A tunnel structure health assessment method and system

Abstract

The embodiment of the application provides a tunnel structure health evaluation method and system, and belongs to the technical field of tunnel artificial intelligence monitoring. The method comprises the following steps: collecting design information and operation information of a target tunnel, and constructing a digital twin model of the target tunnel based on the design information and the operation information; acquiring multi-source monitoring data through an automatic monitoring device arranged in the target tunnel; inputting the multi-source monitoring data into a graph neural network model for spatio-temporal feature fusion processing, so as to output a tunnel structure anomaly probability and a structure health index; feeding back the tunnel structure anomaly probability and the structure health index to the digital twin model, performing residual life prediction on the target tunnel structure, and outputting early warning information when the prediction result meets a preset alarm condition. The application scheme realizes automatic spatio-temporal fusion and artificial intelligence analysis of tunnel multi-source monitoring data, and reliably drives the digital twin model to complete structure residual life prediction and early warning.

Description

Technical Field

[0001] This invention relates to the field of tunnel artificial intelligence monitoring technology, and specifically to a method and system for assessing the structural health of tunnels. Background Technology

[0002] Tunnels are crucial infrastructure for urban transportation and highway transport, and many tunnels have been in use for a long time. Tunnel lining structures are constantly exposed to vehicle loads, groundwater, and ambient temperature, which over time can lead to cracks, leaks, and deformation. If these defects are not detected promptly, they can cause localized structural loosening, and in severe cases, compromise overall safety. Currently, the common practice is manual inspection, where personnel enter the tunnel to check its appearance and then read sensor data to make judgments.

[0003] Manual methods are time-consuming and labor-intensive, and easily influenced by experience, leading to inconsistent conclusions among different personnel. Sensors can collect stress and displacement data, but the data types are diverse, sampling times are not uniform, and spatial locations are scattered, making it difficult to effectively utilize all this information manually.

[0004] Digital twin technology can build a tunnel model on a computer, but most existing models are built offline with fixed parameters, and the differences from the real structure become increasingly apparent after a period of operation. For image data and time-series physical quantity data, existing methods lack unified analysis tools, making it difficult to achieve automatic identification and early warning. With the increasing number of tunnels and their service life, a new technology is needed to entrust multi-source monitoring data to artificial intelligence models for analysis, allowing the system to perform fusion calculations automatically, reducing human intervention, and improving judgment efficiency and reliability.

[0005] Based on this need, an artificial intelligence monitoring and analysis method for actual tunnel scenarios needs to be proposed to continuously track lining defects and predict their lifespan, helping maintenance personnel to keep abreast of the tunnel's health status. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for assessing the health of tunnel structures, so as to at least solve the problems in the prior art where the data of multi-source monitoring are scattered, the sampling time is not uniform, the spatial location is difficult to align and integrate for analysis, and the identification of defects and life prediction rely on human experience.

[0007] To achieve the above objectives, the first aspect of the present invention provides a method for assessing the health of a tunnel structure. The method includes: collecting design and operational information of a target tunnel, and constructing a digital twin model of the target tunnel based on the design and operational information; acquiring multi-source monitoring data through automated monitoring equipment installed in the target tunnel, the multi-source monitoring data including stress data, displacement data, and image data; inputting the multi-source monitoring data into a graph neural network model for spatiotemporal feature fusion processing to output a tunnel structure anomaly probability and a structural health index; feeding back the tunnel structure anomaly probability and structural health index to the digital twin model to predict the remaining life of the target tunnel structure, and outputting a warning message when the prediction result meets preset alarm conditions.

[0008] Optionally, constructing a digital twin model of the target tunnel based on design and operational information includes: discretizing the target tunnel lining structure into multiple twin nodes according to the design information, and establishing three-dimensional geometric connection relationships and boundary constraint relationships between the twin nodes to obtain an initial digital twin topology framework for the target tunnel; extracting the time-series load parameters and time-series response parameters corresponding to each twin node from the operational information, and writing the time-series load parameters and time-series response parameters into the node attribute interface; and incrementally updating and correcting the parameters of the digital twin model based on the deviation analysis rules between the design information and the operational information to obtain a digital twin model that changes with the operational information.

[0009] Optionally, establishing three-dimensional geometric connection relationships and boundary constraint relationships between twin nodes to obtain an initial digital twin topology framework for the target tunnel includes: dividing the target tunnel into cross-sectional units along the route direction based on the tunnel structure segmentation drawings included in the design information, and abstracting each lining ring or lining section into twin node objects; automatically generating a twin node adjacency matrix based on the node spatial location parameters given in the design information using three-dimensional spatial proximity discrimination rules to establish quantifiable geometric connection relationships between nodes; constructing node timing consistency verification rules based on the monitoring equipment installation status and sampling timing status included in the operation information, and performing online screening and reconstruction on the twin node adjacency matrix to form dynamic boundary constraint relationships that match the current operation status; and jointly applying the reconstructed twin node adjacency matrix and dynamic boundary constraint relationships to the initial digital twin topology framework generation process to obtain an initial digital twin topology framework that reflects the true geometric characteristics and actual operating conditions of the target tunnel.

[0010] Optionally, multi-source monitoring data is acquired through automated monitoring equipment installed within the target tunnel. This includes: obtaining time-series stress monitoring data and time-series displacement monitoring data for each key structural component based on stress sensors and laser ranging sensors deployed at multiple pre-defined key structural locations within the target tunnel, according to the sampling control rules given in the operation information; periodically acquiring images of the target tunnel lining structure using a patrol robot, and binding the structural appearance image data to the spatial location of the corresponding key structural component according to the structural coordinate positioning rules included in the design information; employing a unified clock synchronization mechanism for different types of automated monitoring equipment, and aligning the time-series stress monitoring data, time-series displacement monitoring data, and structural appearance image data according to the timestamp generation rules set in the operation information to form a multi-source data set with a consistent time reference; and performing online validity judgment and standardization processing on the multi-source data set according to the data format verification rules agreed upon in the design and operation information to obtain multi-source monitoring data.

[0011] Optionally, the graph neural network model includes: a dual-mode input layer composed of an image encoder and a physical quantity encoder to receive multi-source monitoring data and generate node feature vectors; a spatiotemporal fusion backbone composed of multiple graph convolutional layers and temporal convolutional layers as the conventional processing structure of the graph neural network; wherein, a structural constraint modulation layer is set after the spatiotemporal fusion backbone, the structural constraint modulation layer is composed of a constraint embedding port and a feature modulation unit connected in series, to receive geometric constraint parameters determined by design information and boundary constraint parameters determined by operation information, and to concatenate and fuse them with the node feature vectors, performing channel-level weighted modulation on the channel features of each node; and a dual output head set after the structural constraint modulation layer, the dual output head including a health index regression branch and an anomaly probability output branch, used to output the node-level structural health index and node-level anomaly probability value of the target tunnel respectively according to the spatiotemporally fused node feature vectors.

[0012] Optionally, the multi-source monitoring data is input into a graph neural network model for spatiotemporal feature fusion processing to output the tunnel structure anomaly probability and structural health index. This includes: performing temporal filtering and normalization processing on the stress and displacement data in the multi-source monitoring data at the node feature input end of the graph neural network to obtain a standardized physical quantity feature sequence; simultaneously performing convolutional feature extraction and feature compression processing on the image data in the multi-source monitoring data to obtain an image encoding feature sequence corresponding to the standardized physical quantity feature sequence; performing feature alignment processing on the standardized physical quantity feature sequence and the image encoding feature sequence at the node dimension with the same timestamp to obtain a node-level aligned fused feature sequence; performing multi-round graph convolution aggregation and temporal feature regression processing on the node-level aligned fused feature sequence to form a stable node spatiotemporal joint output result; and generating the structural anomaly probability output value and structural health index output value corresponding to each monitoring location based on the node spatiotemporal joint output result.

[0013] Optionally, based on the spatiotemporal joint output results of the nodes, structural anomaly probability output values ​​and structural health index output values ​​corresponding to each monitoring location are generated, including: performing node-dimensional parsing on the spatiotemporal joint output results of the nodes to obtain node spatiotemporal output vectors corresponding to each monitoring location; inputting the anomaly feature components in the node spatiotemporal output vectors into a probability mapping channel to generate continuous numerical structural anomaly probability output values ​​node by node; inputting the health feature components in the node spatiotemporal output vectors into a regression mapping channel to generate continuous exponential structural health index output values ​​node by node; concatenating the structural anomaly probability output values ​​at multiple consecutive time points along the time dimension, and using the concatenated result as the structural anomaly probability output value; concatenating the structural health index output values ​​at multiple consecutive time points along the time dimension, and using the concatenated result as the structural health index output value.

[0014] Optionally, the tunnel structure anomaly probability and structural health index are fed back to the digital twin model to predict the remaining service life of the target tunnel structure. This includes: inputting the tunnel structure anomaly probability into the cross-sectional risk attribute port of the digital twin model to label the influence weight of the spatial distribution characteristics of the anomaly probability in the remaining service life prediction stage; inputting the structural health index into the material health attribute port of the digital twin model to modulate the lining stiffness degradation characteristic parameters corresponding to each monitoring location; and performing a joint recursive calculation on the input anomaly probability and health index in the digital twin model to obtain the predicted value of the remaining service life of the target tunnel under the current service state.

[0015] Optionally, when the prediction result meets the preset alarm conditions, an early warning message is output, including: when the output value of the structural anomaly probability at any monitoring location continuously increases and exceeds the normal fluctuation level at multiple consecutive moments, the prediction result is determined to meet the preset alarm conditions, and an early warning message is output; when the output value of the structural health index at any monitoring location suddenly decreases and exceeds the average change level of adjacent monitoring locations at multiple consecutive moments, the prediction result is determined to meet the preset alarm conditions, and an early warning message is output; when multiple spatially adjacent monitoring locations simultaneously meet the preset alarm conditions at the same moment, the judgment results of the multiple monitoring locations are combined as the prediction result meeting the preset alarm conditions, and an early warning message is directly output.

[0016] A second aspect of the present invention provides a tunnel structure health assessment system, the system comprising: a model building unit for collecting design and operational information of a target tunnel and constructing a digital twin model of the target tunnel based on the design and operational information; a data acquisition unit for acquiring multi-source monitoring data through automated monitoring equipment installed in the target tunnel, the multi-source monitoring data including stress data, displacement data, and image data; a processing unit for inputting the multi-source monitoring data into a graph neural network model for spatiotemporal feature fusion processing to output a tunnel structure anomaly probability and a structural health index; and an early warning unit for feeding back the tunnel structure anomaly probability and structural health index to the digital twin model to predict the remaining life of the target tunnel structure and outputting early warning information when the prediction result meets preset alarm conditions.

[0017] Through the above technical solution, this invention establishes a digital twin model using design and operational information, providing a calculable virtual benchmark for the tunnel structure. Automated monitoring equipment continuously acquires stress, displacement, and image data, solving the problem of limited information acquisition in traditional manual inspections. A graph neural network model fuses spatiotemporal features of multi-source monitoring data, enabling a unified representation of dispersed heterogeneous data within the same model. The model can simultaneously output anomaly probabilities and health indices, achieving automatic identification and continuous assessment of localized tunnel defects. These two results are then fed back to the digital twin model for remaining life prediction, and timely warning information is output when the predicted conditions meet alarm criteria. The entire process reduces human intervention, improves assessment efficiency and reliability, and helps maintenance personnel detect risks earlier and better understand the true health status of the tunnel structure.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of a tunnel structure health assessment method provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a graph neural network model provided in one embodiment of the present invention; Figure 3 This is a system structure diagram of a tunnel structure health assessment system provided in one embodiment of the present invention; Figure 4 This is an internal structural diagram of a computer device provided in one embodiment of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 As shown, embodiments of the present invention provide a method for assessing the health of tunnel structures, the method comprising: Step S10: Collect the design and operation information of the target tunnel, and construct a digital twin model of the target tunnel based on the design and operation information.

[0022] Specifically, constructing a digital twin model of the target tunnel based on design and operational information includes: discretizing the target tunnel lining structure into multiple twin nodes according to the design information, and establishing three-dimensional geometric connection relationships and boundary constraint relationships between the twin nodes to obtain an initial digital twin topology framework for the target tunnel; extracting the time-series load parameters and time-series response parameters corresponding to each twin node from the operational information, and writing the time-series load parameters and time-series response parameters into the node attribute interface; and incrementally updating and correcting the parameters of the digital twin model based on the deviation analysis rules between the design information and the operational information to obtain a digital twin model that changes with the operational information.

[0023] Furthermore, establishing three-dimensional geometric connection relationships and boundary constraint relationships between twin nodes to obtain an initial digital twin topology framework for the target tunnel includes: dividing the target tunnel into cross-sectional units along the route direction based on the tunnel structure segmentation drawings included in the design information, and abstracting each lining ring or lining section into twin node objects; automatically generating a twin node adjacency matrix based on the node spatial location parameters given in the design information using three-dimensional spatial proximity discrimination rules to establish quantifiable geometric connection relationships between nodes; constructing node timing consistency verification rules based on the monitoring equipment installation status and sampling timing status included in the operation information, and performing online screening and reconstruction on the twin node adjacency matrix to form dynamic boundary constraint relationships that match the current operation status; and jointly applying the reconstructed twin node adjacency matrix and dynamic boundary constraint relationships to the initial digital twin topology framework generation process to obtain an initial digital twin topology framework that reflects the true geometric characteristics and actual operating conditions of the target tunnel.

[0024] In this embodiment of the invention, collecting design and operational information of the target tunnel and constructing a digital twin model of the target tunnel based on this information is the fundamental workflow for subsequent structural health assessment. The design information originates from a complete set of design drawings, construction completion data, and three-dimensional geometric model data of the target tunnel generated during the construction phase, used to express the tunnel's cross-sectional shape, lining type, material parameters, and structural segmentation. The operational information originates from the daily operational load and historical monitoring response data of the target tunnel recorded by the management platform during its current service phase, used to express the actual stress and deformation characteristics of each structural component at different time periods. By simultaneously acquiring design and operational information, the constructed digital twin model can maintain a consistent correspondence with the target tunnel in both geometric shape and service status, thereby avoiding the problem of insufficient representation caused by relying solely on a single design drawing or single monitoring data to build the model.

[0025] In engineering implementation, the target tunnel lining structure is discretized into multiple twin nodes based on the design information. Assume the total length of the target tunnel along the route is... The data, derived from the route mileage in the design information, serves as the overall scale benchmark for the node discretization process. Based on the tunnel structure segmentation drawings included in the design information, the target tunnel is divided along the route direction into... Each discrete analysis segment, derived from the structural partitioning rules in the design information, serves as a parameter for the number of twin nodes. The length of each analysis segment is denoted as... ,in , The mileage range data of the lining rings or lining sections derived from the design information are used to determine the longitudinal distribution characteristics of each corresponding node in the digital twin model, and satisfy the following during the discretization process: ; To further ensure the correspondence between the twin node objects and the actual structural parts, during the node discretization process, each lining ring or lining analysis segment is abstracted into a node object based on the cross-sectional segment drawings in the design information. ,in For the first The parameters of each twin node object are derived from the cross-sectional element division results in the design information and are used as the basic building blocks of the digital twin model.

[0026] After obtaining multiple twin node objects, it is necessary to establish the three-dimensional geometric connection relationships and boundary constraint relationships between the twin nodes based on the node spatial location parameters given in the design information, in order to form the initial digital twin topology framework of the target tunnel. Let each twin node object... The position vector in three-dimensional space is ,in For the first The location parameters of the twin nodes are used to determine the specific placement of the twin nodes in the spatial dimension; This is the lateral position parameter of the node, used to express the node's location in the horizontal direction of the tunnel cross section; This is the node elevation position parameter, used to express the node's location in the vertical direction; This is the longitudinal mileage position parameter of the node, used to express the node's location along the tunnel length.

[0027] In order for the connection relationships between nodes to be quantified and used by the virtual model, based on the position vector The destination is determined by automatically generating a twin node adjacency matrix using three-dimensional spatial proximity rules. Suppose any two twin node objects and The spatial distance parameter between them is ,in The distance between nodes is used to determine whether nodes are geometrically adjacent, and its destination is used to participate in the generation and discrimination of adjacency matrix elements.

[0028] Based on the destination of this distance parameter, construct the adjacency matrix. ,in Parameters for node connection relationships; matrix elements For the adjacency relationship determination result parameters, when When it is less than the nearest threshold, This indicates that nodes have a geometric adjacency relationship in space, and is used to establish quantifiable geometric connections between twin nodes; when When it is greater than the nearest threshold, This indicates that the nodes are not adjacent to each other in space.

[0029] After completing node discretization and constructing static geometric connections, the temporal attribute interface of the digital twin model needs further refinement based on operational information. This operational information includes the actual load changes and historical response changes at each monitoring location over different time periods. This is to allow the node objects in the virtual model to... It can load corresponding service status parameters as they change over time in each twin node object. Interface for setting node attributes ,in This is a time index parameter, used as an index base to distinguish data at different times in the node attribute interface.

[0030] According to the information recorded in the operation information, at the time Within the period, corresponding twin node objects The time-series load parameters are denoted as ,in For nodes At any moment The actual load sampling parameters are used to write to the node attribute interface, and their destination is the node stress state parameters in the modulation virtual model. The corresponding time-series response parameters of each twin node object are denoted as... ,in For nodes At any moment The historical response sampling parameters are used to write to the node attribute interface, and their destination is the node deformation state parameters in the modulation virtual model. By using... and The destination of each node is written into the node attribute interface, which can establish the service status expression relationship of node objects in the time dimension in the model, so that the geometric topology framework of the digital twin model and the actual operation information form a quantitative closed loop consistent relationship.

[0031] In the further stages of model building, based on the deviation analysis rules between design information and operational information, the parameters of the digital twin model need to be incrementally updated and corrected to obtain a digital twin model that changes with operational information. Let the design information provide the first... The design phase parameters for each node object are denoted as follows: The correction parameters at the corresponding time in the runtime information are denoted as ,in Parameters are designed for the nodes to serve as baseline parameters for deviation analysis; These are the node runtime parameters, used as dynamic reference parameters for deviation analysis. During incremental update correction, the deviation vector is defined as follows: ; in For nodes At any moment The parameter deviation vector is used to participate in the online correction of digital twin model parameters, and its destination is to drive the incremental update and correction of model node attribute parameters. This is achieved by mapping the deviation vector of each node object over different time periods. The feedback of the destination to the parameter update process of the digital twin model can enable the parameters of the digital twin model to evolve dynamically with the new operational information, thereby avoiding the problem of fixed parameters and making the model expression closer to the actual service state of the target tunnel.

[0032] During the incremental update and correction process of the digital twin model, node timing consistency verification rules are also constructed based on the monitoring equipment installation status and sampling timing status contained in the operation information, and the adjacency matrix is ​​checked. Perform online filtering and refactoring to form dynamic boundary constraint relationships that match the current operating state. The parameters of the online filtering and refactoring result are denoted as follows. It is used to modulate the boundary constraint relationships between node objects in the virtual model, and its purpose is to maintain the consistent implementability of the model node attribute interface in the time dimension.

[0033] Through the above process, the reconstructed adjacency matrix and dynamic boundary constraints are used together to improve the initial digital twin topology framework, thereby forming an initial digital twin topology framework that is geometrically consistent with the target tunnel and consistent with the actual operating conditions. Based on this, the model parameters are incrementally updated and corrected in combination with the continuous changes in operating information, and finally a target tunnel digital twin model that dynamically evolves with the operating information is obtained.

[0034] In another possible implementation, to further improve the consistency between the digital twin model and the actual tunnel structure, a mobile reference monitoring device is deployed within the target tunnel to temporarily and intensively collect data at key cross-sectional locations. The mobile reference monitoring device moves slowly along the tunnel alignment, continuously acquiring images of the lining appearance and local micro-displacement parameters during its movement. The spatial position vector of the mobile reference monitoring device is used as a dynamic verification index and written into the node attribute interface of the digital twin model in real time. By comparing the differences between the parameters generated by the fixed monitoring points and those generated by the mobile reference monitoring device, a local deviation is obtained for correction. This deviation is then applied inversely to the geometric connection relationship construction process of the corresponding twin nodes, enabling the digital twin model to achieve adaptive reconstruction at some nodes.

[0035] Step S20: Acquire multi-source monitoring data through automated monitoring equipment installed in the target tunnel. The multi-source monitoring data includes stress data, displacement data, and image data.

[0036] Specifically, multi-source monitoring data is acquired through automated monitoring equipment installed within the target tunnel. This includes: obtaining time-series stress monitoring data and time-series displacement monitoring data for each key structural component based on stress sensors and laser ranging sensors deployed at multiple pre-defined key structural locations within the target tunnel, according to the sampling control rules provided in the operational information; periodically acquiring images of the target tunnel lining structure using a patrol robot, and binding the structural appearance image data to the spatial location of the corresponding key structural component according to the structural coordinate positioning rules included in the design information; employing a unified clock synchronization mechanism for different types of automated monitoring equipment, and aligning the time-series stress monitoring data, time-series displacement monitoring data, and structural appearance image data according to the timestamp generation rules set in the operational information to form a multi-source data set with a consistent time reference; and performing online validity judgment and standardization processing on the multi-source data set according to the data format verification rules agreed upon in the design and operational information to obtain multi-source monitoring data.

[0037] In this embodiment of the invention, multi-source monitoring data is acquired through automated monitoring equipment installed within the target tunnel. The target tunnel has multiple pre-defined key structural components during daily operation; the spatial locations and numbers of these key structural components are derived from the lining segment drawings in the design information. To ensure that each monitoring data point has a clear source and destination, multiple key structural components are monitored. Stress sensors and laser rangefinders are deployed at the locations. The stress sensors detect changes over time. The actual stress parameters of the lining structure are continuously collected during the cycle to obtain time-series stress monitoring data. Here For the first Each monitoring location at time The stress sampling parameters are derived from the real-time sampling results of stress sensors installed at the corresponding critical structural parts, and are used to directly reflect the normal stress state and variation of the structural parts.

[0038] Laser rangefinder at the same monitoring position According to time Collect structural displacement parameters to obtain time-series displacement monitoring data. . For the first Each monitoring location at time The displacement sampling parameters are obtained through non-contact measurement of the micro-displacements of the lining structure by a laser rangefinder sensor, reflecting the deformation state of the structure under actual loads. By setting up two types of sensors in parallel at the node dimension, time-series stress monitoring data and time-series displacement monitoring data for each key structural component can be obtained separately.

[0039] At the image data acquisition level, this invention further utilizes an inspection robot to periodically acquire images of the target tunnel lining structure. The inspection robot moves along the route to the spatial locations of multiple key structural parts. In time During the period, images of the lining surface are acquired to obtain structural surface image data. .Should For the first Each monitoring location at time The image sampling parameters are digital records of the lining structure's appearance from the image acquisition device mounted on the inspection robot. These parameters are used to represent structural features such as cracks, leaks, and localized damage. Because image data and physical quantity data may differ in time reference, the structural appearance image data must be bound to the spatial location of the corresponding key structural parts according to the structural coordinate positioning rules contained in the design information.

[0040] Let the spatial position vector of the inspection robot during its movement be... ,in For inspection robots in time The time corresponding to the first The spatial location parameters of key structural parts are used to determine the spatial attribution of image data in the virtual model.

[0041] Regarding timing alignment, this invention employs a unified clock synchronization mechanism for different types of automated monitoring equipment. The unified clock index is denoted as a timestamp parameter. , A common time index parameter is used to generate the timestamps defined in the operational information, serving as a unified time reference for all monitoring data. Time alignment is performed on time-series stress monitoring data, time-series displacement monitoring data, and structural appearance image data, expressed as: ; in For the first Each monitoring location at time The multi-source data set parameters are used to generate a multi-source data set with a consistent time base. Because... The three types of data differ in format. This invention further performs online validity assessment and standardization on the multi-source data set according to the data format verification rules agreed upon in the design and operational information. The standardized image parameters are set as follows: The standardized physical quantity parameters are as follows: ; in The long-term average stress reference parameter given in the operation information is used as a reference mean for stress data standardization. The stress standard deviation benchmark parameter given in the operation information is used as a scale factor for stress data standardization. The long-term average displacement baseline parameter given in the operation information is used as a reference mean for displacement data standardization. The displacement standard deviation benchmark parameter given in the operational information is used as a scale factor for standardizing the displacement data. Through this standardization operation, the processed data... , , A unified and usable feature expression form is formed in both numerical range and timestamp, thus obtaining a truly meaningful: ; in The parameters are the node-level heterogeneous data set parameters after standardization of multi-source monitoring data.

[0042] Step S30: Input the multi-source monitoring data into the graph neural network model for spatiotemporal feature fusion processing to output the tunnel structure anomaly probability and structural health index.

[0043] Specifically, the graph neural network model includes: a dual-mode input layer composed of an image encoder and a physical quantity encoder to receive multi-source monitoring data and generate node feature vectors; a spatiotemporal fusion backbone composed of multiple graph convolutional layers and temporal convolutional layers as the conventional processing structure of the graph neural network; wherein, a structural constraint modulation layer is set after the spatiotemporal fusion backbone, the structural constraint modulation layer is composed of constraint embedding ports and feature modulation units connected in series, to receive geometric constraint parameters determined by design information and boundary constraint parameters determined by operation information, and to concatenate and fuse them with node feature vectors, performing channel-level weighted modulation on the channel features of each node; and a dual output head set after the structural constraint modulation layer, the dual output head including a health index regression branch and an anomaly probability output branch, used to output the node-level structural health index and node-level anomaly probability value of the target tunnel respectively according to the spatiotemporally fused node feature vectors.

[0044] In this embodiment of the invention, multi-source monitoring data is acquired by automated monitoring equipment installed inside the target tunnel. After these data are aligned in both node and time dimensions, they need to be input into a graph neural network model for intelligent computation. Since the multi-source monitoring data contains two independent and heterogeneous types of information: image data and physical quantity data, the graph neural network model first sets up a dual-mode input layer consisting of an image encoder and a physical quantity encoder. The image encoder processes data from monitoring locations... Image data parameters Perform convolutional encoding operation ,in The node embedding parameters, encoded from the image, are a component of the node feature vector, which is then fed into subsequent graph convolutional layers. The physical quantity encoder performs temporal encoding operations on stress and displacement data: ; in The node embedding parameters, which are encoded physical quantities, are derived from the stress parameters collected by the stress sensor. Displacement parameters acquired by laser rangefinders This is a component used to generate the node feature vector, which is then input to subsequent graph convolutional layers. The feature concatenation process after parallel encoding through the dual-mode input layer is as follows: ; Received For the first Each monitoring location in time The initial feature vector of the node.

[0045] In the joint spatial and temporal aggregation stage, the graph neural network model internally sets up a spatiotemporal fusion backbone composed of multiple graph convolutional layers and temporal convolutional layers. Here These are the network layer index parameters, used to distinguish the processing flow of different convolutional mapping stages in the model. The initial feature vector matrix of the spatiotemporal fusion backbone for all nodes. Perform progressive feature mapping operation: ; in For the first Layer in time Node aggregation feature parameters at time t; The twin node adjacency matrix in the design information is used to determine the spatial connectivity in graph convolution operations; For the first The learnable weight parameters of the layer are used to modulate the feature channel mapping. The spatiotemporal joint output of the nodes after multi-layer progressive aggregation is denoted as: ; in For monitoring location In time Spatiotemporal fusion output characteristic parameters at any given time.

[0046] Prior to the output stage, a structural constraint modulation layer is set after the spatiotemporal fusion backbone to perform channel-level weighted modulation operations on the node feature vectors. This structural constraint modulation layer receives geometric constraint parameter vectors from the design information through a constraint embedding port. The boundary constraint parameter vector from the operation information is received through the feature modulation unit. These constraint parameters are then concatenated and fused with the spatiotemporal joint output of the nodes: ; in These are the synthesized node characteristic parameters after incorporating structural constraints. Further channel-level weighted modulation is then performed on the synthesized node characteristics: ; in The modulated node outputs characteristic parameters.

[0047] The dual-output head, positioned after the structural constraint modulation layer, outputs characteristic parameters to the modulated nodes. Perform parallel mapping operations separately: ; in This is a parameter for the structural anomaly probability output value, used as the final output object for the structural anomaly probability output value. This is the output value parameter for the structural health index, used as the final output object for the structural health index.

[0048] In one specific implementation, such as Figure 2 The input layer aggregates data from various monitoring locations. Aligned multi-source monitoring data are fed into the dual-mode input layer. The image encoder receives image data and generates image embedding features. The physical quantity encoder receives stress data and displacement data and generates physical quantity embedded features. The two types of encoding results are concatenated at the node dimension to form a node feature vector. The input is fed into the spatiotemporal fusion backbone for progressive aggregation processing to obtain the spatiotemporal output features of the nodes. Subsequently, geometric constraint parameters and boundary constraint parameters are received through the node feature fusion interface to form a fused node feature vector, which is then fed into the structural constraint modulation layer. The structural constraint modulation layer performs channel-level weighted modulation on the channel features of each node and outputs the results to the dual output head layer. The health index evaluation branch and the health index verification module in the dual output head jointly output the structural health index value, while the anomaly probability output branch outputs the structural anomaly probability value.

[0049] Specifically, the multi-source monitoring data is input into a graph neural network model for spatiotemporal feature fusion processing to output the tunnel structure anomaly probability and structural health index. This includes: performing temporal filtering and normalization processing on the stress and displacement data in the multi-source monitoring data at the node feature input end of the graph neural network to obtain a standardized physical quantity feature sequence; simultaneously performing convolutional feature extraction and feature compression processing on the image data in the multi-source monitoring data to obtain an image encoding feature sequence corresponding to the standardized physical quantity feature sequence; performing feature alignment processing on the standardized physical quantity feature sequence and the image encoding feature sequence at the node dimension with consistent timestamps to obtain a node-level aligned fused feature sequence; performing multi-round graph convolution aggregation and temporal feature regression processing on the node-level aligned fused feature sequence to form a stable node spatiotemporal joint output result; and generating the structural anomaly probability output value and structural health index output value corresponding to each monitoring location based on the node spatiotemporal joint output result.

[0050] In this embodiment of the invention, a graph neural network model is used to receive multi-source monitoring data that has been collected and time-aligned, and to perform a unified analysis of the structural health status of the target tunnel. The multi-source monitoring data includes three types of information: stress data, displacement data, and image data, which are directly collected by automated monitoring equipment deployed inside the tunnel. To facilitate model use, this invention first performs time-series filtering on the stress and displacement data at the node feature input end to remove obvious noise points, and then normalizes the filtered results according to a uniform scale, thereby forming a standardized feature sequence of physical quantities, making the physical parameters at different times comparable.

[0051] Meanwhile, for the image data acquired by the inspection robot, the image encoding channel within the model performs convolutional feature extraction processing and obtains an image encoding feature sequence through feature compression. This sequence corresponds to the standardized feature sequence of physical quantities in terms of timestamps and node numbers. Subsequently, the two types of feature sequences are aligned at the node dimension to ensure that multi-source data from the same monitoring location are fused under the same time reference, resulting in a node-level aligned fused feature sequence. The graph neural network model then performs multi-round graph convolutional aggregation processing on the fused features to complete the feature transfer of spatial information, and combines it with a temporal convolutional layer to perform temporal feature regression processing, achieving coordinated aggregation of space and time to form a stable node spatiotemporal joint output result. Based on the different feature components of the node spatiotemporal joint output result, the dual output head layer generates the structural anomaly probability output value and structural health index output value corresponding to each monitoring location, respectively.

[0052] Specifically, based on the spatiotemporal joint output results of the nodes, structural anomaly probability output values ​​and structural health index output values ​​are generated for each monitoring location. This includes: performing node-dimensional parsing on the spatiotemporal joint output results to obtain node spatiotemporal output vectors for each monitoring location; inputting the anomaly feature components in the node spatiotemporal output vectors into a probability mapping channel to generate continuous numerical structural anomaly probability output values ​​for each node; inputting the health feature components in the node spatiotemporal output vectors into a regression mapping channel to generate continuous exponential structural health index output values ​​for each node; concatenating the structural anomaly probability output values ​​at multiple consecutive time points along the time dimension, and using the concatenated result as the structural anomaly probability output value; and concatenating the structural health index output values ​​at multiple consecutive time points along the time dimension, and using the concatenated result as the structural health index output value.

[0053] In this embodiment of the invention, after the graph neural network model completes the encoding and aggregation processing of multi-source monitoring data, it then... Get the corresponding monitoring location at all times The spatiotemporal joint output results of the nodes .Should The output is a progressive aggregation of multiple graph convolutional layers and temporal convolutional layers in the spatiotemporal fusion backbone, used to parse out the usable health-related feature components and abnormality-related feature components in this process.

[0054] In order to The information forms a clear closed loop. This invention first performs node-dimensional parsing processing on the spatiotemporal joint output results of the nodes. Assume the target tunnel has a total of [number missing] nodes within the current monitoring period. If there are 10 effective monitoring locations, the spatiotemporal joint output of the node can be expressed as: ; in For time The node spatiotemporal output matrix parameters are used for parsing and mapping at the node level, and their destinations are the anomaly probability output branch and the health index output branch, respectively. For a single node... The parsing results can be further abstracted into node spatiotemporal output vectors: ; in For the first Each monitoring location at time The spatiotemporal output vector of the node is parsed into different channel feature components to participate in subsequent mapping.

[0055] After obtaining the spatiotemporal output vector of the node through parsing, this invention further sets up a probability mapping channel to automatically convert the anomalous feature components therein into continuous numerical form of structural anomaly probability output values. Let the anomalous feature components be: ; in For the first Each node in time The anomaly-related channel feature parameters, used to indicate whether the node has an anomalous trend, are input to the probability mapping channel to generate the structural anomaly probability output value. ; in For the first Each monitoring location in time The structural anomaly probability output value parameter is used to participate in the time dimension splicing and form the time series result of the structural anomaly probability output value.

[0056] Simultaneously, this invention also includes a regression mapping channel to automatically map the health feature components in the spatiotemporal output vector of nodes into a continuous exponential form of the structural health index output value. Let the health feature components be: ; in For the first Each node in time The health-related channel feature parameters are used to express the comprehensive characteristics of the node's structural state. Their destination is the regression mapping channel, which generates the structural health index output value. The regression mapping channel performs a linear mapping operation on the health feature components: ; in For the first Each monitoring location in time The structural health index output value parameter is as follows; The weight parameters in the regression mapping channel are used to modulate the mapping scale of health features to the exponential space; This is the bias parameter in the regression mapping channel, used to determine the mapping reference position.

[0057] After generating structural anomaly probability output values ​​and structural health index output values ​​node by node, this invention further performs time-dimensional concatenation processing on the output results from multiple consecutive time points to form a time-series closed loop with long-term tracking significance. Let the concatenation window length be... The time-sequencing process of the structural anomaly probability output value can be expressed as follows: ; in For the first The sequence of structural anomaly probability output values ​​at multiple consecutive time points at a monitoring location. The time-stitching process of the structural health index output values ​​can be represented as: ; in For the first The sequence of structural health index output values ​​at multiple consecutive time points at each monitoring location.

[0058] After the final assembly is completed, the structural anomaly probability output value and the structural health index output value are respectively used as two independent output objects of the graph neural network model. , .

[0059] Step S40: Feed back the tunnel structure anomaly probability and structural health index to the digital twin model to predict the remaining life of the target tunnel structure, and output early warning information when the prediction results meet the preset alarm conditions.

[0060] Specifically, the tunnel structure anomaly probability and structural health index are fed back to the digital twin model to predict the remaining life of the target tunnel structure. This includes: inputting the tunnel structure anomaly probability into the cross-sectional risk attribute port of the digital twin model to indicate the influence weight of the spatial distribution characteristics of the anomaly probability in the remaining life prediction stage; inputting the structural health index into the material health attribute port of the digital twin model to modulate the lining stiffness degradation characteristic parameters corresponding to each monitoring location; and performing a joint recursive calculation on the input anomaly probability and health index in the digital twin model to obtain the predicted value of the remaining life of the target tunnel under the current service state.

[0061] Furthermore, when the prediction result meets the preset alarm conditions, an early warning message is output, including: when the output value of the structural anomaly probability at any monitoring location continuously increases and exceeds the normal fluctuation level at multiple consecutive moments, the prediction result is determined to meet the preset alarm conditions, and an early warning message is output; when the output value of the structural health index at any monitoring location suddenly decreases and exceeds the average change level of adjacent monitoring locations at multiple consecutive moments, the prediction result is determined to meet the preset alarm conditions, and an early warning message is output; when multiple spatially adjacent monitoring locations simultaneously meet the preset alarm conditions at the same moment, the judgment results of the multiple monitoring locations are combined as the prediction result meeting the preset alarm conditions, and an early warning message is directly output.

[0062] In this embodiment of the invention, the preceding steps have completed the construction of a digital twin model of the target tunnel, and a graph neural network model has been used to perform spatiotemporal fusion processing on multi-source monitoring data to obtain the structural anomaly probability output value and structural health index output value corresponding to each monitoring location. To ensure that the output results of the artificial intelligence model can be fully utilized in engineering maintenance and safety management, this invention further feeds these two outputs back to the digital twin model to drive the prediction of the remaining life of the target tunnel structure, and outputs early warning information when the prediction results meet preset alarm conditions. The structural anomaly probability output value is used to express the probability of an abnormal trend appearing in the tunnel lining structure in space, and the structural health index output value is used to express the level of degradation of the structural material performance in the time dimension. Most existing technologies process risk assessment and life estimation separately, or use only a single type of data for calculation, making it difficult to achieve automated fusion and utilization. This invention, by establishing a unified symbolic expression system and a joint recursive calculation mechanism, enables the structural anomaly probability output value and the structural health index output value to be synergistically applied in the same digital twin model, thereby forming a complete closed loop in the execution path.

[0063] Specifically, the digital twin model includes a cross-sectional risk attribute port to receive structural anomaly probability output values. Assume the target tunnel, in its current service state, contains multiple effective monitoring locations. Then the overall form of the structural anomaly probability output value is: ; in For the first The time-series vector of structural anomaly probability output values ​​at multiple consecutive time points at each monitoring location is used to express the change process of the anomaly probability over time. To facilitate the quantification of anomaly probability during lifetime prediction, this invention further uses the structural anomaly probability output value at the current recursive time point... Corresponding component values This is transformed into a cross-sectional risk impact weight parameter. Let the impact weight formed by the cross-sectional risk attribute port be... Its purpose is to label the spatial distribution characteristics of anomaly probabilities within the digital twin model, weighing their impact on remaining life prediction. To ensure that this weight parameter is consistent with the output value of structural anomaly probability in direction, it can be expressed using the following mapping relationship: ; in The preset normal fluctuation level parameter is used to limit the scale of the impact of abnormal probability on lifespan reduction; The data is directed to participate in the joint recursive calculation process within the digital twin model, giving higher-risk monitoring locations a greater weight in lifetime evolution analysis. This process ensures that the structural anomaly probability output value has a clear destination, eliminating reliance on subjective scoring or manual assignment by maintenance personnel.

[0064] Simultaneously, the present invention also feeds back the structural health index output value to the material health attribute port in the digital twin model, which is used to modulate and update the lining stiffness deterioration characteristic parameters corresponding to each monitoring location. Let the overall form of the structural health index output value be: ; in For the first The time-series vector of structural health index output values ​​at multiple consecutive time points from each monitoring location is used to express the long-term change process of structural health status over time. Let the first... Each monitoring location at time The lining stiffness degradation coefficient is The purpose of the material health attribute port is to update the parameters of the digital twin model online based on changes in the structural health index output value. To ensure that the logical direction of increased material degradation due to a decrease in the health index is consistent, this invention can use an exponential modulation relationship to update the stiffness degradation coefficient: ; in For the first The initial stiffness reference parameters corresponding to each monitoring location are used as the initial starting point for the recursive calculation of stiffness degradation. The preset modulation coefficient parameter is used to control the sensitivity of the structural health index output value to the material stiffness degradation coefficient. Used to determine the material degradation status at this monitoring location in lifetime evolution analysis.

[0065] After inputting the structural anomaly probability output value and the structural health index output value into their respective ports, this invention performs joint recursive calculations in the digital twin model to obtain the predicted remaining life of the target tunnel structure under its current service condition. Let the... Each monitoring location in time The predicted remaining lifespan at time t is Then its recursive update rule is: ; in The risk deduction coefficient parameter is used to express the influence intensity of the cross-sectional risk influence weight formed after the structural anomaly probability output value is fed back on the lifetime evolution. This is a material degradation deduction parameter, used to express the intensity of the influence of the material stiffness degradation coefficient formed after the structural health index output value is fed back on the life evolution. It can be used for subsequent predictions or directly read by maintenance personnel, so that the parameters of the digital twin model are updated online as the output value of the structural health index changes.

[0066] For the time continuity determination of the structural anomaly probability output value, the solution of this invention is expressed as follows: ; in For the first Each monitoring location in time The anomaly probability change rate parameter at time t is used to determine the continuous upward trend of the structural anomaly probability output value over time.

[0067] When the following conditions are met, an early warning message is output: If the structural anomaly probability output value at any monitoring location shows a continuous upward trend within a time window and exceeds the preset normal fluctuation level parameter, the alarm condition is met and an early warning message is output; If the structural health index output value at any monitoring location shows a sudden downward trend within a time window and exceeds the average change level parameter of adjacent monitoring locations, the alarm condition is met and an early warning message is output; If multiple spatially adjacent monitoring locations simultaneously meet the anomaly probability alarm condition or the health index alarm condition at the same time, they are combined into one alarm event and an early warning message is output.

[0068] The spatial adjacency set The parameters of the adjacency matrix of twin nodes in the digital twin topology framework are used to ensure the continuous correspondence of early warning information in spatial location.

[0069] like Figure 3As shown, this invention provides a tunnel structure health assessment system, comprising: a model building unit for collecting design and operational information of a target tunnel and constructing a digital twin model of the target tunnel based on the design and operational information; a data acquisition unit for acquiring multi-source monitoring data, including stress data, displacement data, and image data, through automated monitoring equipment installed within the target tunnel; a processing unit for inputting the multi-source monitoring data into a graph neural network model for spatiotemporal feature fusion processing to output a tunnel structure anomaly probability and a structural health index; and an early warning unit for feeding back the tunnel structure anomaly probability and structural health index to the digital twin model to predict the remaining life of the target tunnel structure and outputting early warning information when the prediction results meet preset alarm conditions.

[0070] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor A01, a network interface A02, memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The network interface A02 is used for communication with external terminals via a network connection. When the computer program B02 is executed by the processor A01, it implements a tunnel structure health assessment method.

[0071] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0072] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0073] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A method for assessing the health of tunnel structures, characterized in that, The method includes: Collect design and operational information of the target tunnel, and construct a digital twin model of the target tunnel based on the design and operational information; Multi-source monitoring data is acquired by automated monitoring equipment installed in the target tunnel. The multi-source monitoring data includes stress data, displacement data, and image data. The multi-source monitoring data is input into a graph neural network model for spatiotemporal feature fusion processing to output the tunnel structure anomaly probability and structural health index. The probability of tunnel structure anomaly and the structural health index are fed back to the digital twin model to predict the remaining life of the target tunnel structure, and an early warning message is output when the prediction result meets the preset alarm conditions.

2. The tunnel structure health assessment method according to claim 1, characterized in that, A digital twin model of the target tunnel is constructed based on design and operational information, including: Based on the design information, the target tunnel lining structure is discretized into multiple twin nodes, and the three-dimensional geometric connection relationship and boundary constraint relationship between the twin nodes are established to obtain the initial digital twin topology framework of the target tunnel. Extract the timing load parameters and timing response parameters of each twin node from the operation information, and write the timing load parameters and timing response parameters into the node attribute interface; Based on the deviation analysis rules between the design information and the operational information, the parameters of the digital twin model are incrementally updated and corrected to obtain a digital twin model that changes with the operational information.

3. The tunnel structural health assessment method according to claim 2, characterized in that, Establish the three-dimensional geometric connection relationships and boundary constraint relationships between twin nodes to obtain the initial digital twin topology framework of the target tunnel, including: Based on the tunnel structure segment drawings contained in the design information, the target tunnel is divided into cross-sectional units along the route, and each lining ring or lining section is abstracted into twin node objects. Based on the node spatial location parameters given in the design information, a twin node adjacency matrix is ​​automatically generated using a three-dimensional spatial proximity discrimination rule to establish quantifiable geometric connection relationships between nodes. Based on the monitoring equipment installation status and sampling timing status contained in the operation information, node timing consistency verification rules are constructed, and online filtering and reconstruction are performed on the twin node adjacency matrix to form a dynamic boundary constraint relationship that matches the current operation status. The reconstructed twin node adjacency matrix and dynamic boundary constraint relationship are jointly used in the initial digital twin topology framework generation process to obtain an initial digital twin topology framework that reflects the true geometric characteristics and actual operating conditions of the target tunnel.

4. The tunnel structure health assessment method according to claim 1, characterized in that, Multi-source monitoring data is acquired through automated monitoring equipment installed within the target tunnel, including: Based on stress sensors and laser ranging sensors deployed at multiple pre-set key structural parts of the target tunnel, and in accordance with the sampling control rules given in the operation information, time-series stress monitoring data and time-series displacement monitoring data for each key structural part are obtained respectively. The inspection robot periodically collects images of the target tunnel lining structure. Based on the structural coordinate positioning rules contained in the design information, the structural appearance image data is bound to the spatial position of the corresponding key structural parts. A unified clock synchronization mechanism is adopted for different types of automated monitoring equipment. Based on the timestamp generation rules set in the operation information, the time-series stress monitoring data, time-series displacement monitoring data and structural appearance image data are time-aligned to form a multi-source data set with a consistent time reference. According to the data format verification rules agreed in the design and operation information, online validity judgment and standardization processing are performed on the multi-source data set to obtain multi-source monitoring data.

5. The tunnel structure health assessment method according to claim 1, characterized in that, The graph neural network model includes: A dual-mode input layer consisting of an image encoder and a physical quantity encoder is used to receive multi-source monitoring data and generate node feature vectors. A spatiotemporal fusion backbone composed of multiple graph convolutional layers and temporal convolutional layers serves as a standard processing structure for graph neural networks; among them, A structural constraint modulation layer is set after the spatiotemporal fusion backbone. The structural constraint modulation layer is composed of a constraint embedding port and a feature modulation unit connected in series to receive geometric constraint parameters determined by design information and boundary constraint parameters determined by operation information, and to splice and fuse them with node feature vectors to perform channel-level weighted modulation on the channel features of each node. The dual output head, which is set after the structural constraint modulation layer, includes a health index regression branch and an anomaly probability output branch, and is used to output the node-level structural health index and node-level anomaly probability value of the target tunnel respectively based on the node feature vector after spatiotemporal fusion.

6. The tunnel structural health assessment method according to claim 5, characterized in that, The multi-source monitoring data is input into a graph neural network model for spatiotemporal feature fusion processing to output the tunnel structure anomaly probability and structural health index, including: At the node feature input end of the graph neural network, time-series filtering and normalization processing are performed on the stress data and displacement data in the multi-source monitoring data to obtain a standardized feature sequence of physical quantities. Simultaneously, convolutional feature extraction and feature compression processing are performed on the image data in the multi-source monitoring data to obtain an image coding feature sequence corresponding to the standardized feature sequence of physical quantities. The standardized feature sequence of the physical quantity and the image coding feature sequence are aligned at the node dimension with the same timestamp to obtain a node-level aligned fused feature sequence. Multi-round graph convolution aggregation and temporal feature regression are performed on the node-level aligned and fused feature sequences to form a stable spatiotemporal joint output result of nodes; Based on the spatiotemporal joint output results of the nodes, the structural anomaly probability output value and structural health index output value corresponding to each monitoring location are generated respectively.

7. The tunnel structural health assessment method according to claim 6, characterized in that, Based on the spatiotemporal joint output results of the nodes, the structural anomaly probability output value and structural health index output value for each monitoring location are generated, including: The spatiotemporal joint output results of the nodes are parsed by node dimensions to obtain the spatiotemporal output vectors of the nodes corresponding to each monitoring location; The abnormal feature components in the spatiotemporal output vector of the node are input into the probability mapping channel to generate continuous numerical structural anomaly probability output values ​​for each node. Input the health feature components in the spatiotemporal output vector of the node into the regression mapping channel to generate a continuous exponential structural health index output value for each node. The structural anomaly probability output values ​​at multiple consecutive time points are concatenated along the time dimension, and the concatenation result is used as the structural anomaly probability output value. The structural health index output values ​​at multiple consecutive time points are concatenated along the time dimension, and the concatenated result is used as the structural health index output value.

8. The tunnel structural health assessment method according to claim 7, characterized in that, The probability of tunnel structure anomalies and the structural health index are fed back into the digital twin model to predict the remaining life of the target tunnel structure, including: The tunnel structure anomaly probability is input into the cross-sectional risk attribute port in the digital twin model to label the influence weight of the spatial distribution characteristics of the anomaly probability in the remaining life prediction stage. The structural health index is input into the material health attribute port of the digital twin model to modulate the lining stiffness deterioration characteristic parameters corresponding to each monitoring location. In the digital twin model, a joint recursive calculation is performed on the input anomaly probability and health index to obtain the predicted remaining life of the target tunnel under its current service condition.

9. The tunnel structure health assessment method according to claim 8, characterized in that, When the prediction result meets the preset alarm conditions, an early warning message is output, including: When the output value of the structural anomaly probability at any monitoring location continuously increases and exceeds the normal fluctuation level at multiple consecutive moments, the prediction result is determined to meet the preset alarm conditions, and an early warning message is output. When the output value of the structural health index at any monitoring location suddenly decreases at multiple consecutive times and exceeds the average change level of adjacent monitoring locations, the prediction result is determined to meet the preset alarm conditions, and an early warning message is output. When multiple spatially adjacent monitoring locations simultaneously meet the preset alarm conditions at the same time, the judgment results of the multiple monitoring locations are combined into a prediction result that meets the preset alarm conditions, and a warning message is directly output.

10. A tunnel structure health assessment system, characterized in that, The system includes: The model building unit is used to collect design and operational information of the target tunnel and build a digital twin model of the target tunnel based on the design and operational information. The acquisition unit is used to acquire multi-source monitoring data through automated monitoring equipment installed in the target tunnel. The multi-source monitoring data includes stress data, displacement data, and image data. The processing unit is used to input the multi-source monitoring data into a graph neural network model for spatiotemporal feature fusion processing, so as to output the tunnel structure anomaly probability and structural health index. The early warning unit is used to feed back the probability of tunnel structure anomaly and the structural health index to the digital twin model to predict the remaining life of the target tunnel structure, and output early warning information when the prediction results meet the preset alarm conditions.

Images