Ancient building risk prediction and control method and system based on large model

Abstract

The present invention discloses a method and system for risk prediction and control of ancient buildings based on a large model, and relates to the technical field of ancient building protection. The risk prediction and control system for ancient buildings based on a large model includes: a multi-source data acquisition module, a multimodal data fusion module, a large model risk prediction module, a risk assessment and early warning module, a digital twin dynamic control module, and a restoration decision module. The present invention realizes a comprehensive perception of the state of ancient buildings by fusing structural state data, visual image data, and environmental data; the ViT‑GPT model pre-trained by self-supervised learning is combined with the LSTM time series model to significantly improve the time series analysis ability and multimodal data processing ability of risk prediction, and can accurately predict the future damage probability and risk trend of ancient buildings, filling the deficiencies of the existing technology in long-span risk prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of ancient building protection, and in particular to a method and system for predicting and controlling ancient building risks based on a large model. Background Art

[0002] As an important cultural heritage, the protection and restoration of ancient buildings are of great significance. However, due to the structural complexity of ancient buildings and the diversity of long-term exposure to the natural environment, existing technologies have significant limitations in risk prediction and control.

[0003] Existing ancient building protection technologies usually rely on a single data source for analysis, which makes it difficult to fully reflect the overall status of ancient buildings. Current risk prediction methods are mostly based on traditional statistical models or simple machine learning methods. These models are incapable of processing complex multimodal data and time series predictions, and it is difficult to effectively capture the damage evolution process and future risk trends of ancient buildings. Most of the existing risk assessment mechanisms are based on static threshold algorithms and lack the ability to dynamically respond to changes in ancient building risks. The application of digital twin technology in the protection of ancient buildings is insufficient. Existing digital twin models are usually difficult to dynamically synchronize real-time data of actual ancient buildings, and the simulation capabilities under different environmental pressure conditions are limited, making it impossible to fully evaluate structural stability and formulate optimized reinforcement plans. Existing restoration decisions mostly rely on subjective judgments based on expert experience, and lack data-driven intelligent decision-making assistance methods. Summary of the invention

[0004] The risk prediction and control method of ancient buildings based on a large model includes the following steps:

[0005] Step S1, multi-source data collection: obtaining structural status data of key parts of ancient buildings, visual image data of the surface of ancient buildings, and environmental data of the ancient buildings as a whole;

[0006] Step S2, multimodal data fusion: preprocess the multi-source data obtained in step S1, including wavelet transform denoising, standardization, and principal component analysis dimensionality reduction; extract and fuse features of structural state data, visual image data, and environmental data to obtain a high-dimensional feature vector of multimodal fusion features;

[0007] Step S3, large model risk prediction: Based on the feature extraction and fusion of step S2, the ViT-GPT model pre-trained by self-supervised learning is used in combination with the LSTM time series prediction model to establish an ancient building risk prediction model, predict the damage probability of ancient buildings in the next 3 months, 6 months, and 12 months, and output the predicted risk area and damage level;

[0008] Step S4, risk assessment and early warning: Based on the prediction results of step S3, the Bayesian dynamic threshold algorithm is used to calculate the risk level, and a secondary early warning mechanism is set. When the predicted risk level reaches the threshold, the automatic inspection by the drone is triggered to generate a hyperspectral scan map; if the risk is further escalated, the finite element analysis model is called to calculate the stress distribution, determine the reinforcement plan, and visualize the risk distribution map through WebGIS; the prediction results include risk areas, damage levels, and damage probability distribution;

[0009] Step S5, dynamic management and control of digital twins: Based on the risk data generated in step S4, a digital twin model is constructed, and the stress conditions of the ancient building structure under different environmental pressures are simulated in real time through the coupling technology of Unity-3D modeling and finite element analysis, a structural stability assessment report is generated, and the reinforcement plan is optimized through reinforcement learning; the risk data includes the prediction results of step S3, risk level, drone inspection feedback data, finite element analysis calculation data and risk distribution map;

[0010] Step S6, repair decision: Combined with the structural stability assessment report and optimized reinforcement plan provided in step S5, the deep reinforcement learning algorithm is called to obtain the repair decision, automatically generate the repair material selection plan and construction plan, and implement them after expert review.

[0011] As a preferred technical solution of the present invention, the multi-source data acquisition specifically includes:

[0012] Structural status data: MEMS acceleration sensors, strain gauges, temperature and humidity sensors, and laser rangefinders are deployed at key locations of ancient buildings to collect vibration, stress, temperature and humidity, and crack displacement data in real time; the key locations include beams, columns, walls, foundations, and marked locations;

[0013] Visual image data: Obtain surface image data of ancient buildings containing damage information through hyperspectral cameras and drone remote sensing equipment; the damage information includes cracks, spalling, bioerosion, deformation, material deterioration and external force damage;

[0014] Environmental data: including geological structure change data and environmental factor data; the geological structure change data is obtained through seismic wave monitoring equipment and lidar; the environmental factor data is obtained by calling the weather station API, including wind speed, precipitation, temperature and humidity in the ancient building area.

[0015] As a preferred technical solution of the present invention, the multimodal data fusion specifically includes:

[0016] Data preprocessing:

[0017] Wavelet transform denoising: multi-scale decomposition of vibration data is performed through discrete wavelet transform to remove high-frequency noise; standardization: normalization of the numerical range of different data sources is performed through Min-Max normalization; principal component analysis dimensionality reduction: principal component analysis is performed on hyperspectral images and remote sensing data;

[0018] Feature extraction:

[0019] A bidirectional LSTM network is used to process the structural state data and extract the time series data of historical vibration, stress change, crack propagation rate, and temperature and humidity change trend. Combined with the Transformer-BERT encoder, the global dependency of the time series data is calculated through the self-attention mechanism to generate structural state features.

[0020] A convolutional neural network is used to extract hierarchical features from visual image data, and a feature pyramid network is combined to enhance the disease detection capabilities at different scales, and the disease features are output; the hierarchical features include information about each type of disease;

[0021] Use random forests to perform regression analysis on environmental data and extract the impact of environmental factors on the risk of ancient buildings; combine LSTM networks to model long-term climate trends and extract environmental change characteristics;

[0022] High-dimensional feature vector construction:

[0023] Dynamic time warping is used to align the time scales of time series data and visual image data. A multi-head self-attention mechanism is used to calculate the correlation between structural state features, disease features, and environmental change features to obtain a high-dimensional feature vector of multimodal fusion features, and weights are assigned to different features through weighted feature fusion.

[0024] The feature extraction and fusion are integrated into the ancient building risk prediction model.

[0025] As a preferred technical solution of the present invention, the structure of the ancient building risk prediction model includes:

[0026] Input layer: used to receive multi-source data pre-processed in step S2, including structural state data, visual image data and environmental data, and perform standard formatting on them to meet the model input requirements;

[0027] Feature extraction layer: used to extract structural state features of structural state data, disease features of visual image data, and environmental change features of environmental data;

[0028] Feature fusion layer: used to fuse features through a multi-head self-attention mechanism to obtain a high-dimensional feature vector of multimodal fusion features; and adopt a self-supervised contrastive learning mechanism to enhance the representation ability of different modal features;

[0029] Risk prediction layer: The Seq2Seq-Transformer module in the ViT-GPT model is used to optimize the temporal expression capability of disease characteristics. The LSTM time series prediction model is combined with the fused high-dimensional feature vector to predict the damage probability of ancient buildings in the next 3, 6, and 12 months, and calculate the damage level of each risk area. The damage risk trend is predicted, and the key risk areas are highlighted through the attention mechanism.

[0030] Output layer: Output the predicted risk area, damage level and damage probability distribution; calculate the confidence of the prediction results in combination with Bayesian uncertainty estimation.

[0031] As a preferred technical solution of the present invention, the training of the ancient building risk prediction model includes:

[0032] A1. Construct a multimodal training dataset based on historical building structure status data, visual image data, and environmental data, including input data X and label data Y , where the input data ,in is the structural status data, For visual image data, is environmental data; label data includes, Damage probability labels are generated based on historical inspection data and expert annotations, indicating the probability of damage to the ancient building in the future. Damage level labels, set as discrete levels, determined based on historical maintenance records and expert assessment results;

[0033] The training sample set is expressed as: ,in is the number of training samples, is the sample index;

[0034] A2. Use self-supervised contrastive learning to pre-train the input data, optimize the multimodal feature representation, and set positive samples : The same building history data, negative samples : Different building data, using contrast loss function to optimize feature representation: ,in is the high-dimensional feature vector output by the feature extraction network, is the cosine similarity function, which is used to measure the similarity between two feature vectors; is the temperature parameter, which is used to control the sensitivity of the similarity calculation. Indicates the number of negative samples;

[0035] A3. Training of feature extraction layer:

[0036] Bidirectional LSTM+Transformer training is used to extract structural features and optimize time series feature learning. The training goal is to minimize the mean square error: ,in is the actual damage probability, is the predicted value, is the total duration of the forecast time range;

[0037] The CNN-FPN network is used to train visual image feature extraction, and the loss function is Focal:

[0038] ,in To predict the probability of disease, is the category weight, which is used to prevent category imbalance; is an adjustment factor used to control the influence of difficult samples;

[0039] Random forest + LSTM training environment data feature extraction is used, and the training goal is also to minimize the mean square error;

[0040] A4. Use multi-head self-attention mechanism to calculate the correlation between different modal features: ,in Q, K, V is the query, key, and value matrix, is the corresponding training weight, Represents dimension;

[0041] The training goal is to minimize the feature fusion error: ,in They are the adaptive fusion weights of structural state, visual image, and environmental data respectively; is the target damage probability label;

[0042] A5. Use ViT-GPT+LSTM to predict the time series of damage probability and optimize the risk prediction for the next 3, 6, and 12 months. The loss function is to minimize the cross entropy loss: ,in is the total number of category intervals of damage probability, is the one-hot encoding of the true category, is the predicted damage probability distribution;

[0043] Predicting future risk levels based on Seq2Seq-Transformer: ,in is the real damage level label, is the predicted probability, r Indicates the level of damage;

[0044] A6. Use Adam optimizer to update model parameters: ,in is the parameter of the tth round of training, represents the learning rate, Represents the loss function; the early stopping strategy and K-fold cross validation are used to improve the generalization ability and ensure the performance convergence of the model on the test set.

[0045] As a preferred technical solution of the present invention, the calculation of the risk level includes:

[0046] Extract the damage probability from the prediction result output in step S3 , damage level and risk areas ;

[0047] The Bayesian dynamic threshold algorithm is used to comprehensively consider the predicted damage probability and damage level in different time ranges to calculate the overall risk index: ,in is the time weight coefficient, satisfying , is the total duration of the forecast time range;

[0048] The Bayesian uncertainty estimation method is used to calculate the confidence interval of the damage probability, which includes: obtaining distribution parameters by sampling the ancient building risk prediction model multiple times, including the mean of the output damage probability and variance , using the normal distribution assumption, at the confidence level The confidence interval of the damage probability is:

[0049] ,in is the quantile of the standard normal distribution;

[0050] The risk level is adjusted dynamically according to the risk index and confidence interval. The risk level classification rules are as follows:

[0051] ,in and There are two risk level thresholds set by a dynamic piecewise function.

[0052] As a preferred technical solution of the present invention, the calculation of stress distribution includes:

[0053] Based on the data of step S1 and step S4, a finite element analysis model is established in combination with the geometric structure and material properties of the ancient building; wherein the data of step S1 includes the external geometric shape and size of the ancient building obtained by laser scanning of the laser radar; the distribution information of surface defects of key parts obtained by the hyperspectral camera and the drone inspection; the material degradation parameters derived from the vibration data, crack displacement data and environmental temperature and humidity data collected by the MEMS sensor, temperature and humidity sensor and laser rangefinder, including elastic modulus, Poisson's ratio and density; the data of step S4 is the finite element analysis calculation data, including the high-risk areas in the risk assessment results, the structural stability analysis data and the predicted damage location distribution;

[0054] Set loading and boundary conditions for the finite element model, apply dynamic time-dependent external loads based on seismic wave monitoring data and environmental data obtained from the weather station API; adjust the stress load distribution through drone inspection feedback data, simulate crack propagation and stress concentration effects as feedback stress loads;

[0055] The stress distribution of the ancient building under the above load and boundary conditions is calculated by finite element method, and the calculation formula is: ,in is the stress tensor, is the strain tensor, which comes from the node displacement difference; is the material stiffness matrix, determined by the material properties;

[0056] Combined with the data of high-risk areas, the key parts of the ancient buildings are analyzed, including the maximum principal stress of beams and columns. , the equivalent stress of the wall , the stress concentration position and distribution pattern of the foundation, where the equivalent stress calculation formula is: ,in is the principal stress component;

[0057] Combined with the crack growth rate and stress concentration area of ​​the drone inspection feedback data, the stress distribution results were verified, and the reinforcement learning algorithm was used to optimize the load distribution.

[0058] Output stress distribution diagram and stability assessment report, including the specific location and risk level of high stress areas, trend analysis of stress changes over time and determined reinforcement plan.

[0059] As a preferred technical solution of the present invention, the digital twin dynamic control specifically includes:

[0060] Based on the risk data generated in step S4, a three-dimensional digital twin model of the ancient building is established through Unity-3D modeling technology. The geometric structure of the ancient building, the surface disease distribution information of key parts obtained by laser scanning with laser radar, and the multimodal data collected by MEMS sensors and hyperspectral cameras are combined to build a virtual simulation environment consistent with the actual ancient building; the finite element analysis model is used to simulate the stress behavior of the ancient building structure under different environmental pressures to generate a preliminary structural stability assessment report;

[0061] The UAV inspection feedback data, environmental change characteristic data, crack propagation rate and stress concentration distribution data obtained in step S3 and step S4 are transmitted to the digital twin model in real time, the state parameters and disease distribution information of the structure in the twin environment are updated, and the dynamic synchronization of the digital twin model is realized;

[0062] Based on the digital twin model, the stress distribution of key parts of the ancient building is simulated in real time through finite element analysis technology under different environmental conditions, and a simulation report on the structural stress condition and crack expansion trend is generated; the reinforcement plan is optimized by combining the reinforcement learning algorithm, including stress distribution adjustment and repair suggestions for high-risk parts;

[0063] WebGIS technology is used to visualize the simulation analysis results, including risk distribution maps of high-risk areas, stress concentration areas, and crack extension trend maps.

[0064] As a preferred technical solution of the present invention, the repair decision specifically includes:

[0065] Based on the structural stability assessment report and optimized reinforcement plan provided in step S5, the repair requirements are determined by integrating the drone inspection feedback data, risk level assessment results and historical maintenance records, including the type of disease, damage level, repair priority and repair scope;

[0066] Combining risk data and simulation analysis results, the selection of restoration materials is optimized through deep reinforcement learning algorithms, and the strength, durability, environmental corrosion resistance and compatibility of the materials with the original materials of the ancient buildings are comprehensively considered to generate a list of recommended restoration materials; the list of recommended restoration materials is verified by comparing the parameters of ancient building restoration materials pre-collected in the expert database, which includes material performance data collected through historical restoration cases, experimental data and third-party standards;

[0067] Based on the optimized reinforcement plan and combined with the historical restoration case library, a restoration plan is generated, including construction technology, restoration sequence and construction environment requirements; the restoration plan is dynamically adjusted in combination with drone inspection data and environmental change trends;

[0068] The generated restoration plan is submitted to the expert team for review. The structural stability of the restored ancient building is simulated through finite element simulation to verify the feasibility and safety of the plan. If the stability requirements are not met, the plan is iteratively optimized through the reinforcement learning algorithm and resubmitted for review.

[0069] The approved repair plan enters the implementation phase, and the construction process is tracked by a real-time monitoring system, including stress changes in key parts and the effect of disease repair. After the repair is completed, the monitoring feedback data is updated to the digital twin model for subsequent risk prediction and dynamic management and control.

[0070] After the restoration is completed, a comparison map of the distribution of defects after restoration is generated through drone inspections and hyperspectral scanning, and the stability of the restoration area is verified in combination with the structural status sensor data; the overall improvement rate and risk reduction level of the restoration effect are calculated through the risk assessment module, and a restoration report is formed and stored in the ancient building protection database.

[0071] The ancient building risk prediction and control system based on large models includes the following modules:

[0072] Multi-source data acquisition module: used to obtain the structural status data of key parts of ancient buildings, the visual image data of the surface of ancient buildings, and the overall environmental data of ancient buildings;

[0073] Multimodal data fusion module: used to pre-process the acquired multi-source data, extract and fuse the features of the structural state data, visual image data, and environmental data, and obtain the high-dimensional feature vector of the multimodal fusion feature;

[0074] Large model risk prediction module: It is used to establish an ancient building risk prediction model by using the ViT-GPT model pre-trained by self-supervised learning and combining it with the LSTM time series prediction model. It predicts the probability of damage to ancient buildings in the next 3, 6, and 12 months, and outputs the predicted risk area and damage level.

[0075] Risk assessment and early warning module: used to calculate the risk level using the Bayesian dynamic threshold algorithm and set a secondary early warning mechanism. When the risk level exceeds the threshold, the drone will be triggered to automatically inspect and generate a hyperspectral scan image; if the risk level further exceeds the threshold , the finite element analysis model is called to calculate the stress distribution, determine the reinforcement plan, and visualize the risk distribution map through WebGIS;

[0076] Digital twin dynamic control module: used to build a digital twin model. Through the coupling technology of Unity-3D modeling and finite element analysis, it can simulate the stress of ancient building structures under different environmental pressures in real time, generate a structural stability assessment report, and optimize the reinforcement plan through reinforcement learning.

[0077] Restoration decision module: used to call deep reinforcement learning algorithm to obtain restoration decisions, automatically generate restoration material selection plan and construction plan, and implement them after expert review.

[0078] The present invention has the following advantages:

[0079] The present invention realizes a comprehensive perception of the status of ancient buildings by integrating structural status data, visual image data and environmental data; it combines the ViT-GPT model pre-trained by self-supervised learning with the LSTM time series model, which significantly improves the timing analysis capability and multimodal data processing capability of risk prediction, and can accurately predict the future damage probability and risk trend of ancient buildings, filling the shortcomings of existing technologies in long-term risk prediction.

[0080] The present invention introduces a Bayesian dynamic threshold algorithm and a two-level early warning mechanism, combined with UAV inspections and hyperspectral scanning data, to dynamically respond to risk changes and generate risk assessment results in real time; by calculating stress distribution through finite element analysis, the risk assessment is further refined, providing a scientific basis for the formulation of reinforcement plans.

[0081] The present invention constructs a dynamic and synchronized digital twin model based on Unity-3D modeling and finite element analysis coupling technology, which can simulate in real time the stress conditions of ancient buildings under different environmental pressures; it can not only perform real-time risk monitoring, but also optimize reinforcement plans through reinforcement learning, providing efficient dynamic control capabilities for the protection of ancient buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art description are briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the provided drawings without creative work.

[0083] Figure 1 This is a schematic diagram of the structure of the ancient building risk prediction and control system based on a large model adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0085] Embodiment 1, the ancient building risk prediction and control method based on the large model, comprises the following steps:

[0086] Step S1, multi-source data collection: obtaining structural status data of key parts of ancient buildings, visual image data of the surface of ancient buildings, and environmental data of the ancient buildings as a whole;

[0087] The multi-source data collection specifically includes:

[0088] Structural status data: MEMS acceleration sensors, strain gauges, temperature and humidity sensors, and laser rangefinders are deployed at key locations of ancient buildings to collect vibration, stress, temperature and humidity, and crack displacement data in real time; the key locations include beams, columns, walls, foundations, and marked locations;

[0089] Visual image data: Surface image data of ancient buildings containing disease information is obtained through hyperspectral cameras and drone remote sensing equipment; the disease information includes cracks (surface cracks, deep cracks), peeling (surface peeling, chalking), biological erosion (fungi, moss, termites), deformation (tilting, vault sinking, wooden structure deformation), material deterioration (salting, carbonization, metal rust) and external force damage (human damage, earthquake damage, wind disaster impact);

[0090] Environmental data: including geological structure change data and environmental factor data; the geological structure change data is obtained through seismic wave monitoring equipment and lidar; the environmental factor data is obtained by calling the weather station API, including wind speed, precipitation, temperature and humidity in the ancient building area.

[0091] Step S2, multimodal data fusion: preprocess the multi-source data obtained in step S1, including wavelet transform denoising, standardization, and principal component analysis dimensionality reduction; extract and fuse features of structural state data, visual image data, and environmental data to obtain a high-dimensional feature vector of multimodal fusion features;

[0092] The multimodal data fusion specifically includes:

[0093] Data preprocessing:

[0094] Wavelet transform denoising: multi-scale decomposition of vibration data is performed through discrete wavelet transform to remove high-frequency noise while retaining low-frequency features and enhancing signal stability; standardization: Min-Max normalization is used to normalize the value ranges of different data sources to make their distribution consistent and reduce the impact of different data dimensions; principal component analysis dimensionality reduction: principal component analysis is performed on hyperspectral images and remote sensing data to extract the first n principal components, reduce computational complexity, and retain more than 80% of important information;

[0095] Feature extraction:

[0096] A bidirectional LSTM network is used to process the structural state data, and the time series data of historical vibration, stress change, crack expansion rate, and temperature and humidity change trend are extracted; combined with the Transformer-BERT encoder, the global dependency of the time series data is calculated through the self-attention mechanism to generate structural state features; a convolutional neural network is used to extract hierarchical features from visual image data, and a feature pyramid network is combined to enhance the disease detection capabilities at different scales and output disease features; the hierarchical features include information about each type of disease; random forests are used to perform regression analysis on environmental data to extract the impact pattern of environmental factors on the risk of ancient buildings; the LSTM model is used to model long-term climate trends and extract environmental change features;

[0097] High-dimensional feature vector construction:

[0098] Through dynamic time warping, the time scales of time series data and visual image data are aligned to enhance the matching of different data types. The multi-head self-attention mechanism is used to calculate the correlation between structural state characteristics, disease characteristics, and environmental change characteristics to obtain a high-dimensional feature vector of multimodal fusion features. Weighted feature fusion is used to assign weights to different features to improve the accuracy of risk prediction. The Faiss vector database is used to store the high-dimensional feature vectors corresponding to each feature, and the HNSW index is used for fast retrieval to improve computational efficiency.

[0099] The feature extraction and fusion are integrated into the ancient building risk prediction model.

[0100] Step S3, large model risk prediction: Based on the feature extraction and fusion of step S2, the ViT-GPT model pre-trained by self-supervised learning is used in combination with the LSTM time series prediction model to establish an ancient building risk prediction model, predict the damage probability of ancient buildings in the next 3 months, 6 months, and 12 months, and output the predicted risk area and damage level;

[0101] The structure of the ancient building risk prediction model includes:

[0102] Input layer: used to receive multi-source data pre-processed in step S2, including structural state data, visual image data and environmental data, and perform standard formatting on them to meet the model input requirements;

[0103] Feature extraction layer: used to extract structural state features of structural state data, disease features of visual image data, and environmental change features of environmental data;

[0104] Feature fusion layer: It is used to perform feature fusion through a multi-head self-attention mechanism to obtain a high-dimensional feature vector of multi-modal fusion features; and adopts a self-supervised contrast learning mechanism to enhance the representation ability of different modal features and improve the robustness of feature fusion;

[0105] Risk prediction layer: The Seq2Seq-Transformer module in the ViT-GPT model is used to optimize the temporal expression capability of disease characteristics. The LSTM time series prediction model is combined with the fused high-dimensional feature vector to predict the damage probability of ancient buildings in the next 3, 6, and 12 months, and calculate the damage level of each risk area. The damage risk trend is predicted, and the key risk areas are highlighted through the attention mechanism.

[0106] Output layer: Output the predicted risk area, damage level and damage probability distribution; calculate the confidence of the prediction results in combination with Bayesian uncertainty estimation.

[0107] The training of the ancient building risk prediction model includes the following steps:

[0108] A1. Construct a multimodal training dataset based on historical building structure status data, visual image data, and environmental data, including input data X and label data Y , where the input data ,in is the structural status data, For visual image data, is environmental data; label data includes, Damage probability labels are generated based on historical inspection data and expert annotations, indicating the probability of damage to the ancient building in the future. Damage level labels, set as discrete levels, determined based on historical maintenance records and expert assessment results;

[0109] The training sample set is expressed as: ,in is the number of training samples, is the sample index;

[0110] A2. Use self-supervised contrastive learning to pre-train the input data, optimize the multimodal feature representation, and set positive samples : The same building history data, negative samples : Different building data, using contrast loss function to optimize feature representation: ,in is the high-dimensional feature vector output by the feature extraction network, is the cosine similarity function, which is used to measure the similarity between two feature vectors; is the temperature parameter, which is used to control the sensitivity of the similarity calculation. Indicates the number of negative samples;

[0111] A3. Training of feature extraction layer:

[0112] Bidirectional LSTM+Transformer training is used to extract structural features and optimize time series feature learning. The training goal is to minimize the mean square error: ,in is the actual damage probability, is the predicted value, is the total duration of the forecast time range;

[0113] The CNN-FPN network is used to train visual image feature extraction, and the loss function is Focal:

[0114] ,in To predict the probability of disease, is the category weight, which is used to prevent category imbalance; is an adjustment factor used to control the influence of difficult samples;

[0115] Random forest + LSTM training environment data feature extraction is used, and the training goal is also to minimize the mean square error;

[0116] A4. Use multi-head self-attention mechanism to calculate the correlation between different modal features: ,in Q, K, V is the query, key, and value matrix, is the corresponding training weight, Represents dimension;

[0117] The training goal is to minimize the feature fusion error: ,in They are the adaptive fusion weights of structural state, visual image, and environmental data respectively; is the target damage probability label;

[0118] A5. Use ViT-GPT+LSTM to predict the time series of damage probability and optimize the risk prediction for the next 3, 6, and 12 months. The loss function is to minimize the cross entropy loss: ,in is the total number of category intervals of damage probability, is the one-hot encoding of the true category, is the predicted damage probability distribution;

[0119] Predicting future risk levels based on Seq2Seq-Transformer: ,in is the real damage level label, is the predicted probability, r Indicates the level of damage;

[0120] A6. Use Adam optimizer to update model parameters: ,in is the parameter of the tth round of training, represents the learning rate, Represents the loss function; the early stopping strategy and K-fold cross validation are used to improve the generalization ability and ensure the performance convergence of the model on the test set.

[0121] Step S4, risk assessment and early warning: Based on the prediction results of step S3, the Bayesian dynamic threshold algorithm is used to calculate the risk level, and a secondary early warning mechanism is set. When the predicted risk level reaches the threshold, the automatic inspection by the drone is triggered to generate a hyperspectral scan map; if the risk is further escalated, the finite element analysis model is called to calculate the stress distribution, determine the reinforcement plan, and visualize the risk distribution map through WebGIS; the prediction results include risk areas, damage levels, and damage probability distribution;

[0122] The calculation of the risk level includes:

[0123] Extract the damage probability from the prediction result output in step S3 , damage level and risk areas ;

[0124] The Bayesian dynamic threshold algorithm is used to comprehensively consider the predicted damage probability and damage level in different time ranges to calculate the overall risk index: ,in is the time weight coefficient, satisfying , is the total duration of the forecast time range;

[0125] The Bayesian uncertainty estimation method is used to calculate the confidence interval of the damage probability, which includes: obtaining distribution parameters by sampling the ancient building risk prediction model multiple times, including the mean of the output damage probability and variance , using the normal distribution assumption, at the confidence level The confidence interval of the damage probability is:

[0126] ,in is the quantile of the standard normal distribution;

[0127] The risk level is adjusted dynamically according to the risk index and confidence interval. The risk level classification rules are as follows:

[0128] ,in and There are two risk level thresholds set by a dynamic piecewise function.

[0129] The calculation of stress distribution includes:

[0130] Based on the data of step S1 and step S4, a finite element analysis model is established in combination with the geometric structure and material properties of the ancient building; wherein the data of step S1 includes the external geometric shape and size of the ancient building obtained by laser scanning of the laser radar; the distribution information of surface defects of key parts obtained by the hyperspectral camera and the drone inspection; the material degradation parameters derived from the vibration data, crack displacement data and environmental temperature and humidity data collected by the MEMS sensor, temperature and humidity sensor and laser rangefinder, including elastic modulus, Poisson's ratio and density; the data of step S4 is the finite element analysis calculation data, including the high-risk areas in the risk assessment results, the structural stability analysis data and the predicted damage location distribution;

[0131] Set loading and boundary conditions for the finite element model, apply dynamic time-dependent external loads based on seismic wave monitoring data and environmental data obtained from the weather station API; adjust the stress load distribution through drone inspection feedback data, simulate crack propagation and stress concentration effects as feedback stress loads;

[0132] The stress distribution of the ancient building under the above load and boundary conditions is calculated by finite element method, and the calculation formula is: ,in is the stress tensor, is the strain tensor, which comes from the node displacement difference; is the material stiffness matrix, determined by the material properties;

[0133] Combined with the data of high-risk areas, the key parts of the ancient buildings are analyzed, including the maximum principal stress of beams and columns. , the equivalent stress of the wall , the stress concentration position and distribution pattern of the foundation, where the equivalent stress calculation formula is: ,in is the principal stress component;

[0134] Combined with the crack growth rate and stress concentration area of ​​the drone inspection feedback data, the stress distribution results were verified, and the reinforcement learning algorithm was used to optimize the load distribution.

[0135] The reinforcement learning algorithm includes:

[0136] A reinforcement learning training environment is constructed, in which the state space, action space and reward function are defined as follows: state space, including stress distribution data (maximum principal stress, equivalent stress) of key parts, crack propagation rate, environmental load data and attribute parameters of ancient building materials (elastic modulus and Poisson's ratio); action space: actions represent different reinforcement measures (material replacement, crack filling, external force loading adjustment, etc.) applied to different structural parts, and each action corresponds to a different stress distribution adjustment scheme; reward function, the reward function is defined according to the stress distribution effect after the reinforcement scheme is optimized, specifically including reducing the maximum principal stress value in high-risk areas; low overall stress concentration (calculated by Von-Mises stress); improving structural stability indicators;

[0137] The present invention adopts the deep deterministic policy gradient algorithm in the deep reinforcement learning algorithm to achieve the optimization of the continuous action space. The algorithm model includes: a policy network for generating the optimal reinforcement action; a value network for evaluating the value of the current state-action pair. During the training process, the experience replay mechanism is used to store the sampled data to break the correlation between samples to improve the convergence of the algorithm.

[0138] The multi-source data obtained from steps S1 and S4 (including stress distribution data, crack growth rate, environmental pressure, etc.) are used as inputs of the reinforcement learning algorithm; the stability and effectiveness of the input features are ensured through data standardization and noise reduction processing;

[0139] Initialize the reinforcement learning environment and network parameters; based on the current state, the policy network generates a reinforcement action; execute the action in the simulation environment, update the stress distribution and feedback the new state; calculate the reward value based on the reward function, and update the policy network and value network parameters; iterate the training until the reward function converges.

[0140] After the training is completed, the optimal reinforcement plan output by the reinforcement learning algorithm includes specific reinforcement parts, reinforcement materials and stress distribution adjustment strategies; the optimized stress distribution is verified through finite element simulation, focusing on evaluating the structural stability and risk reduction effect of key parts. The stress distribution map and stability assessment report are output, including the specific location and risk level of high stress areas, trend analysis of stress changes over time and the determined reinforcement plan.

[0141] Step S5, dynamic management and control of digital twins: Based on the risk data generated in step S4, a digital twin model is constructed, and the stress conditions of the ancient building structure under different environmental pressures are simulated in real time through the coupling technology of Unity-3D modeling and finite element analysis, a structural stability assessment report is generated, and the reinforcement plan is optimized through reinforcement learning; the risk data includes the prediction results of step S3, risk level, drone inspection feedback data, finite element analysis calculation data and risk distribution map;

[0142] The digital twin dynamic control specifically includes:

[0143] Based on the risk data generated in step S4, a three-dimensional digital twin model of the ancient building is established through Unity-3D modeling technology. The geometric structure of the ancient building, the surface disease distribution information of key parts obtained by laser scanning with laser radar, and the multimodal data collected by MEMS sensors and hyperspectral cameras are combined to build a virtual simulation environment consistent with the actual ancient building; the finite element analysis model is used to simulate the stress behavior of the ancient building structure under different environmental pressures to generate a preliminary structural stability assessment report;

[0144] The UAV inspection feedback data, environmental change characteristic data, crack propagation rate and stress concentration distribution data obtained in step S3 and step S4 are transmitted to the digital twin model in real time, the state parameters and disease distribution information of the structure in the twin environment are updated, and the dynamic synchronization of the digital twin model is realized;

[0145] Based on the digital twin model, the stress distribution of key parts of the ancient building is simulated in real time through finite element analysis technology under different environmental conditions, and a simulation report on the structural stress condition and crack expansion trend is generated; the reinforcement plan is optimized by combining the reinforcement learning algorithm, including stress distribution adjustment and repair suggestions for high-risk parts;

[0146] WebGIS technology is used to visualize the simulation analysis results, including risk distribution maps of high-risk areas, stress concentration areas, and crack extension trend maps; combined with risk level assessment results, the status changes of risk areas are dynamically displayed to support users' rapid response to reinforcement decisions.

[0147] Step S6, repair decision: Combined with the structural stability assessment report and optimized reinforcement plan provided in step S5, the deep reinforcement learning algorithm is called to obtain the repair decision, automatically generate the repair material selection plan and construction plan, and implement them after expert review.

[0148] The repair decision specifically includes:

[0149] Based on the structural stability assessment report and optimized reinforcement plan provided in step S5, the repair requirements are determined by integrating the drone inspection feedback data, risk level assessment results and historical maintenance records, including the type of disease, damage level, repair priority and repair scope;

[0150] Combining risk data and simulation analysis results, the selection of restoration materials is optimized through deep reinforcement learning algorithms, and the strength, durability, environmental corrosion resistance and compatibility of the materials with the original materials of the ancient buildings are comprehensively considered to generate a list of recommended restoration materials; the list of recommended restoration materials is verified by comparing the parameters of ancient building restoration materials pre-collected in the expert database, which includes material performance data collected through historical restoration cases, experimental data and third-party standards;

[0151] The deep reinforcement learning algorithm includes:

[0152] Construction of reinforcement learning environment: The state space includes the current damage characteristics of ancient buildings, including risk level, disease type (cracks, peeling), damage range, environmental impact (temperature, humidity, wind speed); material properties, including material strength, resistance to environmental corrosion, and compatibility; construction conditions, including construction environment restrictions, time requirements, cost budget, etc.

[0153] Action space: including selection of repair materials (material type and quantity); repair methods (crack filling, surface coating, structural reinforcement, etc.); adjustment of construction sequence.

[0154] The reward function includes:

[0155] Positive rewards: Reduced risk level after repair; Improved stability of the repair area; Reduced material and construction costs; Shortened construction time.

[0156] Negative penalties: High-risk areas still exist after restoration; material selection does not match the original characteristics of the ancient building; the restoration plan does not meet budget or construction constraints.

[0157] The present invention adopts a dominant actor-critic based deep reinforcement learning algorithm, which is applicable to optimization problems in continuous and discrete action spaces.

[0158] Specific algorithm structure: policy network, which generates the optimal action strategy for repair material selection and construction method; critic network, which evaluates the value of the current state-action pair and provides feedback to optimize the output of the policy network.

[0159] The input data include: historical restoration case library (used to generate the initial state space); environmental constraints (such as budget and construction time limit).

[0160] Feature processing: Standardize the input data and use principal component analysis (PCA) to reduce the dimensionality of high-dimensional features to ensure the effectiveness and stability of the algorithm input.

[0161] The training process includes initialization, building a reinforcement learning environment, initializing the state space and action space; strategy generation, the policy network generates repair actions (including material selection and construction plan) according to the current state; state update and reward calculation, after executing the repair action, update the repair state and calculate the reward value; network optimization, use the gradient descent method to update the policy network and critic network parameters, and iterate the optimization until the reward function converges.

[0162] Repair plan generation: The optimal repair plan output by the deep reinforcement learning algorithm includes: repair material selection results; specific construction methods and sequence; and dynamically adjusted construction schedule.

[0163] The generated repair plan is submitted to the expert team for review, and the stability of the repaired structure is verified through simulation. If deficiencies are found, further optimization is carried out through algorithm iteration.

[0164] Based on the optimized reinforcement plan and combined with the historical restoration case library, a restoration plan is generated, including construction technology, restoration sequence and construction environment requirements; the restoration plan is dynamically adjusted in combination with drone inspection data and environmental change trends;

[0165] The generated restoration plan is submitted to the expert team for review. The structural stability of the restored ancient building is simulated through finite element simulation to verify the feasibility and safety of the plan. If the stability requirements are not met, the plan is iteratively optimized through the reinforcement learning algorithm and resubmitted for review.

[0166] The approved repair plan enters the implementation phase, and the construction process is tracked by a real-time monitoring system, including stress changes in key parts and the effect of disease repair. After the repair is completed, the monitoring feedback data is updated to the digital twin model for subsequent risk prediction and dynamic management and control.

[0167] After the restoration is completed, a comparison map of the distribution of defects after restoration is generated through drone inspections and hyperspectral scanning, and the stability of the restoration area is verified in combination with the structural status sensor data; the overall improvement rate and risk reduction level of the restoration effect are calculated through the risk assessment module, and a restoration report is formed and stored in the ancient building protection database.

[0168] Example 2, ancient building risk prediction and control system based on large model, see Figure 1 As shown, it includes the following modules:

[0169] Multi-source data acquisition module: used to obtain the structural status data of key parts of ancient buildings, the visual image data of the surface of ancient buildings, and the overall environmental data of ancient buildings;

[0170] Multimodal data fusion module: used to pre-process the acquired multi-source data, extract and fuse the features of the structural state data, visual image data, and environmental data, and obtain the high-dimensional feature vector of the multimodal fusion feature;

[0171] Large model risk prediction module: It is used to establish an ancient building risk prediction model by using the ViT-GPT model pre-trained by self-supervised learning and combining it with the LSTM time series prediction model. It predicts the probability of damage to ancient buildings in the next 3, 6, and 12 months, and outputs the predicted risk area and damage level.

[0172] Risk assessment and early warning module: used to calculate the risk level using the Bayesian dynamic threshold algorithm and set a secondary early warning mechanism. When the risk level exceeds the threshold, the drone will be triggered to automatically inspect and generate a hyperspectral scan image; if the risk level further exceeds the threshold , the finite element analysis model is called to calculate the stress distribution, determine the reinforcement plan, and visualize the risk distribution map through WebGIS;

[0173] Digital twin dynamic control module: used to build a digital twin model. Through the coupling technology of Unity-3D modeling and finite element analysis, it can simulate the stress of ancient building structures under different environmental pressures in real time, generate a structural stability assessment report, and optimize the reinforcement plan through reinforcement learning.

[0174] Restoration decision module: used to call deep reinforcement learning algorithm to obtain restoration decisions, automatically generate restoration material selection plan and construction plan, and implement them after expert review.

[0175] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The risk prediction and control method of ancient buildings based on large models is characterized by: The following steps are involved: Step S1, multi-source data collection: obtaining structural status data of key parts of ancient buildings, visual image data of the surface of ancient buildings, and environmental data of the ancient buildings as a whole; Step S2, multimodal data fusion: preprocessing the multi-source data obtained in step S1, including wavelet transform denoising, standardization, and principal component analysis dimensionality reduction; Extract and fuse the structural state data, visual image data, and environmental data to obtain a high-dimensional feature vector of multimodal fusion features; Step S3, large model risk prediction: Based on the feature extraction and fusion of step S2, the ViT-GPT model pre-trained by self-supervised learning is used in combination with the LSTM time series prediction model to establish an ancient building risk prediction model, predict the damage probability of ancient buildings in the next 3 months, 6 months, and 12 months, and output the predicted risk area and damage level; Step S4, risk assessment and warning: Based on the prediction results of step S3, the Bayesian dynamic threshold algorithm is used to calculate the risk level, and a secondary warning mechanism is set. When the predicted risk level exceeds the threshold, When the UAV is triggered to automatically inspect, a hyperspectral scan image is generated; If the risk level further exceeds the threshold , the finite element analysis model is called to calculate the stress distribution, determine the reinforcement plan, and visualize the risk distribution map through WebGIS; the prediction results include risk areas, damage levels and damage probability distribution; The calculation of the risk level includes: Extract the damage probability from the prediction result output in step S3 , damage level and risk areas ; The Bayesian dynamic threshold algorithm is used to comprehensively consider the predicted damage probability and damage level in different time ranges to calculate the overall risk index: ,in is the time weight coefficient, satisfying , is the total duration of the forecast time range; The Bayesian uncertainty estimation method is used to calculate the confidence interval of the damage probability, which includes: obtaining distribution parameters by sampling the ancient building risk prediction model multiple times, including the mean of the output damage probability and variance , using the normal distribution assumption, at the confidence level The confidence interval of the damage probability is: ,in is the quantile of the standard normal distribution; The risk level is adjusted dynamically according to the risk index and confidence interval. The risk level classification rules are as follows: ,in and are two risk level thresholds set by a dynamic piecewise function; Step S5, dynamic management and control of digital twins: Based on the risk data generated in step S4, a digital twin model is constructed, and the stress conditions of the ancient building structure under different environmental pressures are simulated in real time through the coupling technology of Unity-3D modeling and finite element analysis, a structural stability assessment report is generated, and the reinforcement plan is optimized through reinforcement learning; the risk data includes the prediction results of step S3, risk level, drone inspection feedback data, finite element analysis calculation data and risk distribution map; Step S6, repair decision: Combined with the structural stability assessment report and optimized reinforcement plan provided in step S5, the deep reinforcement learning algorithm is called to obtain the repair decision, automatically generate the repair material selection plan and construction plan, and implement them after expert review.

2. The method for predicting and controlling ancient building risks based on a large model according to claim 1 is characterized in that: The multi-source data collection specifically includes: Structural status data: MEMS acceleration sensors, strain gauges, temperature and humidity sensors, and laser rangefinders are deployed at key locations of ancient buildings to collect vibration, stress, temperature and humidity, and crack displacement data in real time; the key locations include beams, columns, walls, foundations, and marked locations; Visual image data: Obtain surface image data of ancient buildings containing damage information through hyperspectral cameras and drone remote sensing equipment; the damage information includes cracks, spalling, bioerosion, deformation, material deterioration and external force damage; Environmental data: including geological structure change data and environmental factor data; the geological structure change data is obtained through seismic wave monitoring equipment and lidar; the environmental factor data is obtained by calling the weather station API, including wind speed, precipitation, temperature and humidity in the ancient building area.

3. The method for predicting and controlling ancient building risks based on a large model according to claim 1 is characterized in that: The multimodal data fusion specifically includes: Data preprocessing: Wavelet transform denoising: multi-scale decomposition of vibration data is performed through discrete wavelet transform to remove high-frequency noise; standardization: normalization of the numerical range of different data sources is performed through Min-Max normalization; principal component analysis dimensionality reduction: principal component analysis is performed on hyperspectral images and remote sensing data; Feature extraction: A bidirectional LSTM network is used to process the structural state data and extract the time series data of historical vibration, stress change, crack propagation rate, and temperature and humidity change trend. Combined with the Transformer-BERT encoder, the global dependency of the time series data is calculated through the self-attention mechanism to generate structural state features. A convolutional neural network is used to extract hierarchical features from visual image data, and a feature pyramid network is combined to enhance the disease detection capabilities at different scales, and the disease features are output; the hierarchical features include information about each type of disease; Use random forests to perform regression analysis on environmental data and extract the impact of environmental factors on the risk of ancient buildings; combine LSTM networks to model long-term climate trends and extract environmental change characteristics; High-dimensional feature vector construction: Dynamic time warping is used to align the time scales of time series data and visual image data. A multi-head self-attention mechanism is used to calculate the correlation between structural state features, disease features, and environmental change features to obtain a high-dimensional feature vector of multimodal fusion features, and weights are assigned to different features through weighted feature fusion. The feature extraction and fusion are integrated into the ancient building risk prediction model.

4. The method for predicting and controlling ancient building risks based on a large model according to claim 1 is characterized in that: The structure of the ancient building risk prediction model includes: Input layer: used to receive multi-source data pre-processed in step S2, including structural state data, visual image data and environmental data, and perform standard formatting on them to meet the model input requirements; Feature extraction layer: used to extract structural state features of structural state data, disease features of visual image data, and environmental change features of environmental data; Feature fusion layer: used to fuse features through a multi-head self-attention mechanism to obtain a high-dimensional feature vector of multimodal fusion features; and adopt a self-supervised contrastive learning mechanism to enhance the representation ability of different modal features; Risk prediction layer: The Seq2Seq-Transformer module in the ViT-GPT model is used to optimize the temporal expression capability of disease characteristics. The LSTM time series prediction model is combined with the fused high-dimensional feature vector to predict the damage probability of ancient buildings in the next 3, 6, and 12 months, and calculate the damage level of each risk area. The damage risk trend is predicted, and the key risk areas are highlighted through the attention mechanism. Output layer: Output the predicted risk area, damage level and damage probability distribution; calculate the confidence of the prediction results in combination with Bayesian uncertainty estimation.

5. The method for predicting and controlling ancient building risks based on a large model according to claim 4 is characterized in that: The training of the ancient building risk prediction model includes: A1. Construct a multimodal training dataset based on historical building structure status data, visual image data, and environmental data, including input data X and label data Y , where the input data ,in is the structural status data, For visual image data, is environmental data; label data includes, Damage probability labels are generated based on historical inspection data and expert annotations, indicating the probability of damage to the ancient building in the future. Damage level labels, set as discrete levels, determined based on historical maintenance records and expert assessment results; The training sample set is expressed as: ,in is the number of training samples, is the sample index; A2. Use self-supervised contrastive learning to pre-train the input data, optimize the multimodal feature representation, and set positive samples : The same building history data, negative samples : Different building data, using contrast loss function to optimize feature representation: ,in is the high-dimensional feature vector output by the feature extraction network, is the cosine similarity function, which is used to measure the similarity between two feature vectors; is the temperature parameter, which is used to control the sensitivity of the similarity calculation. Indicates the number of negative samples; A3. Training of feature extraction layer: Bidirectional LSTM+Transformer training is used to extract structural features and optimize time series feature learning. The training goal is to minimize the mean square error: ,in is the actual damage probability, is the predicted value, is the total duration of the forecast time range; The CNN-FPN network is used to train visual image feature extraction, and the loss function is Focal: ,in To predict the probability of disease, is the category weight, which is used to prevent category imbalance; is an adjustment factor used to control the influence of difficult samples; Random forest + LSTM training environment data feature extraction is used, and the training goal is also to minimize the mean square error; A4. Use multi-head self-attention mechanism to calculate the correlation between different modal features: ,in Q, K, V is the query, key, and value matrix, is the corresponding training weight, Represents dimension; The training goal is to minimize the feature fusion error: ,in They are the adaptive fusion weights of structural state, visual image, and environmental data respectively; is the target damage probability label; A5. Use ViT-GPT+LSTM to predict the time series of damage probability and optimize the risk prediction for the next 3, 6, and 12 months. The loss function is to minimize the cross entropy loss: ,in is the total number of category intervals of damage probability, is the one-hot encoding of the true category, is the predicted damage probability distribution; Predicting future risk levels based on Seq2Seq-Transformer: ,in is the real damage level label, is the predicted probability, r Indicates the level of damage; A6. Use Adam optimizer to update model parameters: ,in is the parameter of the tth round of training, represents the learning rate, Represents the loss function; the early stopping strategy and K-fold cross validation are used to improve the generalization ability and ensure the performance convergence of the model on the test set.

6. The method for predicting and controlling ancient building risks based on a large model according to claim 1 is characterized in that: The calculation of stress distribution includes: Based on the data of step S1 and step S4, a finite element analysis model is established in combination with the geometric structure and material properties of the ancient building; wherein the data of step S1 includes the external geometric shape and size of the ancient building obtained by laser scanning of the laser radar; the distribution information of surface defects of key parts obtained by the hyperspectral camera and the drone inspection; the material degradation parameters derived from the vibration data, crack displacement data and environmental temperature and humidity data collected by the MEMS sensor, temperature and humidity sensor and laser rangefinder, including elastic modulus, Poisson's ratio and density; the data of step S4 is the finite element analysis calculation data, including the high-risk areas in the risk assessment results, the structural stability analysis data and the predicted damage location distribution; Set loading and boundary conditions for the finite element model, apply dynamic time-dependent external loads based on seismic wave monitoring data and environmental data obtained from the weather station API; adjust the stress load distribution through drone inspection feedback data, simulate crack propagation and stress concentration effects as feedback stress loads; The stress distribution of the ancient building under the above load and boundary conditions is calculated by finite element method, and the calculation formula is: ,in is the stress tensor, is the strain tensor, which comes from the node displacement difference; is the material stiffness matrix, determined by the material properties; Combined with the data of high-risk areas, the key parts of the ancient buildings are analyzed, including the maximum principal stress of beams and columns. , the equivalent stress of the wall , the stress concentration position and distribution pattern of the foundation, where the equivalent stress calculation formula is: ,in is the principal stress component; Combined with the crack growth rate and stress concentration area of ​​the drone inspection feedback data, the stress distribution results were verified, and the reinforcement learning algorithm was used to optimize the load distribution. Output stress distribution diagram and stability assessment report, including the specific location and risk level of high stress areas, trend analysis of stress changes over time and determined reinforcement plan.

7. The method for predicting and controlling ancient building risks based on a large model according to claim 1 is characterized in that: The digital twin dynamic control specifically includes: Based on the risk data generated in step S4, a three-dimensional digital twin model of the ancient building is established through Unity-3D modeling technology. The geometric structure of the ancient building, the surface disease distribution information of key parts obtained by laser scanning with laser radar, and the multimodal data collected by MEMS sensors and hyperspectral cameras are combined to build a virtual simulation environment consistent with the actual ancient building; the finite element analysis model is used to simulate the stress behavior of the ancient building structure under different environmental pressures to generate a preliminary structural stability assessment report; The UAV inspection feedback data, environmental change characteristic data, crack propagation rate and stress concentration distribution data obtained in step S3 and step S4 are transmitted to the digital twin model in real time, the state parameters and disease distribution information of the structure in the twin environment are updated, and the dynamic synchronization of the digital twin model is realized; Based on the digital twin model, the stress distribution of key parts of the ancient building is simulated in real time through finite element analysis technology under different environmental conditions, and a simulation report on the structural stress condition and crack expansion trend is generated; the reinforcement plan is optimized by combining the reinforcement learning algorithm, including stress distribution adjustment and repair suggestions for high-risk parts; WebGIS technology is used to visualize the simulation analysis results, including risk distribution maps of high-risk areas, stress concentration areas, and crack extension trend maps.

8. The method for predicting and controlling ancient building risks based on a large model according to claim 1 is characterized in that: The repair decision specifically includes: Based on the structural stability assessment report and optimized reinforcement plan provided in step S5, the repair requirements are determined by integrating the drone inspection feedback data, risk level assessment results and historical maintenance records, including the type of disease, damage level, repair priority and repair scope; Combining risk data and simulation analysis results, the selection of restoration materials is optimized through deep reinforcement learning algorithms, and the strength, durability, environmental corrosion resistance and compatibility of the materials with the original materials of the ancient buildings are comprehensively considered to generate a list of recommended restoration materials; the list of recommended restoration materials is verified by comparing the parameters of ancient building restoration materials pre-collected in the expert database, which includes material performance data collected through historical restoration cases, experimental data and third-party standards; Based on the optimized reinforcement plan and combined with the historical restoration case library, a restoration plan is generated, including construction technology, restoration sequence and construction environment requirements; the restoration plan is dynamically adjusted in combination with drone inspection data and environmental change trends; The generated restoration plan is submitted to the expert team for review. The structural stability of the restored ancient building is simulated through finite element simulation to verify the feasibility and safety of the plan. If the stability requirements are not met, the plan is iteratively optimized through the reinforcement learning algorithm and resubmitted for review. The approved repair plan enters the implementation phase, and the construction process is tracked by a real-time monitoring system, including stress changes in key parts and the effect of disease repair. After the repair is completed, the monitoring feedback data is updated to the digital twin model for subsequent risk prediction and dynamic management and control. After the restoration is completed, a comparison map of the distribution of defects after restoration is generated through drone inspections and hyperspectral scanning, and the stability of the restoration area is verified in combination with the structural status sensor data; the overall improvement rate and risk reduction level of the restoration effect are calculated through the risk assessment module, and a restoration report is formed and stored in the ancient building protection database.

9. The ancient building risk prediction and control system based on large models is characterized by: The system applies the ancient building risk prediction and control method based on a large model as described in any one of claims 1 to 8 above, Includes the following modules: Multi-source data acquisition module: used to obtain the structural status data of key parts of ancient buildings, the visual image data of the surface of ancient buildings, and the overall environmental data of ancient buildings; Multimodal data fusion module: used to pre-process the acquired multi-source data, extract and fuse the features of the structural state data, visual image data, and environmental data, and obtain the high-dimensional feature vector of the multimodal fusion feature; Large model risk prediction module: It is used to establish an ancient building risk prediction model by using the ViT-GPT model pre-trained by self-supervised learning and combining it with the LSTM time series prediction model. It predicts the probability of damage to ancient buildings in the next 3, 6, and 12 months, and outputs the predicted risk area and damage level. Risk assessment and early warning module: used to calculate the risk level using the Bayesian dynamic threshold algorithm and set a secondary early warning mechanism. When the UAV is triggered to automatically inspect, a hyperspectral scan image is generated; If the risk level further exceeds the threshold , the finite element analysis model is called to calculate the stress distribution, determine the reinforcement plan, and visualize the risk distribution map through WebGIS; Digital twin dynamic control module: used to build a digital twin model. Through the coupling technology of Unity-3D modeling and finite element analysis, it can simulate the stress of ancient building structures under different environmental pressures in real time, generate a structural stability assessment report, and optimize the reinforcement plan through reinforcement learning. Restoration decision module: used to call deep reinforcement learning algorithm to obtain restoration decisions, automatically generate restoration material selection plan and construction plan, and implement them after expert review.

Images