A deep learning-based damage identification method for shock isolation and vibration reduction components

Abstract

This invention discloses a damage identification method for seismic isolation and damping components based on deep learning, belonging to the field of structural monitoring and identification technology based on deep learning. It includes: S1: Multi-source sensor data acquisition and preprocessing, with five sets of sensors installed on the upper and lower connecting plates of the support and the surface of the rubber layer for multi-source data acquisition; S2: Multi-modal feature extraction with physical information enhancement, extracting features from the five sets of multi-modal data and enhancing features based on constitutive equations; S3: Multi-task deep learning model training, training the deep learning model and fine-tuning it based on a small amount of real data; S4: Damage identification and interpretability analysis, with the model outputting damage type, damage location, and quantified damage degree; S5: Real-time monitoring and early warning through edge-cloud collaboration, deploying a lightweight model at the edge and a high-precision model in the cloud. This method addresses the core technical shortcomings of existing technologies, such as high data dependence, low identification accuracy, poor generalization ability, insufficient interpretability, and weak environmental robustness.

Description

Technical Field

[0001] This invention belongs to the field of structural monitoring and identification technology based on deep learning, specifically relating to a method for identifying damage to seismic isolation and damping components based on deep learning. Background Technology

[0002] Seismic isolation and damping technology is a core technical means to improve the seismic performance of structures in the field of building and bridge engineering. By installing seismic isolation bearings, dampers, and other components in the foundation or between floors, seismic energy can be effectively isolated from the transmission to the upper structure, significantly reducing the structural response under seismic loads. With the continuous expansion of engineering construction in high-intensity seismic zones in my country and the sustained growth in the demand for seismic reinforcement of existing buildings, the number of seismic isolation and damping components used has increased exponentially. However, during long-term service, seismic isolation and damping components are inevitably affected by environmental erosion, load cycles, and seismic loads, resulting in various types of damage such as rubber aging, lead core yielding, steel plate delamination, bolt loosening, and bearing slippage. These damages can lead to degradation of the mechanical properties of the components, and in severe cases, can even cause the failure of the seismic isolation system, resulting in significant loss of life and property. Therefore, real-time and accurate damage identification and condition assessment of seismic isolation and damping components are crucial to ensuring the safe service of seismic isolation structures.

[0003] In recent years, deep learning-based structural damage identification methods have gradually become a research hotspot in this field due to their powerful feature extraction and nonlinear modeling capabilities. Existing technologies have proposed various deep learning-based damage identification schemes for seismic isolation and damping components. These include using multi-input convolutional neural networks to fuse acceleration and displacement signals for damage identification, employing unsupervised anomaly detection methods to address the scarcity of damage data, using computer vision technology to identify apparent damage on support surfaces, and combining piezoelectric admittance technology with transfer learning to achieve cross-domain damage identification. While these methods have achieved certain identification results under ideal laboratory conditions, they still suffer from the following technical shortcomings that urgently need to be addressed in practical engineering applications: 1) The data bottleneck problem is prominent. In actual engineering, the seismic isolation and damping components are designed and calculated with precision and accuracy. The probability of damage events of the components is actually low, resulting in extremely scarce real damage data and high labeling costs. Furthermore, there are inherent model errors and distribution differences between damage data generated based on numerical simulation and real engineering data. Directly using it to train the model will lead to a serious lack of model generalization ability, making it impossible to apply stably in actual engineering. 2) The level of precision in damage identification is insufficient. Most existing methods can only identify the overall stiffness degradation of the component as a comprehensive indicator, and cannot distinguish between different types of damage with different physical mechanisms, such as rubber aging, lead core yielding, and steel plate delamination. It is also difficult to quantify the contribution of each damage when multiple damages exist at the same time. However, the impact of different types of damage on the performance of the component and the corresponding maintenance strategies are completely different. Simply identifying stiffness degradation cannot provide a scientific basis for engineering maintenance decisions. 3) Insufficient utilization of multimodal information: Existing methods mostly integrate monitoring data of only two modes, acceleration and displacement, and fail to make full use of multi-source information that is more sensitive to specific damage, such as force-displacement hysteresis curves, temperature, strain, and acoustic emission. The lack of this information limits the further improvement of damage identification accuracy. Summary of the Invention

[0004] This invention provides a deep learning-based method for damage identification of seismic isolation and damping components. It innovatively adopts an "edge-cloud collaborative" distributed computing architecture, breaking through the inherent bottlenecks of traditional centralized data processing in terms of real-time performance, computing resources, and network bandwidth. This method deeply and organically integrates the physical and mechanical laws of seismic isolation and damping components with data-driven multimodal deep learning methods, effectively solving the core technical defects commonly found in existing technologies, such as high data dependence, low identification accuracy, poor generalization ability, insufficient interpretability, and weak environmental robustness. Ultimately, it achieves accurate differentiation of multiple types of damage in seismic isolation and damping components, precise location of damage, accurate quantification of damage severity, and interpretable reasoning of damage mechanisms, providing scientific and reliable technical support for the safe operation and maintenance of seismic isolation structures throughout their entire lifecycle.

[0005] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution: A deep learning-based method for damage identification of seismic isolation and damping components includes: S1: Multi-source sensor data acquisition and preprocessing: Multi-source sensors are installed on the upper and lower connecting plates of the seismic isolation and damping bearing; Multi-source sensors are also attached to the surface of the rubber layer of the seismic isolation and damping bearing. Collect raw data from multiple sources and preprocess the data; S2: Multimodal feature extraction with enhanced physical information. Feature extraction is performed on acceleration signal data, displacement signal data, force signal data, strain signal data, and temperature signal data respectively. All features are input into the physical information embedding layer and feature enhancement is performed in combination with the constitutive equation of the seismic isolation and damping bearing. S3: Multi-task deep learning model training, constructing a multi-input multi-output deep learning model hierarchical structure, using meta-learning methods for pre-training, fine-tuning based on a small amount of real engineering data, and then using a progressive sample generation method to expand the training set; S4: Damage identification and interpretability analysis. Input the data to be detected into the deep learning model trained in S3, and output the probability distribution of damage type, damage location coordinates and damage degree quantification value (0~100%) respectively. The model displays key signal features of interest and, based on a feature-physical quantity mapping library, transforms abstract features into interpretable changes in physical parameters. It then calls the damage mechanism reasoning module, combining the identification results with physical knowledge to generate a damage analysis report. S5: Real-time monitoring and early warning of edge-cloud collaboration, deploying lightweight models at the edge for real-time data monitoring and preliminary anomaly detection; High-precision models are deployed in the cloud to perform refined damage identification on abnormal data, and graded early warnings are triggered based on the identification results.

[0006] Preferably, the multi-source sensors installed on the upper and lower connecting plates of the vibration isolation and damping bearing include at least: a triaxial accelerometer, a displacement sensor, and a force sensor; The multi-source sensors that are attached to the surface of the rubber layer of the vibration isolation and damping bearing include at least: fiber optic strain sensor and temperature sensor; The raw data of five modes were collected synchronously at a sampling frequency of 100Hz, and the sampling time was 30 seconds after each earthquake or artificial excitation.

[0007] Preferably, the multi-source raw data undergoes preprocessing including at least denoising, standardization, and missing value completion; The denoising process employs a moving average filter, which smooths the data by taking the mean within a local window; the standardization process uses the Min-Max standardization method; and the missing value completion is accomplished using interpolation. For continuous signal data, a sliding window is used to divide the continuous signal into sample segments of length 1024.

[0008] Preferably, wavelet packet transform is performed on the acceleration signal data to extract energy features of different frequency bands; hysteresis curve analysis is performed on the displacement signal data to extract physical features including at least equivalent stiffness, equivalent damping, and energy dissipation capacity; maximum strain and strain distribution uniformity features are extracted from the strain signal data; and temperature gradient and temperature change rate features are extracted from the temperature signal data. All the features extracted above are input into the physical information embedding layer, and feature enhancement is performed by combining the constitutive equation of the seismic isolation and damping bearing.

[0009] Preferably, the feature enhancement method for the extracted four types of features—acceleration, displacement, strain, and temperature—combined with the constitutive equation of the seismic isolation and damping bearing is as follows: The physical enhancement of displacement hysteresis characteristics involves extracting equivalent stiffness, equivalent damping, and energy dissipation capacity from displacement signals, which are macroscopic statistical characteristics. These characteristics can be transformed into constitutive parameter characteristics through constitutive equations. 1) Constitutive parameter identification and feature generation Input the measured displacement-force hysteresis curve, and fit the Bouc-Wen model using the least squares method or genetic algorithm to obtain the values ​​including α, K1, and u. y F y Constitutive parameters of β and γ; Output enhancement features: yield strength degradation rate, stiffness degradation coefficient, hysteresis pinching exponent, and cumulative plastic deformation energy; 2) Feature extraction of internal state variables Using the hysteresis variable z(t) in the Bouc-Wen model, load history features are extracted: Maximum hysteresis value z max : Reflects the maximum degree of plastic deformation experienced by the support Rate of change of hysteretic variables : The rate of energy dissipation The cumulative value of the hysteretic variable ∫∣z(t)∣dt: reflects the total cumulative damage; Physical calibration of acceleration energy characteristics: Physical consistency calibration is performed using constitutive equations based on energy features extracted from different frequency bands through wavelet packet transform. The theoretical dynamic response of the support is calculated based on the constitutive equation, and the theoretical energy distribution of each frequency band is obtained. By comparing the measured energy characteristics with the theoretical energy distribution, errors caused by sensor noise and environmental interference are corrected. Enhanced features include: energy deviation ratio, high-frequency energy ratio correction value, and resonant band energy increment. Multiscale fusion of strain characteristics: Strain signals reflect the local deformation state of the support and can be fused with the macroscopic mechanical characteristics of the constitutive equations at multiple scales. Establish the mapping relationship between strain distribution and overall support stiffness: ,in , These represent the maximum strain and the uniformity of strain distribution, respectively. Correcting the yield displacement in the constitutive equation using the maximum strain characteristic u y : ,in This represents the theoretical yield displacement of the seismic isolation bearing under standard reference conditions; it is a factory-designed inherent parameter. Enhanced features generated: local strain concentration factor, stiffness-strain correlation index; Physical correction of temperature characteristics: Temperature has a significant impact on the mechanical properties of seismic isolation and damping bearings. A temperature-constitutive parameter correction model is used to convert temperature characteristics into performance correction coefficients. Lead core yield strength correction: For every 1°C increase in temperature, the yield strength decreases by 0.4%; among which... This represents the theoretical yield strength of the seismic isolation bearing under standard reference conditions; Rubber stiffness correction: ,in Given prior parameters; Friction coefficient correction: ,in Given prior parameters; The extracted temperature gradient and temperature change rate features are input into the above-mentioned modified model to generate constitutive parameter features after temperature correction, which significantly improves the model's generalization ability under different ambient temperatures. Characteristic regularization of the physical constraint loss function: During the training process of the physical information embedding layer, constitutive equation residual loss is incorporated: 2 ,in F pred The restoring force is calculated based on the constitutive equation and the model parameter θ. F mea It is the measured resilience; by minimizing L phy This guides the model to learn feature representations that conform to physical laws, avoiding unreasonable results such as negative stiffness, negative energy consumption, and sudden stiffness changes.

[0010] Preferably, the hierarchical structure of the deep learning model is as follows: Input layer: Includes feature data for five modalities; Shared feature layer: Composed of multiple 1D convolutional layers and LSTM layers, it extracts temporal features; Task feature layer: corresponding to the three tasks of damage type classification, location localization, and severity quantification; Physical constraint layer: The residuals of the dynamic equations are added to the loss function as regularization terms.

[0011] Preferably, the step of fine-tuning the model based on a small amount of real engineering data and employing domain adaptive technology to reduce the distribution difference between simulated and real data specifically involves: 1) Alignment of source and target domain data preparation and preprocessing 1.1) Dataset Definition and Partitioning Source domain data Numerical simulation datasets containing all different types of seismic isolation bearings that have completed MAML meta-training, such as: lead-core rubber bearing (LRB), high-damping rubber bearing (HDR), and friction pendulum bearing (FPS); covering 5 typical damage types, 10 levels of damage (0%-100%, 10% intervals) and 20 different excitation conditions, with a total sample size of ≥100,000; Target domain data The actual monitoring data collected at the project site to be deployed is divided into: Annotation set : 5-20 labeled samples for each damage type, including damage type, location and severity labels, obtained by professional inspectors through manual inspection; Unlabeled set A large amount of unlabeled normal operating condition and suspected damage condition data, with a sample size of ≥1,000, were continuously collected by field sensors; 1.2) Data Standardization and Physical Feature Alignment The same preprocessing methods, including denoising, sliding window segmentation (window length 1024, step size 256), and missing value imputation, are used for both source and target domain data. Standardization is performed using quantile standardization to eliminate differences in the range, sensitivity, and noise level of different sensors, and the sampling frequency is unified (100Hz). Extract the same set of physical features such as equivalent stiffness, equivalent damping, energy dissipation capacity, and strain distribution uniformity from the source and target domain data respectively; Calculate the mean and variance differences of the physical characteristics of the source domain and the target domain, and perform a linear transformation on the physical characteristics of the source domain to make its statistical distribution consistent with that of the target domain. 2) Construction of Adaptive Modules for Multimodal Adversarial Domains A multimodal domain adaptation module is inserted after the shared feature layer and before the task-specific layer of the original deep learning model. This module adopts a two-level structure of "modality-specific discriminator + global feature discriminator": The modality-specific discriminant is designed as follows: Design independent domain discriminants for each of the five modes: acceleration, displacement, force, temperature, and strain. Each discriminator uses a 3-layer fully connected network: input layer, corresponding to the single-modality feature dimension → hidden layer, 128 neurons, ReLU activation, Dropout=0.2 → output layer, 1 neuron, Sigmoid activation; Function: Align the feature distribution of each modality separately to eliminate inter-domain differences in single-modality data; The global feature domain discriminator is designed as follows: Perform domain discrimination on the global features after multimodal fusion; A 4-layer fully connected network is used: input layer, fused feature dimension 512 → hidden layer 1, 256 neurons, ReLU activation, Dropout=0.2 → hidden layer 2, 128 neurons, ReLU activation, Dropout=0.2 → output layer, 1 neuron, Sigmoid activation; Function: To align the overall feature distribution after multimodal fusion and capture cross-modal interdomain differences; Gradient Reversal Layer (GRL): Insert a gradient inversion layer before the input layer of all domain discriminators; During forward propagation, the input remains unchanged, and during backward propagation, the gradient is multiplied by -λ, where λ is an adaptive weight coefficient that increases linearly from 0 to 1 with the number of training rounds. Function: To enable adversarial training between the feature extractor and the domain discriminator, allowing the feature extractor to learn domain-invariant features; 3) Loss function design The task loss, domain adversarial loss, and physical consistency loss are weighted and combined to form the total loss function: 3.1) Mission Loss : Damage type classification: Cross-entropy loss Damage location: Mean square error loss Damage Quantification: Mean Square Error Loss Using only the target domain annotation set D t−l calculate; Domain confrontation loss : Modality-specific adversarial loss: the sum of the binary cross-entropy losses of the five modality discriminators. Global Feature Adversarial Loss: Binary Cross-Entropy Loss of the Global Discriminator Use source domain data D s and unlabeled set of target domain D t−u calculate; Physical consistency loss : The error between the equivalent stiffness corresponding to the damage level predicted by the computational model and the measured equivalent stiffness extracted from real data. Calculation formula: ,in To predict the degree of damage d i The theoretical equivalent stiffness calculated using the constitutive model of the seismic isolation bearing; To obtain real data x iThe measured equivalent stiffness extracted from it; Purpose: To ensure that the model does not violate the basic physical laws of seismic isolation bearings during domain alignment, and to avoid the problem of "distribution alignment but physical errors"; 3.2) Weighting coefficient settings α: Domain adversarial loss weight, initially 0.1, increasing linearly to 0.5 with each training epoch. β: Physical consistency loss weight, fixed at 0.3; The progressive sample generation method expands the training set by generating continuous damage evolution samples based on normal data and a small amount of damage data; a three-stage progressive training method is used to fine-tune the deep learning model.

[0012] Preferably, the three-stage progressive training and fine-tuning process of the deep learning model is as follows: Phase 1: Domain Adaptive Pre-training Freeze the shared feature layers and task-specific layers of the model, and train only the multimodal domain adaptive module; use source domain data. D s and unlabeled set of target domain D t−u Training was performed; the optimizer used was Adam, the learning rate was 1e-4, the batch size was 32, and the number of training epochs was 20; the feature distributions of the source and target domains were initially aligned, reducing the domain classification accuracy to below 60%; Phase Two: Joint Training Unfreeze task-specific layers while keeping shared feature layers frozen; use target domain annotation sets. D t−l and unlabeled sets D t−u Joint training was performed; the optimizer used was Adam, with a learning rate of 5e-5, a batch size of 16, and 30 training epochs; the model learned the damage features of the target domain while maintaining feature distribution alignment. Phase 3: Fine-tuning of the entire model Unfreeze all layers of the model; use the target domain annotation set. D t−l Supervised training was performed; the optimizer used was Adam, with a learning rate of 1e-5, a batch size of 8, and 20 training epochs; an early stopping strategy was adopted: training was stopped when the validation set loss no longer decreased for 5 consecutive epochs; this further optimized the model's performance in the target domain.

[0013] Preferably, the feature-physical quantity mapping relationship library is divided into three layers, as follows: The first layer: the basic physical parameter library, which stores all measurable physical parameters of seismic isolation and damping components, along with their normal ranges and damage thresholds. It forms the foundation of the entire mapping database. Mechanical performance parameters: equivalent stiffness, equivalent damping ratio, yield strength, ultimate bearing capacity, energy dissipation capacity, stiffness degradation rate Deformation parameters: horizontal displacement, vertical displacement, shear deformation angle, residual deformation, strain distribution Environmental parameters: ambient temperature, temperature gradient, humidity, wind speed Material parameters: rubber hardness, rubber aging degree, lead core yield strength, steel plate elastic modulus Connection parameters: bolt preload, gap between connecting plates, and coefficient of friction of the contact surface; The second layer: the feature-physical quantity association rule base, which stores the quantitative association relationships between abstract features extracted by deep learning and basic physical parameters, is the core of the mapping relationship library. Single-modal feature association rules: Establish the association between the features of five modes (acceleration, displacement, force, temperature, and strain) and their corresponding physical parameters. For example, the energy characteristics of acceleration signals in the 10-20Hz frequency band are positively correlated with the degree of rubber aging. For example, the non-uniformity of strain distribution is linearly correlated with the delamination area of ​​the steel plate. Multimodal fusion feature association rules: Establishing the association between multimodal fusion features and combinations of multiple physical parameters. For example, the pinching effect characteristics of the force-displacement hysteresis curve are correlated with the combination of the lead core yield strength and the rubber aging degree. Feature importance weights: Store the importance weight of each feature for different physical parameters and different damage types; The third layer: a damage mechanism knowledge base, which stores the physical mechanisms, evolutionary laws, and typical characteristic patterns of different types of damage, and is the foundation for damage mechanism reasoning. Correspondence between damage type and changes in physical parameters: Rubber aging: Increased equivalent stiffness, decreased equivalent damping ratio, and reduced energy dissipation capacity. Lead core yielding: equivalent stiffness decreases, equivalent damping ratio increases, and the hysteresis curve shows pinching. Steel plate delamination: uneven strain distribution, increased local strain, and slightly reduced equivalent stiffness. Loose bolts: Increased high-frequency vibration energy and increased relative displacement of connecting plates Support slippage: The hysteresis curve shows a slippage segment, and the residual deformation increases; Damage evolution patterns: the development speed and evolution path of different damage types under different loads and environmental conditions; Damage coupling effect: the mutual influence relationship when multiple damages exist simultaneously; Based on the mapping relationship library, abstract features are transformed into interpretable changes in physical parameters.

[0014] Preferably, the damage mechanism reasoning module and its process for generating the damage analysis report are as follows: The damage mechanism reasoning module includes: The physical consistency verification submodule is used to verify whether the damage identification results output by the deep learning model conform to the basic physical laws of seismic isolation and damping components, and to filter out obviously unreasonable misjudgment results. The standard physical parameter change pattern corresponding to this damage type is queried from the feature-physical quantity mapping relationship library; the cosine similarity between the physical parameter change value obtained in this identification and the standard pattern is calculated; if the similarity is lower than the threshold (0.7), it is judged as a suspected misjudgment, triggering the secondary verification process: Re-extract the physical features of the original data; call the high-precision model in the cloud for secondary identification; if the secondary identification results are still inconsistent, mark it as "awaiting manual review"; If the similarity is higher than the threshold, proceed to the next step of reasoning; The damage type matching and cause inference submodule is used to accurately match damage types and infer the root cause of damage based on the combined characteristics of changes in physical parameters. The implementation process of this module is as follows: 1) Multi-dimensional feature matching: The detected changes in physical parameters are matched with all damage types in the damage mechanism knowledge base in multiple dimensions, and the matching score is calculated. 2) Environmental factor correction: The matching score is corrected by combining historical environmental data such as temperature, humidity, and load. 3) Causal reasoning: Based on the damage type with the highest matching degree, combined with historical monitoring data and engineering information, reason about the root cause of the damage; 4) Output damage causes ranked by confidence level. Causes with a confidence level ≥ 80% are considered primary causes, while those with a confidence level between 50% and 80% are considered secondary causes. The multi-damage coupling decoupling submodule is used to quantify the contribution of each damage to the overall performance degradation when multiple damages coexist, thus addressing the shortcomings of existing technologies in handling complex coupled conditions. The implementation process of this module is as follows: 1) Coupled damage detection: If the changes in physical parameters simultaneously conform to the standard patterns of multiple damage types, and the matching degree of each is ≥60%, then it is determined to be a multi-damage coupled working condition. 2) Contribution quantification: A decoupling algorithm based on partial least squares regression (PLSR) is used to calculate the contribution of each damage type to the overall performance degradation; Establish a regression model between changes in physical parameters and the degree of damage; calculate the contribution of each damage by using the projected importance (VIP) value of the variables; 3) Interaction analysis: Analyze the mutual promotion or inhibition relationships between various injuries; 4) Output the contribution percentage and interaction relationship diagram of each damage type; The damage evolution trend prediction submodule is used to predict the development trend of damage in different time periods in the future based on the damage evolution dynamics model and historical monitoring data; the implementation process of this module is as follows: 1) Model selection: Select the corresponding damage evolution dynamics model according to the damage type. Rubber aging: The Arrhenius model was used, taking into account the effects of temperature and time. Fatigue damage: Miner linear cumulative damage model was used. Bolt loosening: A loosening evolution model is used, taking into account the influence of the number of load cycles; 2) Parameter calibration: Using historical monitoring data of the support, calibrate the parameters of the evolution model; 3) Multi-scenario prediction: Predicts the damage development trend under three different scenarios: Conservative scenario: Under normal operating conditions, with no additional load. General scenario: Minor earthquakes are possible. Extreme scenario: Possibly experiencing a moderate or major earthquake; 4) Remaining useful life (RUL) estimation: Based on the damage evolution trend, estimate the time to reach the damage threshold. Minor injury threshold: 20% of the injury level. Moderate injury threshold: 50% of the injury level Severe injury threshold: 80% of the damage level; 5) Output damage development trend curve and estimated remaining service life; The maintenance decision generation submodule is used to generate differentiated and actionable maintenance decision recommendations based on damage type, severity, cause, evolution trend, and engineering importance. The implementation process of this module is as follows: 1) Risk Level Assessment: Taking into account the extent of damage, the rate of development, and the importance of the structure, the damage risk level is assessed. Low risk: Damage less than 20%, slow progression, does not affect structural safety. Medium risk: Damage level ≤ 20% < 50%, developing rapidly, may affect structural performance. High risk: Damage ≥50%, rapid development, seriously threatening structural safety; 2) Maintenance strategy matching: Match the corresponding maintenance strategy according to the risk level and damage type. Low risk: Regular monitoring, with a follow-up examination every 6 months. Medium risk: Recent repairs, replacement of damaged parts, or reinforcement. High risk: Immediately stop using it and carry out emergency reinforcement or replacement; 3) Detailed maintenance plan: Provide specific maintenance steps, required materials, estimated time and cost; 4) Priority ranking: When multiple supports are damaged at the same time, maintenance priorities are ranked according to risk level and structural importance; The process for generating the analysis report is as follows: 1) Data Collection: Automatically collect all relevant data from this test, including: Basic information: bearing number, location, type, installation time Monitoring data: raw data waveforms, key characteristic values, and changes in physical parameters. Identification results: Damage type, location, severity, and confidence level. Reasoning results: Causes of damage, contribution of coupled damage, evolution trend, and risk level; 2) Template population: Automatically populates the collected data into a preset report template; 3) Chart generation: Automatically generate key charts, including: raw data waveform chart, Grad-CAM++ heat map, physical parameter change comparison chart, damage development trend curve, and coupled damage contribution pie chart; 4) Text generation: Based on the reasoning results, automatically generate natural language descriptions of the analysis conclusions and suggestions; 5) Review and output: Review the format and logic of the generated report, and finally output a PDF report.

[0015] Preferably, the lightweight model structure includes: Input layer: Input channels correspond to: Channel 0 = acceleration, Channel 1 = displacement, Channel 2 = force, Channel 3 = temperature, Channel 4 = strain; Lightweight feature extraction blocks: Employ depthwise separable convolutions, significantly reducing the number of parameters and computational cost; Structure of each lightweight block: Depthwise Convolution: Kernel size = 3, stride = 1, padding = 1 Pointwise Convolution: Kernel size = 1, stride = 1 BatchNorm → Swish Changes in the number of channels: 5→32→64→128; Dynamic channel attention module: We employ a lightweight channel attention mechanism (ECA-Net) to enhance feature representation capabilities with almost no increase in computational cost. Structure: Global average pooling → 1D convolution (kernel size = 3) → Sigmoid → multiply with original features Function: Automatically enhances the characteristics of channels sensitive to damage and suppresses noise in irrelevant channels; Global average pooling layer: compresses a two-dimensional feature map into a one-dimensional feature vector; Dual output head: Anomaly detection head: Structure: Fully connected layer (128→64) → ReLU → Fully connected layer (64→1) → Sigmoid Output: Probability of anomaly (0-1), with a threshold of 0.5; if the value exceeds 0.5, the result is considered an anomaly. Preliminary damage identification head: Structure: Fully connected layer (128→64) → ReLU → Fully connected layer (64→5) → Softmax Output: Preliminary probability distributions for the five damage types; The high-precision model structure includes: Input layer: Input channels correspond to: Channel 0 = acceleration, Channel 1 = displacement, Channel 2 = force, Channel 3 = temperature, Channel 4 = strain; Modal Independent Feature Extraction Block: Design an independent feature extraction network for each modality to fully explore the unique damage features of each modality's data; Multimodal attention fusion module: a. Employing an intermodal cross-attention mechanism, the contribution weights of different modalities under different damage types are automatically learned to achieve adaptive feature fusion: b. Flatten each modal feature into a two-dimensional feature map; c. Calculate the attention weight matrix between modalities: , where F i Let W be the feature of the i-th modality, and W be the learnable attention weight matrix. d. Perform weighted summation of the features from each modality to fuse the features: Shared feature coding blocks: Composed of 3 layers of residual 1D convolutional blocks, it further extracts high-level abstract features from the fused features: Each residual block structure: 1D convolution (3×1) → BatchNorm → Swish → 1D convolution (3×1) → BatchNorm → Residual connection → Swish Changes in the number of channels: 208→256→256→512 Output dimensions: [batch_size, 128, 512] Finally, a global average pooling layer is added, with an output dimension of [batch_size, 512]. Multi-task output head: a. Damage type classification head: Structure: Fully connected layer (512→256)→ReLU→Dropout (0.3)→Fully connected layer (256→5)→Softmax Output: Probability distribution of 5 damage types; b. Damage location positioning head: Structure: Fully connected layer (512→256)→ReLU→Dropout (0.3)→Fully connected layer (256→3) Output: 3D coordinates (x, y, z) of the damage location, in meters; c. Damage Quantization Head: Structure: Fully connected layer (512→256)→ReLU→Dropout (0.3)→Fully connected layer (256→1)→Sigmoid Output: Damage level (0-1), multiplied by 100 to get the percentage; Physical constraint layer: The constitutive relation and dynamic equation of the seismic isolation and damping components are added as regularization terms to the loss function to guide the model to learn characteristics that conform to physical laws. Input: The damage level predicted by the model Calculation: The theoretical equivalent stiffness Kpred is calculated based on the damage level and constitutive model. Loss: Calculate the mean square error between Kpred and the measured equivalent stiffness Kmeas extracted from the original data. Weight: Physical constraint loss accounts for 30% of the total loss.

[0016] Preferably, the hierarchical early warning levels and corresponding strategy settings are as follows: Level 1 warning: Damage level <20%, defined as minor damage, suggesting regular check-ups; Level 2 warning: Damage 20% ≤ Degree of damage < 50%, defined as moderate damage, repair recommended in the near future; Level 3 warning: Damage ≥ 50% is defined as serious damage. Immediately stop using the product and reinforce it.

[0017] Preferably, the cloud periodically updates the model using newly accumulated data and distributes the updated model to the edge.

[0018] The beneficial effects of this invention are: This invention employs a three-tiered data efficiency mechanism: Model-Independent Meta-Learning (MAML) pre-training, adversarial domain adaptive fine-tuning, and progressive damage sample generation. It achieves high-precision identification with only 5-20 real labeled samples per damage type. Simultaneously, the feature learning direction guided by physical constraints effectively eliminates model errors and distribution differences between numerical simulation data and real engineering data, solving the problem of insufficient training data caused by low probability of damage events and high labeling costs in actual engineering. The multi-task learning network simultaneously completes three tasks: damage type classification, location localization, and severity quantification. A damage coupling decoupling algorithm is proposed, which can accurately identify complex working conditions where multiple damages coexist and quantify the contribution of each damage. By performing meta-training on datasets of multiple different types of seismic isolation bearings, the model learns a general representation of damage features. At the same time, through domain adaptation technology, it can quickly adapt to the sensor characteristics, environmental conditions and structural characteristics of different engineering sites, without the need to collect a large amount of data for training for each specific project. This invention can not only accurately identify existing damage, but also predict the future development trend and remaining service life of damage based on a damage evolution dynamics model, and generate differentiated maintenance decision recommendations according to the type, degree and risk level of damage; through a graded early warning system, it realizes dynamic management of the health status of seismic isolation and damping components throughout their entire life cycle, helping engineers to discover potential safety hazards in advance and scientifically formulate maintenance plans. Attached Figure Description

[0019] 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.

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0021] 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.

[0022] Example 1 A deep learning-based method for damage identification of seismic isolation and damping components includes: S1: Multi-source sensor data acquisition and preprocessing: Triaxial accelerometer, displacement sensor and force sensor are installed on the upper and lower connecting plates of the seismic isolation and damping bearing; Fiber optic strain sensor and temperature sensor are also attached to the surface of the rubber layer of the seismic isolation and damping bearing. The sampling frequency of the above 5 sets of multi-source sensors was set to 100Hz, and the sampling duration was set to 30 seconds after each earthquake or artificial excitation; multi-source raw data were collected and the data were preprocessed. Preprocessing includes: using moving average filtering to smooth the data by taking the mean within a local window for noise reduction; using Min-Max standardization to standardize the data; and using sampling interpolation to complete missing values. For continuous signal data, a sliding window is used to divide the continuous signal into sample segments of length 1024; S2: Multimodal feature extraction with enhanced physical information, performing corresponding feature extraction on acceleration signal data, displacement signal data, force signal data, strain signal data, and temperature signal data respectively; Wavelet packet transform was performed on the above acceleration signal data to extract energy characteristics of different frequency bands; hysteresis curve analysis was performed on the displacement signal data to extract physical characteristics of equivalent stiffness, equivalent damping, and energy dissipation capacity; maximum strain and strain distribution uniformity characteristics were extracted from the strain signal data; and temperature gradient and temperature change rate characteristics were extracted from the temperature signal data. All features are then input into the physical information embedding layer; feature enhancement is performed by combining the constitutive equations of the seismic isolation and damping bearings. The physical enhancement of displacement hysteresis characteristics involves extracting equivalent stiffness, equivalent damping, and energy dissipation capacity from displacement signals, which are macroscopic statistical characteristics. These characteristics can be transformed into constitutive parameter characteristics through constitutive equations. 1) Constitutive parameter identification and feature generation Input the measured displacement-force hysteresis curve, and fit the Bouc-Wen model using the least squares method or genetic algorithm to obtain the values ​​including α, K1, and u. y F y Constitutive parameters of β and γ; Output enhancement features: yield strength degradation rate, stiffness degradation coefficient, hysteresis pinching exponent, and cumulative plastic deformation energy; 2) Feature extraction of internal state variables Using the hysteresis variable z(t) in the Bouc-Wen model, load history features are extracted: Maximum hysteresis value z max : Reflects the maximum degree of plastic deformation experienced by the support Rate of change of hysteretic variables : The rate of energy dissipation The cumulative value of the hysteretic variable ∫∣z(t)∣dt: reflects the total cumulative damage; Physical calibration of acceleration energy characteristics: Physical consistency calibration is performed using constitutive equations based on energy features extracted from different frequency bands through wavelet packet transform. The theoretical dynamic response of the support is calculated based on the constitutive equation, and the theoretical energy distribution of each frequency band is obtained. By comparing the measured energy characteristics with the theoretical energy distribution, errors caused by sensor noise and environmental interference are corrected. Enhanced features include: energy deviation ratio, high-frequency energy ratio correction value, and resonant band energy increment. Multiscale fusion of strain characteristics: Strain signals reflect the local deformation state of the support and can be fused with the macroscopic mechanical characteristics of the constitutive equations at multiple scales. Establish the mapping relationship between strain distribution and overall support stiffness: ,in , These represent the maximum strain and the uniformity of strain distribution, respectively. Correcting the yield displacement in the constitutive equation using the maximum strain characteristic u y : ,in This represents the theoretical yield displacement of the seismic isolation bearing under standard reference conditions; it is a factory-designed inherent parameter. Enhanced features generated: local strain concentration factor, stiffness-strain correlation index; Physical correction of temperature characteristics: Temperature has a significant impact on the mechanical properties of seismic isolation and damping bearings. A temperature-constitutive parameter correction model is used to convert temperature characteristics into performance correction coefficients. Lead core yield strength correction: For every 1°C increase in temperature, the yield strength decreases by 0.4%; among which... This represents the theoretical yield strength of the seismic isolation bearing under standard reference conditions; Rubber stiffness correction: ,in Given prior parameters; Friction coefficient correction: ,in Given prior parameters; The extracted temperature gradient and temperature change rate features are input into the above-mentioned modified model to generate constitutive parameter features after temperature correction, which significantly improves the model's generalization ability under different ambient temperatures. Characteristic regularization of the physical constraint loss function: During the training process of the physical information embedding layer, constitutive equation residual loss is incorporated: 2 ,in F pred The restoring force is calculated based on the constitutive equation and the model parameter θ. F mea It is the measured resilience; by minimizing L phy This guides the model to learn feature representations that conform to physical laws, avoiding unreasonable results such as negative stiffness, negative energy consumption, and sudden stiffness changes.

[0023] S3: Multi-task deep learning model training, constructing a multi-input, multi-output deep learning model hierarchical structure as follows: Input layer: includes the feature data of the five modalities collected and preprocessed in the aforementioned steps; Shared feature layer: Composed of multiple 1D convolutional layers and LSTM layers, it extracts temporal features; Task feature layer: corresponding to the three tasks of damage type classification, location localization, and severity quantification; Physical constraint layer: The residuals of the dynamic equations are added as regularization terms to the loss function. A meta-learning method is used for pre-training, and the model is fine-tuned based on a small amount of real engineering data. Domain adaptation techniques are employed to reduce the distributional differences between simulated and real data. Specifically: 1) Alignment of source and target domain data preparation and preprocessing 1.1) Dataset Definition and Partitioning Source domain data Numerical simulation datasets containing all different types of seismic isolation bearings that have completed MAML meta-training, such as: lead-core rubber bearing (LRB), high-damping rubber bearing (HDR), and friction pendulum bearing (FPS); covering 5 typical damage types, 10 levels of damage (0%-100%, 10% intervals) and 20 different excitation conditions, with a total sample size of ≥100,000; Target domain data The actual monitoring data collected at the project site to be deployed is divided into: Annotation set : 5-20 labeled samples for each damage type, including damage type, location and severity labels, obtained by professional inspectors through manual inspection; Unlabeled set A large amount of unlabeled normal operating condition and suspected damage condition data, with a sample size of ≥1,000, were continuously collected by field sensors; 1.2) Data Standardization and Physical Feature Alignment The same preprocessing methods, including denoising, sliding window segmentation (window length 1024, step size 256), and missing value imputation, are used for both source and target domain data; standardization is performed using quantile standardization to eliminate differences in the range, sensitivity, and noise level of different sensors, and sampling is performed at a uniform frequency of 100Hz; Extract the same set of physical features such as equivalent stiffness, equivalent damping, energy dissipation capacity, and strain distribution uniformity from the source and target domain data respectively; Calculate the mean and variance differences of the physical characteristics of the source domain and the target domain, and perform a linear transformation on the physical characteristics of the source domain to make its statistical distribution consistent with that of the target domain. 2) Construction of Adaptive Modules for Multimodal Adversarial Domains A multimodal domain adaptation module is inserted after the shared feature layer and before the task-specific layer of the original deep learning model. This module adopts a two-level structure of "modality-specific discriminator + global feature discriminator": The modality-specific discriminator is designed as follows: Design independent domain discriminants for each of the five modes: acceleration, displacement, force, temperature, and strain. Each discriminator uses a 3-layer fully connected network: input layer, corresponding to the single-modality feature dimension → hidden layer, 128 neurons, ReLU activation, Dropout=0.2 → output layer, 1 neuron, Sigmoid activation; Function: Align the feature distribution of each modality separately to eliminate inter-domain differences in single-modality data; The global feature domain discriminator is designed as follows: Perform domain discrimination on the global features after multimodal fusion; A 4-layer fully connected network is used: input layer, fused feature dimension 512 → hidden layer 1, 256 neurons, ReLU activation, Dropout=0.2 → hidden layer 2, 128 neurons, ReLU activation, Dropout=0.2 → output layer, 1 neuron, Sigmoid activation; Function: To align the overall feature distribution after multimodal fusion and capture cross-modal interdomain differences; Gradient Reversal Layer (GRL): Insert a gradient inversion layer before the input layer of all domain discriminators; During forward propagation, the input remains unchanged, and during backward propagation, the gradient is multiplied by -λ, where λ is an adaptive weight coefficient that increases linearly from 0 to 1 with the number of training rounds. Function: To enable adversarial training between the feature extractor and the domain discriminator, allowing the feature extractor to learn domain-invariant features; 3) Loss function design The task loss, domain adversarial loss, and physical consistency loss are weighted and combined to form the total loss function: 3.1) Mission Loss : Damage type classification: Cross-entropy loss Damage location: Mean square error loss Damage Quantification: Mean Square Error Loss Using only the target domain annotation set D t−l calculate; Domain confrontation loss : Modality-specific adversarial loss: the sum of the binary cross-entropy losses of the five modality discriminators. Global Feature Adversarial Loss: Binary Cross-Entropy Loss of the Global Discriminator Use source domain data D s and unlabeled set of target domain D t−u calculate; Physical consistency loss : The error between the equivalent stiffness corresponding to the damage level predicted by the computational model and the measured equivalent stiffness extracted from real data. Calculation formula: ,in To predict the degree of damage d i The theoretical equivalent stiffness calculated using the constitutive model of the seismic isolation bearing; To obtain real data x i The measured equivalent stiffness extracted from it; Purpose: To ensure that the model does not violate the basic physical laws of seismic isolation bearings during domain alignment, and to avoid the problem of "distribution alignment but physical errors"; 3.2) Weighting coefficient settings α: Domain adversarial loss weight, initially 0.1, increasing linearly to 0.5 with each training epoch. β: Physical consistency loss weight, fixed at 0.3; The progressive sample generation method expands the training set by generating continuous damage evolution samples based on normal data and a small amount of damage data; a three-stage progressive training fine-tuning of the deep learning model is employed, as follows: Phase 1: Domain Adaptive Pre-training Freeze the shared feature layers and task-specific layers of the model, and train only the multimodal domain adaptive module; use source domain data. D s and unlabeled set of target domain D t−uTraining was performed; the optimizer used was Adam, the learning rate was 1e-4, the batch size was 32, and the number of training epochs was 20; the feature distributions of the source and target domains were initially aligned, reducing the domain classification accuracy to below 60%; Phase Two: Joint Training Unfreeze task-specific layers while keeping shared feature layers frozen; use target domain annotation sets. D t−l and unlabeled sets D t−u Joint training was performed; the optimizer used was Adam, with a learning rate of 5e-5, a batch size of 16, and 30 training epochs; the model learned the damage features of the target domain while maintaining feature distribution alignment. Phase 3: Fine-tuning of the entire model Unfreeze all layers of the model; use the target domain annotation set. D t−l Supervised training was performed; the optimizer used was Adam, with a learning rate of 1e-5, a batch size of 8, and 20 training epochs; an early stopping strategy was adopted: training was stopped when the validation set loss no longer decreased for 5 consecutive epochs; this further optimized the model's performance in the target domain.

[0024] S4: Damage identification and interpretability analysis. Input the data to be detected into the deep learning model trained in S3, and output the probability distribution of damage type, damage location coordinates and damage degree quantification value (0~100%) respectively. The model showcases key signal features of interest and, based on a feature-physical quantity mapping library, transforms abstract features into interpretable changes in physical parameters. In this embodiment, the feature-physical quantity mapping relationship library is divided into three layers, as follows: The first layer: the basic physical parameter library, which stores all measurable physical parameters of seismic isolation and damping components, along with their normal ranges and damage thresholds. It forms the foundation of the entire mapping database. Mechanical performance parameters: equivalent stiffness, equivalent damping ratio, yield strength, ultimate bearing capacity, energy dissipation capacity, stiffness degradation rate Deformation parameters: horizontal displacement, vertical displacement, shear deformation angle, residual deformation, strain distribution Environmental parameters: ambient temperature, temperature gradient, humidity, wind speed Material parameters: rubber hardness, rubber aging degree, lead core yield strength, steel plate elastic modulus Connection parameters: bolt preload, gap between connecting plates, and coefficient of friction of the contact surface; The second layer: the feature-physical quantity association rule base, which stores the quantitative association relationships between abstract features extracted by deep learning and basic physical parameters, is the core of the mapping relationship library. Single-modal feature association rules: Establish the association between the features of five modes (acceleration, displacement, force, temperature, and strain) and their corresponding physical parameters. For example, the energy characteristics of acceleration signals in the 10-20Hz frequency band are positively correlated with the degree of rubber aging. For example, the non-uniformity of strain distribution is linearly correlated with the delamination area of ​​the steel plate. Multimodal fusion feature association rules: Establishing the association between multimodal fusion features and combinations of multiple physical parameters. For example, the pinching effect characteristics of the force-displacement hysteresis curve are correlated with the combination of the lead core yield strength and the rubber aging degree. Feature importance weights: Store the importance weight of each feature for different physical parameters and different damage types; The third layer: a damage mechanism knowledge base, which stores the physical mechanisms, evolutionary laws, and typical characteristic patterns of different types of damage, and is the foundation for damage mechanism reasoning. Correspondence between damage type and changes in physical parameters: Rubber aging: Increased equivalent stiffness, decreased equivalent damping ratio, and reduced energy dissipation capacity. Lead core yielding: equivalent stiffness decreases, equivalent damping ratio increases, and the hysteresis curve shows pinching. Steel plate delamination: uneven strain distribution, increased local strain, and slightly reduced equivalent stiffness. Loose bolts: Increased high-frequency vibration energy and increased relative displacement of connecting plates Support slippage: The hysteresis curve shows a slippage segment, and the residual deformation increases; Damage evolution patterns: the development speed and evolution path of different damage types under different loads and environmental conditions; Damage coupling effect: the mutual influence relationship when multiple damages exist simultaneously; The mapping relationship library constructed above transforms abstract features into interpretable changes in physical parameters; The damage mechanism reasoning module is invoked, and a damage analysis report is generated by combining the identification results with physical knowledge. The aforementioned damage mechanism reasoning module includes: The physical consistency verification submodule is used to verify whether the damage identification results output by the deep learning model conform to the basic physical laws of seismic isolation and damping components, and to filter out obviously unreasonable misjudgment results. The standard physical parameter change pattern corresponding to this damage type is queried from the feature-physical quantity mapping relationship library; the cosine similarity between the physical parameter change value obtained in this identification and the standard pattern is calculated; if the similarity is lower than the threshold (0.7), it is judged as a suspected misjudgment, triggering the secondary verification process: Re-extract the physical features of the original data; call the high-precision model in the cloud for secondary identification; if the secondary identification results are still inconsistent, mark it as "awaiting manual review"; If the similarity is higher than the threshold, proceed to the next step of reasoning; The damage type matching and cause inference submodule is used to accurately match damage types and infer the root cause of damage based on the combined characteristics of changes in physical parameters. The implementation process of this module is as follows: 1) Multi-dimensional feature matching: The detected changes in physical parameters are matched with all damage types in the damage mechanism knowledge base in multiple dimensions, and the matching score is calculated. 2) Environmental factor correction: The matching score is corrected by combining historical environmental data such as temperature, humidity, and load. 3) Causal reasoning: Based on the damage type with the highest matching degree, combined with historical monitoring data and engineering information, reason about the root cause of the damage; 4) Output damage causes ranked by confidence level. Causes with a confidence level ≥ 80% are considered primary causes, while those with a confidence level between 50% and 80% are considered secondary causes. The multi-damage coupling decoupling submodule is used to quantify the contribution of each damage to the overall performance degradation when multiple damages coexist, thus addressing the shortcomings of existing technologies in handling complex coupled conditions. The implementation process of this module is as follows: 1) Coupled damage detection: If the changes in physical parameters simultaneously conform to the standard patterns of multiple damage types, and the matching degree of each is ≥60%, then it is determined to be a multi-damage coupled working condition. 2) Contribution quantification: A decoupling algorithm based on partial least squares regression (PLSR) is used to calculate the contribution of each damage type to the overall performance degradation; Establish a regression model between changes in physical parameters and the degree of damage; calculate the contribution of each damage by using the projected importance (VIP) value of the variables; 3) Interaction analysis: Analyze the mutual promotion or inhibition relationships between various injuries; 4) Output the contribution percentage and interaction relationship diagram of each damage type; The damage evolution trend prediction submodule is used to predict the development trend of damage in different time periods in the future based on the damage evolution dynamics model and historical monitoring data; the implementation process of this module is as follows: 1) Model selection: Select the corresponding damage evolution dynamics model according to the damage type. Rubber aging: The Arrhenius model was used, taking into account the effects of temperature and time. Fatigue damage: Miner linear cumulative damage model was used. Bolt loosening: A loosening evolution model is used, taking into account the influence of the number of load cycles; 2) Parameter calibration: Using historical monitoring data of the support, calibrate the parameters of the evolution model; 3) Multi-scenario prediction: Predicts the damage development trend under three different scenarios: Conservative scenario: Under normal operating conditions, with no additional load. General scenario: Minor earthquakes are possible. Extreme scenario: Possibly experiencing a moderate or major earthquake; 4) Remaining useful life (RUL) estimation: Based on the damage evolution trend, estimate the time to reach the damage threshold. Minor injury threshold: 20% of the injury level. Moderate injury threshold: 50% of the injury level Severe injury threshold: 80% of the damage level; 5) Output damage development trend curve and estimated remaining service life; The maintenance decision generation submodule is used to generate differentiated and actionable maintenance decision recommendations based on damage type, severity, cause, evolution trend, and engineering importance. The implementation process of this module is as follows: 1) Risk Level Assessment: Taking into account the extent of damage, the rate of development, and the importance of the structure, the damage risk level is assessed. Low risk: Damage less than 20%, slow progression, does not affect structural safety. Medium risk: Damage level ≤ 20% < 50%, developing rapidly, may affect structural performance. High risk: Damage ≥50%, rapid development, seriously threatening structural safety; 2) Maintenance strategy matching: Match the corresponding maintenance strategy according to the risk level and damage type. Low risk: Regular monitoring, with a follow-up examination every 6 months. Medium risk: Recent repairs, replacement of damaged parts, or reinforcement. High risk: Immediately stop using it and carry out emergency reinforcement or replacement; 3) Detailed maintenance plan: Provide specific maintenance steps, required materials, estimated time and cost; 4) Priority ranking: When multiple supports are damaged at the same time, maintenance priorities are ranked according to risk level and structural importance; The analysis report generation process is as follows: 1) Data Collection: Automatically collect all relevant data from this test, including: Basic information: bearing number, location, type, installation time Monitoring data: raw data waveforms, key characteristic values, and changes in physical parameters. Identification results: Damage type, location, severity, and confidence level. Reasoning results: Causes of damage, contribution of coupled damage, evolution trend, and risk level; 2) Template population: Automatically populates the collected data into a preset report template; 3) Chart generation: Automatically generate key charts, including: raw data waveform chart, Grad-CAM++ heat map, physical parameter change comparison chart, damage development trend curve, and coupled damage contribution pie chart; 4) Text generation: Based on the reasoning results, automatically generate natural language descriptions of the analysis conclusions and suggestions; 5) Review and output: Review the format and logic of the generated report, and finally output the report in PDF format.

[0025] S5: Real-time monitoring and early warning of edge-cloud collaboration, deploying lightweight models at the edge for real-time data monitoring and preliminary anomaly detection; The lightweight model structure includes: Input layer: Input channels correspond to: Channel 0 = acceleration, Channel 1 = displacement, Channel 2 = force, Channel 3 = temperature, Channel 4 = strain; Lightweight feature extraction blocks: Employ depthwise separable convolutions, significantly reducing the number of parameters and computational cost; Structure of each lightweight block: Depthwise Convolution: Kernel size = 3, stride = 1, padding = 1 Pointwise Convolution: Kernel size = 1, stride = 1 BatchNorm → Swish Changes in the number of channels: 5→32→64→128; Dynamic channel attention module: We employ a lightweight channel attention mechanism (ECA-Net) to enhance feature representation capabilities with almost no increase in computational cost. Structure: Global average pooling → 1D convolution (kernel size = 3) → Sigmoid → multiply with original features Function: Automatically enhances the characteristics of channels sensitive to damage and suppresses noise in irrelevant channels; Global average pooling layer: compresses a two-dimensional feature map into a one-dimensional feature vector; Dual output head: Anomaly detection head: Structure: Fully connected layer (128→64) → ReLU → Fully connected layer (64→1) → Sigmoid Output: Probability of anomaly (0-1), with a threshold of 0.5; if the value exceeds 0.5, the result is considered an anomaly. Preliminary damage identification head: Structure: Fully connected layer (128→64) → ReLU → Fully connected layer (64→5) → Softmax Output: Preliminary probability distributions for the five damage types; High-precision models are deployed in the cloud to perform refined damage identification on abnormal data and trigger graded early warnings based on the identification results; The high-precision model structure includes: Input layer: Input channels correspond to: Channel 0 = acceleration, Channel 1 = displacement, Channel 2 = force, Channel 3 = temperature, Channel 4 = strain; Modality-Independent Feature Extraction Blocks: Independent feature extraction networks are designed for each modality to fully exploit the unique impairment features of each modality's data, as shown in the table below: Multimodal attention fusion module: a. Employing an intermodal cross-attention mechanism, the contribution weights of different modalities under different damage types are automatically learned to achieve adaptive feature fusion: b. Flatten each modal feature into a two-dimensional feature map; c. Calculate the attention weight matrix between modalities: , where F i Let W be the feature of the i-th modality, and W be the learnable attention weight matrix. d. Perform weighted summation of the features from each modality to fuse the features: Shared feature coding blocks: Composed of 3 layers of residual 1D convolutional blocks, it further extracts high-level abstract features from the fused features: Each residual block structure: 1D convolution (3×1) → BatchNorm → Swish → 1D convolution (3×1) → BatchNorm → Residual connection → Swish Changes in the number of channels: 208→256→256→512 Output dimensions: [batch_size, 128, 512] Finally, a global average pooling layer is added, with an output dimension of [batch_size, 512]. Multi-task output head: a. Damage type classification head: Structure: Fully connected layer (512→256)→ReLU→Dropout (0.3)→Fully connected layer (256→5)→Softmax Output: Probability distribution of 5 damage types; b. Damage location positioning head: Structure: Fully connected layer (512→256)→ReLU→Dropout (0.3)→Fully connected layer (256→3) Output: 3D coordinates (x, y, z) of the damage location, in meters; c. Damage Quantization Head: Structure: Fully connected layer (512→256)→ReLU→Dropout (0.3)→Fully connected layer (256→1)→Sigmoid Output: Damage level (0-1), multiplied by 100 to get the percentage; Physical constraint layer: The constitutive relation and dynamic equation of the seismic isolation and damping components are added as regularization terms to the loss function to guide the model to learn characteristics that conform to physical laws. Input: The damage level predicted by the model Calculation: The theoretical equivalent stiffness Kpred is calculated based on the damage level and constitutive model. Loss: Calculate the mean square error between Kpred and the measured equivalent stiffness Kmeas extracted from the original data. Weight: Physical constraint loss accounts for 30% of the total loss; The levels of tiered early warning and the corresponding strategy settings are as follows: Level 1 warning: Damage level <20%, defined as minor damage, suggesting regular check-ups; Level 2 warning: Damage 20% ≤ Degree of damage < 50%, defined as moderate damage, repair recommended in the near future; Level 3 warning: Damage ≥ 50% is defined as serious damage. Immediately stop using the product and reinforce it. The cloud periodically updates the model using newly accumulated data and then distributes the updated model to the edge.

Claims

1. A method for damage identification of seismic isolation and damping components based on deep learning, characterized in that, include: S1: Multi-source sensor data acquisition and preprocessing, multi-source sensors are set on the upper and lower connecting plates of the seismic isolation and damping bearing; Multi-source sensors are also attached to the surface of the rubber layer of the vibration isolation and damping bearing; Collect raw data from multiple sources and preprocess the data; S2: Multimodal feature extraction with enhanced physical information. Feature extraction is performed on acceleration signal data, displacement signal data, force signal data, strain signal data, and temperature signal data respectively. All features are input into the physical information embedding layer and feature enhancement is performed in combination with the constitutive equation of the seismic isolation and damping bearing. S3: Multi-task deep learning model training, constructing a multi-input multi-output deep learning model hierarchical structure, using meta-learning methods for pre-training, fine-tuning based on a small amount of real engineering data, and then using a progressive sample generation method to expand the training set; S4: Damage identification and interpretability analysis. Input the data to be detected into the deep learning model trained in S3, and output the probability distribution of damage type, damage location coordinates and damage degree quantification value respectively. The model showcases key signal features of interest and, based on a feature-physical quantity mapping library, transforms abstract features into interpretable changes in physical parameters. The damage mechanism reasoning module is invoked, and a damage analysis report is generated by combining the identification results with physical knowledge. S5: Real-time monitoring and early warning of edge-cloud collaboration, deploying lightweight models at the edge for real-time data monitoring and preliminary anomaly detection; High-precision models are deployed in the cloud to perform refined damage identification on abnormal data, and graded early warnings are triggered based on the identification results.

2. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, The multi-source sensors installed on the upper and lower connecting plates of the vibration isolation and damping bearing include at least: a triaxial accelerometer, a displacement sensor, and a force sensor; The multi-source sensors that are attached to the surface of the rubber layer of the vibration isolation and damping bearing include at least: fiber optic strain sensor and temperature sensor; The raw data of five modes were collected synchronously at a sampling frequency of 100Hz, and the sampling time was 30 seconds after each earthquake or artificial excitation.

3. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, The multi-source raw data shall undergo preprocessing including at least denoising, standardization, and missing value completion. The denoising process employs a moving average filter, which smooths the data by taking the mean within a local window; the standardization process uses the Min-Max standardization method; and the missing value completion is accomplished using interpolation. For continuous signal data, a sliding window is used to divide the continuous signal into sample segments of length 1024.

4. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, Wavelet packet transform is performed on the acceleration signal data to extract energy features of different frequency bands; Hysteresis curve analysis is performed on the displacement signal data to extract physical characteristics including at least equivalent stiffness, equivalent damping, and energy dissipation capacity; maximum strain and strain distribution uniformity characteristics are extracted from the strain signal data. Extract temperature gradient and temperature change rate features from the temperature signal data; All the features extracted above are input into the physical information embedding layer, and feature enhancement is performed by combining the constitutive equation of the seismic isolation and damping bearing.

5. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, The following feature enhancement method is used to enhance the extracted four types of features—acceleration, displacement, strain, and temperature—in conjunction with the constitutive equation of the seismic isolation and damping bearing: The physical enhancement of displacement hysteresis characteristics allows the equivalent stiffness, equivalent damping, and energy dissipation capacity extracted from the displacement signal to be transformed into constitutive parameter characteristics through constitutive equations. 1) Constitutive parameter identification and feature generation Input the measured displacement-force hysteresis curve, and fit the Bouc-Wen model using the least squares method or genetic algorithm to obtain the values ​​including α, K1, and u. y F y Constitutive parameters of β and γ; Output enhancement features: yield strength degradation rate, stiffness degradation coefficient, hysteresis pinching exponent, and cumulative plastic deformation energy; 2) Feature extraction of internal state variables Using the hysteresis variable z(t) in the Bouc-Wen model, load history features are extracted: Maximum hysteresis value z max : Reflects the maximum degree of plastic deformation experienced by the support Rate of change of hysteretic variables : The rate of energy dissipation The cumulative value of the hysteretic variable ∫∣z(t)∣dt: reflects the total cumulative damage; Physical calibration of acceleration energy characteristics: Physical consistency calibration is performed using constitutive equations based on energy features extracted from different frequency bands through wavelet packet transform. The theoretical dynamic response of the support is calculated based on the constitutive equation, and the theoretical energy distribution of each frequency band is obtained. By comparing the measured energy characteristics with the theoretical energy distribution, errors caused by sensor noise and environmental interference are corrected. Enhanced features include: energy deviation ratio, high-frequency energy proportion correction value, and resonant band energy increment. Multiscale fusion of strain characteristics: Strain signals reflect the local deformation state of the support and can be fused with the macroscopic mechanical characteristics of the constitutive equations at multiple scales. Establish the mapping relationship between strain distribution and overall support stiffness: ,in , These represent the maximum strain and the uniformity of strain distribution, respectively. Correcting the yield displacement in the constitutive equation using the maximum strain characteristic u y : ,in This represents the theoretical yield displacement of the seismic isolation bearing under standard reference conditions; it is a factory-designed inherent parameter. Enhanced features generated: local strain concentration factor, stiffness-strain correlation index; Physical correction of temperature characteristics: Temperature has a significant impact on the mechanical properties of seismic isolation and damping bearings. A temperature-constitutive parameter correction model is used to convert temperature characteristics into performance correction coefficients. Lead core yield strength correction: ;in This represents the theoretical yield strength of the seismic isolation bearing under standard reference conditions; Rubber stiffness correction: ,in Given prior parameters; Friction coefficient correction: ,in Given prior parameters; The extracted temperature gradient and temperature change rate features are input into the above-mentioned modified model to generate constitutive parameter features after temperature correction, which significantly improves the model's generalization ability under different ambient temperatures. Characteristic regularization of the physical constraint loss function: During the training process of the physical information embedding layer, constitutive equation residual loss is incorporated: 2 ,in F pred The restoring force is calculated based on the constitutive equation and the model parameter θ. F mea It is the measured resilience; by minimizing L phy This guides the model to learn feature representations that conform to physical laws, avoiding unreasonable results such as negative stiffness, negative energy consumption, and sudden stiffness changes.

6. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, The hierarchical structure of the deep learning model is as follows: Input layer: Includes feature data for five modalities; Shared feature layer: Composed of multiple 1D convolutional layers and LSTM layers, it extracts temporal features; Task feature layer: corresponding to the three tasks of damage type classification, location localization, and severity quantification; Physical constraint layer: The residuals of the dynamic equations are added as regularization terms to the loss function; The process of fine-tuning the model based on a small amount of real engineering data and employing domain adaptation techniques to reduce the distributional differences between simulated and real data involves the following steps: 1) Alignment of source and target domain data preparation and preprocessing 1.1) Dataset Definition and Partitioning Source domain data : A numerical simulation dataset containing all different types of seismic isolation bearings that have completed MAML meta-training, covering 5 typical damage types, 10 levels of damage and 20 different excitation conditions, with a total sample size of ≥100,000; Target domain data The actual monitoring data collected at the project site to be deployed is divided into: Annotation set : 5-20 labeled samples for each damage type, including damage type, location, and severity labels; Unlabeled set A large amount of unlabeled normal operating condition and suspected damage condition data, with a sample size of ≥1,000, were continuously collected by field sensors; 1.2) Data Standardization and Physical Feature Alignment The source and target domain data are processed using the same preprocessing methods, including denoising, sliding window segmentation, and missing value imputation; standardization is performed using quantile standardization; and sampling is performed at a uniform frequency of 100Hz. Extract the same set of physical features from both the source and target domain data; Calculate the mean and variance differences of the physical characteristics of the source domain and the target domain, and perform a linear transformation on the physical characteristics of the source domain to make its statistical distribution consistent with that of the target domain. 2) Construction of Adaptive Modules for Multimodal Adversarial Domains A multimodal domain adaptation module is inserted after the shared feature layer and before the task-specific layer of the original deep learning model. This module adopts a two-level structure of a modality-specific discriminator and a global feature discriminator. The modality-specific discriminant is designed as follows: Design independent domain discriminants for each of the five modes: acceleration, displacement, force, temperature, and strain. Each discriminator uses a 3-layer fully connected network: input layer, hidden layer, and output layer; The global feature domain discriminator is designed as follows: Perform domain discrimination on the global features after multimodal fusion; A 4-layer fully connected network is used: input layer, hidden layer 1, hidden layer 2, and output layer; Gradient inversion layer: Insert a gradient inversion layer before the input layer of all domain discriminators; During forward propagation, the input remains unchanged, and during backward propagation, the gradient is multiplied by -λ, where λ is an adaptive weight coefficient that increases linearly from 0 to 1 with the number of training rounds. 3) Loss function design The task loss, domain adversarial loss, and physical consistency loss are weighted and combined to form the total loss function: 3.1) Mission Loss : Damage type classification: Cross-entropy loss Damage location: Mean square error loss Damage Quantification: Mean Square Error Loss Using only the target domain annotation set D t−l calculate; Domain confrontation loss : Modality-specific adversarial loss: the sum of the binary cross-entropy losses of the five modality discriminators. Global Feature Adversarial Loss: Binary Cross-Entropy Loss of the Global Discriminator Use source domain data D s and unlabeled set of target domain D t−u calculate; Physical consistency loss : The error between the equivalent stiffness corresponding to the damage level predicted by the computational model and the measured equivalent stiffness extracted from real data. Calculation formula: ,in To predict the degree of damage d i The theoretical equivalent stiffness calculated using the constitutive model of the seismic isolation bearing; To obtain real data x i The measured equivalent stiffness extracted from it; 3.2) Weighting coefficient settings α: Domain adversarial loss weight, initially 0.1, increasing linearly to 0.5 with each training epoch. β: Physical consistency loss weight, fixed at 0.3; The progressive sample generation method expands the training set by generating continuous damage evolution samples based on normal data and a small amount of damage data; a three-stage progressive training method is used to fine-tune the deep learning model.

7. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 6, characterized in that, The three-stage progressive training and fine-tuning process of the deep learning model is as follows: Phase 1: Domain Adaptive Pre-training Freeze the shared feature layers and task-specific layers of the model, and train only the multimodal domain adaptive module; use source domain data. D s and unlabeled set of target domain D t−u Conduct training; This achieves initial alignment of the feature distributions of the source and target domains, reducing the domain classification accuracy to below 60%. Phase Two: Joint Training Unfreeze specific layers of the task while keeping shared feature layers frozen; Use target domain annotation set D t−l and unlabeled sets D t−u Conduct joint training; The goal is to enable the model to learn the damage features of the target domain while maintaining the alignment of the feature distribution. Phase 3: Fine-tuning of the entire model Unfreeze all layers of the model; use the target domain annotation set. D t−l Conduct supervised training; An early stopping strategy is adopted: training is stopped when the validation set loss no longer decreases for 5 consecutive rounds; To further optimize the model's performance in the target domain.

8. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, The feature-physical quantity mapping relationship library is divided into three layers, as follows: The first layer: the basic physical parameter library, which stores all measurable physical parameters of seismic isolation and damping components, along with their normal ranges and damage thresholds. It forms the foundation of the entire mapping database. The second layer: the feature-physical quantity association rule base, which stores the quantitative association relationships between abstract features extracted by deep learning and basic physical parameters, is the core of the mapping relationship library. Single-modal feature association rules: Establish the association between the features of five modes (acceleration, displacement, force, temperature, and strain) and their corresponding physical parameters. Multimodal fusion feature association rules: Establishing the association between multimodal fusion features and combinations of multiple physical parameters. Feature importance weights: Store the importance weight of each feature for different physical parameters and different damage types; The third layer: a damage mechanism knowledge base, which stores the physical mechanisms, evolutionary laws, and typical characteristic patterns of different types of damage, and is the foundation for damage mechanism reasoning. Based on the mapping relationship library, abstract features are transformed into interpretable changes in physical parameters.

9. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, The damage mechanism reasoning module and its process for generating damage analysis reports are as follows: The damage mechanism reasoning module includes: The physical consistency verification submodule is used to verify whether the damage identification results output by the deep learning model conform to the basic physical laws of seismic isolation and damping components, and to filter out obviously unreasonable misjudgment results. The system queries the feature-physical quantity mapping database to find the standard physical parameter change pattern corresponding to this damage type; it then calculates the cosine similarity between the identified physical parameter change value and the standard pattern; if the similarity is below the threshold, it is judged as a suspected misjudgment, triggering a secondary verification process. Re-extract the physical features of the original data; call the high-precision model in the cloud for secondary identification; if the secondary identification results are still inconsistent, mark it as "awaiting manual review"; If the similarity is higher than the threshold, proceed to the next step of reasoning; The damage type matching and cause inference submodule is used to accurately match damage types and infer the root cause of damage based on the combined characteristics of changes in physical parameters. The implementation process of this module is as follows: 1) Multi-dimensional feature matching: The detected changes in physical parameters are matched with all damage types in the damage mechanism knowledge base in multiple dimensions, and the matching score is calculated. 2) Environmental factor correction: The matching score is corrected by combining historical temperature, humidity, and load environmental data; 3) Causal reasoning: Based on the damage type with the highest matching degree, combined with historical monitoring data and engineering information, reason about the root cause of the damage; 4) Output damage causes ranked by confidence level. Causes with a confidence level ≥ 80% are considered primary causes, while those with a confidence level between 50% and 80% are considered secondary causes. The multi-damage coupling decoupling submodule is used to quantify the contribution of each damage to the overall performance degradation when multiple damages coexist, thus addressing the shortcomings of existing technologies in handling complex coupled conditions. The implementation process of this module is as follows: 1) Coupled damage detection: If the changes in physical parameters simultaneously conform to the standard patterns of multiple damage types, and the matching degree of each is ≥60%, then it is determined to be a multi-damage coupled working condition. 2) Contribution quantification: A decoupling algorithm based on partial least squares regression is used to calculate the contribution of each damage type to the overall performance degradation; Establish a regression model between changes in physical parameters and the degree of damage; calculate the contribution of each damage by projecting the importance value of the variables; 3) Interaction analysis: Analyze the mutual promotion or inhibition relationships between various injuries; 4) Output the contribution percentage and interaction relationship diagram of each damage type; The damage evolution trend prediction submodule is used to predict the development trend of damage in different time periods in the future based on the damage evolution dynamics model and historical monitoring data; the implementation process of this module is as follows: 1) Model selection: Select the corresponding damage evolution dynamics model according to the damage type. Rubber aging: The Arrhenius model was used, taking into account the effects of temperature and time. Fatigue damage: Miner linear cumulative damage model was used. Bolt loosening: A loosening evolution model is used, taking into account the influence of the number of load cycles; 2) Parameter calibration: Using historical monitoring data of the support, calibrate the parameters of the evolution model; 3) Multi-scenario prediction: Predicts the damage development trend under three different scenarios: Conservative scenario: Under normal operating conditions, with no additional load. General scenario: Minor earthquakes are possible. Extreme scenario: Possibly experiencing a moderate or major earthquake; 4) Remaining service life estimation: Based on the damage evolution trend, estimate the time to reach the damage threshold. Minor injury threshold: 20% of the injury level. Moderate injury threshold: 50% of the injury level Severe injury threshold: 80% of the damage level; 5) Output damage development trend curve and estimated remaining service life; The maintenance decision generation submodule is used to generate differentiated and actionable maintenance decision recommendations based on damage type, severity, cause, evolution trend, and engineering importance. The implementation process of this module is as follows: 1) Risk Level Assessment: Taking into account the extent of damage, the rate of development, and the importance of the structure, the damage risk level is assessed. Low risk: Damage level < 20%, slow development, does not affect structural safety. Medium risk: 20% ≤ damage level < 50%, developing rapidly, may affect structural performance. High risk: Damage ≥50%, rapid development, seriously threatening structural safety; 2) Maintenance strategy matching: Match the corresponding maintenance strategy according to the risk level and damage type. Low risk: Regular monitoring, with a follow-up examination every 6 months. Medium risk: Recent repairs, replacement of damaged parts, or reinforcement. High risk: Immediately stop using it and carry out emergency reinforcement or replacement; 3) Detailed maintenance plan: Provide specific maintenance steps, required materials, estimated time and cost; 4) Priority ranking: When multiple supports are damaged at the same time, maintenance priorities are ranked according to risk level and structural importance; The process for generating the analysis report is as follows: 1) Data Collection: Automatically collect all relevant data from this test, including: Basic information: bearing number, location, type, installation time Monitoring data: raw data waveforms, key characteristic values, and changes in physical parameters. Identification results: Damage type, location, severity, and confidence level. Reasoning results: Causes of damage, contribution of coupled damage, evolution trend, and risk level; 2) Template population: Automatically populates the collected data into a preset report template; 3) Chart generation: Automatically generate key charts, including: raw data waveform chart, Grad-CAM++ heat map, physical parameter change comparison chart, damage development trend curve, and coupled damage contribution pie chart; 4) Text generation: Based on the reasoning results, automatically generate natural language descriptions of the analysis conclusions and suggestions; 5) Review and output: Review the format and logic of the generated report, and finally output the report in PDF format.

10. The method for damage identification of seismic isolation and damping components based on deep learning according to claim 1, characterized in that, The hierarchical early warning levels and corresponding strategy settings are as follows: Level 1 warning: Damage level < 20%, defined as minor damage, suggesting regular check-ups; Level 2 warning: Damage 20% ≤ Degree of damage < 50%, defined as moderate damage, repair recommended in the near future; Level 3 warning: Damage ≥ 50% is defined as serious damage. Immediately stop using the product and reinforce it.

Images