A building structure monitoring and early warning method and system combining digital twinning and machine learning

Abstract

The application provides a kind of building structure monitoring early warning method and system combined with digital twin and machine learning, and relates to the technical field of building structure detection system.The system comprises: data acquisition layer, for deploying multiple types of Internet of Things sensors and laser scanning equipment, collecting building data globally;Data preprocessing and edge computing layer, for wavelet threshold denoising of original monitoring data, while completing edge lightweight model inference;Digital twin modeling layer, for constructing multi-scale dynamic twin model;Machine learning intelligent research and early warning layer, realizes structure explicit risk determination, implicit damage identification, state trend prediction, and triggers graded early warning according to risk level.The system constructs dynamic digital twin modeling, realizes the synchronous mapping of physical building and digital model, takes into account explicit risk rapid determination and implicit damage, intelligent identification of trend degradation, and triggers graded early warning according to risk level.

Description

Technical Field

[0001] This invention belongs to the technical field of building structure detection systems, specifically relating to a building structure monitoring and early warning method and system that combines digital twins with machine learning. Background Technology

[0002] During long-term operation, building structures are subject to the combined effects of multiple factors, including self-weight loads, live loads, temperature and humidity environments, wind and rain erosion, material aging, and residual construction stress. This can easily lead to safety hazards such as deformation, settlement, stress concentration, and component damage. If these hazards are not identified, predicted, and addressed in a timely manner, they can easily cause major safety accidents such as structural failure and collapse. Therefore, conducting all-weather, high-precision, and intelligent health monitoring and early warning systems for building structures is a core technological means to ensure safe building operation, achieve refined operation and maintenance, and extend the service life of buildings.

[0003] Current building structural health monitoring technology still faces challenges, with insufficient overall intelligence, precision, and real-time capabilities. Specific shortcomings include: First, poor model dynamic synchronization and low simulation accuracy. Traditional monitoring systems often employ static finite element models or simplified BIM geometric models, capable of fixed simulations based solely on initial building parameters. They cannot dynamically update to reflect actual building operating conditions, structural aging, or environmental changes. A significant lag exists between the physical building and the digital model, making it difficult to accurately reflect the real-time stress and deformation state of the structure, resulting in low reliability of the monitoring simulation. Second, the risk assessment model is simplistic, leading to high false alarm and missed alarm rates. Existing early warning mechanisms generally employ fixed threshold judgment rules, only providing post-event alerts for explicit exceedances of parameters such as displacement, settlement, and strain. They cannot uncover hidden risks such as accumulated structural fatigue, early micro-damage, and slow performance degradation within time-series monitoring data. Faced with complex conditions involving multiple loads and coupled environments, the accuracy of risk identification is extremely low, lacking the ability to predict risks in advance. Third, low data utilization and poor model generalization. Traditional monitoring often relies on single sensors for independent data acquisition and fragmented data storage, lacking mechanisms for multi-source heterogeneous data fusion and processing. Furthermore, conventional machine learning monitoring models are purely data-driven, lacking constraints from structural physical mechanisms, making them prone to overfitting during training. They also exhibit insufficient stability under extreme weather conditions, sudden loads, and other special operating conditions, with weak generalization and anti-interference capabilities. Fourth, early warnings are delayed and lack risk tracing capabilities. Existing systems primarily rely on reactive, post-event early warnings, failing to predict structural state trends or extrapolate lifespan. They can only provide simple alerts for abnormal exceedances, unable to accurately pinpoint the specific location of damage, quantify damage levels and impact ranges, or trace the causes of anomalies or simulate and reproduce operating conditions, making it difficult to support precise decision-making by maintenance personnel. Fifth, there is an inherent contradiction between monitoring accuracy and real-time performance. High-precision finite element simulation models are computationally intensive and consume significant computing power, making it difficult to meet the demands of real-time edge monitoring and millisecond-level early warning responses. Lightweight, simplified models, on the other hand, significantly lose structural detail features, severely reducing simulation accuracy and failing to balance monitoring accuracy with real-time early warning performance.

[0004] In summary, existing building structure monitoring technologies suffer from technical bottlenecks such as static modeling, mechanized analysis, fragmented data, and delayed early warning, making it difficult to meet the intelligent, precise, and predictive safety monitoring needs of modern buildings throughout their entire life cycle. There is an urgent need for a new monitoring and early warning method and system that integrates physical mechanisms and artificial intelligence. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a building structure monitoring and early warning method and system that combines digital twins with machine learning. This solves the technical problems of poor synchronization, weak intelligent judgment, delayed early warning, and inability to balance accuracy and real-time performance in traditional monitoring models. It enables a fully closed-loop intelligent monitoring system that achieves dynamic mapping of building structures across all dimensions, accurate identification of hidden damage, trend prediction, hierarchical early warning, and risk tracing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] Firstly, this application provides a building structure monitoring and early warning method combining digital twins and machine learning, including:

[0008] S1. Deploy multiple types of IoT sensors and laser scanning equipment to collect data on building structure deformation, stress, settlement, environmental parameters, appearance damage and load across the entire area, and construct the original monitoring dataset;

[0009] S2. Perform wavelet threshold denoising, anomaly removal, time-series alignment, missing data completion and multi-source heterogeneous data fusion on the original monitoring dataset, and at the same time complete lightweight model inference at the edge to reduce the computing load on the cloud.

[0010] S3. Construct a multi-scale dynamic twin model that integrates geometric parameters, physical properties, and structural dynamics mechanisms. Based on measured data and combined with the physical residual loss function, correct the model parameters in real time to achieve a 1:1 dynamic synchronous mapping between the physical building and the digital model.

[0011] S4. Combining the building safety code rule base with multiple machine learning models, and relying on the improved LSTM time series prediction formula and autoencoder anomaly measurement formula, we can realize the determination of structural explicit risks, the identification of hidden damage, the prediction of state trends, and trigger graded early warnings according to the risk level.

[0012] S5 provides a 3D visualization display of the building's real-time twin status, monitoring data, and early warning information, enabling risk tracing, operational condition simulation, lifespan projection, and outputting intelligent operation and maintenance decision-making suggestions.

[0013] Furthermore, step S1 includes accelerometers, strain sensors, displacement sensors, settlement sensors, temperature and humidity sensors, wind pressure sensors, and UAV laser point cloud scanning equipment deployed at key building components; the sensors collect dynamic response data of the structure and environmental condition data in real time, and the laser point cloud scanning equipment periodically collects three-dimensional point cloud data of the building's exterior, realizing full-domain synchronous acquisition of the internal state and external form of the structure.

[0014] Furthermore, the specific processing methods for the data preprocessing and edge computing layer include: using a wavelet threshold denoising algorithm to remove random noise and impulse interference from the sensor; the wavelet threshold denoising formula is as follows:

[0015] ;

[0016] in, These are the original wavelet decomposition coefficients. These are the denoised wavelet coefficients. For adaptive threshold, It is a symbolic function;

[0017] Missing data are filled in using linear interpolation and time-series fitting algorithms. Heterogeneous data with different sampling frequencies and dimensions are fused through time-series alignment and dimensionality normalization. The normalization formula is as follows:

[0018] ;

[0019] in, This is the raw monitoring data. For normalized data, These are the maximum and minimum values ​​of the data sequence, respectively.

[0020] Lightweight early warning models are deployed on edge terminals to complete real-time data inference and preliminary anomaly screening, while complex model training and simulation calculations are uploaded to the cloud for execution.

[0021] Furthermore, in step S3, the model construction and dynamic update methods include:

[0022] A1. Based on architectural BIM design drawings, construction records, and material mechanics parameters, build an initial geometric-physical coupled finite element twin model and assign material properties, constraints, and mechanical boundary conditions to the model components.

[0023] A2. Obtain 3D data of building entities through UAV laser point cloud scanning, complete the accurate mapping of point cloud data with voxels, nodes and elements of finite element model, and correct geometric deformation deviations of model;

[0024] A3. Using the structural dynamics equilibrium equations as physical constraints, construct a model loss function that includes physical residual terms, and iteratively optimize uncertain parameters such as model stiffness, damping, and elastic modulus by combining real-time monitoring data; the structural dynamics constraint equations satisfy:

[0025] ;

[0026] in, For the quality matrix, Here is the damping matrix. Here is the stiffness matrix. These are the acceleration, velocity, and displacement response vectors, respectively. This is the time-series load vector;

[0027] Construct the physical residual loss function:

[0028] ;

[0029] in, The fitting loss between measured and simulated data, For the loss of dynamic physical residuals, These are the physical constraint weight coefficients; minimized Eliminate the error between simulation data and physical response data;

[0030] S4. An incremental update mechanism is adopted to fine-tune the model parameters based on newly added monitoring data at regular intervals, so as to realize the dynamic adaptive update of the digital model of the entire building life cycle.

[0031] Furthermore, in step S4, a dual-drive judgment architecture of rule base + machine learning is adopted, including a bottom-level rule judgment module and an upper-level intelligent prediction and recognition module.

[0032] The underlying rule analysis module presets deformation, settlement, and stress thresholds based on national building structure safety standards to quickly determine structural over-standard risks.

[0033] The upper-layer intelligent prediction and recognition module includes an improved LSTM time-series prediction model and a deep autoencoder anomaly detection model.

[0034] The improved LSTM model captures temporal features based on a gating mechanism, and the cell state update formula is as follows:

[0035] ;

[0036] in, These are the input gate, forget gate, and output gate, respectively. The cell state at time t. Output in hidden state It is the sigmoid activation function. The improved LSTM model is used to mine the temporal correlation features between structural loads, environmental parameters and structural state, thereby achieving 7-30 day structural performance trend prediction.

[0037] The deep autoencoder unsupervised model is used to compare the deviation characteristics between the real-time structural state and the healthy baseline state, and to determine anomalies by reconstruction error. The anomaly measurement formula is as follows:

[0038] ;

[0039] in, Features of the health baseline sample To reconstruct sample features in real time, For feature dimension; when When the error exceeds the preset abnormal threshold, it is determined that there are early hidden micro-damage and slow performance degradation problems.

[0040] Furthermore, in step S4, a three-tiered early warning mechanism is configured, combined with a comprehensive risk index. To achieve tiered and quantitative risk management, the risk index is calculated using the following formula:

[0041] ;

[0042] in, To reconstruct abnormal errors, This is the threshold for abnormal errors. This is the current structural parameter value. To standardize safety thresholds, The trend deterioration coefficient, The weighting coefficients are and satisfy the following conditions: ;

[0043] The criteria for determining the warning level are as follows: Level 1 Warning: Only data recording and continuous monitoring are performed; Level 2 warning: Automatically pushes maintenance and rectification reminders; Level 3 early warning: This triggers the platform's audible and visual alarms, emergency push notifications on mobile devices, and risk locking.

[0044] Meanwhile, relying on the dynamic simulation and deduction function of the twin model, the abnormality causes can be traced in reverse, effectively distinguishing between sensor failure, environmental interference and structural damage, accurately locating the damage coordinates and quantifying the degree of damage.

[0045] Furthermore, the method also includes using a distillation lightweight algorithm to construct a teacher-student model structure, with the distillation loss function being:

[0046] ;

[0047] in, For the student model and the loss of the true label, Output probability distribution loss for both the student and teacher models. This is the distillation weighting coefficient;

[0048] The distillation algorithm is used to compress parameters and simplify the structure of high-precision twin models and machine learning models, significantly reducing model size and computational load while retaining core feature parameters and inference accuracy. An edge-cloud collaborative computing architecture is built, with the edge responsible for real-time data preprocessing and lightweight model inference, and the cloud responsible for high-precision simulation calculation and model iterative training, effectively balancing monitoring accuracy and millisecond-level early warning response speed.

[0049] Furthermore, the method also includes employing an online incremental update strategy and defining an incremental loss function:

[0050] ;

[0051] in, The model fitting loss is for the newly added monitoring data. For historical sample memory loss, This is the historical weight decay coefficient;

[0052] Based on the incremental loss function, the parameters of the machine learning model and twin model can be continuously optimized iteratively based on the newly added monitoring data accumulated over a long period of time. The historical health status characteristics of the structure are retained, and the system can be adaptively adapted to complex working conditions such as aging of building materials, dynamic changes in load, and environmental fluctuations, which significantly improves the generalization and operational stability of the system's long-term monitoring.

[0053] Furthermore, in step S5, a three-dimensional interactive twin visualization platform is constructed to render and display the building structure's operating status, monitoring data change curves, damage distribution locations, and early warning indicators in real time. It supports functions such as historical condition reproduction, risk cause tracing, and structural remaining life estimation. It can automatically generate customized operation and maintenance decision suggestions such as inspection, reinforcement, and rectification based on different risk levels, forming a complete closed-loop management system of monitoring-analysis-early warning-operation and maintenance.

[0054] Secondly, this application provides a building structure monitoring and early warning system combining digital twins and machine learning, including:

[0055] The data acquisition layer is used to deploy various types of IoT sensors and laser scanning equipment to collect data on building structure deformation, stress, settlement, environmental parameters, appearance damage and load across the entire domain, and to build the original monitoring dataset.

[0056] The data preprocessing and edge computing layer is used to perform wavelet threshold denoising, anomaly removal, time-series alignment, missing data completion and multi-source heterogeneous data fusion on the raw monitoring data, while completing lightweight model inference at the edge and reducing the computing load on the cloud.

[0057] The digital twin modeling layer is used to construct a multi-scale dynamic twin model that integrates geometric parameters, physical properties, and structural dynamics mechanisms. Based on measured data and combined with the physical residual loss function, the model parameters are corrected in real time to achieve a 1:1 dynamic synchronous mapping between the physical building and the digital model.

[0058] The machine learning intelligent judgment and early warning layer is used to combine the building safety code rule base with multiple machine learning models, relying on the improved LSTM time series prediction formula and autoencoder anomaly measurement formula to realize the judgment of explicit structural risks, the identification of hidden damage, the prediction of state trends, and to trigger graded early warnings according to the risk level.

[0059] The visualization operation and maintenance and decision output layer is used to display the real-time twin status of the building, monitoring data and early warning information in three dimensions, complete risk tracing, working condition simulation, lifespan extrapolation, and output intelligent operation and maintenance decision suggestions.

[0060] Compared with the prior art, the present invention has the following significant advantages:

[0061] 1. This system constructs a dynamic digital twin model, which integrates physical mechanisms and measured data to achieve synchronous mapping between physical buildings and digital models, improving the accuracy of building simulation models and realistically restoring the real-time stress, deformation and aging state of building structures. At the same time, the system constructs a dual-driven judgment mechanism of rule base and machine learning, which takes into account the rapid judgment of explicit risks and the intelligent identification of hidden damage and trend degradation, and triggers graded early warnings according to the risk level.

[0062] 2. This system is equipped with a multi-source data fusion noise reduction and incremental learning mechanism to effectively solve the problems of sensor data noise, missing data, and fragmentation. At the same time, the model can be autonomously iterated and optimized to adapt to the changes in the working conditions of the entire building life cycle. Its anti-interference ability and generalization ability in complex scenarios are far superior to pure data-driven models. Through model distillation and lightweighting and edge-cloud collaborative computing architecture, the early warning response speed is greatly improved while maintaining simulation accuracy, ensuring the effectiveness of real-time monitoring.

[0063] 3. The visualization operation and maintenance and decision output layer of this system has the capabilities of accurate damage location, risk source tracing, life prediction and intelligent decision output. It realizes full-process visualization control through a three-dimensional visualization twin platform, which greatly reduces the cost of manual inspection, improves the refinement and intelligence level of building structure operation and maintenance, and effectively ensures the safety of building operation throughout its entire life cycle. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the overall layered architecture of the present invention. Detailed Implementation

[0065] 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. The present invention will be further described in detail and completely below with reference to specific embodiments, technical parameters, algorithm implementation process and comparative experiments.

[0066] Example:

[0067] This embodiment discloses a building structure monitoring and early warning system combining digital twins and machine learning. A super high-rise commercial office building in a certain city is selected as the monitoring object. The building has a total height of 148m, a frame-core tube steel structure, and has been in service for 8 years. It is subject to long-term wind loads, temperature stresses, and alternating floor live loads, which can easily lead to hidden dangers such as stress concentration in beams and columns, micro-deformation of the curtain wall, and uneven floor settlement. The system of this invention deploys sensing devices throughout the entire area, integrating digital twin dynamic modeling and multi-machine collaborative analysis to achieve automated monitoring, hidden damage identification, trend prediction, and hierarchical early warning and control of the entire life cycle of the building's main structure, core tube components, steel beams and columns, and curtain wall cladding.

[0068] At the data acquisition layer, a comprehensive sensing network is deployed based on the structural stress characteristics and high-risk areas of the building: strain sensors and acceleration sensors are deployed on the vertical load-bearing components of the core tube, large-span steel beams, and key sections of floor slabs; high-precision settlement and displacement sensors are deployed at the four corners of each floor and settlement observation points; and wind pressure sensors and temperature and humidity sensors are deployed at key points on the facade and curtain wall. All sensors are uniformly set to a 100Hz sampling frequency to ensure complete capture of dynamic load data. Simultaneously, a high-definition lidar drone is used to complete a comprehensive point cloud scan of the building facade every 7 days, generating a raw 3D point cloud set. Simultaneously, internal structural mechanical response data, external environmental data, and appearance morphology data are collected to construct a spatiotemporally aligned multi-source raw monitoring dataset.

[0069] Data preprocessing and edge computing layer perform standardized data purification and fusion processing: Wavelet threshold denoising algorithm is used to reduce noise in the original data. In this embodiment, db4 wavelet basis and 5-level decomposition are selected. The adaptive threshold T is dynamically solved according to the real-time noise variance, effectively removing pulse noise and high-frequency noise caused by wind and rain interference and electromagnetic interference. For data loss caused by short-term equipment offline and signal fluctuation, time-series piecewise linear interpolation is used to complete the data. Then, the extreme value normalization formula is used to uniformly map data of different dimensions such as strain, settlement, acceleration, and wind pressure to the [0,1] interval to complete the fusion of multi-source heterogeneous data.

[0070] Lightweight inference models are deployed on edge terminals to perform initial screening of real-time data, uploading only valid abnormal fragments and full data at regular intervals to the cloud, thereby reducing cloud computing power consumption.

[0071] The digital twin modeling layer performs full-process dynamic modeling and parameter iterative updates, using the following methods:

[0072] S1. First, based on the building as-built BIM drawings, steel structure material parameters, concrete strength grade, and construction reserved settlement parameters, an initial geometric-physical coupled finite element twin model is built, structural solid elements, shell elements and beam elements are divided, and mechanical properties such as elastic modulus, Poisson's ratio, damping ratio, and mass density are assigned to each component. Boundary conditions such as ground fixed support and node hinge are set.

[0073] S2. Next, import UAV laser point cloud data and use voxel registration algorithm to achieve accurate mapping between point cloud and finite element model nodes, correcting geometric deviations such as building facade deformation and component offset;

[0074] S3. Subsequently, physical constraints are constructed based on the structural dynamic equilibrium equations. The physical weight coefficient λ = 0.6 is set, and a total loss function is constructed that includes data fitting loss and dynamic residual loss. By minimizing To eliminate the error between simulation data and physical response data, measured acceleration, strain, and settlement data are used as true values. The gradient descent algorithm is used to iteratively optimize key unknown parameters such as model stiffness matrix, damping matrix, and elastic modulus, so that the error between the simulated response of the twin model and the response of the physical building is controlled within 2%.

[0075] S4. The system performs incremental updates every 24 hours to continuously adapt to structural aging and changes in environmental conditions, ensuring the model's fidelity throughout its entire lifecycle.

[0076] The machine learning-based intelligent judgment and early warning layer adopts a dual-drive architecture of "standard rule base + dual-model intelligent recognition": the rule base presets hard thresholds for settlement, displacement, and strain based on the "Code for Design of Building Structures" and "Standard for Steel Structure Design" to quickly determine explicit risks of exceeding standards. The intelligent recognition module uses an improved LSTM model and a deep autoencoder model working together: the improved LSTM model is built according to the recorded gating update formula, with an input dimension of 12, a hidden layer dimension of 64, an iteration batch of 32, and an iteration round of 100. It takes 72 hours of continuous historical monitoring time series data as input, learns the temporal coupling law of temperature, wind pressure, load, and structural deformation, and outputs a structural state prediction curve for the next 15 days, which can accurately capture trend degradation characteristics such as slow settlement and fatigue accumulation. The depth autoencoder adopts a 4-layer encoding-decoding structure. It uses steady-state data from the early stage of building health service as a benchmark sample. It calculates the deviation between real-time data and health benchmark through the reconstruction error formula. When the reconstruction error Erec exceeds the preset threshold, it can identify hidden anomalies such as early micro-damage, bolt loosening, and component stress shift that cannot be observed by the naked eye, thus making up for the omission defects of traditional threshold monitoring.

[0077] The early warning layer employs a quantitative risk index grading mechanism, calculating the risk level according to the published comprehensive risk index formula, with weighting coefficients α=0.4, β=0.4, and γ=0.2, ensuring a weight sum of 1. It comprehensively considers three core indicators: abnormal error in model reconstruction, the magnitude of real-time parameter exceedances, and the degree of trend deterioration. The system executes tiered control logic: 0 < R ≤ 0.3 is a Level 1 warning, where the system only retains data and continuously tracks it; 0.3 < R ≤ 0.7 is a Level 2 warning, automatically generating maintenance work orders and pushing rectification suggestions such as component inspection, fastening checks, and settlement monitoring; R > 0.7 is a Level 3 warning, immediately triggering platform audible and visual alarms, sending pop-up notifications to administrators' mobile devices, locking the risk area, and suspending high-risk operations. Simultaneously, relying on twin model simulation backtracking, it compares sensor data waveforms, environmental data fluctuations, and structural simulation responses to accurately distinguish between equipment failures, instantaneous environmental interference, and structural damage, outputting the damaged component number, damage location coordinates, and damage quantification level.

[0078] This embodiment employs a lightweight model distillation scheme, constructing a high-precision cloud-based teacher model and a lightweight edge student model. The distillation weight α=0.7 is set, and model compression is achieved by combining hard label loss and soft label distribution loss. This results in a 75% reduction in model size and a 4x increase in inference speed, while retaining over 95% of simulation and recognition accuracy. Coupled with an edge-cloud collaborative architecture, the edge handles millisecond-level real-time inference, while the cloud handles high-precision simulation, model training, and lifetime extrapolation, addressing the pain point of traditional models' inability to balance accuracy and real-time performance. Simultaneously, the system utilizes an incremental learning mechanism, setting a historical decay coefficient η=0.3. This retains historical health characteristics while learning from new operating data, preventing model forgetting and overfitting, and adapting to complex operating conditions such as long-term building aging, seasonal temperature variations, and extreme winds, continuously improving the model's generalization ability.

[0079] The visualized operation and maintenance layer establishes a 3D interactive twin platform that renders the overall building structure, subdivided components, monitoring points, data curves, and early warning indicators in real time. It supports the reproduction of operating conditions at any time period, the tracing of risk causes, and the quantitative extrapolation of the remaining structural lifespan. Based on different risk levels, it automatically outputs customized operation and maintenance solutions such as reinforcement, component replacement, periodic monitoring, and load limiting, forming a closed-loop intelligent management system encompassing "perception-preprocessing-modeling-judgment-early warning-operation and maintenance." Long-term field testing in this project has verified that the system achieves a 99.2% accuracy rate in identifying explicit structural risks and a 95.7% accuracy rate in identifying early-stage latent micro-damage. The system's early warning response delay is ≤50ms. Compared to traditional fixed-threshold monitoring systems, the false alarm rate is reduced by 68%, and the missed alarm rate is reduced by 72%. It effectively achieves predictive safety management of building structures and possesses strong engineering applicability and industry promotion value.

[0080] In one embodiment, this application provides a building structure monitoring and early warning method combining digital twins and machine learning, including:

[0081] S1. Deploy multiple types of IoT sensors and laser scanning equipment to collect data on building structure deformation, stress, settlement, environmental parameters, appearance damage and load across the entire area, and construct the original monitoring dataset;

[0082] S2. Perform wavelet threshold denoising, anomaly removal, time-series alignment, missing data completion and multi-source heterogeneous data fusion on the original monitoring dataset, and at the same time complete lightweight model inference at the edge to reduce the computing load on the cloud.

[0083] S3. Construct a multi-scale dynamic twin model that integrates geometric parameters, physical properties, and structural dynamics mechanisms. Based on measured data and combined with the physical residual loss function, correct the model parameters in real time to achieve a 1:1 dynamic synchronous mapping between the physical building and the digital model.

[0084] S4. Combining the building safety code rule base with multiple machine learning models, and relying on the improved LSTM time series prediction formula and autoencoder anomaly measurement formula, we can realize the determination of structural explicit risks, the identification of hidden damage, the prediction of state trends, and trigger graded early warnings according to the risk level.

[0085] S5 provides a 3D visualization display of the building's real-time twin status, monitoring data, and early warning information, enabling risk tracing, operational condition simulation, lifespan projection, and outputting intelligent operation and maintenance decision-making suggestions.

[0086] The beneficial effects of the digital twin combined with machine learning in the building structure monitoring and early warning system implementation are not elaborated here.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring and early warning of building structures using digital twins combined with machine learning, characterized in that, include: S1. Deploy multiple types of IoT sensors and laser scanning equipment to collect data on building structure deformation, stress, settlement, environmental parameters, appearance damage and load across the entire area, and construct the original monitoring dataset; S2. Perform wavelet threshold denoising, anomaly removal, time-series alignment, missing data completion and multi-source heterogeneous data fusion on the original monitoring dataset, and at the same time complete lightweight model inference at the edge to reduce the computing load on the cloud. S3. Construct a multi-scale dynamic twin model that integrates geometric parameters, physical properties, and structural dynamics mechanisms. Based on measured data and combined with the physical residual loss function, correct the model parameters in real time to achieve a 1:1 dynamic synchronous mapping between the physical building and the digital model. S4. Combining the building safety code rule base with multiple machine learning models, and relying on the improved LSTM time series prediction formula and autoencoder anomaly measurement formula, it realizes the determination of structural explicit risks, identification of hidden damage, and prediction of state trends, and triggers graded early warnings according to the risk level. S5 provides a 3D visualization display of the building's real-time twin status, monitoring data, and early warning information, enabling risk tracing, operational condition simulation, lifespan projection, and outputting intelligent operation and maintenance decision-making suggestions.

2. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: Step S1 includes accelerometers, strain sensors, displacement sensors, settlement sensors, temperature and humidity sensors, wind pressure sensors, and UAV laser point cloud scanning equipment deployed at key building components; the sensors collect dynamic response data of the structure and environmental conditions in real time, and the laser point cloud scanning equipment periodically collects three-dimensional point cloud data of the building's exterior, realizing full-domain synchronous acquisition of the internal state and external form of the structure.

3. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: The specific processing methods for the data preprocessing and edge computing layer include: using a wavelet threshold denoising algorithm to remove random noise and impulse interference from the sensor. The wavelet threshold denoising formula is as follows: ； in, These are the original wavelet decomposition coefficients. These are the denoised wavelet coefficients. For adaptive threshold, It is a symbolic function; Missing data are filled in using linear interpolation and time-series fitting algorithms. Heterogeneous data with different sampling frequencies and dimensions are fused through time-series alignment and dimensionality normalization. The normalization formula is as follows: ； in, This is the raw monitoring data. For normalized data, These are the maximum and minimum values ​​of the data sequence, respectively. Lightweight early warning models are deployed on edge terminals to complete real-time data inference and preliminary anomaly screening, while complex model training and simulation calculations are uploaded to the cloud for execution.

4. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: In step S3, the model construction and dynamic update methods include: A1. Based on architectural BIM design drawings, construction records, and material mechanics parameters, build an initial geometric-physical coupled finite element twin model and assign material properties, constraints, and mechanical boundary conditions to the model components. A2. Obtain 3D data of building entities through UAV laser point cloud scanning, complete the accurate mapping of point cloud data with voxels, nodes and elements of finite element model, and correct geometric deformation deviations of model; A3. Using the structural dynamics equilibrium equations as physical constraints, construct a model loss function that includes physical residual terms, and iteratively optimize uncertain parameters such as model stiffness, damping, and elastic modulus by combining real-time monitoring data; the structural dynamics constraint equations satisfy: ； in, For the quality matrix, Here is the damping matrix. Here is the stiffness matrix. These are the acceleration, velocity, and displacement response vectors, respectively. This is the time-series load vector; Construct the physical residual loss function: ； in, The fitting loss between measured and simulated data, For the loss of dynamic physical residuals, These are the physical constraint weight coefficients; minimized Eliminate the error between simulation data and physical response data; A4. An incremental update mechanism is adopted to fine-tune the model parameters based on newly added monitoring data at regular intervals, so as to realize the dynamic adaptive update of the digital model of the entire building life cycle.

5. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: In step S4, a dual-drive judgment architecture of rule base + machine learning is adopted, including a bottom-level rule judgment module and an upper-level intelligent prediction and recognition module. The underlying rule analysis module presets deformation, settlement, and stress thresholds based on national building structure safety standards to quickly determine structural over-standard risks. The upper-layer intelligent prediction and recognition module includes an improved LSTM time-series prediction model and a deep autoencoder anomaly detection model. The improved LSTM model captures temporal features based on a gating mechanism, and the cell state update formula is as follows: ； in, These are the input gate, forget gate, and output gate, respectively. The cell state at time t. Output in hidden state It is the sigmoid activation function. The improved LSTM model is used to mine the temporal correlation features between structural loads, environmental parameters and structural state, thereby achieving 7-30 day structural performance trend prediction. The deep autoencoder unsupervised model is used to compare the deviation characteristics between the real-time structural state and the healthy baseline state, and to determine anomalies by reconstruction error. The anomaly measurement formula is as follows: ； in, Features of the health baseline sample To reconstruct sample features in real time, For feature dimension; when When the error exceeds the preset abnormal threshold, it is determined that there are early hidden micro-damage and slow performance degradation problems.

6. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: In step S4, a three-level early warning mechanism is configured, combined with a comprehensive risk index. To achieve tiered and quantitative risk management, the risk index is calculated using the following formula: ； in, To reconstruct abnormal errors, This is the threshold for abnormal errors. This is the current structural parameter value. To standardize safety thresholds, The trend deterioration coefficient, The weighting coefficients are and satisfy the following conditions: ; The criteria for determining the warning level are as follows: Level 1 Warning: Only data recording and continuous monitoring are performed; Level 2 warning: Automatically pushes maintenance and rectification reminders; Level 3 early warning: This triggers the platform's audible and visual alarms, emergency push notifications on mobile devices, and risk locking. Meanwhile, relying on the dynamic simulation and deduction function of the twin model, the abnormality causes can be traced in reverse, effectively distinguishing between sensor failure, environmental interference and structural damage, accurately locating the damage coordinates and quantifying the degree of damage.

7. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: The method also includes using a distillation lightweight algorithm to construct a teacher-student model structure, with the distillation loss function being: ； in, For the student model and the loss of the true label, Output probability distribution loss for both the student and teacher models. This is the distillation weighting coefficient; The distillation algorithm is used to compress parameters and simplify the structure of high-precision twin models and machine learning models, significantly reducing model size and computational load while retaining core feature parameters and inference accuracy. An edge-cloud collaborative computing architecture is built, with the edge responsible for real-time data preprocessing and lightweight model inference, and the cloud responsible for high-precision simulation calculation and model iterative training, effectively balancing monitoring accuracy and millisecond-level early warning response speed.

8. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: The method also includes employing an online incremental update strategy and defining an incremental loss function: ； in, The model fitting loss is for the newly added monitoring data. For historical sample memory loss, This is the historical weight decay coefficient; Based on the incremental loss function, the parameters of the machine learning model and twin model can be continuously iteratively optimized based on the newly added monitoring data accumulated over a long period of time. The historical health status characteristics of the structure are retained, and the system can be adaptively adapted to complex working conditions such as aging of building materials, dynamic changes in load, and environmental fluctuations, which significantly improves the generalization and operational stability of the system's long-term monitoring.

9. The building structure monitoring and early warning method combining digital twins and machine learning according to claim 1, characterized in that: In step S5, a three-dimensional interactive twin visualization platform is constructed to render and display the building structure's operating status, monitoring data change curves, damage distribution locations, and early warning indicators in real time. It supports functions such as historical condition reproduction, risk cause tracing, and structural remaining life estimation. It can automatically generate customized operation and maintenance decision suggestions such as inspection, reinforcement, and rectification based on different risk levels, forming a complete closed-loop management system of monitoring-analysis-early warning-operation and maintenance.

10. A building structure monitoring and early warning system combining digital twins and machine learning, characterized in that, include: The data acquisition layer is used to deploy various types of IoT sensors and laser scanning equipment to collect data on building structure deformation, stress, settlement, environmental parameters, appearance damage and load across the entire domain, and to build the original monitoring dataset. The data preprocessing and edge computing layer is used to perform wavelet threshold denoising, anomaly removal, time-series alignment, missing data completion and multi-source heterogeneous data fusion on the raw monitoring data, while completing lightweight model inference at the edge and reducing the computing load on the cloud. The digital twin modeling layer is used to construct a multi-scale dynamic twin model that integrates geometric parameters, physical properties, and structural dynamics mechanisms. Based on measured data and combined with the physical residual loss function, the model parameters are corrected in real time to achieve a 1:1 dynamic synchronous mapping between the physical building and the digital model. The machine learning intelligent judgment and early warning layer is used to combine the building safety code rule base with multiple machine learning models, relying on the improved LSTM time series prediction formula and autoencoder anomaly measurement formula to realize the judgment of explicit structural risks, the identification of hidden damage, the prediction of state trends, and to trigger graded early warnings according to the risk level. The visualization operation and maintenance and decision output layer is used to display the real-time twin status of the building, monitoring data and early warning information in three dimensions, complete risk tracing, working condition simulation, lifespan extrapolation, and output intelligent operation and maintenance decision suggestions.

Images