A civil engineering structure health monitoring system

Abstract

The application relates to the technical field of civil engineering monitoring and discloses a civil engineering structure health monitoring system. A high-density sensor network of the system collects structure strain field distribution, vibration frequency spectrum and environmental corrosion parameters through a distributed multi-modal sensor array; a time-space alignment engine performs clock drift compensation and space coordinate normalization on asynchronous sampling data to generate original feature tensors of a unified time-space reference; a semantic modeling unit maps the original feature tensors to component semantic space in combination with design drawings and a material parameter library to output hierarchical structure features with topological markers; an adaptive fusion core performs dynamic weight distribution to eliminate sensor conflict data and generate anti-interference fusion diagnosis indexes; an edge computing node runs a lightweight damage detection model according to the fusion diagnosis indexes to output a local component health state grade; and a cloud-side collaborative analyzer aggregates multi-edge node results to predict the overall remaining life of the structure in combination with historical degradation data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of civil engineering monitoring, in particular to a civil engineering structure health monitoring system. BACKGROUND

[0002] In the field of civil engineering, large structures such as bridges, tunnels, and high-rise buildings may be affected by various factors such as load, environmental erosion, and material aging during long-term service, leading to gradual degradation of structural performance and potential safety hazards or accidents. Therefore, effective health monitoring of civil engineering structures to timely grasp the changes in structural state has become an important means to ensure the safe operation of structures.

[0003] In existing civil engineering structure health monitoring technology, sensors are usually arranged in a single point or sparse distribution, making it difficult to fully capture the overall state information of the structure. Some monitoring systems only use a single type of sensor, which can only obtain single-dimensional data such as strain or vibration, and cannot comprehensively reflect the health status of the structure under the action of multiple factors. At the same time, due to the differences in sampling frequency and data format of different types of sensors, the asynchronous sampling data collected lack a unified time and space reference, making it difficult to effectively correlate the data, and subsequent data processing and analysis are difficult.

[0004] In the data processing link, existing technologies often directly analyze the original monitoring data without fully combining the design information and material properties of the structure, making it difficult to accurately reflect the health status of the specific component. In addition, the data collected by different sensors may have conflicts or errors, and existing fusion algorithms mostly use fixed weight distribution methods, which are difficult to dynamically adjust according to data reliability, resulting in weak anti-interference ability of the fused diagnostic indicators and difficulty in accurately identifying structural damage.

[0005] In terms of monitoring result application, existing systems mostly focus on local component damage detection and lack the ability to predict the overall remaining life of the structure. Some systems attempt to perform overall analysis, but do not fully integrate the results of multiple local monitoring nodes, nor do they establish a scientific prediction model based on historical degradation data, making the evaluation of the overall safety state of the structure not comprehensive enough and difficult to meet the needs of long-term safety management of structures in engineering practice. SUMMARY

[0006] The present application aims to provide a civil engineering structure health monitoring system to solve the problems raised in the background.

[0007] To achieve the above-mentioned purpose, the present application provides a civil engineering structure health monitoring system, which comprises:

[0008] A high-density sensor network is used to collect the strain field distribution, vibration spectrum and environmental corrosion parameters of the structure through a distributed multi-modal sensor array;

[0009] A space-time alignment engine is used to compensate for clock drift and normalize spatial coordinates of the asynchronous sampling data of the multi-modal sensor array, so as to generate an original feature tensor with a unified space-time reference;

[0010] A semantic modeling unit is used to map the original feature tensor to a preset component semantic space according to design drawings of the structure and a material parameter library, and output hierarchical structure features with topological markers;

[0011] An adaptive fusion core is used to receive the hierarchical structure features and perform dynamic weight distribution, eliminate conflicting data between sensors through an iterative optimization algorithm, and generate an anti-interference fusion diagnosis index;

[0012] An edge computing node is used to run a lightweight damage detection model in real time according to the fusion diagnosis index, and output a local component health state grade;

[0013] A cloud collaborative analyzer is used to aggregate output results of multiple edge computing nodes, and predict the overall remaining life of the structure in combination with historical degradation data.

[0014] Preferably, the high-density sensor network comprises:

[0015] A strain field collection subnetwork is used to measure micro-strain on the surface of the structure by using a fiber Bragg grating sensor array, and calculate the principal stress direction by using a Bragg wavelength shift;

[0016] A vibration perception subnetwork is used to capture the natural frequency and damping ratio of the structure by using a piezoelectric accelerometer array, and extract frequency band energy features by using a wavelet packet decomposition;

[0017] A corrosion monitoring subnetwork is used to detect the chloride ion concentration in concrete and the steel reinforcement polarization resistance by using an electrochemical impedance sensor.

[0018] Preferably, the space-time alignment engine performs:

[0019] Lagrange interpolation compensation is performed on the wavelength sampling sequence of the fiber Bragg grating sensor array to eliminate time offset caused by transmission delay;

[0020] A sensor spatial position relationship diagram is established, and a local coordinate system of the piezoelectric accelerometer array is converted to a global reference system through Lie algebra transformation;

[0021] Sliding window Kalman filtering is used to perform time synchronization calibration on slow-varying signals of the electrochemical impedance sensor.

[0022] Preferably, the semantic modeling unit comprises:

[0023] A component segmentation module divides the original feature tensor into feature blocks corresponding to physical components according to beam-column joint information in the design drawing;

[0024] A material property binding module extracts concrete grade and steel bar specification from a material parameter library, and adds elastic modulus and yield strength labels to each feature block;

[0025] A topological relationship encoder generates a connection stiffness matrix between components according to a structural mechanics model and embeds it into the hierarchical structural features.

[0026] Preferably, the adaptive fusion core executes:

[0027] A conflict detection model based on Mahalanobis distance is constructed to identify abnormal data blocks in the hierarchical structural features that exceed the statistical confidence interval;

[0028] The reliability weight of multi-modal sensor data is estimated by a Gaussian mixture model to weight and reconstruct the abnormal data blocks;

[0029] A variational autoencoder is used to project the reconstructed features onto a low-dimensional manifold to generate a fusion diagnostic index.

[0030] Preferably, the edge computing node includes:

[0031] A micro-crack detection unit analyzes the strain concentration coefficient in the fusion diagnostic index through a deep separable convolutional network;

[0032] A stiffness degradation evaluation unit calculates the equivalent stiffness decay rate of the component according to the vibration spectral features and the connection stiffness matrix;

[0033] A corrosion risk quantification unit predicts the steel bar corrosion rate based on the chloride ion diffusion equation and the polarization resistance value.

[0034] Preferably, the cloud-side collaborative analyzer executes:

[0035] A global finite element model of the structure is constructed, and the local component health state level is input as a boundary condition;

[0036] A time series prediction network is used to simulate the stress redistribution process under different load conditions;

[0037] The structural system reliability index is calculated by Monte Carlo sampling to output the residual life confidence interval.

[0038] Preferably, the system further includes:

[0039] A dynamic update interface automatically corrects the material constitutive relationship of the finite element model according to newly collected hierarchical structural features;

[0040] A feedback regulation loop compares the residual error between the predicted strain field and the actual measured value to adjust the convolution kernel parameters of the depth-wise separable convolution network.

[0041] Preferably, the dynamic updating interface comprises:

[0042] A parameter sensitivity analysis module identifies the material parameters that have the greatest impact on structural reliability.

[0043] An incremental learning engine updates the Young's modulus matrix of the finite element model through an online backpropagation algorithm.

[0044] Preferably, the feedback regulation loop comprises:

[0045] A residual error distribution statistics module calculates the root mean square value of the prediction error of each component node.

[0046] An adaptive learning rate scheduler dynamically adjusts the update step size of the convolution kernel parameters according to the root mean square value.

[0047] Compared with the prior art, the present application has the following advantages:

[0048] The high-density sensor network uses a distributed multi-modal sensor array, which can simultaneously collect the strain field distribution, vibration spectrum, and environmental corrosion parameters of the structure, breaking the limitations of traditional monitoring systems with single sensors or sparse arrangement. It can obtain the state information of the structure from multiple dimensions and comprehensively reflect the health status of the structure under the influence of different factors, making the monitoring data cover a wider range and provide more comprehensive data sources for subsequent accurate analysis of the structure health status.

[0049] The space-time alignment engine compensates for clock drift and normalizes spatial coordinates for the asynchronous sampling data of the multi-modal sensor array, generating an original feature tensor with a unified space-time reference. This process solves the problem of data correlation caused by differences in sampling frequency and data format between different sensors, making the originally scattered and asynchronous data form a unified standard feature tensor and eliminating the deviation of data in the time and space dimensions. It lays a good foundation for data transmission and collaborative analysis between subsequent modules, making the data processing of the entire monitoring process more coherent and accurate.

[0050] The semantic modeling unit maps the original feature tensor to the preset component semantic space according to the design drawings of the structure and the material parameter library, and outputs the hierarchical structure features with topological labels. This unit fully combines the design information and material properties of the structure, corresponds the abstract original feature tensor to the specific structural component, gives the monitoring data a clear physical meaning and component attribution, enables the subsequent damage detection and health assessment to be accurately positioned to the specific component, avoids the problem of disconnection between data and components in traditional monitoring, makes the monitoring results more targeted and practical, and facilitates engineers to intuitively understand the health status of each component.

[0051] The adaptive fusion core receives the hierarchical structure features and performs dynamic weight distribution, eliminates the conflicting data between sensors through an iterative optimization algorithm, and generates an anti-interference fusion diagnosis index. Compared with the traditional fixed weight fusion algorithm, the core can dynamically adjust the weight according to the reliability of different sensor data, effectively identify and eliminate conflicting data, minimize the influence of data errors and interference on the diagnosis results, and generate a fusion diagnosis index that more truly reflects the health status of the structure, thereby improving the credibility of the diagnosis results and providing high-quality data support for subsequent damage detection of the edge computing node.

[0052] The edge computing node runs a lightweight damage detection model in real time according to the fusion diagnosis index, and outputs the health status grade of the local component. The application of the lightweight model ensures the real-time performance of damage detection, can quickly respond to changes in the structure state, and timely output the health grade of the local component, so that engineers can first understand the health status of each local component, facilitate timely investigation and processing of components with hidden dangers, effectively avoid safety risks caused by delayed detection, and improve the timeliness and efficiency of structure safety management.

[0053] The cloud-side collaborative analyzer aggregates the output results of multiple edge computing nodes, and predicts the overall residual life of the structure in combination with historical degradation data. This module realizes the global integration of local monitoring results, is no longer limited to the evaluation of a single local component, but starts from a global perspective, combines the long-term historical degradation data of the structure, and predicts the overall residual life of the structure, thereby providing comprehensive reference for engineers to develop long-term structure maintenance and repair plans, helping to reasonably arrange resources, prolong the service life of the structure, ensure the long-term safe and stable operation of the structure, and reduce economic losses and safety accident risks caused by blind maintenance or neglect of hidden dangers. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A timing diagram of the civil engineering structure health monitoring system described in the present application;

[0055] Figure 2 A workflow diagram of the high-density sensor network;

[0056] Figure 3 Workflow diagram for the spatio-temporal alignment engine. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the 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 work fall within the protection scope of the present application.

[0058] Please refer to Figure 1 The present application provides a civil engineering structure health monitoring system based on a collaborative work of a high-density sensor network, a spatio-temporal alignment engine, a semantic modeling unit, an adaptive fusion core, an edge computing node and a cloud collaborative analyzer.

[0059] A high-density sensor network collects the strain field distribution, vibration spectrum and environmental corrosion parameters of the structure through a distributed multi-modal sensor array, which includes fiber Bragg grating sensors, piezoelectric accelerometers and electrochemical impedance sensors to cover the key physical parameters of the structure. A space-time alignment engine receives the asynchronous sampling data of the multi-modal sensor array, performs clock drift compensation and spatial coordinate normalization processing, where the clock drift compensation aligns the time series using the Lagrange interpolation method, and the spatial coordinate normalization maps the local coordinate system to the global reference system through Lie algebra transformation, finally generating the original feature tensor with a unified space-time reference. A semantic modeling unit maps the original feature tensor to the pre-set component semantic space according to the design drawings and material parameter library of the structure; this unit divides the original feature tensor into feature blocks corresponding to physical components by analyzing the geometric information in the design drawings, and extracts attributes such as concrete grade and steel specification from the material parameter library, adds labels such as elastic modulus and yield strength to each feature block, and generates the connection stiffness matrix between components according to the structural mechanics model, and outputs hierarchical structural features with topological labels. An adaptive fusion core receives hierarchical structural features and performs dynamic weight distribution to eliminate conflicting data between sensors; the core detects abnormal data blocks through the Mahalanobis distance model, estimates the reliability weight of each sensor data using the Gaussian mixture model, and then uses the variational autoencoder for feature reconstruction to generate anti-interference fusion diagnostic indicators. An edge computing node runs real-time lightweight damage detection models, including micro-crack detection, stiffness degradation evaluation and corrosion risk quantification modules, based on the fusion diagnostic indicators, these models analyze the local component state based on deep separable convolutional networks and physical equations, and output the health status level. A cloud-side collaborative analyzer aggregates the results of multiple edge computing nodes, constructs a global finite element model of the structure combined with historical degradation data, simulates stress redistribution through a time series prediction network, and calculates the system reliability using Monte Carlo sampling, finally outputs the confidence interval of the remaining life. The system realizes adaptive optimization through a dynamic update interface and a feedback adjustment loop, the dynamic update interface corrects the constitutive relationship of the finite element model according to new data, and the feedback adjustment loop adjusts the model parameters through residual comparison to ensure long-term monitoring accuracy.

[0060] Example 1: see Figure 2, the strain field acquisition subnetwork is deployed with fiber grating sensor arrays, which are attached to the concrete surface or embedded in the structure surface in a grid mode. Each sensing unit is made of a Bragg grating by ultraviolet laser inscription technology, and the grating period is optimized according to the preset measurement range. The array wiring adopts a hybrid architecture of star topology and bus topology, with the backbone fiber laid along the main stress path of the structure, and the branch nodes connected to individual sensing units to reduce signal attenuation. During data acquisition, the demodulation equipment emits a broadband light source, and the wavelength shift of the reflected Bragg wave by the sensor is linearly related to the strain change. The wavelength demodulator records the shift and converts it into a micro-strain value. The principal stress direction calculation combines a flower-shaped sensor group composed of strain gauges in multiple directions to analyze the plane stress state through the Mohr circle principle, and finally generate a strain cloud map covering the structure surface. The piezoelectric accelerometer array of the vibration perception subnetwork selects a shear type three-axis sensor, and the installation position is determined through modal analysis to cover the main mode nodes of the structure. The sensor base is fixed with epoxy resin glue to ensure synchronous vibration with the structure, and the signal cable is shielded from electromagnetic interference by a metal tube. The acquisition system is set to an adjustable sampling frequency, usually 200 Hz for bridge structures to capture modal information in the range of 0-100 Hz. The wavelet packet decomposition algorithm uses the Mallat fast algorithm to realize eight-layer decomposition of the acceleration signal to obtain 256 sub-bands, and calculates the energy value of each sub-band and normalizes it to form a feature vector. The correlation between the frequency band energy features and the structure damage is mapped through the historical database, for example, a local stiffness decrease will cause the high-frequency band energy proportion to rise. This system identifies abnormal vibration patterns by tracking these feature changes in real time.

[0061] The electrochemical impedance sensor of the corrosion monitoring subnetwork adopts a three-electrode system, with a pre-embedded stainless steel rod as the working electrode, an Ag / AgCl electrode as the reference electrode, and a platinum wire electrode as the counter electrode. The sensor is pre-embedded in the protective layer at different depths before concrete pouring, and the electrode spacing is set to 2-5 cm according to the size of the reinforcement grid. The excitation signal is generated by a constant potential instrument, with a frequency scanning range from 10 mHz to 100 kHz. The Nyquist plot obtained by measurement is fitted with an equivalent circuit model by ZView software. The chloride ion concentration is calculated according to the diameter of the diffusion impedance arc in the impedance spectrum, and the polarization resistance is derived from the real part value of the impedance in the low frequency region. The temperature compensation module integrates a PT100 temperature sensor to correct the temperature effect on electrochemical parameters in real time. The power supply system of the sensor network uses a hybrid scheme of solar panels and super capacitors, and the wireless transmission module uses a mixed communication protocol of 4G / LTE network and LoRa. The data acquisition terminal is equipped with edge computing capability, which can perform preliminary filtering and compression processing on raw data. The network management system includes an automatic diagnosis function, which periodically checks the sensor connection state and battery power, and abnormal nodes will trigger an alarm signal. Clock synchronization uses the IEEE1588 precision time protocol, and all sensor nodes achieve microsecond-level time synchronization through the master clock. The wavelength demodulation equipment of the fiber Bragg sensor array uses tunable Fabry-Perot filter technology, with a scanning frequency of 100 Hz and a wavelength resolution of 1 pm. Each sensing channel is equipped with an optical switch for round-robin detection, and the optical path attenuation is monitored in real time to ensure data reliability. Temperature compensation is achieved by setting a reference grating, which is packaged separately in a stress-free area. Strain transfer efficiency is optimized through finite element analysis, and the sensor packaging structure adopts a multi-layer protection design to reduce shear hysteresis effect. The signal conditioning circuit of the piezoelectric accelerometer includes a charge amplifier and an anti-aliasing filter, and the output voltage range is adjusted to ±5V to adapt to the range of the acquisition card. The sensor sensitivity coefficient is stored in the embedded chip, and the system automatically reads and converts it to physical quantities. The array layout adopts a strategy of combining uniform point distribution with increased density at key positions, such as around the opening and near the support. The modal identification algorithm uses the stochastic subspace method to automatically extract the first 10 order frequency and damping ratio parameters.

[0062] The potentiostat circuit of the electrochemical impedance sensor is designed with a zero resistance ammeter to eliminate the influence of wire resistance on measurement results. Shielded twisted pair is used for electrode leads to reduce AC interference. Changes in contact impedance caused by concrete drying shrinkage are compensated by regular calibration. The monitoring period is set according to environmental aggressiveness, with daily measurements in splash zones and weekly measurements in atmospheric zones. Data validity verification is achieved through repeated measurements, and automatic retesting is performed when the coefficient of variation exceeds 5%. The network communication protocol uses a custom binary format to compress data packets, with the packet header containing a timestamp, node ID, and data type identifier. The routing algorithm dynamically selects the transmission path based on signal strength, and switches communication modes when the packet loss rate exceeds the threshold. The data security mechanism uses the AES encryption algorithm, and each sensor node has an independent key. The remote configuration function allows adjustment of sampling parameters, and the system maintains a double buffer area to enable parallel processing of acquisition and transmission. The self-calibration mechanism is implemented through a built-in standard signal source, which periodically injects a known amplitude electrical signal to verify channel gain. Sensor health status monitoring includes impedance detection and baseline drift analysis, and abnormal sensor data is marked and excluded from fusion calculations. Network coverage optimization is based on ray tracing algorithms, and relay nodes are deployed inside complex steel structures to eliminate signal shadows. Electromagnetic compatibility design includes feedthrough filters and shielded housings to ensure stable operation in harsh industrial environments.

[0063] Example 2: see Figure 3, the spatio-temporal alignment engine and the semantic modeling unit component constitute the key bridge from raw data to engineering semantics, the spatio-temporal alignment engine processes heterogeneous data streams from high-density sensor networks, its primary task is to time-align the asynchronous sampling sequence of the fiber Bragg grating sensor array; the built-in crystal oscillator in each sensor node has a slight frequency offset, resulting in clock drift, the engine takes the standard time provided by the Beidou timing module as the reference, and reconstructs the equally spaced sampling points through the Lagrange interpolation method: the interpolation algorithm adopts a three-point interpolation mode, constructs a quadratic polynomial function with the adjacent three original data points before and after the target time, and dynamically fits the theoretical wavelength value at that time, this processing effectively eliminates the millisecond-level time jitter caused by transmission path differences. For the problem of spatial coordinate unification of the piezoelectric accelerometer array, the engine reads the pre-stored sensor layout coordinate table, which records the three-dimensional coordinates of each accelerometer in the global coordinate system and the orientation angle of the local coordinate system; through the exponential mapping relationship in the Lie group Lie algebra theory, the acceleration vector in the local coordinate system is rotated to the global direction, and the transformation matrix is calculated from the rotation vector corresponding to the sensor installation posture, ensuring that the vibration data at different positions can be vector superimposed and compared. The slow-varying signal of the electrochemical impedance sensor is synchronized using a sliding window Kalman filter, the filter state variable is set to a two-element group of chloride ion concentration and polarization resistance, the observation equation is established based on the sensor output, and the process noise covariance is adaptively adjusted according to the historical data statistics; the filter advances at a fixed time step, and the prediction-update cycle is independently run for each sensor channel, finally outputting the parameter sequence with strictly aligned timestamps.

[0064] The work of the semantic modeling unit begins with semantic analysis of the original feature tensor, the component segmentation module loads the BIM model or CAD design drawing of the structure, and analyzes the geometric information of the entity components therein; the module extracts the geometric boundaries of components such as beams, columns and plates through the boundary representation method, generates a three-dimensional bounding box as a spatial index, and automatically clusters and merges the sensor measurement points falling within the same bounding box to the corresponding component. The material attribute binding module connects the material parameter database, which stores parameters such as concrete strength grade, steel type, and protective layer thickness in the form of a relational table; the module associates with the database through component numbers, and adds material constitutive parameters to each component feature block. For example, add elastic modulus, Poisson's ratio and compressive strength labels to concrete components, and add yield strength and fatigue limit labels to steel components. The topological relationship encoder constructs the connection relationship based on the stiffness method principle in structural mechanics, which generates a topological graph between components according to the node connection information in the design drawing; the nodes in the graph represent the centers of gravity of the components, the edges represent the connection relationships, and the weights of the edges are assigned by the inverse of the connection stiffness, finally forming a sparse stiffness matrix and converting it into an adjacency list format embedded feature data.

[0065] The implementation of the spatio-temporal alignment engine contains detailed exception handling mechanisms. When a sensor data stream interruption is detected, the engine initiates a data reconstruction procedure based on the spatio-temporal Kriging interpolation method to estimate the missing data using the measurements of the surrounding normal sensors. The interpolation weights are determined by the distance function and the correlation coefficient. For the coordinate deviation caused by sensor layout errors, the system provides a manual calibration interface that allows engineers to input actual measurement coordinates to override the initial settings. The clock synchronization protocol uses an improved IEEE 1588v2 version, which reduces the delay introduced by the software stack through hardware timestamping. The master clock and each acquisition node periodically exchange bidirectional time messages, calculate and compensate for transmission delays.

[0066] The material parameter library of the semantic modeling unit uses a version management mechanism. When the structure is reinforced and repaired, a new version of the material parameters can be created and the effective time is marked. The component segmentation algorithm supports multi-scale division. For large continuous bodies such as box girder structures, they can be further divided into several virtual sub-components to improve analysis accuracy. The topology encoder also considers nonlinear connection cases. For special components such as seismic isolation bearings, their connection stiffness is expressed as a function of displacement rather than a fixed value. The unit output uses the standard JSON-LD format, and each feature block contains spatial coordinates, material properties, and topological relationships. The triple semantic labels form a machine-readable engineering information model. In the specific data processing pipeline, the original sensor data packet first enters the time alignment queue. The queue manager sorts the data packets according to the timestamps in the data packets, and triggers the interpolation recalculation process for late data. The spatial coordinate transformation module uses quaternions instead of Euler angles to avoid gimbal lock problems. The transformed data is uniformly converted to the commonly used North-East-Down coordinate system. The semantic mapping process uses a parallel processing architecture. Different component feature extraction tasks are assigned to multiple computing cores for synchronous execution. The data consistency check module verifies the geometric relationship logic, such as the coordinates of the common nodes of connected components must be consistent. When conflicts are found, an automatic manual review process is triggered.

[0067] The system also contains a metadata management function to record the parameter configuration and data traceability information of each processing step. The spatio-temporal alignment engine can configure different interpolation strategies, using spline interpolation for vibration data to maintain waveform smoothness, and linear interpolation for slowly varying parameters to reduce computational overhead. The semantic modeling unit supports incremental update mode, when new sensor nodes are added during monitoring, the system can dynamically expand the component feature block without reinitializing the entire model. The stiffness matrix output by the topology encoder uses a block storage format, which facilitates direct calling by subsequent finite element analysis software. When the coordinate information of some sensors is missing, the system will automatically estimate their approximate positions based on signal correlation and prompt for confirmation. The material parameter library has a completeness verification mechanism, and the missing mandatory parameters will prevent the feature block from passing downstream. The time synchronization system establishes a hierarchical time correction structure, with regional master clocks synchronizing a number of sub-domain clocks, and sensors within the sub-domain using a simpler NTP protocol to reduce communication overhead. The data timestamp contains time zone information and leap second flags to ensure the time continuity of long-term monitoring data. The semantic model version and monitoring data establish a bidirectional index, supporting state backtracking analysis at any point in history.

[0068] In embodiment 3, the adaptive fusion core receives hierarchical structural features from the semantic modeling unit, which includes spatio-temporally aligned sensor readings and additional component material properties and topological relationships. The first step of core processing is to build a conflict detection model based on Mahalanobis distance. This model uses a historical data set collected under the reference state of structural health as the reference distribution, calculates the Mahalanobis distance between each current feature vector and the mean of the reference distribution, and compares the distance value weighted by the inverse covariance matrix with a pre-set threshold. When the distance value of a certain feature vector exceeds the critical value determined by the chi-square distribution, the data block is marked as potentially abnormal data. A Gaussian mixture model then estimates the weights of multi-modal sensor data reliability. The model treats each sensor's data stream as a Gaussian distribution component, iteratively solves the mean, covariance, and mixing coefficient of each component using the expectation-maximization algorithm, and determines the weight of each data point based on its posterior probability of belonging to each component. Abnormal data points are assigned lower weight coefficients due to their greater deviation from the main distribution components. A variational autoencoder performs dimensionality reduction on the weighted reconstructed features. The encoder part consists of fully connected layers to compress high-dimensional features into low-dimensional latent variables. The decoder part attempts to reconstruct the denoised features from the latent variables, and its loss function consists of reconstruction error and KL divergence between latent variable distribution and standard normal distribution:

[0069]

[0070] where: represents the variable in the latent space, is the input feature vector, ​is the approximate posterior distribution of the encoder output, is the prior distribution (standard normal distribution), is the conditional likelihood defined by the decoder, denotes the KL divergence, is the expectation operator. By minimizing this loss function, the model learns the low-dimensional manifold structure of the data essence and finally outputs the anti-interference fusion diagnostic indicator.

[0071] The edge computing node is deployed in the cabinet near the structure in the monitoring site, and the micro-crack detection unit adopts a deep separable convolutional network architecture. The network input layer receives the strain field distribution map in the fusion diagnostic indicator. First, the spatial features of each channel of strain data are extracted by a deep convolutional layer, and then the cross-channel information is fused point by point. The network outputs the crack risk probability of each grid unit. The stiffness degradation evaluation unit is calculated based on the vibration frequency spectrum features and the connection stiffness matrix provided by the semantic modeling unit. This unit obtains the theoretical vibration frequency of the structure by solving the generalized eigenvalue problem, compares the measured frequency with the theoretical frequency, and inversely calculates the equivalent stiffness reduction coefficient of each component combined with the mode shape participation coefficient. The corrosion risk quantification unit integrates electrochemical monitoring data and environmental parameters. This unit solves the non-steady-state diffusion equation to simulate the transport process of chloride ions in concrete. The boundary conditions of the equation are determined by the measured surface concentration of the sensor, and the corrosion current density is derived by the Stern-Gerlier formula using the polarization resistance value. Finally, the spatial and temporal distribution prediction of the steel reinforcement corrosion rate is output. The implementation of the adaptive fusion core includes a dynamic weight updating mechanism. After the monitoring system runs for a period of time, the parameters of the Gaussian mixture model need to be recalibrated. The system regularly collects new health state data, updates the distribution parameters of each component through an online learning algorithm, and avoids false alarms caused by natural degradation of the structure performance. The conflict detection module has an adaptive threshold adjustment function. When the environmental temperature changes dramatically, the system temporarily relaxes the threshold condition of Mahalanobis distance to reduce false alarms. The training process of the variational autoencoder uses the mini-batch gradient descent method. Each time the edge computing node is dormant, the network parameters are fine-tuned using local historical data to maintain the adaptability of the model to changes in the structure state.

[0072] The micro-crack detection network of the edge computing node adopts a transfer learning strategy, with initial weights pre-trained on a large number of strain field data generated by finite element simulation. After on-site deployment, it adapts to the actual structure characteristics through online learning. The stiffness degradation evaluation algorithm considers the material nonlinear effect, and automatically switches to the tangent stiffness-based iterative algorithm to improve accuracy when detecting large deformation. The corrosion prediction model integrates environmental temperature and humidity sensor data, dynamically adjusts the temperature correction factor in the diffusion coefficient, and improves the prediction accuracy in different seasons. The data flow is processed on the edge side in strict time sequence control, with a complete data fusion cycle including data quality check, conflict detection, weight calculation and feature reconstruction performed every five minutes. The three evaluation units of the edge computing node work in a pipeline parallel mode, with the micro-crack detection unit passing the data to the stiffness evaluation unit immediately after processing, while receiving the next batch of data to maximize the use of computing resources. All intermediate results are time-stamped and stored in the local ring buffer for on-demand retrieval by the cloud-side collaborative analyzer.

[0073] The system design takes into account the constraints of computing resources, with the latent variable dimension of the variational autoencoder dynamically adjusted according to the complexity of the component, using an 8-dimensional latent space for simple components and a 32-dimensional latent space for complex node regions. The number of layers of the depth separable convolutional network is also configurable, with a lightweight 6-layer network used for regular monitoring and automatically switched to a detailed 12-layer network for detailed analysis when an anomaly is detected. The grid division accuracy of the corrosion prediction model matches the sensor density, ensuring that each grid contains at least one monitoring point. In terms of fault tolerance, when a sensor fails for a long time, the adaptive fusion core automatically adjusts the number of components of the Gaussian mixture model and reallocates the weight coefficients. The edge computing node has a degraded operation mode, which can still make state judgments based on the output of the remaining units when one evaluation unit fails. All algorithm modules are equipped with watchdog timers to monitor execution time, and tasks that exceed the time limit will be terminated and restarted to ensure system real-time performance. The data consistency check module verifies the physical reasonableness of the fused diagnostic indicators, such as the strain concentration coefficient should not exceed the theoretical value corresponding to the material yield limit, and triggers data re-fusion process when abnormal values are found.

[0074] Example 4: Taking a certain long-span steel box girder bridge as the application background, the cloud platform receives the component health state data uploaded from the edge computing nodes of the bridge tower, box girder, and suspender, etc. These data are continuously updated at a five-minute cycle. The cloud collaborative analyzer first builds a finite element model of the overall bridge. The model meshing uses a mixed scheme of tetrahedron and shell elements. The steel box girder part is discretized by shell elements, and the bridge tower is simulated by beam elements. The grid size is set to 0.5 meters according to the calculation accuracy requirement. The local component health state level is input as the boundary condition of the model. For example, when the edge node detects that the stiffness degradation level of a certain deck slab is level two, the elastic modulus of the corresponding region is reduced by 15%; when the corrosion level of the suspender reaches level three, the cross-sectional area of the element is reduced by 10%. The time series prediction network uses the Gated Recurrent Unit architecture to process historical monitoring data. The network input includes the time series of stress, strain, vibration frequency, etc. in the past 30 days, and the output is the stress redistribution under different load cases in the next 7 days. The load case combination includes constant load, highway live load, wind load, and temperature action according to the specification requirements. The network training uses the health monitoring data of the first three years after the bridge is built as the training set, and uses the data enhancement strategy of randomly discarding part of the sensor channels to improve the generalization ability. The Monte Carlo sampling process estimates the structural reliability by generating a random parameter set. In each sampling, random variables such as material strength and load size are generated, and the structural response is calculated through finite element analysis. The frequency of stress exceeding the limit in one hundred thousand samplings is counted, and the confidence interval estimate of the remaining life is finally obtained.

[0075] The dynamic updating interface automatically corrects the finite element model according to the newly collected hierarchical structural features. When systematic deviations are found between the model prediction values and the actual measured values, the interface starts the parameter sensitivity analysis program. This program observes the output changes by perturbing the input parameters of the finite element model, identifies the most sensitive constitutive relationship parameters such as concrete elastic modulus and steel yield strength, etc. The incremental learning engine uses an online learning algorithm to adjust the model parameters. After receiving new monitoring data each time, the engine calculates the residual error between the predicted strain field and the actual measured value, and adjusts the material parameter matrix of the finite element model by gradient descent method. The adjustment amplitude is proportional to the residual error size. Referring to Table 1, the component health state data uploaded by the edge node is shown:

[0076] Table 1: Steel box girder bridge component health state monitoring data record

[0077]

[0078] The feedback regulation loop continuously compares the difference between the finite element model predicted strain distribution and the actual sensor measurements, and the residual distribution statistics module calculates the root mean square error at each measurement point, which is rolled up in units of weeks. When the residual of a certain area is consistently large, the adaptive learning rate scheduler increases the material parameter update step of the corresponding unit in that area, accelerating the model correction process; otherwise, for well-fitted areas, the learning rate is reduced to maintain stability. This regional differentiated learning strategy enables the model to quickly adapt to local changes in structural performance.

[0079] In the specific operation process, the cloud platform performs a complete life prediction analysis every morning. First, it aggregates the component health data from the edge nodes for the past 24 hours, updating the boundary conditions of the finite element model. The time series prediction network is then run to simulate stress redistribution under standard fleet loads, strong wind conditions, and other conditions, focusing on the trend of stress concentration areas. Monte Carlo sampling is performed in parallel on a cloud computing cluster, with each computing node handling one thousand sampling analyses. Finally, the failure probability statistics results from all nodes are aggregated. The dynamic update interface is activated after each life prediction is completed, using the latest monitoring data to correct the model parameters. The corrected model is used for the next round of prediction to form a closed-loop optimization.

[0080] The system has fault-tolerant processing mechanisms for abnormal data. When the data uploaded by a certain edge node is significantly inconsistent with other nodes, the cloud platform will mark the data and start a review process. During the review process, the system temporarily uses historical mean values instead of abnormal values for analysis, and sends a review notification to the operation and maintenance personnel. The model version management function records the detailed information of each parameter adjustment, supporting rollback to any historical version for analysis and comparison. The data visualization interface displays the stress distribution and remaining life prediction results of the structure in the form of a three-dimensional color cloud map, with red areas indicating high-risk areas that need to be focused on. During the operation of the bridge, the system detected stress concentration at the junction of the box girder web and the bottom plate. The finite element model predicted stress levels that were 8% lower than the actual measured values. The dynamic update interface found through parameter sensitivity analysis that the material constitutive relationship in the weld area had a significant impact on the calculation results, so the elastic-plastic model parameters for the units in this area were adjusted. After the adjustment, the deviation between the predicted values and the measured values was reduced to within 2%, and the fatigue life of the area was shortened by about 15% compared to the initial estimate. This dynamic correction capability enables the system to track gradual changes in structural performance, providing more accurate basis for maintenance decisions.

[0081] The system also integrates environmental monitoring data, taking temperature gradient and humidity change as additional loads. During the period of seasonal transition, the temperature change induced structural deformation can significantly affect the stress distribution, and the model considers this effect through coupled thermo-mechanical analysis. Long-term monitoring found that the stress of the suspender generally increases by 3-5% during the high-temperature period in summer, and the system automatically incorporates a temperature compensation factor in the prediction to eliminate the interference of environmental factors on the life assessment. All analysis results generate structured reports and are transmitted to the information system of the maintenance management department through a secure interface. The cloud collaborative analyzer adopts a microservice architecture, with finite element analysis, time series prediction, and reliability calculation modules deployed independently and exchanging data through a message queue. This design allows dynamic expansion of resources according to the computing load and automatically starts more computing nodes when large-scale Monte Carlo sampling is needed. The data storage uses a hybrid scheme of time series database and relational database, with monitoring data stored in the time series database for fast query and analysis results and model parameters stored in the relational database to ensure transaction consistency. The system generates a health diagnosis report every week, automatically labeling components with performance degradation exceeding the limit and their impact on overall safety.

[0082] Example 5: The parameter sensitivity analysis module of the dynamic updating interface employs the variance-based Sobol method for global sensitivity analysis, which generates input parameter combinations through Monte Carlo sampling and observes the change in output response variance; in specific implementation, the module sets the material parameters of the finite element model such as the elastic modulus of concrete, the yield strength of steel reinforcement, and the Poisson's ratio as input variables, each parameter is uniformly sampled within its possible value range, the output variance of the structural reliability index is calculated through multiple finite element analysis, then the first-order sensitivity index and total-order sensitivity index of each parameter are calculated, the closer the index value is to 1, the greater the influence of the parameter on the output result. The Saltelli sampling sequence is used to generate input samples in the analysis process, this sampling strategy can obtain stable sensitivity index estimation with fewer samples, the system automatically performs a full parameter sensitivity analysis periodically (such as every month), and identifies the key material parameters that have the greatest impact on the current structural system reliability. The incremental learning engine updates the Young's modulus matrix of the finite element model through the online backpropagation algorithm, the engine constructs the finite element analysis process as a differentiable computation graph, so that the mapping relationship from material parameters to displacement response has a computable gradient; when new monitoring data arrives, the engine first runs the finite element analysis to obtain the predicted displacement field, then calculates the loss function between the predicted value and the actual sensor measurement value, which usually takes the form of mean square error, then calculates the gradient of each element value in the Young's modulus matrix with respect to the loss function through the backpropagation algorithm, finally updates the modulus value using the stochastic gradient descent algorithm with momentum. The updating process imposes physical constraints to ensure that the Young's modulus is always within the reasonable range allowed by material science, at the same time the update amplitude is limited by the trust region strategy, the change of each element modulus in a single iteration does not exceed 5% of the current value, preventing over-adjustment from causing model instability.

[0083] The residual distribution statistics module of the feedback regulation loop calculates the prediction error of each component node using a sliding time window strategy. The window size is dynamically adjusted according to the characteristics of the data. For data that changes rapidly, such as vibration frequency, a shorter 24-hour window is used. For slowly changing parameters, such as corrosion rate, a 30-day window is used. The module calculates the residual between the finite element model prediction and the actual sensor measurement at each time step. The residual sequence of each component node within the window period is calculated, and the root mean square value is calculated and normalized to the 0-1 interval. The statistical process considers the spatio-temporal correlation of the data. Cluster analysis is performed on adjacent nodes in space. Nodes with similar residual patterns are divided into the same region. Representative statistics are calculated for each region. The adaptive learning rate scheduler dynamically adjusts the update step of the convolution kernel parameters based on the residual root mean square value. The scheduler maintains a learning rate matrix corresponding to the finite element grid nodes. The initial value is set to the global learning rate. At the beginning of each training period, the scheduler reads the normalized root mean square value of each node output by the residual statistics module. For nodes with larger residuals, the learning rate is increased proportionally to accelerate the parameter convergence speed in that region. For nodes with smaller residuals, the learning rate is decreased accordingly to improve parameter stability. The learning rate adjustment uses a smooth transition strategy. The learning rate of adjacent nodes is limited within a preset range to avoid sudden oscillation. The upper and lower limits of the learning rate are set as hard boundaries to prevent gradient explosion or disappearance.

[0084] In practical application of bridge monitoring, the system found that the stress prediction residual of the box girder web area was consistently higher than other parts. Parameter sensitivity analysis showed that the material constitutive parameters of the welding nodes in this area had a significant impact on the calculation results. The incremental learning engine adjusted the Young's modulus value of the units in this area, gradually reducing the difference between prediction and measurement through multiple iterations. The feedback regulation loop also monitored that the residual root mean square value of the web area was 2.3 times that of the bridge deck area. Therefore, the learning rate of this area was automatically increased to 1.8 times the baseline value, while the well-fitted bridge deck area learning rate was decreased to 0.7 times the baseline value. The system also has a learning rate decay mechanism. As the number of training rounds increases, the global base learning rate gradually decays exponentially to avoid parameter oscillation around the optimal solution due to excessive learning rate in the later training period. The decay rate is adaptively adjusted according to the overall residual trend. When the residual is monitored to decline into a plateau, the system will accelerate the decay rate to promote convergence. In special cases, such as after structural repair and reinforcement, the system will reset the learning rate to the initial value and start adapting to the changed structure state again.

[0085] The dynamic updating interface is provided with a parameter updating review mechanism. When the parameter modification suggestion proposed by the incremental learning engine exceeds the historical variation range, the system will pause the automatic updating and generate a task to be reviewed, which can be executed after being confirmed by the engineer. The review threshold is dynamically set according to the historical variation amplitude of the parameter, and is usually twice the standard deviation of the variation range of the parameter in the past year. This review mechanism prevents parameter error updating caused by sensor abnormalities or model defects, and improves the system reliability. The feedback regulation loop includes an abnormal residual detection function. When the residual of a certain node suddenly and sharply increases, the system will mark the node and check the sensor working state to exclude false residuals caused by equipment failure. At the same time, the loop analyzes the spatial distribution pattern of the residual. If adjacent nodes show similar residual characteristics, it is judged that the local model parameters need to be adjusted; if the residual distribution shows a random scatter pattern, it is more likely to be caused by measurement noise, and the learning rate adjustment amplitude is accordingly reduced.

[0086] During system operation, the parameter update history is recorded and visualized, and engineers can view the variation curve of each material parameter over time to determine the structural performance degradation trend. The learning rate adjustment log is also recorded in detail to provide a data analysis basis for system optimization. Long-term operation data shows that this adaptive updating strategy gradually improves the model prediction accuracy as the monitoring time increases, especially when the structure experiences seasonal temperature changes and traffic load fluctuations, it can quickly track the changes in structural response characteristics. In implementation, the optimization of computational efficiency is also considered. The parameter sensitivity analysis adopts a parallel computing strategy, which assigns different parameter combinations to multiple computing cores to perform finite element analysis simultaneously. The incremental learning process uses small-batch gradient descent, and each update is based on data within a time window rather than a single point, improving the stability of parameter updating. The residual statistics module uses an incremental calculation algorithm to avoid recalculating the statistics of the entire time window each time, reducing computational overhead. These optimization measures enable the system to efficiently process monitoring data for large structures on conventional server hardware.

[0087] Example 6: A cross-river immersed tunnel is taken as an application scenario. The tunnel is composed of multiple sections of reinforced concrete immersed tubes that are connected, and the structural integrity of the tube sections, the sealing state of the joints, and the influence of surrounding water and soil pressure need to be monitored for a long time. A high-density sensor network is arranged according to the tube sections, with fiber Bragg grating sensors installed in a matrix along the circumferential and longitudinal directions on the inner wall of the immersed tube, covering key areas such as the middle, end, and joints of the tube section, for collecting strain field distribution data; piezoelectric accelerometer arrays are arranged on the water-facing side and the bottom of the tube section to capture vibration spectra caused by ship navigation and water flow impact; electrochemical impedance sensors are embedded at the bottom of the immersed tube and the joints to monitor the chloride ion concentration in the concrete and the polarization resistance of the steel reinforcement, and environmental temperature and humidity sensors are installed at the top of the tunnel to assist in analyzing the influence of the environment on the structure.

[0088] After receiving the asynchronous sampling data of each sensor, the space-time alignment engine compensates the time offset caused by the difference in signal transmission distance in the immersed tube by using the Lagrange interpolation method for the wavelength sampling sequence of the fiber grating sensor; according to the coordinate system in the tunnel design drawing, a sensor spatial position relationship diagram is established, and the local coordinate system of the piezoelectric accelerometer array is converted to the global reference system of the tunnel through Lie algebra transformation, ensuring the spatial consistency of vibration data at different positions; for the slowly varying signals output by the electrochemical impedance sensor, sliding window Kalman filtering is used for time synchronization calibration to eliminate the interference of environmental temperature fluctuations on the sampling time sequence, and finally the original feature tensor with a unified space-time reference is generated.

[0089] The semantic modeling unit calls the BIM design drawing and material parameter library of the immersed tube tunnel, the component segmentation module divides the original feature tensor into corresponding feature blocks such as single-section immersed tube and pipe joint according to the pipe section segmentation and joint structure information in the drawing; the material attribute binding module extracts parameters such as immersed tube concrete grade, reinforcement diameter and protective layer thickness from the parameter library, and adds material attribute tags such as elastic modulus and yield strength to each feature block; the topological relationship encoder calculates the connection stiffness matrix of the pipe joint according to the tunnel structure mechanics model, and embeds it into the hierarchical structure feature, clearly defining the mechanical correlation of each pipe section and joint.

[0090] After receiving the hierarchical structure feature, the adaptive fusion core constructs a conflict detection model based on Mahalanobis distance to identify abnormal data blocks in the feature that exceed the statistical confidence interval, such as abnormal readings of chloride ion concentration caused by sensor moisture; the reliability of multi-modal sensor data is analyzed through a Gaussian mixture model, and dynamic weights are assigned to different types of data such as strain, vibration and corrosion for weighted reconstruction of abnormal data blocks; the reconstructed features are projected onto a low-dimensional manifold using a variational autoencoder to generate an anti-interference fusion diagnostic index, focusing on reflecting the strain concentration degree of the pipe section, vibration frequency change and reinforcement corrosion risk.

[0091] The edge computing node is deployed in the monitoring room along the tunnel, and a lightweight damage detection model is run according to the fusion diagnostic index. The micro-crack detection unit analyzes the strain concentration coefficient of the pipe section through a deep separable convolutional network to identify whether there are micro-cracks on the inner wall; the stiffness degradation evaluation unit combines the vibration spectrum features and the joint connection stiffness matrix to calculate the equivalent stiffness attenuation rate of the pipe section and the joint, and judge the change of the structure bearing capacity; the corrosion risk quantification unit predicts the reinforcement corrosion rate based on the chloride ion concentration and polarization resistance value measured by the electrochemical impedance sensor through the chloride ion diffusion equation, and outputs the local health status grade of each pipe section and joint.

[0092] The cloud collaborative analyzer aggregates the local component health status grades uploaded by each edge computing node, combines the historical degradation data since the completion of the immersed tunnel, constructs a whole tunnel finite element model, and inputs the health status grades of each pipe section and joint as boundary conditions into the model. A time series prediction network is used to simulate stress redistribution under different working conditions, such as high water level during flood season and stress changes of the tunnel structure when ships pass through under heavy load. The Monte Carlo sampling is used to calculate the reliability index of the tunnel structure system, analyze the risk occurrence probability of pipe section cracking and joint leakage, and finally output the confidence interval of the whole tunnel residual life.

[0093] The dynamic updating interface in the system regularly corrects the material constitutive relation of the finite element model according to the newly collected hierarchical structure features. The parameter sensitivity analysis module identifies the material parameters that have the greatest impact on the reliability of the tunnel structure, such as the elastic modulus of concrete and the corrosion resistance of steel bars. The incremental learning engine updates the Young's modulus matrix of the corresponding pipe section in the finite element model through an online back propagation algorithm, so that the model can reflect the gradual changes in the performance of the structure. The feedback regulation loop continuously compares the residual error between the strain field predicted by the finite element model and the actual measured value, and the residual error distribution statistical module calculates the root mean square value of the prediction error of each pipe section node. The adaptive learning rate scheduler dynamically adjusts the convolution kernel parameter update step size of the depth separable convolution network according to the root mean square value. When the residual error at the pipe joint is large, the parameter adjustment amplitude is increased to improve the detection accuracy of micro-cracks.

[0094] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0095] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A health monitoring system for civil engineering structures, characterized in that, include: High-density sensor networks are used to collect strain field distribution, vibration spectrum and environmental corrosion parameters of structures through distributed multimodal sensor arrays; The spatiotemporal alignment engine performs clock drift compensation and spatial coordinate normalization on the asynchronous sampling data of the multimodal sensor array, generating an original feature tensor with a unified spatiotemporal reference. The semantic modeling unit maps the original feature tensors to a preset component semantic space based on the structural design drawings and material parameter library, and outputs hierarchical structural features with topological tags. The adaptive fusion core receives hierarchical structural features and performs dynamic weight allocation. It eliminates conflicting data between sensors through iterative optimization algorithms and generates anti-interference fusion diagnostic indicators. Edge computing nodes run lightweight damage detection models in real time based on fused diagnostic indicators and output the health status level of local components. The cloud-based collaborative analyzer aggregates the output results of multiple edge computing nodes and combines them with historical degradation data to predict the overall remaining lifespan of the structure. The adaptive fusion core execution: Construct a conflict detection model based on Mahalanobis distance to identify anomalous data blocks that exceed the statistical confidence interval in hierarchical structural features; The reliability weights of multimodal sensor data are estimated using a Gaussian mixture model, and abnormal data blocks are reconstructed using weighted methods. A variational autoencoder is used to perform low-dimensional manifold projection on the reconstructed features to generate fusion diagnostic indicators.

2. The civil engineering structure health monitoring system according to claim 1, characterized in that, The high-density sensor network includes: The strain field acquisition subnetwork uses a fiber optic grating sensor array to measure the micro-strain on the surface of the structure and calculates the principal stress direction by Bragg wavelength offset. The vibration sensing subnet captures the structure’s natural frequency and damping ratio through a piezoelectric accelerometer array and extracts the frequency band energy features using wavelet packet decomposition. The corrosion monitoring subnetwork uses electrochemical impedance sensors to detect the chloride ion concentration and steel reinforcement polarization resistance in concrete.

3. The civil engineering structure health monitoring system according to claim 2, characterized in that, The spatiotemporal alignment engine executes: Lagrange interpolation compensation is performed on the wavelength sampling sequence of the fiber Bragg grating sensor array to eliminate the time offset caused by transmission delay; Establish a spatial position relationship diagram of the sensors, and transform the local coordinate system of the piezoelectric accelerometer array to the global reference system through Lie algebra transformation; A sliding window Kalman filter was used to perform time synchronization calibration on the slowly varying signal of the electrochemical impedance sensor.

4. The civil engineering structure health monitoring system according to claim 3, characterized in that, The semantic modeling unit includes: The component segmentation module divides the original feature tensor into feature blocks corresponding to the physical components based on the beam-column node information in the design drawings. The material property binding module extracts concrete grade and steel reinforcement specifications from the material parameter library and adds elastic modulus and yield strength labels to each feature block. The topology encoder generates the connection stiffness matrix between components based on the structural mechanics model and embeds it into the hierarchical structural features.

5. The civil engineering structure health monitoring system according to claim 1, characterized in that, The edge computing nodes include: The microcrack detection unit analyzes and fuses the strain concentration coefficient in diagnostic indicators using a depth-separable convolutional network. The stiffness degradation assessment unit calculates the equivalent stiffness attenuation rate of the component based on the vibration spectrum characteristics and the connection stiffness matrix. A corrosion risk quantification unit predicts the corrosion rate of steel bars based on the chloride ion diffusion equation and polarization resistance value.

6. The civil engineering structure health monitoring system according to claim 5, characterized in that, The cloud-based collaborative analyzer executes: Construct a finite element model of the entire structure and use the health status level of local components as boundary condition input; A time series prediction network was used to simulate the stress redistribution process under different loading conditions. The reliability index of the structural system is calculated by Monte Carlo sampling, and the confidence interval of the remaining lifetime is output.

7. The civil engineering structure health monitoring system according to claim 6, characterized in that, Also includes: The interface is dynamically updated to automatically correct the material constitutive relations of the finite element model based on newly acquired hierarchical structural features; The feedback adjustment loop compares the residuals between the predicted strain field and the actual measured value, and adjusts the convolution kernel parameters of the depth-separable convolutional network.

8. The civil engineering structure health monitoring system according to claim 7, characterized in that, The dynamic update interface includes: The parameter sensitivity analysis module identifies the material parameters that have the greatest impact on structural reliability. The incremental learning engine updates the Young's modulus matrix of the finite element model through an online backpropagation algorithm.

9. The civil engineering structure health monitoring system according to claim 8, characterized in that, The feedback adjustment loop includes: The residual distribution statistics module calculates the root mean square value of the prediction error for each component node; An adaptive learning rate scheduler dynamically adjusts the update stride of the convolution kernel parameters based on the root mean square value.

Images