High-rise building deformation early warning method and system based on time sequence deep learning

Abstract

The invention relates to the technical field of building deformation early warning, in particular to a high-rise building deformation early warning method and system based on time sequence deep learning, and the method comprises the steps: obtaining multi-source heterogeneous monitoring data, carrying out the timestamp calibration, and sequentially carrying out the data cleaning and feature extraction and fusion operation based on the calibrated monitoring data, obtaining a standardized time sequence feature vector; and constructing an LSTM early warning model composed of a mechanism constraint network and an improved long short-term memory network by taking the time sequence feature vector as an input, correcting a prediction result, and performing hierarchical judgment according to the corrected deformation prediction result in combination with a preset multi-stage dynamic early warning threshold, according to the invention, through multi-source data acquisition and fusion, LSTM hybrid drive modeling, dynamic grading early warning and low-delay pushing, accurate prediction and early safety early warning of building deformation are realized, the safety and durability of a building structure are improved, and the risk of safety accidents is reduced.

Description

Technical Field

[0001] This invention relates to the field of building deformation early warning technology, and in particular to a method and system for early warning of deformation of high-rise buildings based on temporal deep learning. Background Technology

[0002] With the rapid advancement of urbanization, the number and height of super high-rise buildings continue to rise. Their structural systems are complex and significantly affected by environmental factors. The coupled effects of multiple factors such as temperature, wind load, rainfall, and foundation seepage can easily cause structural deformation. If timely monitoring and early warning are not provided, it will seriously threaten the structural safety of the building and even cause major safety accidents such as collapse. Therefore, accurate monitoring and early warning of deformation of high-rise buildings have become the core key to ensuring the structural safety and durability of buildings.

[0003] Current technologies for monitoring and early warning of deformation in high-rise buildings still suffer from numerous technical shortcomings, failing to meet the practical needs of super high-rise buildings for comprehensive, high-precision, and highly adaptable monitoring and early warning. Firstly, data acquisition dimensions are limited and environmental correlation is insufficient. Traditional monitoring often relies on single satellite positioning data, ignoring the coupled effects of environmental factors such as temperature, wind speed, and rainfall on building deformation. Monitoring data cannot comprehensively reflect the causes of deformation, easily leading to misjudgments or omissions. Secondly, data processing accuracy is limited and spatiotemporal synchronization is poor. Some monitoring methods do not employ precise positioning calculation techniques; issues such as orbital errors and ionospheric delays make it difficult to achieve millimeter-level positioning accuracy. Furthermore, the adaptability and response efficiency of existing early warning models and mechanisms urgently need improvement. Traditional early warning systems often use linear models, unable to handle the temporal correlation and nonlinear characteristics of deformation data, making early and accurate warnings difficult. Moreover, the use of uniform, fixed standards for early warning thresholds without considering dynamic thresholds designed for different deformation causes easily leads to over-warning or delayed warnings, failing to meet the practical engineering needs of deformation monitoring and early warning for super high-rise buildings. Therefore, a high-rise building deformation early warning method and system based on temporal deep learning is currently needed. Summary of the Invention

[0004] To address the problems of poor adaptability and low prediction accuracy in existing high-rise building deformation monitoring and early warning technologies, this invention provides a high-rise building deformation early warning method and system based on temporal deep learning.

[0005] In a first aspect, the present invention provides a method for early warning of deformation of high-rise buildings based on temporal deep learning, which adopts the following technical solution: A method for early warning of deformation in high-rise buildings based on temporal deep learning includes: Acquire multi-source heterogeneous monitoring data and perform timestamp calibration. The multi-source heterogeneous monitoring data includes BDS positioning data, accelerometer data, temperature data, wind speed data, rainfall data, and water level data. Based on the calibrated monitoring data, data cleaning, feature extraction and fusion operations are performed sequentially to obtain a standardized time-series feature vector; An LSTM early warning model consisting of a mechanism constraint network and an improved long short-term memory network is constructed using temporal feature vectors as input, and the prediction results are corrected using the structural dynamic stiffness matrix method. The mechanism network is stripped of temperature-related deformation to obtain residual sequences, and the residual sequences are trained and deformation is predicted using an improved long short-term memory network. Based on the corrected deformation prediction results, and combined with the preset multi-level dynamic early warning thresholds, a graded judgment is made, and safety early warning information is output through visualization and graded push mechanism.

[0006] Furthermore, the data cleaning includes cleaning the BDS positioning data using an RTK / PPK fusion algorithm, specifically as follows: For static measurement points, a precise single-point positioning post-processing mode is adopted. The precise single-point positioning post-processing mode performs post-processing on static observation data based on post-precise ephemeris to obtain high-precision positioning results. For dynamic measuring points, a real-time dynamic differential mode is adopted. The real-time dynamic differential mode is based on the real-time transmitted observation data and the base station data to perform real-time calculation to obtain dynamic positioning results. Satellite orbit errors are corrected based on precise ephemeris data. Ionospheric delay errors are eliminated by linear combination of dual-frequency observations, and gross errors are removed by sliding window filtering combined with statistical criteria.

[0007] Furthermore, the RTK / PPK fusion algorithm is based on a dual-differential carrier phase observation model. This model includes constructing dual-differential observation equations, performing dual-differential operations on carrier phase observations between the rover and the base station, and between the target satellite and the reference satellite, to eliminate satellite clock errors and receiver clock errors. The model then performs integer ambiguity fixing on the carrier phase observations after dual-differential processing, converting the floating ambiguity solution into a fixed integer solution. The dual-differential observation equations are as follows: , in, These are dual differential carrier phase observations. For the carrier phase observations of the target satellite by the rover station, For the carrier phase observations of the target satellite from the reference station, For the rover's carrier phase observations of the reference satellite, The carrier phase observations of the reference satellite from the base station.

[0008] Furthermore, the data clarification also includes denoising the accelerometer data using discrete wavelet transform, performing multi-layer wavelet decomposition using a preset wavelet basis to filter out high-frequency noise, extracting the vibration dominant frequency and damping ratio of the building structure, and when the vibration dominant frequency change rate exceeds a preset threshold, it is included as a new feature in subsequent feature extraction. The temperature data is decomposed into daily temperature difference and hourly temperature; the wind speed data is calculated into average wind speed and maximum gust wind speed; the rainfall data is cumulatively summed over multiple time scales; and the water level data is processed by moving average to eliminate high-frequency fluctuations and extract trend terms.

[0009] Further, the feature extraction and fusion includes constructing an initial feature vector containing displacement increment, displacement rate, cumulative rainfall at multiple time scales, daily temperature difference, and average wind speed; calculating the correlation coefficient between each feature and the deformation target variable using Spearman correlation analysis; screening target features whose absolute correlation coefficient is greater than a preset correlation threshold; determining multicollinearity when the correlation coefficients among multiple features in the target feature set are greater than the preset correlation threshold; performing dimensionality reduction using principal component analysis; retaining principal components whose cumulative variance contribution rate is greater than a preset contribution rate threshold; and calculating a comprehensive feature value as a standardized time-series feature vector by weighting the variance contribution rates of each principal component. The formula for calculating the comprehensive feature value is as follows: , in, For comprehensive eigenvalues, For the i-th principal component, Let be the i-th eigenvalue, and m be the number of principal components retained.

[0010] Furthermore, the construction of the LSTM early warning model, composed of a mechanistic constraint network and an improved long short-term memory network, includes establishing a temperature-deformation linear model based on the theory of thermal expansion of materials, calculating temperature-related deformation through the temperature-deformation linear model, and using the difference between the actual monitored deformation and the temperature-related deformation as a residual sequence. The temperature-deformation linear model is as follows: , in, For temperature-related vertical deformation, is the coefficient of linear expansion of the material. For daily temperature difference, For constant terms; Furthermore, an LSTM early warning model composed of a mechanistic constraint network and an improved long short-term memory network is constructed. This model also includes a residual sequence input mechanism, using the residual sequence as the network input sample to train the network's sensitivity to nonlinear deformation patterns. A two-layer stacked structure is employed to capture long-term temporal dependencies, and a regularization strategy is introduced into the network to prevent overfitting. Finally, a dynamic inertia weight adjustment mechanism is introduced during the forget gate state update process, adaptively adjusting the retention degree of historical states through dynamic adjustment coefficients. The dynamic update formula for the forget gate is: , in, Output the forget gate at the current moment. It is the sigmoid activation function. This is the weight matrix. This is the hidden state from the previous moment. For the current input features, For bias terms, This is the dynamic inertia weight adjustment coefficient.

[0011] Furthermore, the method of correcting the prediction results using the structural dynamics stiffness matrix includes triggering a catastrophic correction model when the monitoring data meets the extreme working condition judgment conditions, constructing the overall stiffness matrix of the building structure, converting extreme wind loads and rainfall loads into equivalent additional load vectors, solving the additional displacement response of the structure under extreme working conditions based on the overall stiffness matrix and the equivalent additional load vectors, and vector superimposing the additional displacement response with the prediction results of the improved long short-term memory network to obtain the corrected final deformation prediction value.

[0012] Furthermore, the method of correcting the prediction results using the structural dynamics stiffness matrix includes triggering a catastrophic correction model when the monitoring data meets the extreme working condition judgment conditions, constructing the overall stiffness matrix of the building structure, converting extreme wind loads and rainfall loads into equivalent additional load vectors, solving the additional displacement response of the structure under extreme working conditions based on the overall stiffness matrix and the equivalent additional load vectors, and vector superimposing the additional displacement response with the prediction results of the improved long short-term memory network to obtain the corrected final deformation prediction value.

[0013] Furthermore, the multi-level dynamic early warning thresholds include foundation deformation thresholds, temperature-sensitive early warning thresholds, wind load anomaly early warning thresholds, and foundation seepage early warning thresholds, specifically: Basic thresholds: cumulative horizontal displacement ≥30mm, cumulative vertical displacement ≥20mm, deformation rate ≥3mm / d; Temperature-sensitive warning threshold: ΔT > 10℃ / 6h and vertical displacement rate > 1.5mm / h; Wind load anomaly warning threshold: wind speed > 15 m / s and horizontal displacement rate > 2 mm / h, while BDS multipath effect > 1 m; Foundation seepage warning threshold: rainfall > 100 mm / 24 h and foundation settlement rate > 1 mm / h.

[0014] Secondly, a high-rise building deformation early warning system based on temporal deep learning includes: The multi-source data acquisition module is configured to acquire multi-source heterogeneous monitoring data and perform timestamp calibration. The multi-source heterogeneous monitoring data includes BDS positioning data, accelerometer data, temperature data, wind speed data, rainfall data, and water level data. The feature fusion module is configured to perform data cleaning, feature extraction and fusion operations sequentially based on the calibrated monitoring data to obtain a standardized time-series feature vector. The deformation prediction and correction module is configured to: construct an LSTM early warning model consisting of a mechanism constraint network and an improved long short-term memory network using the temporal feature vector as input, and correct the prediction results using the structural dynamic stiffness matrix method; The mechanism network is stripped of temperature-related deformation to obtain residual sequences, and the residual sequences are trained and deformation is predicted using an improved long short-term memory network. The early warning information output module is configured to: make a graded judgment based on the corrected deformation prediction results and the preset multi-level dynamic early warning thresholds, and output safety early warning information through visualization and graded push mechanism.

[0015] In summary, the present invention has the following beneficial technical effects: 1. This invention achieves millimeter-level positioning accuracy and effective filtering of high-frequency noise by performing timestamp calibration, RTK / PPK fusion calculation, dual differential carrier phase observation model error elimination, and wavelet transform denoising on multi-source heterogeneous monitoring data. It overcomes the technical defects of traditional monitoring methods, such as limited data processing accuracy and weak error control capability, and provides a high-precision data foundation for subsequent deformation prediction.

[0016] 2. This invention isolates temperature-related deformation by constructing a mechanism constraint network containing a linear temperature-deformation model, and captures nonlinear deformation patterns by training an improved long short-term memory network based on residual sequences. This achieves the organic integration of physical mechanisms and deep learning, overcomes the technical defects of traditional early warning models such as poor adaptability and insufficient prediction accuracy, and improves the model's ability to characterize complex deformation patterns.

[0017] 3. This invention introduces a dynamic inertial weight adjustment mechanism during the forget gate state update process, thereby achieving adaptive adjustment of the sensitivity to sequence state changes. This overcomes the technical deficiency of traditional long short-term memory networks in terms of insufficient dynamic nonlinear response capability and enhances the model's adaptability to data mutations.

[0018] 4. This invention triggers a catastrophic correction model under extreme working conditions, calculates additional deformation using the structural dynamic stiffness matrix method, and superimposes the results of neural network prediction to correct it. This achieves the prediction accuracy guarantee under extreme load conditions, overcomes the technical defects of traditional models in predicting inaccurate results under extreme working conditions such as strong storms and heavy rain, and expands the working condition adaptability range of the early warning method.

[0019] 5. This invention achieves differentiated responses to different deformation causes by using a multi-level dynamic early warning threshold system that presets a basic deformation threshold, a temperature-sensitive early warning threshold, a wind load anomaly early warning threshold, and a foundation seepage early warning threshold. This overcomes the technical defects of traditional fixed early warning thresholds, which are prone to over-warning or delayed warnings, and improves the accuracy and timeliness of early warnings. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the steps of the intelligent early warning method for multi-source monitoring deformation of high-rise buildings that integrates mechanistic constraints and temporal deep learning, according to an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating the workflow of the intelligent early warning method for multi-source monitoring of deformation in high-rise buildings, which integrates mechanistic constraints and temporal deep learning, according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the improved LSTM network structure based on residual input and the early warning push workflow according to an embodiment of the present invention.

[0023] Figure 4 This is a timing diagram of relative displacement in each direction according to an embodiment of the present invention.

[0024] Figure 5 These are time-series data diagrams of an embodiment of the present invention, wherein (a) is a time-series diagram of temperature data changes, (b) is a time-series diagram of wind speed data changes, (c) is a time-series diagram of rainfall data changes, and (d) is a time-series diagram of groundwater level changes. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings. Example

[0026] Reference Figure 1 This embodiment of a high-rise building deformation early warning method based on temporal deep learning includes: Acquire multi-source heterogeneous monitoring data and perform timestamp calibration. The multi-source heterogeneous monitoring data includes BDS positioning data, accelerometer data, temperature data, wind speed data, rainfall data, and water level data. Based on the calibrated monitoring data, data cleaning, feature extraction and fusion operations are performed sequentially to obtain a standardized time-series feature vector; An LSTM early warning model consisting of a mechanism constraint network and an improved long short-term memory network is constructed using temporal feature vectors as input, and the prediction results are corrected using the structural dynamic stiffness matrix method. The mechanism network is stripped of temperature-related deformation to obtain residual sequences, and the residual sequences are trained and deformation is predicted using an improved long short-term memory network. Based on the corrected deformation prediction results, and combined with the preset multi-level dynamic early warning thresholds, a graded judgment is made, and safety early warning information is output through visualization and graded push mechanism.

[0027] Specifically, a method for early warning of deformation in high-rise buildings based on temporal deep learning includes the following steps: like Figure 1 , Figure 2 As shown, S1 involves acquiring multi-dimensional monitoring data that comprehensively reflects the influencing factors of high-rise building deformation. The collected multi-source heterogeneous monitoring data includes BDS positioning data, accelerometer data, temperature data, wind speed data, rainfall data, and water level data. The acquisition methods for each data are as follows: BDS Positioning Data: The BDS positioning data is acquired with millimeter-level accuracy through a ground-based augmentation system. BDS reference stations are deployed in areas with stable geological conditions around the building. These stations are equipped with choke coil antennas to suppress multipath effects and are connected to high-precision receivers to continuously track BeiDou satellite constellation signals. A network of miniature receivers is deployed on the building roof and in the core area. These miniature receivers employ a low-power design and integrate MEMS inertial measurement units, acquiring three-dimensional coordinate information through carrier phase differential positioning. The BDS positioning data uses a dual-frequency observation mode, simultaneously receiving signals at either B1I / B3I or B1C / B2a frequencies to eliminate first-order ionospheric delay errors through dual-frequency combination. Accelerometer data: acquired through structural dynamic response monitoring. Triaxial force-balanced accelerometers are deployed at the bottom, middle, and top key floors of the building's core tube. These accelerometers have a sensitivity of at least 1 μg, a dynamic range covering ±2g, and a sampling frequency of 100 Hz to capture the structural acceleration response under wind and seismic loads. The accelerometers are rigidly connected to the shear wall surface of the core tube via bolts to ensure coordinated vibration between the sensor and the structure, avoiding local modal coupling distortion. Temperature data: The data is collected using a platinum resistance temperature sensor. The platinum resistance sensor has the advantages of high measurement accuracy and good stability, and can accurately obtain changes in ambient temperature. Platinum resistance temperature sensors are deployed on the roof sunshade, the middle standard floor and around the reference station. The sensor adopts a four-wire connection method to eliminate the influence of lead resistance, and the temperature measurement accuracy is better than ±0.1℃, with a time constant of less than 60 seconds, so as to accurately track the driving effect of sudden temperature changes on the thermal expansion and contraction of the structure. Wind speed data: Measured using the ultrasonic time-of-flight method. Ultrasonic anemometers are deployed at different heights on the building roof and windward side. Each anemometer uses three pairs of orthogonally arranged ultrasonic probes. The wind speed vector is calculated by measuring the time difference between propagation with and against the wind. The measurement range covers 0–60 m / s, with a resolution of 0.01 m / s, and can acquire average wind speed, maximum gust wind speed, and wind direction information. Rainfall data: The data is collected using tipping bucket rain gauges. The tipping bucket rain gauges are deployed in open areas around the building. The rain gauges have a rain collection diameter of 200mm. Each tipping bucket rotation corresponds to 0.1mm or 0.2mm of rainfall. Automatic cumulative measurement is achieved by generating pulse signals through a magnetic switch. The measurement range covers 0 to 4mm / min to monitor the impact of rainfall intensity and duration on foundation seepage. Water level data: acquired through the principle of hydrostatic level measurement. Submersible water level sensors are deployed at different locations around the building foundation pit. These sensors convert hydrostatic pressure into electrical signals based on the piezoresistive effect, with a measurement range of 0–10 m and an accuracy of ±0.1%FS. By deploying multiple points and taking the arithmetic mean, local hydrogeological differences are eliminated to track foundation settlement caused by groundwater level fluctuations.

[0028] To ensure the continuity and reliability of monitoring data, this embodiment employs a hardware redundancy and cross-validation strategy. Two sets of temperature sensors, two sets of wind speed sensors, and two sets of rainfall sensors are deployed within a 10m radius of the building. The redundant sensors are covered with radiation-proof Stevenson screens to avoid the influence of solar radiation heat, and data cross-validation is achieved through a 5G wireless transmission link. An anomaly alarm is triggered when the deviation between the observation values ​​of the main sensor and the redundant sensor exceeds a preset threshold. Three water level sensors are deployed at 120° angles around the building's foundation pit, and the arithmetic mean of the monitoring data from these three points during the same period is taken as the final water level monitoring value to eliminate the interference of single-point anomalies on the overall assessment.

[0029] Due to inherent differences between the sampling frequencies of various monitoring data and the clock reference, this embodiment employs a BDS timing module and a local high-precision clock drift compensation mechanism for timestamp calibration. The BDS timing module receives the standard time signal broadcast by the BeiDou Navigation Satellite System and outputs a pulse signal with an accuracy better than 20 nanoseconds as a unified time trigger source. The local high-precision clock uses a temperature-compensated crystal oscillator with a frequency stability better than ±0.1ppm. It tracks the BDS timing signal through a phase-locked loop circuit and maintains the time reference based on its own timekeeping capability when the BDS signal is lost, with a timekeeping accuracy better than 1 millisecond / 24 hours. Various sensor data acquisition units integrate the BDS timing module, triggering the sampling time marker with the rising edge of the pulse signal to ensure that BDS positioning data, accelerometer data, and environmental monitoring data have a unified time reference.

[0030] To address the difference between the second-level sampling frequency of BDS positioning data and the minute-level sampling frequency of environmental monitoring data, this embodiment employs forward interpolation for time series alignment. The forward interpolation method establishes a linear trend prediction model based on historical observation data, extrapolates the trend of observations at missing times, and then generates an equally spaced fused dataset through cubic spline interpolation. The cubic spline interpolation formula is as follows: , in, Let be the interpolation function, and t be the interpolation time. and For adjacent sampling times, , , and These are the interpolation coefficients determined by the function values ​​and first derivative continuity conditions of adjacent data points. Through the aforementioned timestamp calibration and sampling frequency alignment processes, a fused time-series dataset with equal intervals of 1 minute is finally generated.

[0031] S2. Based on the fused time series dataset with a unified timestamp generated in step S1, perform data cleaning, targeted preprocessing, and feature extraction and fusion operations in sequence to eliminate observation noise, gross errors, and information redundancy. The data cleaning process employs an RTK / PPK fusion solution mode for BDS positioning data. This RTK / PPK fusion solution is based on a dual-differential carrier phase observation model to achieve precise positioning. The dual-differential carrier phase observation model performs dual-differential operations on carrier phase observations between the rover and the base station, and between the target satellite and the reference satellite, constructing dual-differential observation equations to eliminate common error terms from satellite clock errors, receiver clock errors, and atmospheric delays. The dual-differential observation equations are as follows: , in, These are dual differential carrier phase observations. For the carrier phase observations of the target satellite by the rover station, For the carrier phase observations of the target satellite from the reference station, For the rover's carrier phase observations of the reference satellite, For the carrier phase observations of the reference satellite from the base station; For static measurement points, a precise single-point positioning post-processing mode is adopted. This mode performs post-processing on static observation data based on post-precise ephemeris to obtain high-precision positioning results. For dynamic measurement points, a real-time dynamic differential mode is adopted. This mode performs real-time calculation based on real-time transmitted observation data and base station data to obtain dynamic positioning results. To achieve millimeter-level positioning accuracy, the key lies in the integer solution of carrier phase ambiguity. This embodiment adopts an improved integer ambiguity fixing algorithm. This algorithm combines historical time series constraints with conventional RTK to improve the stability and convergence speed of integer solution solving. The fixed integer ambiguity can improve the positioning solution from a floating solution to a fixed solution, significantly improving positioning accuracy. The floating ambiguity solution is converted into an integer fixed solution. The carrier phase observation equation is: , in, For carrier phase observations, The geometric distance to the station star. For carrier phase integer ambiguity, For carrier wavelength, As residual errors, the fixed integer ambiguities can improve the positioning solution from a floating solution to a fixed solution, significantly improving positioning accuracy. Furthermore, satellite orbit errors are corrected based on IGS precise ephemeris data, ionospheric delay errors are eliminated by linear combination of dual-frequency observations, and sliding window filtering is used in conjunction with... The criteria for eliminating gross errors, the aforementioned The exclusion criterion is that the absolute value of the difference between the observed value and the mean of the data in the window is greater than 3 times the standard deviation.

[0032] The data cleaning process employs a discrete wavelet transform algorithm to denoise the accelerometer data. A db4 wavelet basis is used for three-level wavelet decomposition to filter out high-frequency noise. Based on the denoised acceleration signal, the vibration dominant frequency and damping ratio of the building structure are extracted using Hilbert transform, and the vibration dominant frequency is included in the initial feature vector. When the rate of change of the vibration dominant frequency exceeds a preset threshold, it indicates that the structure may be damaged. At this time, the feature monitoring process is triggered, and the rate of change of the vibration dominant frequency is added as a new feature and included in subsequent feature extraction.

[0033] The data cleaning process employs differentiated preprocessing strategies for environmental monitoring data to extract trend terms: Temperature data is decomposed into daily and hourly temperature differences, where the daily temperature difference is the difference between the highest and lowest temperatures of the day, highlighting the driving effect of temperature changes on structural thermal expansion and contraction; Wind speed data is calculated to determine average wind speed and maximum gust speed, where the average wind speed reflects the continuous effect of wind load, and the maximum gust speed reflects the instantaneous impact of wind load; Rainfall data is subjected to multi-timescale cumulative sliding summation, calculating 1-hour, 12-hour, and 24-hour cumulative rainfall to reflect the intensity and duration of rainfall; Water level data is processed using a moving average to eliminate high-frequency fluctuations caused by pumping and tides and to highlight trend changes in water level.

[0034] The feature extraction and fusion are based on preprocessed multi-source data to construct an initial feature vector, which includes displacement increment, displacement rate, cumulative rainfall at multiple time scales, daily temperature difference, and average wind speed. To improve the effectiveness of feature selection and the robustness of model prediction, a progressive feature fusion and weight adjustment mechanism is adopted: First, Spearman rank correlation analysis is used to calculate the correlation coefficient between each feature and the deformation target variable, and target features with an absolute correlation coefficient greater than a preset correlation threshold are selected. For cases where the correlation of some features fluctuates at different time periods, adaptive window correlation analysis is introduced to evaluate the dynamic correlation of features at different time periods, so as to avoid misjudgment caused by a single static correlation. For the feature set after Spearman selection, the relative contribution value of each feature to the target prediction is calculated by combining information gain and variance contribution, and the relative contribution value is mapped to the feature weight coefficient to improve the influence of high-contribution features in the model input. When the correlation coefficients among multiple features in the target feature are greater than a preset correlation threshold, multicollinearity is determined to exist. Principal component analysis is used for dimensionality reduction, retaining principal components whose cumulative variance contribution rate is greater than a preset contribution rate threshold. The comprehensive feature value is calculated as a standardized time series feature vector by weighting the variance contribution rates of each principal component. The formula for calculating the comprehensive feature value is as follows: , in, For comprehensive eigenvalues, For the i-th principal component, Let be the i-th feature value, and m be the number of principal components retained. The resulting standardized time-series feature vector will be used as the input feature of the LSTM early warning model, thereby effectively integrating multi-source information from time-series data and improving the model's nonlinear pattern recognition capability.

[0035] like Figure 3As shown in Figure S3, based on standardized time-series feature vectors, a fusion model combining a mechanistic model and an improved long short-term memory neural network is used for deformation prediction. A correction mechanism is designed for extreme working conditions. The mechanistic model is constructed based on the theory of thermal expansion of materials to establish a temperature-deformation linear model, quantifying the influence of temperature on vertical deformation. The temperature-deformation linear model is as follows: , in, For temperature-related vertical deformation, The coefficient of linear expansion is the material's linear expansion coefficient, which is determined based on the material properties of the main building structure. For concrete structures, it is taken as... steel structure , For daily temperature difference, The constant term is determined by least-squares fitting of historical monitoring data. Temperature-related deformation is calculated using the temperature-deformation linear model, and the difference between the actual monitored deformation and the temperature-related deformation is used as a residual sequence. The residual sequence contains nonlinear deformation information caused by wind load, rainfall, and foundation settlement.

[0036] The improved Long Short-Term Memory (LSTM) neural network is designed for time-series prediction of residual sequences. It employs a residual sequence input mechanism, using the residual sequence, stripped of temperature influence, as the primary input. This allows the neural network to focus on modeling nonlinear factors, enhancing its nonlinear pattern recognition capabilities. The improved LSM uses a two-layer stacked structure, with each layer containing 64 neurons and an activation function of tanh. The network input is an 8-dimensional temporal feature vector representing 12 historical periods, and the output is the predicted displacement value for the next 3 periods. This two-layer stacked structure allows the second layer to further abstract the output of the first layer, capturing dependencies across longer time spans and enhancing its ability to learn long-term trends. To improve the model's response to dynamic nonlinearities, a dynamic inertia weight adjustment mechanism is introduced during the forget gate state update process. The dynamic update formula for the forget gate is: , in, Output the forget gate at the current moment. It is the sigmoid activation function. This is the weight matrix. This is the hidden state from the previous moment. For the current input features, For bias terms, This refers to the dynamic inertia weight adjustment coefficient, which is automatically adjusted during training based on historical error trends, increasing as the error increases. To enhance the preservation of historical states, the error is reduced as the error decreases. To accelerate new state learning and enhance the sensitivity of the forgetting gate to changes in sequence states, a Dropout layer and an early stopping mechanism are introduced into the network to prevent overfitting. The Dropout layer has a dropout rate of 0.2, randomly disabling 20% ​​of neuron connections. The early stopping mechanism has a patience value of 10, automatically stopping training when the root mean square error (RMSE) of the validation set fails to decrease for 10 consecutive iterations. The model training uses the Adam optimizer with an initial learning rate of 0.001, combined with an adaptive learning rate decay mechanism to dynamically adjust the learning rate based on the training error. The maximum number of iterations is set to 500, and the upper limit for convergence is set by controlling the training error to a RMSE less than 2 mm. To balance large error suppression with overall fitting stability, a joint evaluation index of RMSE and mean absolute error (MAO) is used as the loss function. The loss function is: , in, The total loss value, The root mean square error, The mean absolute error, This is the overall weighting coefficient, with a value of 0.6.

[0037] The extreme condition correction addresses the additional deformation caused by extreme wind loads and heavy rainfall. The criteria for determining extreme conditions are a wind speed of 24.5 m / s or a daily rainfall exceeding 50 mm. When monitoring data meets these criteria, the system automatically switches to the catastrophic correction model. This catastrophic correction model uses the structural dynamics stiffness matrix method to calculate the additional deformation response under extreme loads. Specifically, it involves discretizing the building into node and component elements and establishing an overall stiffness matrix. Satisfying the static equilibrium equations ,in, Let be the nodal displacement vector. This represents the nodal load vector; the monitored wind load and rainfall conditions are converted into equivalent additional load vectors. , The equivalent component of wind load. The equivalent components of the rain load are used; the additional displacement response is solved using the global stiffness matrix. The additional displacement response is vector-superimposed with the original prediction result of the improved long short-term memory neural network to obtain the corrected deformation prediction value. The correction formula is as follows: , in, The corrected displacement prediction value. The original prediction values ​​for the improved Long Short-Term Memory Neural Network. By adding displacement to catastrophic conditions, and through the above-mentioned mechanism-LSTM fusion modeling and extreme condition correction, the unity of physical interpretability and nonlinear modeling capability is achieved, as well as the prediction accuracy under extreme load conditions is guaranteed.

[0038] Step S4: Multi-level dynamic early warning and information output Based on the corrected prediction results of the LSTM early warning model, combined with preset multi-level dynamic early warning thresholds, the early warning level is determined, and the safety early warning information is output through visualization and tiered push notifications, as follows: Considering that different influencing factors cause different levels of deformation risk, this invention sets multi-level dynamic early warning thresholds, covering four typical scenarios: foundation deformation, temperature-sensitive deformation, abnormal wind load deformation, and foundation seepage deformation. The specific thresholds are as follows: Basic thresholds: cumulative horizontal displacement ≥30mm, cumulative vertical displacement ≥20mm, deformation rate ≥3mm / d; these thresholds are set by statistical analysis of historical monitoring data and fitting of structural response regression model, and are applicable to basic deformation early warning under normal working conditions, reflecting the overall deformation state of the structure. Temperature-sensitive warning threshold: ΔT>10℃ / 6h and vertical displacement rate>1.5mm / h; this threshold is for structural deformation caused by sudden temperature changes, which is common in seasonal transitions or scenarios with large diurnal temperature differences. Wind load anomaly warning threshold: wind speed > 15 m / s and horizontal displacement rate > 2 mm / h, while BDS multipath effect > 1 m; this threshold is for structural horizontal deformation caused by strong wind load and needs to be comprehensively judged in combination with the reliability of BDS positioning data (multipath effect). Foundation seepage warning threshold: rainfall > 100 mm / 24 h and foundation settlement rate > 1 mm / h; this threshold is for foundation seepage and settlement caused by heavy rainfall and is a key early warning indicator for preventing foundation instability.

[0039] To ensure that early warning information is accurately delivered to the relevant management entities and that efficient handling is achieved, a tiered push mechanism is adopted based on the different thresholds that trigger the early warning levels: Warning messages triggered by basic thresholds: pushed to the project's technical manager and construction management personnel, and promptly verified and handled by the on-site management team; Warning information triggered by temperature-sensitive warning thresholds: pushed to the project technical manager, construction management personnel and liaison officer of the supervision unit, so that the construction party and the supervision party can jointly verify the structural temperature deformation; Warning information triggered by the abnormal wind load warning threshold: pushed to the project management team, supervision unit, and liaison officer of the local housing and construction department, requiring the intervention and supervision of the industry competent authority; Warning information triggered by the foundation seepage warning threshold: pushed to the project management team, supervision unit, liaison officer of local housing and construction department and local emergency management department. It is a high-risk warning and requires the emergency management department to coordinate and carry out emergency response work.

[0040] like Figure 4 , Figure 5 As shown, the pushed early warning information must include key information, specifically the location of deformation monitoring points, predicted displacement values, warning levels, and disposal suggestions. These suggestions are pre-set based on different warning scenarios; for example, a foundation threshold warning suggests "verify the accuracy of monitoring equipment and strengthen monitoring of key areas," while a foundation seepage warning suggests "immediately stop on-site construction and activate the foundation reinforcement emergency plan," providing clear guidance for disposal work. Simultaneously, a visualization platform displays monitoring data, deformation trend curves, and warning levels, allowing management to intuitively grasp the structural safety status. Example

[0041] The difference between this embodiment and Embodiment 1 is that this embodiment provides a high-rise building deformation early warning system based on temporal deep learning, including: The multi-source data acquisition module is configured to acquire multi-source heterogeneous monitoring data and perform timestamp calibration. The multi-source heterogeneous monitoring data includes BDS positioning data, accelerometer data, temperature data, wind speed data, rainfall data, and water level data. The feature fusion module is configured to perform data cleaning, feature extraction and fusion operations sequentially based on the calibrated monitoring data to obtain a standardized time-series feature vector. The deformation prediction and correction module is configured to: construct an LSTM early warning model consisting of a mechanism constraint network and an improved long short-term memory network using the temporal feature vector as input, and correct the prediction results using the structural dynamic stiffness matrix method; The mechanism network is stripped of temperature-related deformation to obtain residual sequences, and the residual sequences are trained and deformation is predicted using an improved long short-term memory network. The early warning information output module is configured to: make a graded judgment based on the corrected deformation prediction results and the preset multi-level dynamic early warning thresholds, and output safety early warning information through visualization and graded push mechanism.

[0042] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of deformation in high-rise buildings based on temporal deep learning, characterized in that, include: Acquire multi-source heterogeneous monitoring data and perform timestamp calibration. The multi-source heterogeneous monitoring data includes BDS positioning data, accelerometer data, temperature data, wind speed data, rainfall data, and water level data. Based on the calibrated monitoring data, data cleaning, feature extraction and fusion operations are performed sequentially to obtain a standardized time-series feature vector; An LSTM early warning model consisting of a mechanism constraint network and an improved long short-term memory network is constructed using temporal feature vectors as input, and the prediction results are corrected using the structural dynamic stiffness matrix method. The mechanism network is stripped of temperature-related deformation to obtain residual sequences, and the residual sequences are trained and deformation is predicted using an improved long short-term memory network. Based on the corrected deformation prediction results, and combined with the preset multi-level dynamic early warning thresholds, a graded judgment is made, and safety early warning information is output through visualization and graded push mechanism.

2. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 1, characterized in that, The data cleaning includes cleaning the BDS positioning data using an RTK / PPK fusion algorithm, specifically as follows: For static measurement points, a precise single-point positioning post-processing mode is adopted. The precise single-point positioning post-processing mode performs post-calculation on static observation data based on post-precise ephemeris to obtain high-precision positioning results. For dynamic measuring points, a real-time dynamic differential mode is adopted. The real-time dynamic differential mode is based on the real-time transmitted observation data and the base station data to perform real-time calculation to obtain dynamic positioning results. Satellite orbit errors are corrected based on precise ephemeris data. Ionospheric delay errors are eliminated by linear combination of dual-frequency observations, and gross errors are removed by sliding window filtering combined with statistical criteria.

3. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 2, characterized in that, The RTK / PPK fusion algorithm is based on a dual-differential carrier phase observation model. This model includes constructing dual-differential observation equations, performing dual-differential operations on carrier phase observations between the rover and the base station, and between the target satellite and the reference satellite, to eliminate satellite clock errors and receiver clock errors. The model then performs integer ambiguity fixing on the carrier phase observations after dual-differential processing, converting the floating ambiguity solution into a fixed integer solution. The dual-differential observation equations are as follows: ， in, These are dual differential carrier phase observations. For the carrier phase observations of the target satellite by the rover station, For the carrier phase observations of the target satellite from the reference station, For the rover's carrier phase observations of the reference satellite, The carrier phase observations of the reference satellite from the base station.

4. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 1, characterized in that, The data clarity also includes denoising the accelerometer data by using discrete wavelet transform, performing multi-layer wavelet decomposition using a preset wavelet basis to filter out high-frequency noise, extracting the vibration dominant frequency and damping ratio of the building structure, and when the vibration dominant frequency change rate exceeds a preset threshold, it is included as a new feature in subsequent feature extraction. The temperature data is decomposed into daily temperature difference and hourly temperature; the wind speed data is calculated into average wind speed and maximum gust wind speed; the rainfall data is cumulatively summed over multiple time scales; and the water level data is processed by moving average to eliminate high-frequency fluctuations and extract trend terms.

5. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 1, characterized in that, The feature extraction and fusion process includes constructing an initial feature vector containing displacement increment, displacement rate, cumulative rainfall across multiple time scales, daily temperature difference, and average wind speed; calculating the correlation coefficient between each feature and the deformation target variable using Spearman correlation analysis; screening target features whose absolute correlation coefficient is greater than a preset correlation threshold; determining multicollinearity when the correlation coefficients among multiple target features are greater than the preset correlation threshold; performing dimensionality reduction using principal component analysis; retaining principal components whose cumulative variance contribution rate is greater than a preset contribution rate threshold; and calculating a comprehensive feature value as a standardized time-series feature vector by weighting the variance contribution rates of each principal component. The formula for calculating the comprehensive feature value is as follows: ， in, For comprehensive eigenvalues, For the i-th principal component, Let be the i-th eigenvalue, and m be the number of principal components retained.

6. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 1, characterized in that, The construction of the LSTM early warning model, consisting of a mechanism-constrained network and an improved long short-term memory network, includes establishing a temperature-deformation linear model based on the theory of material thermal expansion, calculating temperature-related deformation through the temperature-deformation linear model, and using the difference between the actual monitored deformation and the temperature-related deformation as a residual sequence. The temperature-deformation linear model is as follows: ， in, For temperature-related vertical deformation, is the coefficient of linear expansion of the material. For daily temperature difference, This is a constant term.

7. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 6, characterized in that, An LSTM early warning model consisting of a mechanistic constraint network and an improved long short-term memory network is constructed. This model also includes a residual sequence input mechanism, using the residual sequence as input samples to train the network's sensitivity to nonlinear deformation patterns. A two-layer stacked structure is employed to capture long-term temporal dependencies, and a regularization strategy is introduced into the network to prevent overfitting. Finally, a dynamic inertia weight adjustment mechanism is introduced during the forget gate state update process, adaptively adjusting the retention degree of historical states through dynamic adjustment coefficients. The dynamic update formula for the forget gate is as follows: ， in, Output the forget gate at the current moment. It is the sigmoid activation function. This is the weight matrix. This is the hidden state from the previous moment. For the current input features, For bias terms, This is the dynamic inertia weight adjustment coefficient.

8. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 7, characterized in that, The method of correcting the prediction results using the structural dynamics stiffness matrix method includes triggering a catastrophic correction model when the monitoring data meets the extreme working condition judgment conditions, constructing the overall stiffness matrix of the building structure, converting extreme wind loads and rainfall loads into equivalent additional load vectors, solving the additional displacement response of the structure under extreme working conditions based on the overall stiffness matrix and the equivalent additional load vectors, and vector superimposing the additional displacement response with the prediction results of the improved long short-term memory network to obtain the corrected final deformation prediction value.

9. The method for early warning of deformation of high-rise buildings based on temporal deep learning according to claim 1, characterized in that, The multi-level dynamic early warning thresholds include foundation deformation thresholds, temperature-sensitive early warning thresholds, wind load anomaly early warning thresholds, and foundation seepage early warning thresholds, specifically: Basic thresholds: cumulative horizontal displacement ≥30mm, cumulative vertical displacement ≥20mm, deformation rate ≥3mm / d; Temperature-sensitive warning threshold: ΔT > 10℃ / 6h and vertical displacement rate > 1.5mm / h; Wind load anomaly warning threshold: wind speed > 15 m / s and horizontal displacement rate > 2 mm / h, while BDS multipath effect > 1 m; Foundation seepage warning threshold: rainfall > 100 mm / 24 h and foundation settlement rate > 1 mm / h.

10. A high-rise building deformation early warning system based on temporal deep learning, executed according to the method of claim 1, characterized in that, include: The multi-source data acquisition module is configured to acquire multi-source heterogeneous monitoring data and perform timestamp calibration. The multi-source heterogeneous monitoring data includes BDS positioning data, accelerometer data, temperature data, wind speed data, rainfall data, and water level data. The feature fusion module is configured to perform data cleaning, feature extraction and fusion operations sequentially based on the calibrated monitoring data to obtain a standardized time-series feature vector. The deformation prediction and correction module is configured to: construct an LSTM early warning model consisting of a mechanism constraint network and an improved long short-term memory network using the temporal feature vector as input, and correct the prediction results using the structural dynamic stiffness matrix method; The mechanism network is stripped of temperature-related deformation to obtain residual sequences, and the residual sequences are trained and deformation is predicted using an improved long short-term memory network. The early warning information output module is configured to: classify and determine the level of early warning based on the corrected deformation prediction results and the preset multi-level dynamic early warning thresholds, and output safety early warning information through visualization and a hierarchical push mechanism.

Images