GSSSM includes two parts: the state space selection mechanism parameter calculation and update module and the graph selective scanning module. gsssm Multiply element-wise with the secondary sequence activated by SiLU, and then pass through the linear layer to obtain H out : H out =W out (H gsssm ⊙SiLU(H sec ))+b out (3) Where W out and b out Respectively represent the weight and bias parameters to be optimized; H out and layer normalized X r Residual connections are performed to obtain the output of the graph selective state space module of the spatiotemporal correlation diagnosis model of the bridge structure service status.

2. The method according to claim 1, characterized in that The step 1 specifically includes: Step 1: Collection and preprocessing of multi-source perception data on the service status of multiple types of bridge components; (1) Multi-source perception data collection of the service status of multiple types of bridge components: collecting historical service status data of key components with pseudo-static and dynamic responses; (2) Preprocessing of multi-source perception data on the service status of multiple types of bridge components: Cleaning the collected data, removing outliers and missing values, and normalizing the multi-type data to ensure data consistency and accuracy; Step 1 and 2: Establish a single bridge graph structural model that considers the spatiotemporal correlation of responses of different components; (1) Definition of structural nodes in the single-body bridge graph: Each key component structure on the single-body bridge is regarded as a node in the graph, and the node attributes are pseudo-static response and dynamic response; (2) Definition of structural edges in a single bridge graph: Based on the spatial connection relationship between multiple types of components in the bridge structure, the edges between nodes are preliminarily defined. The edges are undirected. (3) Generation of monomer bridge graph structure: The monomer bridge graph structure is represented by G = (V; E), where V is the set of component nodes and E is the set of edges; based on the definitions of nodes and edges, the monomer bridge graph structure is generated; based on the definitions of nodes and edges, the adjacency matrix A of the monomer bridge graph structure is determined.

3. The method according to claim 2, characterized in that The second step specifically includes: Step 21: For the bridge structure service condition diagnosis task, the historical monitoring data of pseudo-static response and dynamic response are used as input, and the future pseudo-static response and dynamic response time history of multiple types of bridge components are predicted as output; Step 22: Embed the input dataset into Where N is the number of key components of a single bridge, T is the time step of the input data, and F is the feature dimension, including the pseudo-static response and dynamic response of the structure; the corresponding output data set embedding is expressed as Where T p is the time step of the predicted output data; Step 2 and 3: For the spatiotemporal data of the service status of multi-type, multi-scale and multi-granular bridge structures and multi-type components, including the pseudo-static response and dynamic response of the structure, three different time granularities are used as input data, including the recent data X r , periodic data X c and trend data X q ; Periodic data consists of segments whose number of days in the past in the same time period as the forecast period is an integer multiple of α, and trend data consists of segments whose number of days in the past in the same time period as the forecast period is an integer multiple of β.

4. The method according to claim 3, characterized in that The establishment of the spatiotemporal correlation diagnosis model for the service status of the bridge structure specifically includes: Step 31: The three time granularity data X of the service status of multiple types of bridge structure components r 、X c and X q The outputs of N spatiotemporal graph convolution blocks are weightedly fused; Step 32: Enter X r The outputs of the M graph-selective state-space modules are again weightedly fused; Step 33: Get the final output after passing the fully connected layer 5. The method according to claim 4, characterized in that The state space selection mechanism parameter calculation and update module of the graph state space selection mechanism GSSSM has the following specific steps: (1) Calculate the input-independent parameters A' and D', A log Each row of the matrix is ​​initialized to the logarithm of a continuous integer sequence from 1 to the number of columns, and the elements in D' are initialized to 1. A' is the state transfer matrix, and D' is the direct gain; (2) Use the linear layer to process the input H′ core , and get H″ core ; (3) Split H″ core The parameters Δ, B', and C related to the service condition monitoring data input of multi-type components of bridge structures are obtained. The dimensions of B' and C are the same, B' and C are the input matrix and output matrix respectively, and Δ represents the conversion parameter from the continuous state space model to the discrete state space model; (4) Use the linear layer and SoftPlus activation function to adjust Δ to obtain Δ′; (5) Call the graph selective scanning module GraphSelectiveScan and pass in H′ core ,Δ′,A',B',C,D' is updated and the output H is obtained gsssm ; The parameter calculation and update formula of the state space selection mechanism GSSSM is:

6. The method according to claim 5, characterized in that The graph selective scanning module GraphSelectiveScan of the graph state space selection mechanism GSSSM has the following specific steps: (1) Obtaining the dynamically adjusted monomer bridge graph structure adjacency matrix from the dynamic filter graph convolution layer Initialize the filling matrix and convert the adjacency matrix Copy to the upper left corner of the filled matrix to get P pad_adj Filling matrix, P pad_adj Perform matrix multiplication matmul on the filling matrix and Δ′ to obtain the updated parameter Δ″; (2) Use Δ″ to discretize and update A′ and B′ to obtain ΔA and ΔB, where einsum represents the Einstein summation convention; (3) Initialize the state Z in the state space selection mechanism to 0, and perform state Z from 1 to T for the time step i in the service status monitoring data of multiple types of bridge structure components. i Update and output H i Calculation of (4) Output H i Use stack operation to get H stack , and finally H′ core Add the weight coefficient and direct gain D' to get H gsssm ; The graph selective scanning calculation formula of the graph state space selection mechanism GSSSM is:

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images