An in-situ detection method for seismic isolation and damping systems under complex multi-field coupling environment

Abstract

This invention discloses an in-situ detection method for seismic isolation and damping systems under complex multi-field coupling environments, belonging to the field of seismic isolation and damping system detection technology; it includes: S1: sensor array deployment, deploying multi-source data acquisition sensors on the seismic isolation and damping bearings; S2: data processing, preprocessing and feature extraction of the multi-source raw data acquired in S1; S3: construction of a quantitative analysis model for multi-field coupling effects, establishing a mechanical performance analysis model for the seismic isolation system considering at least the coupling effects of multiple fields including temperature field, humidity field, stress field, seismic field, and wind field; S4: deep fusion of multi-source heterogeneous data, constructing a three-level fusion architecture of data layer - feature layer - decision layer; S5: performance evolution and damage prediction, establishing a deep learning-based performance evolution and damage prediction model for the seismic isolation system; S6: adaptive safety early warning, establishing a multi-level early warning mechanism.

Description

Technical Field

[0001] This invention belongs to the field of vibration isolation and damping system testing technology, specifically relating to an in-situ testing method for vibration isolation and damping systems under complex multi-field coupling environments. Background Technology

[0002] Seismic isolation and damping technology is one of the most effective technical means to improve the seismic performance of civil engineering structures. By setting a seismic isolation layer between the foundation and the superstructure, the natural period of the structure can be significantly extended and the impact of earthquakes on the superstructure can be reduced. It has been widely used in important projects such as high-rise buildings, long-span bridges, nuclear power facilities, and rail transit hubs. As the first line of defense for structural seismic resistance, the working state and performance integrity of the seismic isolation and damping system are directly related to the safety of the entire structure.

[0003] Traditional methods for testing seismic isolation and damping systems mainly rely on regular manual inspections and laboratory sampling tests. Manual inspections have drawbacks such as high labor intensity, long inspection cycles, and strong subjectivity. They can only detect obvious surface damage and cannot detect potential internal performance degradation and damage in a timely manner. Laboratory sampling tests require cutting bearing samples for destructive testing, which can cause irreversible damage to the seismic isolation system. Moreover, the sampling results cannot fully reflect the actual working status of all seismic isolation bearings in the entire project.

[0004] With the development of sensor and computer technologies, online monitoring technology has been gradually applied to the health monitoring of seismic isolation and damping systems. However, most existing online monitoring systems can only monitor a few dynamic parameters such as displacement and acceleration. Although some systems integrate temperature sensors, they only perform simple single-point temperature measurements and cannot obtain information on the distribution of multiple physical fields such as temperature, humidity, and stress fields inside the seismic isolation bearings. More importantly, existing technologies largely ignore the significant impact of the coupling effects of multiple fields such as temperature, humidity, corrosion, wind load, and seismic load on the performance of seismic isolation systems in real engineering environments. The established performance evaluation and damage identification models are based on ideal laboratory conditions, which suffer from low damage identification accuracy, high false alarm rate, and poor reliability in complex and variable engineering field environments. In addition, most existing monitoring systems can only monitor the current state and lack the ability to predict the long-term performance evolution trend of the seismic isolation system, thus failing to achieve predictive maintenance. Summary of the Invention

[0005] This invention provides an in-situ detection method for seismic isolation and damping systems under complex multi-field coupling environments. By deploying a distributed multi-physics field sensor array inside and around the seismic isolation bearing, synchronous in-situ monitoring of multiple physical fields such as temperature, humidity, stress, displacement, and acceleration is achieved. A quantitative relationship model of the mechanical performance changes of the seismic isolation system under multi-field coupling is established. Deep fusion technology of multi-source heterogeneous data is used to achieve real-time evaluation of the performance of the seismic isolation system and accurate identification of damage. Combined with a performance evolution prediction model, the remaining service life of the seismic isolation system can be predicted and safety warnings can be provided.

[0006] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution: An in-situ testing method for seismic isolation and damping systems under complex multi-field coupling environments includes: S1: Sensor array deployment. At least the following data acquisition sensors shall be deployed on the seismic isolation and damping bearing: temperature sensor, humidity sensor, strain monitoring sensor, displacement field sensor, acceleration sensor, and chemical corrosion sensor. S2: Data processing, which involves preprocessing and feature extraction of the multi-source raw data collected in S1; S3: Construction of a quantitative analysis model for multi-field coupling effects. This involves establishing a mechanical performance analysis model for a seismic isolation system that considers the coupling effects of at least the temperature field, humidity field, stress field, seismic field, and wind field. This model includes: Material constitutive relation models are used to establish and analyze the constitutive relations between the materials of vibration isolation and damping bearings and multi-field coupling; Multi-field coupling control equations are used to consider the influence of performance changes of rubber materials under temperature and humidity coupled field environments; Performance parameter mapping is used to establish a quantitative mapping relationship between various performance parameters of the vibration isolation and damping bearing and the influencing factors under multi-field coupling. S4: Deep fusion of multi-source heterogeneous data to build a three-level fusion architecture of data layer - feature layer - decision layer. The data layer fusion performs time synchronization and spatial registration of different types of sensors, and performs noise reduction and missing value completion on the data. Feature layer fusion extracts multi-dimensional features from different types of sensors, including at least time domain, frequency domain, and time-frequency domain features; deep learning algorithms are used to fuse and reduce the dimensionality of the extracted features, and establish a mapping relationship between the features and the performance parameters and damage state of the seismic isolation system; The decision-making layer integrates multiple damage identification results and combines them with the prediction results of a multi-field coupling analysis model to improve the accuracy and reliability of damage identification; it enables accurate identification of damage location, damage type, and damage degree. S5: Performance Evolution and Damage Prediction. A deep learning-based model for the performance evolution and damage prediction of seismic isolation systems is established, including: The data-driven prediction model uses a long short-term memory network (LSTM) combined with an attention mechanism to establish a prediction model of the performance parameters of the seismic isolation system over time, which is used to predict the predicted values ​​and confidence intervals of the performance parameters at different future time points. The physical information fusion prediction model integrates the physical constraints of the multi-field coupling analysis model into the deep learning model and adopts the physical information neural network (PINN) technology to improve the generalization ability and reliability of the prediction model, and realize the performance evolution prediction under small sample and high uncertainty conditions. Establish failure criteria for seismic isolation systems, clarify the performance thresholds corresponding to different degrees of damage, and combine the performance evolution prediction model to predict the time when the seismic isolation system reaches the failure criteria, and give the probability distribution and confidence interval of the remaining service life. S6: Adaptive safety early warning, which combines real-time environmental factors and system status to set dynamic early warning thresholds and establish a multi-level early warning mechanism. Different early warning mechanisms adopt different response measures.

[0007] Preferably, the sensor array is deployed in the following manner: The temperature sensor is a distributed fiber optic temperature sensor, which is arranged radially and axially within the rubber layer of the vibration damping support to form a three-dimensional temperature field monitoring network with a spatial resolution of not less than 5 mm and a temperature measurement accuracy of ±0.1℃. The humidity sensor is a capacitive humidity sensor, which is installed inside and around the vibration damping support to monitor the humidity distribution inside and outside the rubber environment. The measurement range is 0~100%RH, and the accuracy is ±2%RH. The strain monitoring sensor is a fiber optic strain sensor, which is deployed on the surface of the steel plate layer of the vibration isolation bearing to monitor the stress distribution of the steel plate layer. The strain measurement accuracy is ±1 microstrain. The displacement field sensor uses a combination of machine vision and laser displacement sensor to monitor the three-dimensional displacement of the seismic isolation bearing. The horizontal displacement measurement accuracy is ±0.1mm, and the vertical displacement measurement accuracy is ±0.05mm. The acceleration sensor is a triaxial acceleration sensor, which is installed on the upper and lower connecting plates and around the vibration isolation support. The sampling frequency is not less than 1000Hz and the measurement range is ±10g. The chemical corrosion sensor is installed on the surface of the vibration damping support to monitor corrosive chemicals in the environment, including chloride ions and sulfur dioxide.

[0008] Preferably, the edge nodes for data processing in S2 are deployed at the data acquisition site, and the acquired raw data is immediately preprocessed, including missing value imputation, noise reduction, and normalization. The objectives of the feature extraction include: Acceleration features are extracted first, prioritizing the extraction of time-domain features with the highest real-time performance; then frequency-domain features are extracted to reflect changes in the inherent characteristics of the structure; time-frequency domain features are used to capture non-stationary seismic responses. Displacement data characteristics, including static and quasi-static characteristics such as maximum horizontal displacement, maximum vertical displacement, and residual displacement; hysteresis loop characteristics, a core indicator used to reflect the energy dissipation capacity of seismic isolation bearings; spatial distribution characteristics; Strain data characteristics, including basic mechanical characteristics, fatigue damage characteristics, and spatial distribution characteristics, are used to reflect the internal stress state of the support. Temperature and humidity data features include temperature field features, humidity field features, and temperature and humidity coupling features; these serve as key input features for multi-field coupling effects. Corrosion monitoring data features are extracted, with a focus on features related to corrosion rate, to reflect the degree of environmental erosion.

[0009] Preferably, the material constitutive relation model includes at least: 1) Multi-factor coupled hyperelastic constitutive model of rubber materials Based on the modified Yeoh hyperelastic model, and incorporating three correction factors—temperature, humidity, and aging degree—this model accurately describes the large deformation mechanical behavior of vibration isolation rubber in environments ranging from -40℃ to 80℃ and 0% to 100% RH. Temperature correction factor: ,in This is the temperature influence coefficient. For reference temperature, set to 25℃; Humidity correction factor: ,in Humidity influence coefficient For reference humidity, set to 50%RH; Aging correction factor: ,in A represents the aging rate, and A represents the degree of aging, set to 0~1; The calibration method for the above parameter factors is as follows: a. Prepare rubber samples with different aging degrees (obtained through thermo-oxidative accelerated aging test); b. Perform uniaxial tensile, biaxial tensile, and planar tensile tests under different temperature (-40℃, -20℃, 0℃, 25℃, 40℃, 60℃, 80℃) and different humidity (30%RH, 50%RH, 70%RH, 90%RH) conditions; c. The least squares method was used to fit the Yeoh model parameters C10, C20, and C30 under different working conditions; d. Regression analysis yields the influence coefficients α_T, α_H, and α_A of temperature, humidity, and aging degree on the model parameters; 2) Temperature-Cyclic Coupled Elastoplastic Constitutive Model of Lead Core Material By adopting the modified Ramberg-Osgood elastoplastic model and introducing temperature correction and cyclic cumulative damage correction, the mechanical behavior of lead core under different temperatures and cyclic loading cycles can be accurately described. Temperature correction factor: ,in This is the temperature softening coefficient; Cyclic cumulative damage correction factor: ,in The cyclic damage coefficient, The damage index is N, and the number of loading cycles is N. The above parameter factor calibration method is as follows: a. Preparation of standard lead core samples b. Conduct monotonic tensile tests under different temperatures (-20℃, 0℃, 25℃, 40℃, 60℃) to determine σy0, n0, and E00 at different temperatures; c. Conduct cyclic loading tests at different temperatures for different numbers of cycles (10, 50, 100, 200, 500 times); d. Regression analysis yielded the temperature softening coefficient β_T and the cyclic damage coefficients β_D and γ; 3) Temperature-corrosion coupled elastic constitutive model of steel plate material By adopting a linear elastic constitutive model and introducing temperature correction and corrosion damage correction, the mechanical behavior of steel plates under different temperatures and corrosion levels can be accurately described. in The elastic modulus is taken as a reference temperature and under non-corrosive conditions. This is the temperature influence coefficient; For corrosion depth, This represents the critical corrosion depth. This is the corrosion influence coefficient; The calibration method for the above parameters is as follows: a. Prepare steel plate samples with different corrosion levels (obtained through salt spray accelerated corrosion test); b. Conduct tensile tests under different temperature conditions (-20℃, 0℃, 25℃, 40℃, 60℃); c. Regression analysis yields the temperature influence coefficient k_T and the corrosion influence coefficients k_d and m.

[0010] Preferably, the multi-field coupled control equations establish a set of fully coupled thermal-humidity-mechanical control equations, considering the bidirectional coupling effect between various physical fields, and are solved numerically using the finite element method; 1) Heat conduction equation: The density of the material; Specific heat capacity; Let be the thermal conductivity, a function of temperature and humidity. ; Heat generated by mechanical deformation: Heat generated by viscoelastic dissipation: v; 2) Wet diffusion equation: ,in Humidity concentration; Let be the moisture diffusivity, a function of temperature and stress. ; Here, R is the diffusion activation energy, R is the gas constant, and s is the humidity source term. 3) Equilibrium equations: For volume forces, For acceleration; It is the elastic stiffness tensor; For temperature strain, For humidity strain, For plastic deformation, only lead core material; The solution process for the coupled control equations is as follows: The sequential coupling iterative solution method is adopted, in which each physical field is solved alternately in each time step until convergence: Initial step: Input the initial temperature field T0, humidity field H0, stress field σ0, and displacement field u0; nth time step: Solve the heat conduction equation to obtain the temperature field Tn; Solving the moisture diffusion equation yields the humidity field Hn; Update the material constitutive parameters (based on Tn and Hn); Solving the mechanical equilibrium equations yields the displacement field un and the stress field σn; Calculate the heat generated by mechanical deformation, Q_mech, and the heat dissipation due to viscoelastic dissipation, Q_visc. Check the convergence conditions: ||Tn - Tn-1|| < ε_T and ||Hn - Hn-1|| < ε_H and ||un - un-1|| < ε_u; If convergence fails, return to step one and recalculate; if convergence occurs, proceed to the next time step.

[0011] Preferably, the performance parameter mapping adopts a hybrid modeling method of "physical model + data-driven" to establish a quantitative mapping relationship between the key performance parameters of the seismic isolation bearing and multiple influencing factors; 1) Definition and calculation of key performance parameters Obtain the hysteresis curves of the seismic isolation bearings through finite element simulation or experiments, and calculate the following key performance parameters: Horizontal stiffness K_h: K_h = (F_max - F_min) / (Δ_max - Δ_min) Vertical stiffness K_v: K_v = (V_max - V_min) / (δ_max - δ_min) Equivalent damping ratio ξ_eq: ξ_eq = W_d / (4πW_e), where W_d is the hysteresis loop area and W_e is the elastic strain energy; 2) Quantitative mapping relationship between performance parameters and influencing factors The response surface methodology was used to establish the mapping relationship between performance parameters and temperature T, humidity H, stress σ, aging degree A, and number of cyclic loading N. P = a0 + Σai·xi + Σaij·xi·xj + Σaijk·xi·xj·xk Where P is a performance parameter (K_h, K_v, or ξ_eq); xi represents the standardized influencing factors (x1=T, x2=H, x3=σ, x4=A, x5=N); a0, ai, aij, and aijk are response surface coefficients; Parameter determination method: Design an orthogonal experimental design, considering 5 influencing factors, with 5 levels for each factor; Using the above multi-field coupled finite element model, performance parameters under different working conditions are calculated; The response surface coefficients were fitted using the least squares method. The errors in the response surface model were corrected through additional verification experiments and simulations; 3) Performance parameter degradation rate model: A modified Arrhenius equation is used to establish a quantitative relationship between the performance parameter degradation rate and environmental factors and load history. dP / dt = A·exp(-E_a / RT)·f(H)·f(σ)·f(N) Where A is the pre-exponential factor; E_a is the activation energy; f(H) is the humidity effect function: f(H) = 1 + k_H·H; f(σ) is the stress influence function: f(σ) = 1 + k_σ·σ; f(N) is the effect function of cyclic loading: f(N) = 1 + k_N·N; Parameter determination method: Long-term accelerated aging tests were conducted, and the performance parameters of the seismic isolation bearings were tested periodically under different temperature, humidity, stress and cyclic loading conditions. The model parameters A, E_a, k_H, k_σ, and k_N were determined using regression analysis. Establish a predictive model for the evolution of performance parameters over time: P(t) = P0·exp(-dP / dt·t).

[0012] Preferably, the data layer fusion method for multi-source data time synchronization for different types of sensors is as follows: a. Hardware-level time synchronization: All sensors and edge computing nodes are connected to the GPS / BeiDou timing system, providing an absolute time reference at the 1μs level; b. Data labeling: Each data point collected by the sensor is accompanied by a precise timestamp (accuracy 1μs). c. Establishing a unified time axis: using the time axis of the sensor with the highest sampling frequency (accelerometer) as the reference; d. Interpolation resampling: Low-frequency data (temperature, humidity, corrosion): resampled to 1Hz using cubic spline interpolation; Intermediate frequency data (strain, displacement): resampled to 100Hz using linear interpolation; High-frequency data (acceleration): Maintain the original sampling frequency (1000Hz); e. Time alignment: Align all sensor data to a unified point in time to form a time-synchronized multi-source data matrix; The method for spatial registration of multi-source data is as follows: a. Coordinate system definition: The origin is the center of the connecting plate under the seismic isolation bearing, the horizontal direction is the XY axis, and the vertical direction is the Z axis; b. Sensor spatial coordinate calibration: Accurately measure the coordinates (xi, yi, zi) of each sensor in the three-dimensional coordinate system of the support; c. Spatial interpolation mapping: Distributed fiber optic data: directly forms a three-dimensional spatial distribution field; Point sensor data: A continuous spatial distribution field is generated using Kriging interpolation. d. Spatial alignment: Align all physical field data to a uniform spatial grid (grid resolution 5mm×5mm×5mm); The multi-source data denoising employs an adaptive denoising algorithm combining wavelet transform and Kalman filtering, targeting the noise characteristics of different types of sensor data for specific denoising purposes. For data loss issues caused by sensor failures, transmission interruptions, etc., a hybrid completion method combining "spatiotemporal correlation completion" and "machine learning completion" is used. Short-term missing data (<10 data points): Linear interpolation is used to fill in the missing data. Missing data points (10-100 points): Kriging interpolation based on spatiotemporal correlation was used to fill in the missing data. For long-term missing data points (>100 data points): LSTM neural network prediction and completion are used, with historical data from adjacent sensors as input; The deep learning algorithm in the feature layer fusion uses a CNN-LSTM hybrid neural network architecture to achieve deep fusion and dimensionality reduction of multi-source heterogeneous features. This architecture can simultaneously capture the spatial and temporal correlations of features. The training method is as follows: 1) Dataset Construction: A large amount of simulation data (100,000+ samples) was generated using a multi-field coupled finite element model. Combining laboratory test data and actual engineering monitoring data (10,000+ samples); Standardize the data: x_norm = (x - mean) / std; 2) Loss function design: A multi-task loss function is adopted to simultaneously optimize performance parameters for regression and damage classification tasks; 3) Optimizer: Adam optimizer is used, with a learning rate of 0.001, a batch size of 32, and 100 training epochs; 4) Regularization: Dropout (0.2) and L2 regularization (1e-4) are used to prevent overfitting; The process of establishing the mapping between the features and the performance parameters and damage state of the seismic isolation system is as follows: By training the CNN-LSTM hybrid neural network, a nonlinear mapping relationship is automatically established between the fused features and the performance parameters of the seismic isolation system, and between the fused features and the damage state of the seismic isolation system. Performance parameter mapping: Network output layer 1 directly outputs the predicted values ​​of horizontal stiffness, vertical stiffness, and equivalent damping ratio; Damage state mapping: The network output layer 2 outputs the probability distribution of different damage states, including: No damage; rubber aging; lead core yielding; steel plate corrosion; interface debonding; support slippage; The decision layer integrates multiple independent damage identification results and prediction results from multi-field coupled physical models to make the final and most reliable damage decision: 1) Acquisition of multi-source damage identification results Multiple independent damage identification results are obtained and used as input for fusion in the decision layer: Data-driven damage identification results: damage classification probability output from the aforementioned CNN-LSTM network; Damage identification results from the physical model: The monitored performance parameters (displacement, acceleration, strain) are input into the multi-field coupling analysis model; The residuals between the performance parameters predicted by the calculation model and the actual monitored values ​​are calculated. The damage status and degree are determined based on the magnitude of the residual; Single sensor damage identification results: Accelerometer damage identification results (based on frequency domain feature changes); Damage identification results from strain sensor (based on stress concentration characteristics); Temperature sensor damage identification results (based on hotspot features); 2) Fusion decision-making based on improved DS evidence theory An improved DS evidence theory is used to fuse multiple damage identification results for decision-making, addressing the failure of the traditional DS evidence theory when evidence conflicts occur. a. Recognition framework establishment: Θ = {D0, D1, D2, D3, D4, D5}, corresponding to 6 damage states respectively; b. Construction of Basic Probability Assignment (BPA): For data-driven results: directly use the probability output by the network as the BPA; For the physical model results: construct the BPA based on the residual size. The larger the residual, the higher the BPA of the corresponding damage. For single-sensor results: construct a BPA based on the degree of feature change; c. Measurement of Evidence Conflict: The degree of conflict between pieces of evidence is measured using the Jousselme distance. d(m1, m2) = sqrt(0.5·||m1 - m2||²); d. Calculation of Evidence Weights: Calculate the weight of each piece of evidence based on its reliability and the degree of conflict. w_i = (1 - d_i) / Σ(1 - d_j), where d_i is the average distance between the i-th piece of evidence and all other pieces of evidence; e. Weighted average evidence synthesis: m_avg(A) = Σ(w_i·m_i(A)) f. Final decision: Using the principle of maximum basic probability allocation, the damage state with the highest BPA is selected as the final decision result; 3) Fusion of multi-field coupling models and data-driven results a. Using a multi-field coupling analysis model, the performance parameters of the seismic isolation bearing under no-damage conditions are predicted based on current environmental factors such as temperature, humidity, and stress. b. Calculate the deviation between the actual monitored performance parameters and the model predictions; c. Use this deviation as a correction term to adjust the data-driven damage identification results. m_final(D) = m_data(D)·(1+ k·ΔP), where ΔP is the performance parameter deviation and k is the correction coefficient; d. Normalize the corrected BPA to obtain the final damage probability distribution; 4) Precise identification of the location, type, and extent of damage. Based on determining the type of damage, further precise identification of the damage location and extent is achieved: a. Damage location identification: For distributed fiber optic strain data: determine the location of damage based on the location of strain anomaly areas; For acceleration data: the location of the damage is determined based on the differences in frequency domain characteristics of sensors at different locations; By fusing positional information from multiple sensors, the precise location of the damage is determined using triangulation. b. Damage assessment: Establish a feature database for different degrees of damage; The Support Vector Regression (SVR) algorithm is used to establish a mapping relationship between fused features and the degree of damage; Input the current fusion feature and output a quantified value of the degree of damage, 0~1, where 0 represents no damage and 1 represents complete failure.

[0013] Preferably, the data-driven prediction model employs a network architecture combining bidirectional LSTM and multi-head self-attention mechanism, which can simultaneously capture the long-term dependencies of performance evolution and the impact of key time nodes; the model input feature vector is: 32-dimensional historical performance parameter sequence: daily average values ​​of key performance parameters such as horizontal stiffness, vertical stiffness, equivalent damping ratio, maximum displacement, and residual displacement over the past 30 days; 48-dimensional environmental factor sequence: daily average values ​​of environmental parameters such as temperature field characteristics (average temperature, maximum temperature, temperature gradient), humidity field characteristics (average humidity, humidity gradient), and corrosion rate over the past 30 days. 32-dimensional load history sequence: daily statistical values ​​of load parameters such as wind load amplitude, peak ground acceleration, number of cyclic loading, and stress amplitude over the past 30 days; 16-dimensional current state characteristics: current damage level, aging level, cumulative fatigue damage, and other state parameters; Output result: Predicted values ​​of horizontal stiffness, vertical stiffness, and equivalent damping ratio for the next 7 days, 30 days, 90 days, 180 days, and 365 days; 95% confidence interval for each predicted value; The confidence interval estimation of the predicted value is achieved using the quantile regression method, which avoids the problem that traditional point prediction cannot quantify uncertainty. The performance prediction evolution model of the physical information fusion adopts a physical information neural network (PINN) to incorporate the control equations of the multi-field coupling analysis model as soft constraints into the loss function, so that the model can fit the observed data and satisfy the physical laws at the same time during the training process, which significantly improves the generalization ability and reliability of the model. The multi-field coupling control equations established in S3 are transformed into PINN loss function terms, as follows: The heat conduction equation constraint is: L_heat = MSE(ρc·∂T / ∂t -∇·[k(T,H)·∇T] -Q_mech - Q_visc); Moisture diffusion equation constraint: L_moisture = MSE(∂C / ∂t -∇·[D(T,σ)·∇C] - S); Mechanical equilibrium equation constraint: L_mech = MSE(∇·σ+ρb -ρü); Performance degradation equation constraint: L_degradation = MSE(dP / dt - A·exp(-E_a / RT)·f(H)·f(σ)·f(N)); The PINN network structure is as follows: Input layer: Input time t, temperature T, humidity H, stress σ, number of cycles N; Output layer: Outputs horizontal stiffness K_h, vertical stiffness K_v, and equivalent damping ratio ξ_eq; Total loss function: L_total = w_data·L_data + w_physics·(L_heat + L_moisture + L_mech +L_degradation) + w_boundary·L_boundary L_data = MSE(NN_pred, observed_data), which is the loss for data fitting; L_boundary = MSE(NN_pred_boundary, boundary_conditions), which is the loss for boundary conditions; w_data = 1.0, w_physics = 0.1, w_boundary = 0.5, which are the weight coefficients; The failure criterion of the seismic isolation system is established according to the "Standard for Seismic Isolation Design of Buildings" GB 51416 - 2021 and relevant industry codes; 1) Calculation of remaining useful life (RUL) based on performance evolution prediction The remaining useful life is calculated using the probabilistic threshold crossing method: a. Prediction of performance evolution trajectory: Using the trained PINN model, predict the evolution trajectory of the key performance parameters of the seismic isolation bearings in the next T years; Considering the uncertainty of model parameters, generate N possible performance evolution trajectories, N = 1000; b. Calculation of failure time: For each performance evolution trajectory, find the time t_f when the failure threshold is first crossed; If the failure threshold is not crossed within the prediction time T, then t_f = T; c. Calculation of RUL probability distribution: Statistically analyze the distribution of all t_f to obtain the probability density function (PDF) and cumulative distribution function (CDF) of RUL; Calculate the statistical characteristics of RUL: mean, median, standard deviation, 95% confidence interval; 2) Predictive maintenance decision based on RUL According to the RUL prediction results and failure probability, formulate a hierarchical maintenance strategy: RUL > 10 years, failure probability < 5%: Normal maintenance, detect once a year; 5 years < RUL ≤ 10 years, 5% ≤ failure probability < 20%: Strengthen monitoring, detect once every six months; 2 years < RUL ≤ 5 years, 20% ≤ failure probability < 50%: Planned maintenance, formulate a replacement plan; RUL ≤ 2 years, failure probability ≥ 50%: Emergency maintenance, replace the support immediately.

[0014] Preferably, the method for setting the dynamic early warning threshold is as follows: Establishment of a multi-field coupled benchmark performance model A three-dimensional benchmark mapping relationship between environment, state, and performance is established as the basis for calculating dynamic thresholds. Based on the previously calibrated multi-field coupled quantitative analysis model, a benchmark performance database covering the entire working condition range is generated. A continuous benchmark performance surface is obtained by fitting using the response surface method: P0 = f (T,H,σ,N). The environmental parameters (T,H), stress state (σ), and load history (N) at the current moment are input, and the benchmark performance surface function is called to calculate the theoretical benchmark performance parameter P0 (t) under the current working condition. The relative deviation of the performance parameter is calculated: ΔP (t) = (P_measured (t) - P0 (t)) / P0 (t)×100%. All warning thresholds are set based on relative deviation rather than absolute value. By employing a Bayesian online learning framework, the threshold distribution is continuously updated using real-time monitoring data and historical early warning feedback data, thereby achieving adaptive optimization of the threshold. Based on the basic Bayesian framework, three key correction factors are introduced to further improve the adaptability of the threshold: Aging correction factor: f_aging = 1 + k_aging·t, where k_aging is the annual aging rate coefficient and t is the number of years in service; Corrected threshold: T' = T·f_aging; Post-earthquake temporary correction factor: f_earthquake = 1.5, valid for 30 days after the earthquake; Corrected threshold: T' = T·f_earthquake; Sensor reliability correction factor: f_sensor = 1 + k_sensor·(1 - R); where R is the sensor health status, 0~1; k_sensor is the correction coefficient; the corrected threshold: T' = T·f_sensor.

[0015] Preferably, a multi-level early warning mechanism is established, consisting of four levels: abnormal status, performance degradation, damage confirmation, and failure risk, with each level corresponding to specific triggering conditions and risk levels. Use composite triggering logic: A single parameter exceeding the threshold only triggers an initial warning. If two or more related parameters exceed the threshold simultaneously, the warning level will be automatically upgraded to level one. If damage is confirmed by combining the damage identification results from multi-source data fusion, the warning level is directly upgraded to warning or above. Based on the performance evolution prediction results, if the predicted damage will develop rapidly, the warning level will be upgraded in advance. Different response measures are set according to different early warning mechanisms, as follows: Blue Alert Level Response Process: When an alert is triggered, the system automatically records the alert information and sends a notification to the maintenance personnel. The maintenance personnel check the relevant monitoring data and historical trends within 24 hours. The system determines whether the alert is caused by environmental factors or measurement errors. If it is confirmed to be a false alarm, the system records the information and adjusts the threshold. If not, the system increases the monitoring frequency of the support. The system observes the system for 7 days. If the performance parameters return to normal, the alert is lifted. If the situation continues to deteriorate, the alert is upgraded to the warning level. Yellow Warning Level Response Process: Upon triggering the warning, the system automatically records and generates a preliminary analysis report, simultaneously sending SMS, email, and app notifications to the operations and maintenance manager. The manager must organize a specialized on-site inspection (including visual inspection, non-destructive testing, and performance testing) within 48 hours. If no obvious damage is found: strengthen monitoring and develop a preventative maintenance plan. If early damage is found: assess the impact of the damage, develop a repair plan, and implement it within 30 days. After repair, conduct performance acceptance testing; if successful, the warning is lifted; otherwise, it is upgraded to a hazardous level. Red Danger Level Response Procedure: Upon triggering the warning, the system immediately sends emergency notifications via SMS, telephone, and APP pop-up to all relevant personnel, activates the emergency plan, restricts the use of the structure, and evacuates personnel if necessary; within 24 hours, organizes experts to conduct an on-site assessment, formulates an emergency response plan, and implements it immediately. For structures that can be temporarily reinforced, temporary reinforcement measures are taken, and the load is restricted; for structures that cannot be reinforced, use is immediately stopped, and the supports are replaced; after the handling is completed, a comprehensive safety assessment is conducted, and use can only be resumed after passing the assessment.

[0016] The beneficial effects of this invention are: This invention provides an in-situ testing method for seismic isolation and damping systems under complex multi-field coupling environments. It considers the coupling effects of multiple factors such as temperature, humidity, stress, earthquake, and wind load in real engineering environments, and establishes a quantitative analysis model of the mechanical performance of seismic isolation systems under multi-field coupling. This model can accurately reflect the influence of environmental factors and load history on the performance of seismic isolation systems, and eliminates the interference of environmental changes on the test results from the root. By adopting a three-level deep fusion architecture of "data layer-feature layer-decision layer" and combining physical models with data-driven methods, the effective fusion of multi-source data of different types, accuracies and sampling frequencies is achieved. This enables precise location of damage, identification of damage type and assessment of damage degree, and significantly reduces the occurrence of false alarms and false negatives. By integrating a deep learning model with physical constraints, we have achieved accurate prediction of the long-term performance evolution trend of seismic isolation systems and scientific assessment of their remaining service life. This enables us to anticipate performance degradation and potential failure risks in advance, promoting a shift in seismic isolation system maintenance from passive post-event repair to proactive predictive maintenance. It also helps to optimize maintenance strategies and reduce the total life-cycle maintenance cost. An adaptive safety early warning system based on multi-field coupling effect was established, which can dynamically adjust the early warning threshold according to real-time environmental conditions and system operating status. This effectively solves the problem of high false alarm rate of traditional fixed threshold early warning in complex and ever-changing environments, and provides timely and accurate decision-making basis for engineering safety management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 An in-situ testing method for seismic isolation and damping systems under complex multi-field coupling environments includes: S1: Sensor array deployment. At least the following data acquisition sensors shall be deployed on the seismic isolation and damping bearing: temperature sensor, humidity sensor, strain monitoring sensor, displacement field sensor, acceleration sensor, and chemical corrosion sensor. The aforementioned sensors are installed on the vibration isolation and damping bearings in the following manner: The temperature sensor adopts a distributed fiber optic temperature sensor, which is arranged radially and axially within the rubber layer of the vibration damping support to form a three-dimensional temperature field monitoring network with a spatial resolution of not less than 5mm and a temperature measurement accuracy of ±0.1℃. The humidity sensor is a capacitive humidity sensor, which is installed inside and around the vibration damping support to monitor the humidity distribution inside and outside the rubber environment. The measurement range is 0~100%RH, and the accuracy is ±2%RH. The strain monitoring sensor uses a fiber optic strain sensor, which is deployed on the surface of the steel plate layer of the vibration isolation bearing to monitor the stress distribution of the steel plate layer. The strain measurement accuracy is ±1 microstrain. The displacement field sensor combines machine vision and laser displacement sensor to monitor the three-dimensional displacement of the seismic isolation bearing. The horizontal displacement measurement accuracy is ±0.1mm, and the vertical displacement measurement accuracy is ±0.05mm. The accelerometer is a triaxial accelerometer, which is installed on the upper and lower connecting plates and around the vibration damping support. The sampling frequency is not less than 1000Hz and the measurement range is ±10g. Chemical corrosion sensors are installed on the surface of vibration damping supports to monitor corrosive chemicals in the environment, including chloride ions and sulfur dioxide. S2: Data processing, which preprocesses and extracts features from the multi-source raw data collected in S1; the edge nodes of data processing are deployed at the data collection site, and the raw data collected is immediately preprocessed, including missing value imputation, denoising and normalization; existing common data missing value imputation, denoising and normalization methods are adopted for data preprocessing. The goal of feature extraction is: Acceleration features are extracted first, prioritizing the extraction of time-domain features with the highest real-time performance; then frequency-domain features are extracted to reflect changes in the inherent characteristics of the structure; time-frequency domain features are used to capture non-stationary seismic responses. Displacement data characteristics, including static and quasi-static characteristics such as maximum horizontal displacement, maximum vertical displacement, and residual displacement; hysteresis loop characteristics, a core indicator used to reflect the energy dissipation capacity of seismic isolation bearings; spatial distribution characteristics; Strain data characteristics, including basic mechanical characteristics, fatigue damage characteristics, and spatial distribution characteristics, are used to reflect the internal stress state of the support. Temperature and humidity data features include temperature field features, humidity field features, and temperature and humidity coupling features; these serve as key input features for multi-field coupling effects. Corrosion monitoring data features are extracted, with a focus on features related to corrosion rate, to reflect the degree of environmental erosion. S3: Construction of a quantitative analysis model for multi-field coupling effects. This involves establishing a mechanical performance analysis model for a seismic isolation system that considers the coupling effects of at least the temperature field, humidity field, stress field, seismic field, and wind field. This model includes: Material constitutive relation models are used to establish and analyze the constitutive relations between the materials of vibration isolation and damping bearings and multi-field coupling; Material constitutive relation models include: 1) Multi-factor coupled hyperelastic constitutive model of rubber materials Based on the modified Yeoh hyperelastic model, and incorporating three correction factors—temperature, humidity, and aging degree—this model accurately describes the large deformation mechanical behavior of vibration isolation rubber in environments ranging from -40℃ to 80℃ and 0% to 100% RH. Temperature correction factor: ,in This is the temperature influence coefficient. For reference temperature, set to 25℃; Humidity correction factor: ,in Humidity influence coefficient For reference humidity, set to 50%RH; Aging correction factor: ,in A represents the aging rate, and A represents the degree of aging, set to 0~1; The calibration method for the above parameter factors is as follows: a. Prepare rubber samples with different aging degrees (obtained through thermo-oxidative accelerated aging test); b. Perform uniaxial tensile, biaxial tensile, and planar tensile tests under different temperature (-40℃, -20℃, 0℃, 25℃, 40℃, 60℃, 80℃) and different humidity (30%RH, 50%RH, 70%RH, 90%RH) conditions; c. The least squares method was used to fit the Yeoh model parameters C10, C20, and C30 under different working conditions; d. Regression analysis yields the influence coefficients α_T, α_H, and α_A of temperature, humidity, and aging degree on the model parameters; 2) Temperature-Cyclic Coupled Elastoplastic Constitutive Model of Lead Core Material By adopting the modified Ramberg-Osgood elastoplastic model and introducing temperature correction and cyclic cumulative damage correction, the mechanical behavior of lead core under different temperatures and cyclic loading cycles can be accurately described. Temperature correction factor: ,in This is the temperature softening coefficient; Cyclic cumulative damage correction factor: ,in The cyclic damage coefficient, The damage index is N, and the number of loading cycles is N. The above parameter factor calibration method is as follows: a. Preparation of standard lead core samples b. Conduct monotonic tensile tests under different temperatures (-20℃, 0℃, 25℃, 40℃, 60℃) to determine σy0, n0, and E00 at different temperatures; c. Conduct cyclic loading tests at different temperatures for different numbers of cycles (10, 50, 100, 200, 500 times); d. Regression analysis yielded the temperature softening coefficient β_T and the cyclic damage coefficients β_D and γ; 3) Temperature-corrosion coupled elastic constitutive model of steel plate material By adopting a linear elastic constitutive model and introducing temperature correction and corrosion damage correction, the mechanical behavior of steel plates under different temperatures and corrosion levels can be accurately described. in The elastic modulus is taken as a reference temperature and under non-corrosive conditions. This is the temperature influence coefficient; For corrosion depth, This represents the critical corrosion depth. This is the corrosion influence coefficient; The calibration method for the above parameters is as follows: a. Prepare steel plate samples with different corrosion levels (obtained through salt spray accelerated corrosion test); b. Conduct tensile tests under different temperature conditions (-20℃, 0℃, 25℃, 40℃, 60℃); c. Regression analysis yields the temperature influence coefficient k_T and the corrosion influence coefficients k_d and m; Multi-field coupling control equations are used to consider the influence of performance changes of rubber materials under temperature and humidity coupled field environments; A set of fully coupled thermo-humidity-mechanical control equations was established, considering the bidirectional coupling between various physical fields, and numerical solutions were obtained using the finite element method. 1) Heat conduction equation: The density of the material; Specific heat capacity; Let be the thermal conductivity, a function of temperature and humidity. ; Heat generated by mechanical deformation: Heat generated by viscoelastic dissipation: v; 2) Wet diffusion equation: ,in Humidity concentration; Let be the moisture diffusivity, a function of temperature and stress. ; Here, R is the diffusion activation energy, R is the gas constant, and s is the humidity source term. 3) Equilibrium equations: For volume forces, For acceleration; It is the elastic stiffness tensor; For temperature strain, For humidity strain, For plastic deformation, only lead core material; Solution process for coupled control equations: The sequential coupling iterative solution method is adopted, in which each physical field is solved alternately in each time step until convergence: Initial step: Input the initial temperature field T0, humidity field H0, stress field σ0, and displacement field u0; nth time step: Solve the heat conduction equation to obtain the temperature field Tn; Solving the moisture diffusion equation yields the humidity field Hn; Update the material constitutive parameters (based on Tn and Hn); Solving the mechanical equilibrium equations yields the displacement field un and the stress field σn; Calculate the heat generated by mechanical deformation, Q_mech, and the heat dissipation due to viscoelastic dissipation, Q_visc. Check the convergence conditions: ||Tn - Tn-1|| < ε_T and ||Hn - Hn-1|| < ε_H and ||un - un-1|| < ε_u; If the calculation fails to converge, return to step one and recalculate; if the calculation converges, proceed to the next time step. Performance parameter mapping is used to establish a quantitative mapping relationship between various performance parameters of the vibration isolation and damping bearing and the influencing factors under multi-field coupling. The performance parameter mapping adopts a hybrid modeling method of "physical model + data-driven" to establish a quantitative mapping relationship between the key performance parameters of the seismic isolation bearing and multiple influencing factors; 1) Definition and calculation of key performance parameters Obtain the hysteresis curves of the seismic isolation bearings through finite element simulation or experiments, and calculate the following key performance parameters: Horizontal stiffness K_h: K_h = (F_max - F_min) / (Δ_max - Δ_min) Vertical stiffness K_v: K_v = (V_max - V_min) / (δ_max - δ_min) Equivalent damping ratio ξ_eq: ξ_eq = W_d / (4πW_e), where W_d is the hysteresis loop area and W_e is the elastic strain energy; 2) Quantitative mapping relationship between performance parameters and influencing factors The response surface methodology was used to establish the mapping relationship between performance parameters and temperature T, humidity H, stress σ, aging degree A, and number of cyclic loading N. P = a0 + Σai·xi + Σaij·xi·xj + Σaijk·xi·xj·xk Where P is a performance parameter (K_h, K_v, or ξ_eq); xi represents the standardized influencing factors (x1=T, x2=H, x3=σ, x4=A, x5=N); a0, ai, aij, and aijk are response surface coefficients; Parameter determination method: Design an orthogonal experimental design, considering 5 influencing factors, with 5 levels for each factor; Using the above multi-field coupled finite element model, performance parameters under different working conditions are calculated; The response surface coefficients were fitted using the least squares method. The errors in the response surface model were corrected through additional verification experiments and simulations; 3) Performance parameter degradation rate model: A modified Arrhenius equation is used to establish a quantitative relationship between the performance parameter degradation rate and environmental factors and load history. dP / dt = A·exp(-E_a / RT)·f(H)·f(σ)·f(N) Where A is the pre-exponential factor; E_a is the activation energy; f(H) is the humidity effect function: f(H) = 1 + k_H·H; f(σ) is the stress influence function: f(σ) = 1 + k_σ·σ; f(N) is the effect function of cyclic loading: f(N) = 1 + k_N·N; Parameter determination method: Long-term accelerated aging tests were conducted, and the performance parameters of the seismic isolation bearings were tested periodically under different temperature, humidity, stress and cyclic loading conditions. The model parameters A, E_a, k_H, k_σ, and k_N were determined using regression analysis. Establish a predictive model for the evolution of performance parameters over time: P(t) = P0·exp(-dP / dt·t); S4: Deep fusion of multi-source heterogeneous data to build a three-level fusion architecture of data layer - feature layer - decision layer. The data layer fusion performs time synchronization and spatial registration of different types of sensors, and performs noise reduction and missing value completion on the data. The following are methods for multi-source data time synchronization for different types of sensors in data layer fusion: a. Hardware-level time synchronization: All sensors and edge computing nodes are connected to the GPS / BeiDou timing system, providing an absolute time reference at the 1μs level; b. Data labeling: Each data point collected by the sensor is accompanied by a precise timestamp with an accuracy of 1μs; c. Establishing a unified time axis: using the time axis of the sensor with the highest sampling frequency, such as an accelerometer, as the reference; d. Interpolation resampling: For low-frequency data such as temperature, humidity, and corrosion: use cubic spline interpolation to resample to 1Hz; For mid-frequency data such as strain and displacement: resample to 100Hz using linear interpolation; For high-frequency data such as acceleration: maintain the original sampling frequency, 1000Hz; e. Time alignment: Align all sensor data to a unified point in time to form a time-synchronized multi-source data matrix; The method for spatial registration of multi-source data is as follows: a. Coordinate system definition: The origin is the center of the connecting plate under the seismic isolation bearing, the horizontal direction is the XY axis, and the vertical direction is the Z axis; b. Sensor spatial coordinate calibration: Accurately measure the coordinates (xi, yi, zi) of each sensor in the three-dimensional coordinate system of the support; c. Spatial interpolation mapping: Distributed fiber optic data: directly forms a three-dimensional spatial distribution field; Point sensor data: A continuous spatial distribution field is generated using Kriging interpolation. d. Spatial alignment: Align all physical field data to a uniform spatial grid with a grid resolution of 5mm×5mm×5mm; Multi-source data denoising employs an adaptive denoising algorithm combining wavelet transform and Kalman filtering, targeting the noise characteristics of different types of sensor data for specific denoising purposes. For data loss issues caused by sensor failures, transmission interruptions, etc., a hybrid completion method combining spatiotemporal correlation completion and machine learning completion is used. Short-term missing data (<10 data points): Linear interpolation is used to fill in the missing data. Missing data points (10-100 points): Kriging interpolation based on spatiotemporal correlation was used to fill in the missing data. For long-term missing data points (>100 data points): LSTM neural network prediction and completion are used, with historical data from adjacent sensors as input; Feature layer fusion extracts multi-dimensional features from different types of sensors, including at least time domain, frequency domain, and time-frequency domain features; deep learning algorithms are used to fuse and reduce the dimensionality of the extracted features, and establish a mapping relationship between the features and the performance parameters and damage state of the seismic isolation system; The deep learning algorithm in feature layer fusion uses a CNN-LSTM hybrid neural network architecture to achieve deep fusion and dimensionality reduction of multi-source heterogeneous features. This architecture can simultaneously capture the spatial and temporal correlations of features. The training method is as follows: 1) Dataset Construction: A large amount of simulation data (100,000+ samples) was generated using a multi-field coupled finite element model. Combining laboratory test data and actual engineering monitoring data (10,000+ samples); Standardize the data: x_norm = (x - mean) / std; 2) Loss function design: A multi-task loss function is adopted to simultaneously optimize performance parameters for regression and damage classification tasks; 3) Optimizer: Adam optimizer is used, with a learning rate of 0.001, a batch size of 32, and 100 training epochs; 4) Regularization: Dropout (0.2) and L2 regularization (1e-4) are used to prevent overfitting; The process of establishing the mapping between characteristics and the performance parameters and damage state of the seismic isolation system is as follows: By training a CNN-LSTM hybrid neural network, a nonlinear mapping relationship is automatically established between fused features and the performance parameters of the seismic isolation system, and between fused features and the damage state of the seismic isolation system. Performance parameter mapping: Network output layer 1 directly outputs the predicted values ​​of horizontal stiffness, vertical stiffness, and equivalent damping ratio; Damage state mapping: The network output layer 2 outputs the probability distribution of different damage states, including: No damage; rubber aging; lead core yielding; steel plate corrosion; interface debonding; support slippage; The decision-making layer integrates multiple damage identification results and combines them with the prediction results of a multi-field coupling analysis model to improve the accuracy and reliability of damage identification; it enables accurate identification of damage location, damage type, and damage degree. The decision-making layer integrates and synthesizes multiple independent damage identification results and prediction results from multi-field coupled physical models to make the final and most reliable damage decision: 1) Acquisition of multi-source damage identification results Multiple independent damage identification results are obtained and used as input for fusion in the decision layer: Data-driven damage identification results: damage classification probability output from the aforementioned CNN-LSTM network; Damage identification results from the physical model: The displacement, acceleration, and strain performance parameters obtained from monitoring are input into the multi-field coupling analysis model; The residuals between the performance parameters predicted by the calculation model and the actual monitored values ​​are calculated. The damage status and degree are determined based on the magnitude of the residual; Single sensor damage identification results: Damage identification results based on accelerometer features changing; Damage identification results based on strain sensor characteristics of stress concentration; Damage identification results of temperature sensors based on hotspot features; 2) Fusion decision-making based on improved DS evidence theory An improved DS evidence theory is used to fuse multiple damage identification results for decision-making, addressing the failure of the traditional DS evidence theory when evidence conflicts occur. a. Recognition framework establishment: Θ = {D0, D1, D2, D3, D4, D5}, corresponding to 6 damage states respectively; b. Construction of Basic Probability Assignment (BPA): For data-driven results: directly use the probability output by the network as the BPA; For the physical model results: construct the BPA based on the residual size. The larger the residual, the higher the BPA of the corresponding damage. For single-sensor results: construct a BPA based on the degree of feature change; c. Measurement of Evidence Conflict: The degree of conflict between pieces of evidence is measured using the Jousselme distance. d(m1, m2) = sqrt(0.5·||m1 - m2||²); d. Calculation of Evidence Weights: Calculate the weight of each piece of evidence based on its reliability and the degree of conflict. w_i = (1 - d_i) / Σ(1 - d_j), where d_i is the average distance between the i-th piece of evidence and all other pieces of evidence; e. Weighted average evidence synthesis: m_avg(A) = Σ(w_i·m_i(A)) f. Final decision: Using the principle of maximum basic probability allocation, the damage state with the highest BPA is selected as the final decision result; 3) Fusion of multi-field coupling models and data-driven results a. Using a multi-field coupling analysis model, the performance parameters of the seismic isolation bearing under no-damage conditions are predicted based on current environmental factors such as temperature, humidity, and stress. b. Calculate the deviation between the actual monitored performance parameters and the model predictions; c. Use this deviation as a correction term to adjust the data-driven damage identification results. m_final(D) = m_data(D)·(1+ k·ΔP), where ΔP is the performance parameter deviation and k is the correction coefficient; d. Normalize the corrected BPA to obtain the final damage probability distribution; 4) Precise identification of the location, type, and extent of damage. Based on determining the type of damage, further precise identification of the damage location and extent is achieved: a. Damage location identification: For distributed fiber optic strain data: determine the location of damage based on the location of strain anomaly areas; For acceleration data: the location of the damage is determined based on the differences in frequency domain characteristics of sensors at different locations; By fusing positional information from multiple sensors, the precise location of the damage is determined using triangulation. b. Damage assessment: Establish a feature database for different degrees of damage; The Support Vector Regression (SVR) algorithm is used to establish a mapping relationship between fused features and the degree of damage; Input the current fusion feature and output a quantitative value of the degree of damage, 0~1, where 0 represents no damage and 1 represents complete failure; S5: Performance Evolution and Damage Prediction. A deep learning-based model for the performance evolution and damage prediction of seismic isolation systems is established, including: The data-driven prediction model uses a long short-term memory network (LSTM) combined with an attention mechanism to establish a prediction model of the performance parameters of the seismic isolation system over time, which is used to predict the predicted values ​​and confidence intervals of the performance parameters at different future time points. The model structure and implementation details are as follows: The data-driven prediction model employs a network architecture combining bidirectional LSTM and a multi-head self-attention mechanism, capable of simultaneously capturing long-term dependencies in performance evolution and the impact of key time points; the model input feature vector is: 32-dimensional historical performance parameter sequence: daily average values ​​of key performance parameters such as horizontal stiffness, vertical stiffness, equivalent damping ratio, maximum displacement, and residual displacement over the past 30 days; 48-dimensional environmental factor sequence: temperature field characteristics over the past 30 days, including average temperature, maximum temperature, and temperature gradient; humidity field characteristics, including average humidity and humidity gradient; daily average values ​​of environmental parameters such as corrosion rate; 32-dimensional load history sequence: daily statistical values ​​of load parameters such as wind load amplitude, peak ground acceleration, number of cyclic loading, and stress amplitude over the past 30 days; 16-dimensional current state characteristics: current damage level, aging level, cumulative fatigue damage, and other state parameters; Output result: Predicted values ​​of horizontal stiffness, vertical stiffness, and equivalent damping ratio for the next 7 days, 30 days, 90 days, 180 days, and 365 days; 95% confidence interval for each predicted value; The confidence interval of the predicted value is estimated using the quantile regression method, which avoids the problem that traditional point prediction cannot quantify uncertainty. The physical information fusion prediction model integrates the physical constraints of the multi-field coupling analysis model into the deep learning model and adopts the physical information neural network (PINN) technology to improve the generalization ability and reliability of the prediction model, and realize the performance evolution prediction under small sample and high uncertainty conditions. The performance prediction evolution model of physical information fusion adopts physical information neural network (PINN) to incorporate the control equations of multi-field coupling analysis model as soft constraints into the loss function, so that the model can fit the observed data and satisfy the physical laws at the same time during the training process, which significantly improves the generalization ability and reliability of the model. The multi-field coupling control equations established in S3 are transformed into PINN loss function terms, as follows: The heat conduction equation constraint is: L_heat = MSE(ρc·∂T / ∂t -∇·[k(T,H)·∇T] -Q_mech - Q_visc); Moisture diffusion equation constraint: L_moisture = MSE(∂C / ∂t -∇·[D(T,σ)·∇C] - S); Mechanical equilibrium equation constraint: L_mech = MSE(∇·σ+ρb -ρü); Performance degradation equation constraint: L_degradation = MSE(dP / dt - A·exp(-E_a / RT)·f(H)·f(σ)·f(N)); The PINN network structure is as follows: Input layer: Input time t, temperature T, humidity H, stress σ, number of cycles N; Output layer: Outputs horizontal stiffness K_h, vertical stiffness K_v, and equivalent damping ratio ξ_eq; Total loss function: L_total = w_data·L_data + w_physics·(L_heat + L_moisture + L_mech +L_degradation)+ w_boundary·L_boundary L_data = MSE(NN_pred, observed_data), which is the data fitting loss; L_boundary = MSE(NN_pred_boundary, boundary_conditions), which is the boundary condition loss; w_data=1.0, w_physics=0.1, w_boundary=0.5 are the weighting coefficients; Establish failure criteria for seismic isolation systems, clarify the performance thresholds corresponding to different degrees of damage, and combine the performance evolution prediction model to predict the time when the seismic isolation system reaches the failure criteria, and give the probability distribution and confidence interval of the remaining service life. The failure criteria for the seismic isolation system are established based on the "Standard for Seismic Isolation Design of Buildings" GB 51416-2021 and relevant industry standards, and are as follows: 1) Calculation of Remaining Useful Life (RUL) based on Performance Evolution Prediction The remaining useful life is calculated using the probability threshold crossing method: a. Performance evolution trajectory prediction: Using a trained PINN model, the evolution trajectory of key performance parameters of seismic isolation bearings in the next T years is predicted; Considering the uncertainty of the model parameters, generate N possible performance evolution trajectories, N=1000; b. Failure time calculation: For each performance evolution trajectory, find the time t_f when the failure threshold is first crossed; If the failure threshold is not crossed within the predicted time T, then t_f = T; c. Calculation of RUL probability distribution: By statistically analyzing the distribution of all t_f, we obtain the probability density function (PDF) and cumulative distribution function (CDF) of RUL. Calculate the statistical characteristics of RUL: mean, median, standard deviation, and 95% confidence interval; 2) Predictive maintenance decision-making based on RUL Based on the RUL prediction results and failure probability, a graded maintenance strategy is formulated: RUL > 10 years, failure probability < 5%: Normal maintenance, detect once a year; 5 years < RUL ≤ 10 years, 5% ≤ failure probability < 20%: Strengthen monitoring, detect once every six months; 2 years < RUL ≤ 5 years, 20% ≤ failure probability < 50%: Scheduled maintenance, formulate replacement plan; RUL ≤ 2 years, failure probability ≥ 50%: Emergency maintenance, replace the bearing immediately; S6: Adaptive safety warning, combine real-time environmental factors and system status, set dynamic warning thresholds, and establish a multi-level warning mechanism. Different warning mechanisms adopt different response measures for handling; The method for setting dynamic warning thresholds is as follows: 1) Establishment of multi-field coupling benchmark performance model Establish a three-dimensional benchmark mapping relationship of environment-state-performance as the calculation basis for dynamic thresholds. Based on the previously calibrated multi-field coupling quantitative analysis model, generate a benchmark performance database covering the full operating condition range, and use the response surface method to fit to obtain a continuous benchmark performance surface: P0 = f (T, H, σ, N); Input the environmental parameters T, H, stress state σ, and load history N at the current moment, call the benchmark performance surface function, calculate the theoretical benchmark performance parameter P0 (t) under the current operating condition, and calculate the relative deviation of the performance parameter: ΔP (t)= (P_measured (t) - P0 (t)) / P0 (t)×100%. All warning thresholds are set based on relative deviation rather than absolute value; Adopt the Bayesian online learning framework, and continuously update the threshold distribution using real-time monitoring data and historical warning feedback data to achieve adaptive optimization of the thresholds; On the basis of the basic Bayesian framework, introduce three key correction factors to further improve the adaptability of the thresholds: Aging correction factor: f_aging = 1 + k_aging·t, where k_aging is the annual aging rate coefficient and t is the service life; Corrected threshold: T' = T·f_aging; Temporary correction factor after earthquake: f_earthquake = 1.5, valid within 30 days after the earthquake; Corrected threshold: T' = T·f_earthquake; Sensor reliability correction factor: f_sensor = 1 + k_sensor·(1 - R); where R is the sensor health, 0~1; k_sensor is the correction coefficient; Corrected threshold: T' = T·f_sensor; In this embodiment, the multi-level early warning mechanism consists of four levels: abnormal status, performance degradation, damage confirmation, and failure risk. Each level corresponds to specific triggering conditions and risk levels; the details are shown in the table below. Warning level Color Code Triggering conditions Risk Description normal green ΔP <T_attention The system is working normally, and performance parameters fluctuate within the baseline range. Notice blue T_attention ≤ ΔP < T_warning or it is predicted that the warning threshold will be reached within 30 days Performance parameters are fluctuating abnormally, but have not yet reached the degradation criteria; or performance degradation is predicted to occur in the future. warn yellow T_warning ≤ ΔP < T_danger or early damage features are detected or it is predicted that the danger threshold will be reached within 7 days Significant performance degradation has occurred, or early signs of damage have been detected; the system may still function normally, but requires enhanced monitoring and planned maintenance. Danger red ΔP ≥ T_danger or severe damage characteristics are detected or failure is predicted within 24 hours. Performance has severely degraded, or there is a risk of immediate failure; immediate emergency measures are required. Use composite triggering logic: A single parameter exceeding the threshold only triggers an initial warning. If two or more related parameters exceed the threshold simultaneously, the warning level will be automatically upgraded to level one. If damage is confirmed by combining the damage identification results from multi-source data fusion, the warning level is directly upgraded to warning or above. Based on the performance evolution prediction results, if the predicted damage will develop rapidly, the warning level will be upgraded in advance. Different early warning mechanisms require different response measures: Blue Alert Level Response Process: When an alert is triggered, the system automatically records the alert information and sends a notification to the maintenance personnel. The maintenance personnel check the relevant monitoring data and historical trends within 24 hours. The system determines whether the alert is caused by environmental factors or measurement errors. If it is confirmed to be a false alarm, the system records the information and adjusts the threshold. If not, the system increases the monitoring frequency of the support. The system observes the system for 7 days. If the performance parameters return to normal, the alert is lifted. If the situation continues to deteriorate, the alert is upgraded to the warning level. Yellow Warning Level Response Process: Upon triggering the warning, the system automatically records and generates a preliminary analysis report, simultaneously sending SMS, email, and app notifications to the operations and maintenance manager. The manager must organize a specialized on-site inspection (including visual inspection, non-destructive testing, and performance testing) within 48 hours. If no obvious damage is found: strengthen monitoring and develop a preventative maintenance plan. If early damage is found: assess the impact of the damage, develop a repair plan, and implement it within 30 days. After repair, conduct performance acceptance testing; if successful, the warning is lifted; otherwise, it is upgraded to a hazardous level. Red Danger Level Response Procedure: Upon triggering the warning, the system immediately sends emergency notifications via SMS, telephone, and APP pop-up to all relevant personnel, activates the emergency plan, restricts the use of the structure, and evacuates personnel if necessary; within 24 hours, organizes experts to conduct an on-site assessment, formulates an emergency response plan, and implements it immediately. For structures that can be temporarily reinforced, temporary reinforcement measures are taken, and the load is restricted; for structures that cannot be reinforced, use is immediately stopped, and the supports are replaced; after the handling is completed, a comprehensive safety assessment is conducted, and use can only be resumed after passing the assessment.

Claims

1. An in-situ testing method for seismic isolation and damping systems under complex multi-field coupling environments, characterized in that, include: S1: Sensor array deployment. At least the following data acquisition sensors shall be deployed on the seismic isolation and damping bearing: temperature sensor, humidity sensor, strain monitoring sensor, displacement field sensor, acceleration sensor, and chemical corrosion sensor. S2: Data processing, which involves preprocessing and feature extraction of the multi-source raw data collected in S1; S3: Construction of a quantitative analysis model for multi-field coupling effects. This involves establishing a mechanical performance analysis model for a seismic isolation system that considers the coupling effects of at least the temperature field, humidity field, stress field, seismic field, and wind field. This model includes: Material constitutive relation models are used to establish and analyze the constitutive relations between the materials of vibration isolation and damping bearings and multi-field coupling; Multi-field coupling control equations are used to consider the influence of performance changes of rubber materials under temperature and humidity coupled field environments; Performance parameter mapping is used to establish a quantitative mapping relationship between various performance parameters of the vibration isolation and damping bearing and the influencing factors under multi-field coupling. S4: Deep fusion of multi-source heterogeneous data to build a three-level fusion architecture of data layer - feature layer - decision layer. The data layer fusion performs time synchronization and spatial registration of different types of sensors, and performs noise reduction and missing value completion on the data. Feature layer fusion extracts multi-dimensional features from different types of sensors, including at least time domain, frequency domain, and time-frequency domain features; deep learning algorithms are used to fuse and reduce the dimensionality of the extracted features, and establish a mapping relationship between the features and the performance parameters and damage state of the seismic isolation system; The decision-making layer integrates multiple damage identification results and combines them with the prediction results of a multi-field coupling analysis model to improve the accuracy and reliability of damage identification; it enables accurate identification of damage location, damage type, and damage degree. S5: Performance Evolution and Damage Prediction. A deep learning-based model for the performance evolution and damage prediction of seismic isolation systems is established, including: The data-driven prediction model uses a long short-term memory network combined with an attention mechanism to establish a prediction model for the evolution of the performance parameters of the seismic isolation system over time, which is used to predict the predicted values ​​and confidence intervals of the performance parameters at different future time points. The physical information fusion prediction model integrates the physical constraints of the multi-field coupling analysis model into the deep learning model and adopts a physical information neural network to predict the performance evolution under small sample and high uncertainty conditions. Establish failure criteria for seismic isolation systems, clarify the performance thresholds corresponding to different degrees of damage, and combine the performance evolution prediction model to predict the time when the seismic isolation system reaches the failure criteria, and give the probability distribution and confidence interval of the remaining service life. S6: Adaptive safety early warning, which combines real-time environmental factors and system status to set dynamic early warning thresholds and establish a multi-level early warning mechanism. Different early warning mechanisms adopt different response measures.

2. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, The sensor array is deployed in the following manner: The temperature sensors are arranged radially and axially within the rubber layer of the vibration damping support, forming a three-dimensional temperature field monitoring network. The humidity sensor is installed inside and around the vibration damping support to monitor the humidity distribution inside and outside the rubber. The strain monitoring sensors are arranged on the surface of the steel plate layer of the vibration isolation bearing to monitor the stress distribution of the steel plate layer; The displacement field sensor monitors the three-dimensional displacement of the seismic isolation bearing; The acceleration sensor is installed on the upper and lower connecting plates and around the vibration damping support; The chemical corrosion sensor is installed on the surface of the vibration damping support to monitor corrosive chemicals in the environment, including chloride ions and sulfur dioxide.

3. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, In S2, the edge nodes for data processing are deployed at the data acquisition site. The raw data acquired is immediately preprocessed, including missing value imputation, noise reduction, and normalization. The objectives of the feature extraction include: Acceleration features are extracted first, prioritizing the extraction of time-domain features with the highest real-time performance; then frequency-domain features are extracted to reflect changes in the inherent characteristics of the structure; time-frequency domain features are used to capture non-stationary seismic responses. Displacement data characteristics, including static and quasi-static characteristics such as maximum horizontal displacement, maximum vertical displacement, and residual displacement; hysteresis loop characteristics, a core indicator used to reflect the energy dissipation capacity of seismic isolation bearings; spatial distribution characteristics; Strain data characteristics, including basic mechanical characteristics, fatigue damage characteristics, and spatial distribution characteristics, are used to reflect the internal stress state of the support. Temperature and humidity data features include temperature field features, humidity field features, and temperature and humidity coupling features; these serve as key input features for multi-field coupling effects. Corrosion monitoring data features are extracted, with a focus on features related to corrosion rate, to reflect the degree of environmental erosion.

4. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, The material constitutive relation model includes at least the following: 1) Multi-factor coupled hyperelastic constitutive model of rubber materials Based on the modified Yeoh hyperelasticity model, three correction factors—temperature, humidity, and degree of aging—are introduced, among which: Temperature correction factor: ,in This is the temperature influence coefficient. For reference temperature, set to 25℃; Humidity correction factor: ,in Humidity influence coefficient For reference humidity, set to 50%RH; Aging correction factor: ,in A represents the aging rate, and A represents the degree of aging, set to 0~1; 2) Temperature-Cyclic Coupled Elastoplastic Constitutive Model of Lead Core Material By adopting the modified Ramberg-Osgood elastoplastic model and introducing temperature correction and cyclic cumulative damage correction, the mechanical behavior of lead core under different temperatures and cyclic loading cycles can be accurately described. Temperature correction factor: ,in This is the temperature softening coefficient; Cyclic cumulative damage correction factor: ,in The cyclic damage coefficient, The damage index is N, and the number of loading cycles is N. 3) Temperature-corrosion coupled elastic constitutive model of steel plate material By adopting a linear elastic constitutive model and introducing temperature correction and corrosion damage correction, the mechanical behavior of steel plates under different temperatures and corrosion levels can be accurately described. in The elastic modulus is taken as a reference temperature and under non-corrosive conditions. This is the temperature influence coefficient; For corrosion depth, This represents the critical corrosion depth. This represents the corrosion impact coefficient.

5. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, The multi-field coupled control equations establish a set of fully coupled thermal-humid-mechanical control equations, considering the bidirectional coupling effect between various physical fields, and are numerically solved using the finite element method. 1) Heat conduction equation: The density of the material; Specific heat capacity; Let be the thermal conductivity, a function of temperature and humidity. ; Heat generated by mechanical deformation: Heat generated by viscoelastic dissipation: v; 2) Wet diffusion equation: ,in Humidity concentration; Let be the moisture diffusivity, a function of temperature and stress. ; Here, R is the diffusion activation energy, R is the gas constant, and s is the humidity source term. 3) Equilibrium equations: For volume forces, For acceleration; It is the elastic stiffness tensor; For temperature strain, For humidity strain, For plastic deformation, only lead core material; The solution process for the coupled control equations is as follows: The sequential coupling iterative solution method is adopted, in which each physical field is solved alternately in each time step until convergence: Initial step: Input the initial temperature field T0, humidity field H0, stress field σ0, and displacement field u0; nth time step: Solve the heat conduction equation to obtain the temperature field Tn; Solving the moisture diffusion equation yields the humidity field Hn; Update the material constitutive parameters based on Tn and Hn; Solving the mechanical equilibrium equations yields the displacement field un and the stress field σn; Calculate the heat generated by mechanical deformation, Q_mech, and the heat dissipation due to viscoelastic dissipation, Q_visc. Check the convergence conditions: ||Tn - Tn-1|| < ε_T and ||Hn - Hn-1|| < ε_H and ||un - un-1|| < ε_u; If convergence fails, return to step one and recalculate; if convergence occurs, proceed to the next time step.

6. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, The performance parameter mapping adopts a hybrid modeling method of "physical model + data-driven" to establish a quantitative mapping relationship between the key performance parameters of the seismic isolation bearing and multiple influencing factors. 1) Definition and calculation of key performance parameters Obtain the hysteresis curves of the seismic isolation bearings through finite element simulation or experiments, and calculate the following key performance parameters: Horizontal stiffness K_h: K_h = (F_max - F_min) / (Δ_max - Δ_min) Vertical stiffness K_v: K_v = (V_max - V_min) / (δ_max - δ_min) Equivalent damping ratio ξ_eq: ξ_eq = W_d / (4πW_e), where W_d is the hysteresis loop area and W_e is the elastic strain energy; 2) Quantitative mapping relationship between performance parameters and influencing factors The response surface methodology was used to establish the mapping relationship between performance parameters and temperature T, humidity H, stress σ, aging degree A, and number of cyclic loading N. P = a0 + Σai·xi + Σaij·xi·xj + Σaijk·xi·xj·xk Where P is a performance parameter, which can be K_h, K_v, or ξ_eq; xi represents the standardized influencing factors; a0, ai, aij, and aijk are response surface coefficients; 3) Performance parameter degradation rate model: A modified Arrhenius equation is used to establish a quantitative relationship between the performance parameter degradation rate and environmental factors and load history. dP / dt = A·exp(-E_a / RT)·f(H)·f(σ)·f(N) Where A is the pre-exponential factor; E_a is the activation energy; f(H) is the humidity effect function: f(H) = 1 + k_H·H; f(σ) is the stress influence function: f(σ) = 1 + k_σ·σ; f(N) is the effect function of cyclic loading: f(N) = 1 + k_N·N; Establish a predictive model for the evolution of performance parameters over time: P(t) = P0·exp(-dP / dt·t).

7. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, The data layer fusion method for multi-source data time synchronization for different types of sensors is as follows: a. Hardware-level time synchronization: All sensors and edge computing nodes are connected to the GPS / BeiDou timing system, providing an absolute time reference at the 1μs level; b. Data labeling: Each data point collected by the sensor is accompanied by a precise timestamp; c. Establishing a unified time axis: using the time axis of the sensor with the highest sampling frequency as the reference; d. Interpolation resampling: Low-frequency data: resampled to 1Hz using cubic spline interpolation; Intermediate frequency data: resampled to 100Hz using linear interpolation; High-frequency data: Maintain the original sampling frequency of 1000Hz; e. Time alignment: Align all sensor data to a unified point in time to form a time-synchronized multi-source data matrix; The method for spatial registration of multi-source data is as follows: a. Coordinate system definition: The origin is the center of the connecting plate under the seismic isolation bearing, the horizontal direction is the XY axis, and the vertical direction is the Z axis; b. Sensor spatial coordinate calibration: Accurately measure the coordinates (xi, yi, zi) of each sensor in the three-dimensional coordinate system of the support; c. Spatial interpolation mapping: Distributed fiber optic data: directly forms a three-dimensional spatial distribution field; Point sensor data: A continuous spatial distribution field is generated using Kriging interpolation. d. Spatial alignment: Align all physical field data to a unified spatial grid; The multi-source data denoising adopts an adaptive denoising algorithm that combines wavelet transform and Kalman filtering to perform targeted denoising based on the noise characteristics of different types of sensor data; to address the data missing problem, a hybrid completion method combining spatiotemporal correlation completion and machine learning completion is adopted. The deep learning algorithm in the feature layer fusion uses a CNN-LSTM hybrid neural network architecture to achieve deep fusion and dimensionality reduction of multi-source heterogeneous features. This architecture can simultaneously capture the spatial and temporal correlations of features. The training method is as follows: 1) Dataset Construction: A large amount of simulation data was generated using a multi-field coupled finite element model; Combining laboratory test data and actual engineering monitoring data; Standardize the data: x_norm = (x - mean) / std; 2) Loss function design: A multi-task loss function is adopted to simultaneously optimize performance parameters for regression and damage classification tasks; 3) Optimizer: Adam optimizer is used, with a learning rate of 0.001, a batch size of 32, and 100 training epochs; 4) Regularization: Dropout (0.2) and L2 regularization (1e-4) are used to prevent overfitting; The process of establishing the mapping between the features and the performance parameters and damage state of the seismic isolation system is as follows: By training the CNN-LSTM hybrid neural network, a nonlinear mapping relationship is automatically established between the fused features and the performance parameters of the seismic isolation system, and between the fused features and the damage state of the seismic isolation system. Performance parameter mapping: Network output layer 1 directly outputs the predicted values ​​of horizontal stiffness, vertical stiffness, and equivalent damping ratio; Damage state mapping: The network output layer 2 outputs the probability distribution of different damage states, including: No damage; rubber aging; lead core yielding; steel plate corrosion; interface debonding; support slippage; The decision layer integrates multiple independent damage identification results and prediction results from multi-field coupled physical models to make the final and most reliable damage decision: 1) Acquisition of multi-source damage identification results Multiple independent damage identification results are obtained and used as input for fusion in the decision layer: Data-driven damage identification results: damage classification probability output from the aforementioned CNN-LSTM network; Damage identification results from the physical model: Input the monitored performance parameters into the multi-field coupling analysis model; The residuals between the performance parameters predicted by the calculation model and the actual monitored values ​​are calculated. The damage status and degree are determined based on the magnitude of the residual; Single sensor damage identification results: Accelerometer damage identification results; Damage identification results from strain sensor; Temperature sensor damage identification results; 2) Fusion decision-making based on improved DS evidence theory An improved DS evidence theory is used to fuse multiple damage identification results for decision-making. a. Recognition framework establishment: Θ = {D0, D1, D2, D3, D4, D5}, corresponding to 6 damage states respectively; b. Construction of basic probability assignment: For data-driven results: directly use the probability output by the network as the BPA; For the physical model results: construct the BPA based on the residual size. The larger the residual, the higher the BPA of the corresponding damage. For single-sensor results: construct a BPA based on the degree of feature change; c. Measurement of Evidence Conflict: The degree of conflict between pieces of evidence is measured using the Jousselme distance. d(m1, m2) = sqrt(0.5·||m1 - m2||²); d. Calculation of Evidence Weights: Calculate the weight of each piece of evidence based on its reliability and the degree of conflict. w_i = (1 - d_i) / Σ(1 - d_j), where d_i is the average distance between the i-th piece of evidence and all other pieces of evidence; e. Weighted average evidence synthesis: m_avg(A) = Σ(w_i·m_i(A)) f. Final decision: Using the principle of maximum basic probability allocation, the damage state with the highest BPA is selected as the final decision result; 3) Fusion of multi-field coupling models and data-driven results a. Using a multi-field coupling analysis model, predict the performance parameters of the seismic isolation bearing under no-damage conditions based on current environmental factors; b. Calculate the deviation between the actual monitored performance parameters and the model predictions; c. Use this deviation as a correction term to adjust the data-driven damage identification results. m_final(D) = m_data(D)·(1+ k·ΔP), where ΔP is the performance parameter deviation and k is the correction coefficient; d. Normalize the corrected BPA to obtain the final damage probability distribution; 4) Precise identification of the location, type, and extent of damage. Based on determining the type of damage, further precise identification of the damage location and extent is achieved: a. Damage location identification: For distributed fiber optic strain data: determine the location of damage based on the location of strain anomaly areas; For acceleration data: the location of the damage is determined based on the differences in frequency domain characteristics of sensors at different locations; By fusing positional information from multiple sensors, the precise location of the damage is determined using triangulation. b. Damage assessment: Establish a feature database for different degrees of damage; The support vector regression algorithm is used to establish a mapping relationship between fused features and damage level; Input the current fusion feature and output a quantified value of the degree of damage, 0~1, where 0 represents no damage and 1 represents complete failure.

8. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, The data-driven prediction model employs a bidirectional LSTM network architecture combined with a multi-head self-attention mechanism, capable of simultaneously capturing long-term dependencies in performance evolution and the impact of key time points; the model input feature vector is: 32-dimensional historical performance parameter sequence: daily average values ​​of key performance parameters over the past 30 days; 48-dimensional environmental factor sequence: daily average values ​​of temperature field characteristics, humidity field characteristics, and corrosion rate environmental parameters over the past 30 days; 32-dimensional load history sequence: daily statistical values ​​of load parameters over the past 30 days; 16-dimensional current state features: current state parameters; Output result: Predicted values ​​of horizontal stiffness, vertical stiffness, and equivalent damping ratio for the next 7 days, 30 days, 90 days, 180 days, and 365 days; 95% confidence interval for each predicted value; The confidence interval estimation of the predicted value is achieved using the quantile regression method; The performance prediction evolution model of the physical information fusion adopts a physical information neural network to incorporate the control equations of the multi-field coupling analysis model as soft constraints into the loss function, so that the model can fit the observed data and satisfy the physical laws at the same time during the training process, which significantly improves the generalization ability and reliability of the model. The multi-field coupling control equations established in S3 are transformed into PINN loss function terms, as follows: The heat conduction equation constraint is: L_heat = MSE(ρc·∂T / ∂t -∇·[k(T,H)·∇T] - Q_mech - Q_visc); Moisture diffusion equation constraint: L_moisture = MSE(∂C / ∂t -∇·[D(T,σ)·∇C] - S); Mechanical equilibrium equation constraint: L_mech = MSE(∇·σ+ρb -ρü); Performance degradation equation constraint: L_degradation = MSE(dP / dt - A·exp(-E_a / RT)·f(H)·f(σ)·f(N)); The PINN network structure is as follows: Input layer: Input time t, temperature T, humidity H, stress σ, number of cycles N; Output layer: Outputs horizontal stiffness K_h, vertical stiffness K_v, and equivalent damping ratio ξ_eq; Total loss function: L_total = w_data·L_data + w_physics·(L_heat + L_moisture + L_mech + L_degradation) + w_boundary·L_boundary L_data = MSE(NN_pred, observed_data), which is the data fitting loss; L_boundary = MSE(NN_pred_boundary, boundary_conditions), which is the boundary condition loss; w_data=1.0, w_physics=0.1, w_boundary=0.5 are the weighting coefficients; The failure criteria for the seismic isolation system are established based on the "Standard for Seismic Isolation Design of Buildings" GB 51416-2021 and relevant industry standards. The following operations are performed: 1) Calculation of remaining useful life based on performance evolution prediction The remaining useful life is calculated using the probability threshold crossing method: a. Performance evolution trajectory prediction: Using a trained PINN model, the evolution trajectory of key performance parameters of seismic isolation bearings in the next T years is predicted; Considering the uncertainty of the model parameters, generate N possible performance evolution trajectories, N=1000; b. Failure time calculation: For each performance evolution trajectory, find the time t_f when the failure threshold is first crossed; If the failure threshold is not crossed within the predicted time T, then t_f = T; c. Calculation of RUL probability distribution: By statistically analyzing the distribution of all t_f, we can obtain the probability density function and cumulative distribution function of RUL. Calculate the statistical characteristics of RUL: mean, median, standard deviation, and 95% confidence interval; 2) Predictive maintenance decision-making based on RUL Based on the RUL prediction results and failure probability, a graded maintenance strategy is formulated: RUL > 10 years, failure probability < 5%: normal maintenance, annual inspection; 5 years < RUL ≤ 10 years, 5% ≤ failure probability < 20%: Strengthen monitoring, conduct tests every six months; 2 years < RUL ≤ 5 years, 20% ≤ failure probability < 50%: Planned maintenance, develop a replacement plan; RUL ≤ 2 years, failure probability ≥ 50%: Emergency maintenance, replace the support immediately.

9. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, The method for setting the dynamic early warning threshold is as follows: Establishment of a multi-field coupled benchmark performance model A three-dimensional benchmark mapping relationship between environment, state, and performance is established as the basis for calculating dynamic thresholds. Based on the previously calibrated multi-field coupled quantitative analysis model, a benchmark performance database covering the entire working condition range is generated. A continuous benchmark performance surface is obtained by fitting using the response surface method: P0 = f (T,H,σ,N). The environmental parameters, stress state, and load history at the current moment are input, and the benchmark performance surface function is called to calculate the theoretical benchmark performance parameter P0 (t) under the current working condition. The relative deviation of the performance parameter is calculated: ΔP (t) = (P_measured (t) - P0 (t)) / P0 (t)×100%. All warning thresholds are set based on relative deviation rather than absolute value. By employing a Bayesian online learning framework, the threshold distribution is continuously updated using real-time monitoring data and historical early warning feedback data, thereby achieving adaptive optimization of the threshold. Based on the basic Bayesian framework, three key correction factors are introduced to further improve the adaptability of the threshold: Aging correction factor: f_aging = 1 + k_aging·t, where k_aging is the annual aging rate coefficient and t is the number of years in service; Corrected threshold: T' = T·f_aging; Post-earthquake temporary correction factor: f_earthquake = 1.5, valid for 30 days after the earthquake; Corrected threshold: T' = T·f_earthquake; Sensor reliability correction factor: f_sensor = 1 + k_sensor·(1 - R); where R is the sensor health status, 0~1; k_sensor is the correction coefficient; the corrected threshold: T' = T·f_sensor.

10. The in-situ testing method for a seismic isolation and damping system under complex multi-field coupling environment according to claim 1, characterized in that, Establish a multi-level early warning mechanism with four levels: abnormal status, performance degradation, damage confirmation, and failure risk. Each level corresponds to a clear triggering condition and risk level. Use composite triggering logic: A single parameter exceeding the threshold only triggers an initial warning. If two or more related parameters exceed the threshold simultaneously, the warning level will be automatically upgraded by one level. If damage is confirmed by combining the damage identification results from multi-source data fusion, the warning level will be directly upgraded to warning or above. Based on the performance evolution prediction results, if the predicted damage will develop rapidly, the warning level will be upgraded in advance. Different response measures are set according to different early warning mechanisms, as follows: Blue Alert Level Response Process: When an alert is triggered, the system automatically records the alert information and sends a notification to the maintenance personnel. The maintenance personnel check the relevant monitoring data and historical trends within 24 hours. The system determines whether the alert is caused by environmental factors or measurement errors. If it is confirmed to be a false alarm, the system records the information and adjusts the threshold. If not, the system increases the monitoring frequency of the support. The system observes the system for 7 days. If the performance parameters return to normal, the alert is lifted. If the situation continues to deteriorate, the alert is upgraded to the warning level. Yellow warning level response process: When an alert is triggered, the system automatically records and generates a preliminary analysis report, and simultaneously sends SMS, email, and APP notifications to the operations and maintenance manager. The operations and maintenance manager organizes on-site special testing within 48 hours. If no obvious damage is found: Strengthen monitoring and develop a preventative maintenance plan; If early damage is detected: assess the impact of the damage, develop a repair plan, and implement it within 30 days; after the repair is completed, conduct a performance acceptance test, and remove the warning if the acceptance test is successful. The substandard level is upgraded to a hazardous level; Red Danger Level Response Procedure: Upon triggering the warning, the system immediately sends emergency notifications via SMS, telephone, and APP pop-up to all relevant personnel, activates the emergency plan, restricts the use of the structure, and evacuates personnel if necessary; within 24 hours, organizes experts to conduct an on-site assessment, formulates an emergency response plan, and implements it immediately. For structures that can be temporarily reinforced, temporary reinforcement measures are taken, and the load is restricted; for structures that cannot be reinforced, use is immediately stopped, and the supports are replaced; after the handling is completed, a comprehensive safety assessment is conducted, and use can only be resumed after passing the assessment.

Images