Building deformation feature decoupling method and system based on PS-InSAR time series data

Abstract

The application provides a building deformation feature decoupling method and system based on PS-InSAR time series data, and is applied to the technical field of building structure monitoring and three-dimensional field reconstruction. Based on building PS-InSAR monitoring data, meteorological and building space multi-source data are synchronously integrated, after effective monitoring points are screened according to building boundaries, various features are processed by using differential standardization, the deformation trend and seasonal initial features are stripped by using STL / VMD time series decomposition, and the non-stationarity of signals is weakened, then the LSTM standard input tensor is generated by time series sliding window processing and fusion of building static attributes. A double-branch parallel LSTM network is built, the thermal elastic component is extracted by relying on the temperature lag law and deformation conservation constraint in the seasonal branch, and the consolidation and construction settlement features are captured by means of the attention mechanism in the trend branch; finally, the outputs of the double branches are fused, the structure sudden deformation is identified by relying on 3sigma residual error, the deformation curve is deduced by combining meteorological prediction, and building settlement risk quantitative prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of building structure monitoring and three-dimensional field reconstruction technology, and in particular to a method and system for decoupling building deformation features based on PS-InSAR time series data. Background Technology

[0002] In my country, with the acceleration of urbanization, a large number of old buildings are gradually entering the aging stage. At the same time, extreme weather, geological disasters such as earthquakes also pose a serious threat to the safety of buildings. Traditional building risk monitoring and assessment methods have many limitations, such as limited monitoring scope, insufficient accuracy, and poor real-time performance, making it difficult to comprehensively and accurately grasp the risk status of buildings.

[0003] InSAR (Inductively Coupled Aperture Radar Interferometry), as an advanced remote sensing technology, boasts advantages such as large area coverage, high precision, all-weather capability, and immunity to significant influences from lighting and weather conditions, enabling the acquisition of minute deformation information. However, LOS-oriented deformation based on InSAR observations only provides relative positional deformation relationships according to the line-of-sight direction. Changes in relative deformation are typically caused by multiple factors, and single InSAR data cannot decouple and provide corresponding deformation variables under different factors. Geological factors causing building deformation are usually long-term, trend-based deformations, while seasonal factors such as temperature and hydrology typically result in high-frequency, periodic, and nonlinear deformations. LSTM, a special type of recurrent neural network, is specifically designed to address the gradient vanishing problem encountered by standard RNNs when processing long-sequence data, thus offering significant advantages in capturing both trend-based deformations and rapid periodic fluctuations. However, PS-point observation signals in urban environments exhibit strong non-stationarity and multi-component coupling characteristics. Geological evolution trends and the thermal expansion and contraction of building structures intertwine in the time domain. Since the original deformation signals often contain random noise and step fluctuations, directly using conventional neural network modeling makes it difficult to accurately capture the characteristic boundaries of different physical components. Existing single decoupling methods have bottlenecks in handling signal non-stationarity, and there is an urgent need for a composite modeling architecture that first extracts modal features and then performs deep physical mapping.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0006] According to one aspect of this application, a method for decoupling building deformation features based on PS-InSAR time-series data is provided, comprising: acquiring PS-InSAR vector files of the target building area and parsing them to obtain PS-InSAR time-series displacement data; simultaneously collecting temperature and cumulative precipitation meteorological data; and extracting spatial attributes such as building height, type, and construction year; using the building vector frame to complete the spatial scale constraint screening of single building PS points; using Z-score standardization to eliminate dimensions of numerical features; performing one-hot encoding on building use category features; converting the building construction year into observation start interval time-series features and then standardizing it; using STL or VMD decomposition to remove trend and seasonal preliminary estimation features from the time-series displacement data, reducing the non-stationary coupling of deformation signals, alleviating network overfitting, and accelerating convergence; setting a fixed window length and sliding step size to cut deformation and meteorological long-term time-series data, generating continuous samples with temporal causal correlation. This paper flattens and broadcasts the static spatial attributes of buildings in the time dimension, and concatenates them with time-series sample channels to construct a three-dimensional standard data tensor. The constructed data tensor is then input into a dual-branch LSTM model to extract deformation features in parallel. The physical-induced seasonal LSTM branch learns the lag effect of temperature on the building, and introduces building height and coherence into the input layer to correct the thermoelastic response of high-rise buildings. The physical constraint of total deformation conservation is added to the loss function. The adaptive attention trend LSTM branch receives the trend estimate features obtained from pre-decomposition, and introduces a scaling dot product attention mechanism to adaptively locate the time window of settlement acceleration and step change, fitting the nonlinear long-term settlement trend components caused by soil consolidation and tectonic movement. The weighted fusion of the two branch outputs completes the separation of deformation components, and the 3σ residual threshold is used to identify sudden structural deformation anomalies. The long-term deformation curve is derived from the predicted meteorological sequence to generate advanced quantitative assessment information of building settlement risk.

[0007] Another aspect of this application is a building deformation feature decoupling system based on PS-InSAR time series data, used to execute executable instructions to implement the above-described building deformation feature decoupling method based on PS-InSAR time series data.

[0008] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described decoupling method for building deformation features based on PS-InSAR time series data by executing the executable instructions.

[0009] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for decoupling building deformation features based on PS-InSAR time-series data.

[0010] The beneficial technical effects of this application are as follows: It collects multi-source data on building PS time-series displacement, meteorological data, and building attributes; selects effective monitoring points based on building vector frames; and completes differentiated standardization through Z-score and one-hot encoding. It employs STL or VMD algorithms for pre-decomposition of deformation time series, stripping away trend and early seasonal features, and weakening non-stationary coupling of signals. A sliding window is then set to segment the time series, and static attribute broadcasts are spliced ​​to generate a three-dimensional tensor for the adaptive network. A parallel dual-branch LSTM is constructed; the physical-induced branch combines temperature hysteresis and physical constraints to extract thermoelastic seasonal components, while the attention-based adaptive branch captures nonlinear long-term settlement caused by soil consolidation and tectonic activity. Finally, the two types of components are weighted and fused, and a 3σ residual threshold is used to identify sudden deformations caused by construction and structural damage. The predicted meteorological sequence is then input to extrapolate long-term deformation curves, achieving the decomposition of building deformation causes and the quantitative prediction of settlement risk. This method is suitable for monitoring the structural safety of old urban buildings and buildings.

[0011] This application aims to reduce the coupling interference of the original signal, alleviate the overfitting of deep learning, accelerate model convergence, and improve the separation accuracy of deformation components by using pre-processed STL / VMD temporal decomposition; it uses a dual-branch LSTM to perform modeling separately to fit seasonal thermal expansion and long-term settlement, and combines building height correction and attention mechanism to solve the problem that traditional models cannot distinguish between the two types of deformation.

[0012] Furthermore, this application introduces deformation conservation physical constraints, making the model output physically interpretable and avoiding the distortion problem of pure black-box network fitting. It also eliminates seasonal noise based on 3σ residuals, accurately identifying unconventional sudden deformations and improving the reliability of safety early warning. It can extrapolate medium- and long-term deformations, compensating for the deficiency of linear models in ignoring temperature cycles, and providing quantitative evidence for long-term building safety assessments.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0014] Figure 1 The flowchart illustrates a method for decoupling building deformation features based on PS-InSAR time series data according to an embodiment of this application. Figure 2 This is a spatial distribution map of the study area and the selected PS points; Figure 3 This is a verification graph of a 2×4 subgraph matrix for STL mode decomposition. Figure 4 A comparison of the decoupling effects of a dual-branch LSTM; Figure 5 A comparison chart showing the prediction effects of the original deformation and the model reconstruction. Figure 6This is a graph showing the abnormal detection of residuals and the 3σ risk warning curve. Figure 7 This is a comparative chart showing the extrapolation and prediction of deformation trends over the next year. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] In one implementation, Figure 1 A schematic flowchart of a method for decoupling building deformation features based on PS-InSAR time series data according to an embodiment of this application is shown.

[0017] S101: Obtain the PS-InSAR vector file of the target building area and parse it to obtain PS-InSAR time-series displacement data. Simultaneously collect meteorological data such as temperature and cumulative precipitation, and extract the spatial attributes of building height, type, and construction year.

[0018] In one implementation, a PS-InSAR vector file of the target building area is read. The file internally stores the average deformation rate along the radar line-of-sight (LOS) and the cumulative deformation value corresponding to each interferometric monitoring period. The original temporal displacement records of all permanent scattering points are extracted by traversing each vector element. For scenarios where the vector file lacks a direct cumulative deformation field, an interferometric phase conversion formula is used to calculate the cumulative deformation per PS point for a single time period. The formula is as follows: In the formula, the radar wavelength λ = 0.031m (typical X-band wavelength of COSMO-SkyMed satellite). Let be the deformation phase difference value corresponding to the i-th interferometry. Substituting this into the measured value of the monitored phase yields the cumulative deformation at LOS for a single time interval. After calculating the deformation at all single time intervals, the overall average deformation rate at each PS point is solved using the least squares fitting formula, as follows: , Indicates the average deformation rate. Indicates the time of the i-th observation. and These represent the observation time and the average value of the corresponding deformation, respectively.

[0019] Taking the satellite observation data from period 42 of this embodiment as an example, N is taken as 41. The 42 observation timestamps are from 2018-01-31, 2018-04-02 to 2021-10-06. This represents the cumulative deformation value calculated for each period. , The average deformation rate at each point is obtained by fitting the values ​​of all observation times and deformation sequences. These values ​​are then concatenated to generate a complete and continuous LOS time-series displacement sequence. In the spatial constraint screening process, a building boundary vector polygon is imported, and a 5m buffer zone is set. Spatial overlay judgment is performed on the coordinates of all PS points, retaining only scattering points inside the polygon, those falling on the building roof, and those on the exterior wall structure. PS monitoring points irrelevant to the building's perimeter bare ground and road surfaces are removed. Taking the Yuquan Building project as an example, there were originally 247 PS points; after spatial constraint screening, 189 effective building PS points were identified. All subsequent calculations only use the deformation data of the screened points, ensuring a one-to-one correspondence between deformation samples and the main building structure.

[0020] The meteorological data acquisition time perfectly matches the PS-InSAR satellite's transit time. On each satellite observation day, the average daily temperature value from regional ground meteorological stations is retrieved simultaneously, and the daily precipitation monitoring value is stored concurrently. A 30-day sliding cumulative precipitation index is calculated based on the daily precipitation sequence. Taking the March 2018 observation window as an example, the sliding cumulative precipitation is the sum of precipitation values ​​from February 1, 2018 to March 31, 2018, a continuous 30-day period. The temperature time series is input into a subsequent physically induced LSTM branch to train the temporal lag response law of temperature transmission to buildings; the 30-day cumulative precipitation time series characterizes the long-term saturation degree of pore water in the soil, corresponding to soil consolidation and settlement disturbance factors. The deformation, temperature, and cumulative precipitation time series are aligned using a unified observation time index, eliminating time-off samples and ensuring temporal dimension matching during subsequent modal decomposition and tensor splicing, preventing data misalignment and coupling issues.

[0021] Three types of fixed attribute parameters are extracted based on the architectural planning vector layer. These three parameters remain unchanged throughout the observation period. Specifically, for the building height parameter: In this embodiment, the target building has 28 floors and a total height of 96m. This value is used as an auxiliary weight in the LSTM input layer to participate in the correction of the thermoelastic response of high-rise buildings. The greater the height, the higher the weight of the thermal expansion and contraction deformation correction. For the building use category parameter: This building is classified as a commercial hotel. Unique thermal encoding is used to complete the feature conversion, distinguishing the differences in base deformation among residential, industrial, and commercial buildings. For the building completion year parameter: The building was completed in 1980, and the observation started in 2018, with a conversion time interval of 38 years. This interval value is used as a temporal evolution feature. The Z-score standardization formula is used to uniformly eliminate dimensions. The standardized calculation formula is as follows: ,in This represents the standardized data value. The original unstandardized data is represented by μ, σ, and x, where μ and σ represent the mean and standard deviation of the selected data, respectively. The standardized building height, coding category, and static features of the building interval are flattened during the time-series tensor construction phase, broadcast and copied along the time axis, matching all time-series samples within each sliding window. These features are then concatenated with the deformation and meteorological dynamic features along the channel dimension and integrated into the same three-dimensional feature tensor.

[0022] S102 uses the building vector frame to complete the spatial scale constraint screening of single building PS points, uses Z-score standardization to eliminate dimensions for numerical features, performs one-hot encoding on building use category features, and converts the building construction year into the observation start interval time series feature before standardization.

[0023] In one implementation, a vector file storing the radar line-of-sight (LOS) deformation rate and the cumulative deformation value for each interferometric monitoring period is read. All permanent scattering point elements within the file are traversed to extract the complete temporal displacement record for each point. If the cumulative deformation for each interferometric period is not directly recorded in the vector layer fields, the displacement value for each point in a single time period is calculated using the interferometric phase conversion formula. In the formula, the wavelength of the COSMO-SkyMed satellite X-band radar is selected as λ=0.031m. The deformation phase difference value corresponding to the i-th interferometric period can be obtained by substituting the measured phase of the satellite interferometric image into the formula.

[0024] After all the deformation variables for a single time period have been calculated, the overall annual average deformation rate for each PS point is solved using the least squares fitting formula: Taking the Yuquan Building monitoring dataset as an example, the total number of satellite observation periods N is 41, corresponding to 42 sets of observation timestamps, with the time series from January 31, 2018 to October 6, 2021. For single-period observation dates, For the current period, the cumulative deformation is converted. , The average annual settlement rate at each point was obtained by fitting the mean of the deformation sequence over the entire observation period. All periodic values ​​were then concatenated to generate a continuous and complete LOS time-series displacement curve. After completing the time-series calculations, spatial range constraint screening was performed. The target building's closed vector polygon was imported, and coordinate overlay matching was performed using a 5m spatial buffer. Only PS monitoring points falling on the building's roof and exterior wall structure were retained, while scattering points from roads, bare ground, green spaces, and other building perimeter surfaces were removed. The original total number of PS points in this project was 247. After boundary filtering, 189 effective scattering points corresponding to buildings were identified. Subsequent feature calculations only used the time-series data of these filtered points to avoid interference from cross-building deformation signals affecting modeling accuracy.

[0025] The meteorological data acquisition time is perfectly matched with the satellite overpass observation time. The raw monitoring values ​​of the daily average temperature from ground meteorological stations are retrieved synchronously. Simultaneously, based on daily precipitation records, a 30-day sliding cumulative precipitation sequence is calculated. Taking the observation window of March 2018 as an example, the sliding cumulative precipitation statistical interval is the sum of all precipitation over 30 consecutive days from February 1st to March 31st of that year. The temperature time series will serve as the external input for the physically induced LSTM branch, used to train the temporal lag deformation law of buildings affected by temperature fluctuations; the 30-day cumulative precipitation sequence is used to reflect changes in the pore water saturation of the soil, characterizing the long-term settlement disturbance factors caused by soil consolidation. The deformation time series, temperature time series, and cumulative precipitation time series are all aligned using satellite observation time as an index, eliminating time series misalignment problems and ensuring the matching of various feature time dimensions during subsequent modal decomposition and tensor splicing processes.

[0026] Based on the urban building planning vector layer, three types of fixed spatial parameters that do not change over time are extracted. The processing methods for these three types of parameters and the subsequent tensor fusion logic are as follows: **Building height parameter:** Yuquan Building has a total height of 96m, belonging to a 28-story commercial high-rise building. This height value will be used as an auxiliary weight in the input layer of the LSTM network for model computation. The higher the building height value, the higher the weight for thermal expansion and contraction deformation correction, distinguishing the difference in elastic deformation amplitude between high-rise and low-rise buildings. **Building use category parameter:** Buildings are divided into three discrete categories: residential, commercial hotel, and industrial plant. Yuquan Building is classified as a commercial hotel. One-hot encoding is used to complete the feature transformation. Each building type is assigned a separate feature channel, with a single-category channel assigned a value of 1 and the remaining channels set to zero, eliminating the model weight offset caused by the size of the category values.

[0027] The building's completion year is 1980, and satellite observations began in 2018. The difference of 38 years serves as the basic interval characteristic. Subsequent values ​​are normalized using the Z-score standardization formula. In the formula, x represents the original construction interval of 38, and μ and σ are the mean and standard deviation of the construction interval samples of all buildings in the area, respectively. The standardized interval features eliminate differences in numerical magnitude. Building height, coding category, and standardized construction interval are all static constants and do not have temporal fluctuations. When constructing a three-dimensional tensor in the subsequent temporal sliding window, the static features are flattened, broadcast and copied across the entire time axis, and merged with each deformation and meteorological dynamic sample through channel splicing and fusion, and then uniformly input into the network for computation.

[0028] S103 uses STL or VMD to decompose and remove trend and seasonal preliminary estimation features from time-series displacement data, reducing the non-stationary coupling of deformation signals, alleviating network overfitting, and accelerating convergence.

[0029] In one implementation, one of the algorithms, STL Local Weighted Regression or VMD Variational Mode Decomposition, is selected to perform pre-decomposition processing on the standardized time-series displacement data. The original input received by the unit is a single-building-specific LOS time-series displacement sequence after complete processing by the preceding multidimensional data standardization unit. This sequence must meet four pre-processing criteria. First, all PS monitoring points have been spatially scaled based on the target building vector polygon, and external scattering points such as roads and bare ground are removed using a 5m boundary buffer. Taking the Yuquan Building monitoring project as an example, only 189 valid points are retained after screening out of the original 247 PS points, completely avoiding cross-building deformation signal interference. Second, all numerical deformation data are uniformly standardized using Z-score to complete the dimensionless conversion, and the difference in deformation numerical magnitude is eliminated through centralized scaling. Third, the time-series observation period perfectly matches the satellite image acquisition period. This project uses 42 consecutive interferometric images from the COSMO-SkyMed satellite, with an observation time interval from January 31, 2018 to October 6, 2021, totaling 42 sets of time-series timestamps. The entire decomposition operation uses the complete 42 consecutive time series, without truncation or segmentation of the original observation sequence. Fourth, each set of timestamps in the displacement sequence corresponds strictly one-to-one with the synchronously acquired daily average temperature and 30-day sliding cumulative precipitation sequences. There is no offset between the time indices of meteorological data and deformation data. The trend and early seasonal characteristics output by the decomposition can be directly paired with meteorological characteristics. The downstream time-series sliding window mapping unit and dual-branch LSTM network complete the channel stitching input.

[0030] The unit is set up with two independent time series pre-decomposition operation paths that can be selected from one to the other. Operators can use either the STL local weighted regression decomposition path or the VMD variational mode decomposition path separately, depending on the geological and climatic characteristics of the monitoring area. The two paths will eventually output trend preliminary estimation features and seasonal preliminary estimation features in a unified format. The output data structure is fully compatible with the two modeling branches of the subsequent dual-branch long short-term memory network.

[0031] The core computation relies on the Loess Local Weighted Regression Smoothing (LSS) smoother to perform multi-layer time series filtering and decomposition, using the annual climate cycle as the benchmark parameter for the smoothing window to match the regional annual temperature and precipitation cycle variation patterns. The decomposition outputs three types of time series components: the first layer strips away the long-term geological trend component formed by soil consolidation and tectonic movement; the second layer extracts the periodic seasonal component brought about by annual temperature fluctuations; and the third layer separates the irregular random noise residual component. Taking the two sampling monitoring points PS#33 and PS#6 of the project as examples, after 42 periods of time series smoothing and decomposition, the trend component shows a slow and continuous downward curve, while the seasonal component shows a periodic up-and-down oscillation curve corresponding to the annual cold and hot cycles, intuitively distinguishing between the two types of physically driven deformation signals.

[0032] The component splitting rules are as follows: the preliminary long-term geological trend features obtained by stripping are directed to the adaptive attention trend LSTM modeling branch, the preliminary annual seasonal cycle features are directed to the physical induced seasonal LSTM modeling branch, and the random noise components are directly discarded and do not participate in subsequent network training, thereby reducing the interference of invalid signals on model fitting.

[0033] A variational constraint objective function is constructed, and an iterative convergence criterion threshold is set to carry out multiple rounds of iterative optimization. The center frequencies of each intrinsic mode function are continuously adjusted iteratively until the frequency difference between adjacent iterations reaches the convergence criterion, at which point the operation stops.

[0034] After iterative convergence, all intrinsic modes are divided according to frequency range. After solving the 42-period time series of this project, a total of 8 independent IMF components are generated. Among them, 2 low-frequency components are merged as the initial estimation features of long-term trends, 3 periodic high-frequency modes are aggregated as the initial estimation features of seasonality, and the remaining 3 stray high-frequency modes are determined to be random noise and are removed.

[0035] The component splitting rules are consistent with the STL algorithm. Low-frequency trend initial estimation features are fed into the adaptive trend LSTM branch, and periodic high-frequency seasonal initial estimation features are fed into the physically induced seasonal LSTM branch. The dimensions and time series lengths of the two types of initial features are completely consistent with the ST decomposition output, ensuring that the input format of the downstream network remains unchanged.

[0036] The original LOS deformation time series is superimposed with two coupled physical signals: long-term ground subsidence and annual temperature expansion and contraction. These signals exhibit strong non-stationarity, and directly inputting them into an LSTM network can easily lead to weight oscillations, slow iterative convergence, and overfitting. This unit uses STL or VMD pre-decomposition to extract the two coupled physical features in advance, reducing the non-stationarity of the original time series. This provides preliminary estimated features with physical priors for the subsequent dual-branch network, simplifying the network fitting process, shortening the model training iterations, and improving the decoupling accuracy of the soil consolidation and building thermoelastic deformation components. The two sets of preliminary estimated features are simultaneously output to the time series sliding window mapping unit, where they are combined with standardized air temperature, cumulative precipitation, and building static attributes to generate the input tensor for the three-dimensional neural network.

[0037] When using STL decomposition, a local weighted regression smoother is used to split the time series, outputting three initial estimated components: long-term geological trend, annual and seasonal cycle, and random residual. When using VMD decomposition, the variational model is solved iteratively to extract the low-frequency zero-center frequency (IMF) as the initial trend feature and the periodic high-frequency IMF as the initial seasonal feature. The TL algorithm uses the Loess local weighted regression smoother as the core computational carrier for time series decomposition. For the complete LOS deformation time series of a single building after Z-score standardization, smoothing filtering decomposition is performed in layers, which is divided into three-layer component extraction process. The first layer filters out high-frequency fluctuations through a long-period smoothing window and extracts the initial estimated geological trend component formed by the combined effects of long-term consolidation of the stratum soil and slow regional tectonic movement. The second layer uses annual-scale smoothing filtering to remove the initial estimated seasonal cycle component that follows the annual cycle of regional temperature. The third layer removes the trend and seasonal components, leaving irregular small-amplitude oscillation signals, which are divided into random residual components. The three types of components—trend, seasonal, and residual—are output simultaneously, and their physical meanings are independent, with no signal crossover or mixing.

[0038] Taking the Yuquan Building monitoring dataset as an example, the entire time series contains 42 COSMO-SkyMed satellite interferometric records, spanning from January 31, 2018 to October 6, 2021. The STL smoothing window parameters are matched to the local annual temperature and precipitation cycle length, and the smoothing iterations use industry-standard convergence values. After layered filtering, the component variation patterns can be intuitively observed: the seasonal component fluctuation cycle is completely synchronized with the regional 30-day sliding precipitation and daily temperature variation cycles, forming a regular up-and-down oscillation curve; the trend component maintains a continuous and slow linear decreasing pattern, clearly distinguishing between two independent physical driving deformation signals: long-term ground subsidence and building thermal expansion and contraction.

[0039] The annual seasonal preliminary features obtained from STL decomposition are separately fed into the physical-induced seasonal LSTM modeling branch in the dual-branch network, while the geological trend preliminary features are directionally transmitted to the adaptive attention trend LSTM modeling branch. This achieves precise one-to-one matching between the decomposed output features and the corresponding network input channels, and the two types of preliminary features are directly used as the network's prior physical materials.

[0040] The first step of the VMD algorithm is to construct a variational constraint objective function, set an iterative convergence threshold, and conduct multiple rounds of iterative calculations. In each round of iteration, the center frequency values ​​corresponding to all intrinsic mode functions (IMFs) are continuously adjusted. The iteration is terminated when the frequency difference between two adjacent iterations is lower than the convergence threshold. After iteration, all generated eigenmodes are classified and grouped according to frequency intervals: low-frequency IMFs with center frequencies close to zero are merged into preliminary long-term geological trend features; high-frequency IMFs with stable annual cycles are grouped into preliminary temperature-driven seasonal features; and the remaining stray modes without fixed periods are uniformly determined as random noise components and no longer participate in subsequent network modeling calculations.

[0041] After iterative VMD solution, the standardized deformation time series of Phase 42 of Yuquan Building was decomposed into a total of 8 independent intrinsic modes. Among them, 2 low-frequency components were integrated into the trend branch input, 3 periodic high-frequency components were fused into the seasonal branch input, and the remaining 3 sets of stray high-frequency noise were directly discarded to filter out worthless random disturbance signals. The trend and early seasonal feature splitting logic of VMD decomposition output is completely consistent with the STL algorithm. The feature dimension and time series length of the output of the two algorithms are completely consistent. The downstream network can directly read and use it without adjusting the input interface. Relying on the physical prior information provided by the pre-decomposition, the learning difficulty of the network fitting coupled deformation signals is reduced, and the number of training iterations is reduced.

[0042] Both decomposition algorithms use single-building-specific LOS time series data as input material, processed by pre-processed multi-dimensional data standardization units. The input data must meet unified pre-processing specifications: PS point selection has been completed using building vector frames and a 5m buffer, eliminating external scattering points such as roads and bare ground; all deformation values ​​undergo Z-score standardization, with the standardization calculation formula as follows: In the formula, x represents the original cumulative deformation of a single period. This represents the average temporal displacement of a single building. The displacement sequence standard deviation is used; the time series length is fixed to match the complete observation cycle of 42 satellites. Each set of time series timestamps corresponds strictly one-to-one with the synchronously collected daily temperature and 30-day sliding cumulative precipitation sequences, ensuring that the decomposed output features can be directly paired with meteorological time series and seamlessly connected to the downstream time series sliding window mapping unit.

[0043] The original LOS deformation time series is simultaneously superimposed with two types of coupled deformation signals: long-term subsidence of the strata and annual thermal expansion and contraction of temperature. The signals have prominent non-stationary characteristics, and directly inputting them into a dual-branch LSTM network can easily lead to problems such as continuous oscillation of model weights, slow iterative convergence speed, and sample overfitting. By pre-decomposing the data using STL or VMD algorithms, the two intertwined core physical deformation features are separated in advance, which significantly reduces the non-stationarity of the original time series, weakens the fitting interference caused by multi-feature coupling during the network training stage, effectively alleviates the overfitting phenomenon of deep learning models, and accelerates the overall convergence speed. The decomposed trend and seasonal preliminary estimation features are used as core time series materials, and together with standardized temperature, cumulative precipitation, and building static attributes, they are fed into the time series sliding window mapping unit. After window cutting and static feature broadcasting and splicing, a three-dimensional standard data tensor adapted to the LSTM network is generated, which completely connects the entire business processing link of data acquisition, standardization, time series decomposition, and tensor construction.

[0044] By pre-decomposing and removing the two core physical features coupled within the deformation signal, the non-stationarity of the original time series is reduced, the coupling interference during network training is weakened, the overfitting phenomenon of the deep learning model is alleviated, and the model convergence speed is accelerated. The original PS-InSAR standardized LOS time series simultaneously superimposes two independent deformation signals: long-term consolidation and settlement of the strata and thermal expansion and contraction of buildings caused by annual temperature cycles. These two types of deformation are completely intertwined and coupled along the time axis, resulting in a strongly non-stationary overall time series. Taking the COSMO-SkyMed satellite observation sequence of Yuquan Building Phase 42 as an example, the complete observation period is from January 31, 2018 to October 6, 2021. A single original time series curve simultaneously contains the trend of a slowly sinking basement year by year and the periodic up-and-down oscillation of alternating hot and cold temperatures each year. There is no clear boundary between these two patterns of change. When directly fed into the neural network for training, the model needs to simultaneously fit both sets of patterns of change. Feature coupling interference will continuously cause repeated oscillations in the model parameter weights, significantly increasing the number of training convergence iterations. This easily leads to overfitting due to excessive learning of local fluctuations, resulting in a decrease in the model's generalization and recognition capabilities.

[0045] After pre-decoupling using either STL local weighted regression or VMD variational mode analysis, the coupled composite time series is split into two independent time series components: preliminary trend estimation features and preliminary seasonal estimation features. Each type of component corresponds to a single physical deformation cause. Taking the decomposition results of this project as an example, the trend component output by the STL algorithm retains only the slow settlement changes caused by soil and tectonics, while the seasonal component retains only the periodic elastic deformation driven by temperature. After solving the problem, the VMD algorithm generates 8 IMF components. Two low-frequency modes are integrated into a trend sequence, and three periodic high-frequency modes are aggregated into a seasonal sequence. The remaining irregular stray modes are directly discarded, completely achieving decoupling and separation of the two types of physical signals.

[0046] The two sets of initial estimated features after splitting are respectively directed to the independent modeling branches of the dual-branch LSTM. The physical-induced seasonal LSTM only learns the temperature lag deformation law corresponding to the seasonal component, while the adaptive attention trend LSTM only fits the long-term settlement evolution corresponding to the trend component. Each branch of the network only needs to fit a single physical change pattern and no longer processes intertwined composite signals synchronously. Comparing the training process with and without pre-decomposition of the same dataset, the undecomposed original time series training averaged 1200 iterations, and the loss curve continued to oscillate in the later stages of training. After STL / VMD pre-decomposition, the model only needs 520 iterations to achieve stable convergence. The fluctuation of weight updates is significantly reduced, and overfitting is significantly suppressed. The fitting error for the long-term settlement amplitude and seasonal elastic deformation amplitude of buildings decreases by an average of 37%.

[0047] The preliminary trend and seasonal features output by the temporal decomposition unit serve as core temporal data, and are uniformly transferred downstream to the temporal sliding window mapping unit along with standardized daily temperature, 30-day sliding cumulative precipitation, and processed building static attributes. The unit synchronously segments all temporal data according to the set window length and sliding step size to generate sample fragments with temporal causality. Static features such as building height, coding type, and construction interval are flattened in the time dimension and broadcast globally. These features are then concatenated with the decomposed temporal data and meteorological samples in the feature channel dimension to generate a standard three-dimensional tensor with dimensions of (number of samples, window length, feature dimension). This completes the data processing chain, linking multi-source data acquisition, multi-dimensional feature standardization, temporal modality decomposition, and sliding window tensor construction, providing standardized input data for subsequent dual-branch LSTM parallel feature extraction.

[0048] S104 sets a fixed window length and sliding step size to cut deformation and long-term meteorological data, generates continuous samples with temporal causal relationships, flattens and broadcasts the static spatial attributes of buildings in the time dimension, and splices them with the time-series sample channels to construct a three-dimensional standard data tensor.

[0049] In one implementation, a fixed observation sub-interval length W and an independent sliding step size S are set. The trend estimation features, seasonal estimation features, and long-term meteorological data obtained from the decomposition of the standardized deformation sequence are simultaneously segmented to generate continuous sample segments with temporal causal relationships. The time-series sliding window mapping unit receives four types of time-series materials as unified input: trend estimation feature sequences, seasonal estimation feature sequences, standardized daily temperature long series, and standardized 30-day sliding cumulative precipitation long series, all output by STL or VMD algorithms. All four types of time-series materials have undergone pre-modal decomposition and Z-score standardization, and their time indices are completely aligned with 42 satellite observation records, with no temporal offsets or missing points. The unit is configured with two sets of fixed segmentation parameters: the observation sub-interval length W and the time-series sliding step size S. The four types of time-series materials simultaneously perform sliding truncation operations using the exact same two sets of parameters, ensuring complete matching of the time boundaries of different features.

[0050] The entire satellite interferometric observation data acquisition period was from January 31, 2018 to October 6, 2021, recording 42 consecutive time-series samples. Each sample corresponds to a satellite transit interferometric imaging moment. In this embodiment, the length W of the observation sub-interval is set to be equivalent to the number of satellite observations corresponding to 3 months, and the sliding step size S is taken as the satellite observation interval of a single period. Each window only moves forward one set of observation records. Taking the complete 42-period time series as an example, the first window contains the observation data of periods 1 to 3. After completing one sample generation, the window slides forward 1 period to generate the second set of samples consisting of periods 2 to 4. The sliding continues cyclically until the end of the window reaches the observation point of period 42. All sample segments generated by sliding strictly follow the actual satellite imaging time sequence. The time sequence within the sample is arranged in an orderly manner from early to late, without any time reversal or random recombination across time periods. This satisfies the basic constraints of causal modeling of long short-term memory networks. The network can only use historical time series to deduce subsequent deformation states, eliminating the data leakage problem of future data participating in training in advance.

[0051] The four time series data types—preliminary trend features, preliminary seasonal features, daily temperature, and 30-day cumulative precipitation—are synchronously cropped using the same set of W and S parameters. Within any window sample, the observation timestamps corresponding to the four types of data correspond one-to-one. For example, if a window covers nine observation records from February to April 2019, the third time step of the trend component, the third time step of the seasonal component, the temperature during the same period, and the cumulative precipitation during the same period all correspond to the same satellite observation time on March 5, 2019. There is no misalignment in the time dimensions of various features, and the subsequent channel stitching process does not require additional time series alignment correction, directly completing the multi-feature fusion operation.

[0052] All four types of timing materials have undergone standardized processing before segmentation. The standardization calculation formula is as follows: In the formula, x represents the original time series value. This represents the average value over all time periods for a single building. To eliminate numerical differences caused by different physical dimensions of deformation, temperature, and precipitation, the standard deviation of the time series is used. After segmentation, the feature values ​​within each sample segment maintain a uniform scale, avoiding interference from numerical fluctuations between windows that could affect network weight updates. All time series samples after segmentation are transferred to the static feature fusion process. They are combined with static parameters such as building height, uniquely coded building type, and standardized construction interval. The static features are flattened in the time dimension and broadcast and copied globally. The time series length of each window is matched, and finally, a three-dimensional data tensor adapted to the LSTM network is generated by splicing the features.

[0053] For static spatial attributes including building height, building type, and construction interval, dimensional flattening and time-axis global broadcast copying operations are performed sequentially to match the duration of all time-series samples. All static parameters of a single building are integrated into a single one-dimensional feature vector, eliminating the dimensional fragmentation problem caused by multiple independent sets of parameters. Taking the Yuquan Building sample as an example, this building has three sets of static indicators: the original building height is 96m, which is standardized using Z-score conversion. The standardized calculation formula is as follows: In the formula, x is the original building height of 96. The average building height in the area is 72 meters. The standard deviation of building height in the area is 13.6. Substituting this into the calculation, we obtain the standardized height characteristic value.

[0054] The building is designated as a commercial hotel. A three-category one-hot coding rule is used, dividing it into three channels: residential, commercial hotel, and industrial. Commercial channels are assigned a value of 1, while residential and industrial channels are assigned a value of 0, generating a coding vector of length 3. The building was completed in 1980, satellite observations began in 2018, and the original time interval is 38 years. These values ​​are standardized using the same Z-score formula to obtain the standardized construction interval value. The standardized height scalar, the three-dimensional use coding vector, and the standardized construction interval scalar are sequentially concatenated and merged into a single one-dimensional fixed feature vector. This flattens the dimensions, eliminating the dimensional fragmentation caused by scattered independent parameters, resulting in a static one-dimensional feature array of fixed length.

[0055] The time-series sliding clipping uniformly sets the observation sub-interval length W to correspond to a 3-month satellite observation duration. A single dynamic sample segment contains 9 consecutive satellite observation time steps. The static one-dimensional vector has no time dimension and needs to be copied and filled in batches along the time axis. The total number of copies is exactly equal to the number of time steps within the window. Taking a single 3-month window sample of this project as an example, the flattened one-dimensional static vector will be completely copied 9 times. The 1st to 9th observation time steps within the window all reuse the same set of height, purpose, and construction interval feature values. The static parameter values ​​will not change as the window slides or the time sequence progresses.

[0056] The synchronization constraint rules are as follows: for all sample fragments generated by sliding cuts, regardless of changes in the start and end observation dates of the window, the static one-dimensional vector used for broadcast replication remains completely consistent, and the building's basic attributes will not be modified for different windows. When dynamic temporal series undergo translational clipping, static features are only refilled in the temporal domain, and the numerical values ​​and internal channel order remain constant.

[0057] After broadcast replication, the static features form a two-dimensional feature matrix with the exact same length as the single-window time series. The number of rows in the matrix equals the number of time steps in the window, and the number of columns equals the total dimension of the flattened static features. This matrix is ​​directly concatenated with the trend time series, seasonal time series, standardized temperature, and 30-day cumulative precipitation dynamic matrices within the same window at the feature dimension level, unifying them into a complete single-window feature matrix. After all windows have completed feature concatenation, they are stacked to finally generate a three-dimensional standard tensor with dimensions of total number of samples, window length, and total number of fused features. This directly adapts to the input specifications of a dual-branch LSTM network, ensuring complete alignment of the static and dynamic feature time indices, with no dimensional misalignment or time series mismatch issues.

[0058] The segmented trend features, seasonal features, and meteorological dynamic samples are concatenated and fused with the broadcasted static building features along the channel dimension to generate a standardized 3D data tensor with dimensions of sample number, window length, feature dimension, and adapted to an LSTM temporal neural network. Taking the Yuquan Building Phase 42 satellite observation dataset as an example, the observation sub-interval W is uniformly set to correspond to a 3-month satellite observation duration. A single window contains 9 consecutive satellite observation time steps. The complete features of a single window consist of four dynamic time series and one set of broadcast static features. The dynamic time series consists of: a complete initial trend estimation time series for a single window, with 9 time step values; an initial seasonal estimation time series for the same window, with 9 time step values; a standardized daily average temperature sequence for the corresponding period of the window, with 9 time step values; and a standardized 30-day sliding cumulative precipitation sequence for the corresponding period, with 9 time step values. Static features consist of: the target building's static parameters are processed to generate a one-dimensional fixed vector. The building height and construction interval are first converted using Z-score standardization. The standardized calculation formula is as follows: The formula is as follows: the original building height is 96, the average building height in the area is 72, and the standard deviation is 13.6; the original building interval is 38, and the average and standard deviation of the building intervals in the area are taken from the sample statistics. Building uses are divided into three categories: residential, commercial / hotel, and industrial plants. A one-hot encoding method is used to generate a three-dimensional vector. The standardized height scalar, the three-dimensional encoded vector, and the standardized building interval scalar are merged to obtain a one-dimensional static vector of length 5. This vector is completely copied along the window for 9 time steps, generating a static feature matrix with 9 rows and 5 columns, ensuring that the same set of inherent building attributes is reused at each observation time step.

[0059] Within the same sliding window, each of the four dynamic time series is a 9-row, single-column sequence. These are concatenated end-to-end according to channel order to form a dynamic feature sub-matrix. The total number of dynamic feature channels equals the sum of the number of time series for trend, season, temperature, and precipitation. In this example, there are a total of 4 dynamic channels. The 9x4 dynamic matrix and the 9x5 static broadcast matrix are then horizontally concatenated along the feature dimensions to generate a single-window total feature matrix. The number of rows in this matrix equals the window time series length of 9, and the total number of feature channels is 4 for dynamic and 5 for static, totaling 9 feature dimensions. The concatenation process only merges the feature dimensions; the time step dimension remains unchanged, preserving the causal order of satellite observation times. There are no issues of time series misalignment or feature time series mismatch, and all dynamic and static sample time indices correspond perfectly.

[0060] The process involves progressively slicing and cropping the data according to a set sliding step size, generating multiple sets of independent window 2D feature matrices. All window samples are stacked along the sample dimensions in the order of observation, ultimately generating a standard 3D tensor with three dimensions. The dimension definition rules are as follows: the first dimension is the total number of samples, the second dimension is the length of a single window time series, and the third dimension is the total number of features after fusion. Taking the complete time series of Yuquan Building Phase 42 as an example, with a window length of 9 and a sliding step of 1, 34 sets of independent window samples can be generated, corresponding to a tensor with a first dimension value of 34; a fixed second dimension value of 9; and a third dimension with a total fused feature count of 9, ultimately generating a 3D data tensor with dimensions 34, 9, and 9.

[0061] The generated 3D tensor serves as the sole unified input material for the dual-branch Long Short-Term Memory (LSTM) network. The Physically Induced Seasonal LSTM branch and the Adaptive Attention Trend LSTM branch share the same complete tensor, eliminating the need to split the input data into two independent sets. The network's internal channel routing rules are as follows: Seasonal time series, standardized temperature, and 30-day cumulative precipitation data are extracted from all feature channels of the tensor and fed into the Physically Induced Seasonal LSTM branch to learn the periodic thermoelastic deformation of buildings caused by temperature lag; Trend estimation time series data are extracted and fed into the Adaptive Attention Trend LSTM branch, which adaptively identifies settlement abrupt change windows using a scaling dot product attention formula to fit long-term soil and structural settlement. Static feature channels for building height, building type, and construction interval are simultaneously transmitted to both branches, serving as auxiliary weights for the input layer of each branch to complete the deformation correction of high-rise buildings.

[0062] S105 inputs the constructed data tensor into the dual-branch LSTM model to extract deformation features in parallel. The physical-induced seasonal LSTM branch learns the hysteresis effect of temperature on building action, and introduces building height and coherence into the input layer to correct the thermoelastic response of high-rise buildings. The physical constraint of total deformation conservation is added to the loss function. The adaptive attention trend LSTM branch receives the trend estimate features obtained from pre-decomposition, and introduces a scaling dot product attention mechanism to adaptively locate the settlement acceleration and step change time window, and fits the nonlinear long-term settlement trend components caused by soil consolidation and tectonic movement.

[0063] In one implementation, a dual-branch parallel LSTM feature extraction computing architecture is built based on standardized 3D data tensors. The two branches simultaneously complete independent modeling operations for seasonal and long-term deformation components. The 3D tensor output by the temporal sliding window mapping unit is fully imported into the feature input layer. The tensor has three dimensions: total number of samples, single-window temporal length, and total fused feature dimension. Taking the Yuquan Building measured dataset as an example, the entire observation sequence is segmented to generate 34 independent samples. A single window contains 9 satellite observation time steps, with a total of 9 feature channels including dynamic temporal series and static attributes. The final input tensor dimension is 9, 9. The input layer only performs data caching operations, without any channel reduction, temporal truncation, or feature scaling. All samples, the complete window duration, and the original values ​​of all 9 feature channels are fully preserved. The tensor contains all data for four types of dynamic temporal series and three types of building static parameters, with no information loss. The channels within the tensor are arranged in a fixed order: the first to fourth channels are the initial trend estimate time series, the initial seasonal estimate time series, the standardized daily temperature, and the 30-day sliding cumulative precipitation, respectively; the fifth to ninth channels are the standardized building height, the three-dimensional vector of the unique thermal encoding of commercial hotels, and the standardized construction interval, respectively. The input layer retains the order of these channels completely and does not disrupt the arrangement of features.

[0064] After the input layer caches the complete tensor, it flows to the branching layer. The branching layer divides the channels according to the physical deformation causes corresponding to the features, automatically completing feature grouping and parallel distribution in two paths, without modifying the original feature values ​​and temporal arrangement. The flow direction is based on the physical-induced seasonal modeling branch channel set: the seasonal preliminary time series channel, the daily temperature channel, and the 30-day moving cumulative precipitation channel. At the same time, all building static features from the fifth to the ninth channels are distributed as correction parameters.

[0065] Adaptive attention trend modeling branch channel set: Only the trend initial estimation time series channel is retained, and all building static features from the fifth to the ninth channel are distributed simultaneously as correction parameters; taking the number of channels in this example as an example, the seasonal branch receives 4 time series channels plus 5 static channels, totaling 9 sets of features, and the trend branch receives 1 time series channel plus 5 static channels, totaling 6 sets of features. The static feature vectors obtained by the two branches are completely consistent, and only the dynamic time series materials are distinguished from each other.

[0066] The two sets of feature data distributed by the split layer are fed into two separate sets of disconnected LSTM hidden computation units. Each set of units has its own weight matrix and bias vector, and the parameters are not shared. Updates during training do not affect each other. The standard update formula for a single LSTM neuron is as follows: ; ; ; ; ; In the formula, Let t be the forgetting gate vector; The input gate vector is at time t; Output the gate vector at time t; The candidate cell state at time t; initial , equal to 0; operator This represents the Hadamard product; the subscript 't' indicates the time step index. The variables... , represents the LSTM input vector; , represents the activation vector of the forget gate; , indicates the input / update gate activation vector; , represents the output gate activation vector; , which represents the hidden state vector, also known as the output vector of the Long Short-Term Memory network unit; , represents the cell input activation vector; , represents the unit state vector; , representing the weight matrix and bias vector parameters. The two branches simultaneously initiate forward propagation operations, calculating their respective loss values ​​within the same training iteration, and then performing reverse gradient descent to update their own W, U, and b parameters. The weight update of the seasonal branch is only constrained by the seasonal deformation loss, while the weight update of the trend branch is driven only by the long-term settlement loss. The gradient update processes are completely isolated, preventing parameter overlap and feature cross-interference. Static features such as building height, building type, and construction interval continuously participate in gating weight correction at each time step of the two branches. Taking a 96m standardized height as an example, the gating weight scaling factor for high-rise buildings is automatically increased to correct for greater thermoelastic deformation fitting bias under high-temperature conditions.

[0067] To address the mining of environmental thermoelastic periodic deformation, a physically induced LSTM branch input parsing logic is configured. Relying on network gating units, the hysteresis response law of temperature transmission to buildings is learned. Building height and coherence auxiliary weights are added to the input layer to correct the thermoelastic response characteristics of high-rise buildings. The model loss function embeds the physical constraint formula of total deformation conservation, and the constraint output conforms to the periodic deformation law of temperature, capturing the temperature-driven seasonal deformation component. After feature segmentation by the branching layer, all feature materials sent to this branch are divided into two categories: dynamic temporal materials and static building auxiliary features. Dynamic materials include three types of time-series data: seasonal preliminary time-series data obtained from STL / VMD decomposition within the window, standardized daily temperature sequences, and 30-day sliding cumulative precipitation sequences. Static materials are building spatial attribute features after dimensional flattening and full-domain time-axis broadcast replication. Both types of materials are simultaneously input into each layer of the branch network to participate in all iterative calculations. Taking the Yuquan Building monitoring dataset as an example, the time-series sliding window is uniformly set to correspond to 9 satellite observation durations, and the length of a single window time-series is equal to 9 observation time steps.

[0068] The channel division rules are as follows: A separate time series channel is used for the initial seasonal estimate, another for standardized daily temperature, and a third for 30-day sliding cumulative precipitation. Static feature channels include a standardized building height scalar, a 3D unique thermal encoding vector for residential / commercial / hotel / industrial buildings, and a standardized construction interval scalar, totaling five static feature channels. The total number of single-window input channels for the entire branch is eight. Throughout the branch calculation, the standardized building height value is used as a correction auxiliary weight in the calculation of each gated unit. The larger the standardized building height value, the greater the scaling correction weight of the corresponding gated layer, adapting to the significant thermal expansion and contraction deformation response of high-rise concrete buildings after prolonged solar radiation. In this case, the original building height is 96m, the average building height in the area is 72m, and the standard deviation is 13.6m. These values ​​are substituted into the Z-score standardization formula. The standardized height feature is calculated, and this value is simultaneously passed to the forget gate, input gate, and output gate weight matrices to complete the offset correction, distinguishing the differences in temperature deformation amplitude between high-rise and low-rise buildings. The unique hot coding of building use and the synchronous numerical values ​​of standardized construction intervals are used as global static constraint parameters, participating in the feature calculation of each time step throughout the process, and are used to fix the inherent differences in the deformation of the building base.

[0069] The branch internally constructs a standard three-layer gated computational structure for long short-term memory, sequentially configuring a forget gate, an input update gate, and an output gate. All gated units employ the neuron iterative computation formula: ; ; ; ; ; The initial unit state in the formula Initial hidden state All values ​​are assigned to 0. This represents the element-wise product operation of Adama. W and U are branch-specific weight matrices, and b is the branch-independent bias vector. The parameters are only updated during the backpropagation process of this branch and do not share weights with the adaptive trend branch.

[0070] The entire gating iterative process is used to automatically fit the time lag response law of temperature changes transmitted to the building structure, without the need for manually preset fixed delay periods. Taking the annual climate change pattern of Yuquan Building as an example, the temperature value continuously rises in the satellite observation time series every spring, but the thermal expansion of the building's concrete structure does not cause deformation synchronously, and the deformation peak lags by 2 to 3 satellite observation cycles. The network automatically learns this lag interval through multiple rounds of gradient descent iterations, relying on gating weights, and adaptively matches the regional temperature-deformation time difference correlation, without relying on manually given fixed delay parameters to constrain the model.

[0071] In each round of forward and backward iterations of the network, the total deformation conservation formula is embedded as an additional constraint term in the loss function. The conservation relationship is as follows: ,in, Represents the total time-series displacement. This represents the total displacement over time. This represents the total amount of seasonal deformation over time. This represents the residual displacement term. After a single iteration completes the predicted output of the seasonal components of the branch, the long-term settlement component output by the trend branch in the same period and the seasonal predicted component of this branch are summed, and the difference is calculated with the original measured deformation of the corresponding time period in the window. The difference result is superimposed on the overall loss value and participates in gradient backpropagation. This constraint will forcibly reduce the proportion of low-frequency signals of stratum settlement in the branch output, filter out the slow settlement interference caused by soil consolidation and tectonic movement, and retain only the periodic deformation signal caused by the annual cold and warm cycle. This achieves active separation of the two types of coupled deformation signals, seasonal and trend, during the training phase, and eliminates the fitting bias caused by the mutual mixing of features.

[0072] After the model iteration reaches the convergence criterion, the branch output contains only the pure seasonal deformation sequence driven by temperature periodicity. The output time series length is completely consistent with the nine observation time steps of the input window, and only a single seasonal deformation output channel is retained, with no trend, noise, or miscellaneous values ​​mixed in. Taking the Yuquan single-group three-month window sample as an example, the one-dimensional time series curve with a branch output length of nine has a curve fluctuation period that perfectly matches the regional annual temperature and thirty-day cumulative precipitation change period, which can intuitively reflect the complete evolution law of building thermal expansion and contraction. The seasonal output components corresponding to all windows are uniformly cached in an independent feature storage area, without being mixed with the trend branch output data. Subsequently, they are synchronously sent to the component weighted fusion unit and the residual difference calculation unit as dedicated seasonal feature materials for decoupling, anomaly identification, and medium- and long-term deformation inference.

[0073] An adaptive LSTM branching logic is configured for long-term settlement analysis. The branch inputs the trend estimate feature tensor from the pre-decomposition output of STL / VMD, and introduces a scaling dot product attention calculation formula. It adaptively identifies three types of local time windows: settlement acceleration, deceleration, and step abrupt changes, accurately fitting the nonlinear long-term settlement trend components caused by soil consolidation and tectonic movement. This branch only receives the trend estimate time series features from the separately divided branch flow layer. Three types of static features—standardized building height, uniquely coded building type, and standardized construction interval—are simultaneously input into the branch along with the trend time series and participate in the network calculations at each layer. Temperature sequences, 30-day moving average precipitation sequences, and seasonal estimate time series are not input into this branch, isolating environmental periodic fluctuation signals from the input source and avoiding interference from elastic deformation caused by temperature and precipitation in the long-term settlement fitting process.

[0074] Taking the Yuquan Building measured dataset as an example, the entire COSMO-SkyMed satellite observation time series covers 42 observation records from January 31, 2018 to October 6, 2021. The observation sub-window length is uniformly set to correspond to 9 satellite observation steps, and the sliding step size is the single-period observation interval. After complete time series sliding pruning, a total of 34 independent time series samples are generated. The trend components of all 34 samples are processed by STL or VMD pre-modal decomposition. The decomposition process removes high-frequency fluctuations in annual temperature cycles, retaining only low-frequency slow evolution signals corresponding to stratigraphic consolidation and tectonic movement. All 34 samples are completely sent to this branch for model iteration training, without sample truncation or discarding. Example of branch input feature channel composition: The single-window input includes one trend time series channel, coupled with five static auxiliary channels: standardized building height, 3D commercial hotel unique thermal encoding, and standardized construction interval. No temperature, precipitation, or seasonally related feature channels flow in throughout the process, completely separating the modeling input data sources for the two types of physical deformation. The standardized building height conversion uses the Z-score formula. The original building height is 96m, the average building height in the area is μ=72, and the standard deviation is σ=13.6. The calculated standardized height value is continuously used in each round of gate weight correction for the branch road to distinguish the differences in foundation settlement response of buildings with different floor heights. The original building interval of 38 years is also substituted into the formula to complete the standardization and used as a fixed constraint parameter in the calculation.

[0075] The terminal of the multi-layer LSTM hidden layer in the branch is fixedly integrated with a scaled dot product attention calculation module. The module is pre-configured with three sets of independent feature mapping weights: query matrix Q, key matrix K, and value matrix. The attention weight calculation formula is used to complete the solution of all temporal weights. ,in, This indicates that it covers m queries; , This indicates that they collectively contain a set of unordered key-value pairs. Indicates the activation function: ,in, Indicates the first The original scores for each category; Represents the natural constant; This represents the sum of the index scores of all n time-series nodes.

[0076] During module computation, all hidden features within the window's time steps are read. Three sets of feature vectors—query, key, and value—are generated through matrix mapping. After inner product scaling, softmax normalization generates weight coefficients ranging from 0 to 1. The network autonomously assigns weights based on settlement change characteristics. Taking a single 9-step window sample from the Yuquan Building as an example, abrupt changes and accelerated settlement occur in the 3rd and 7th observation periods within the window, automatically assigning high weights of 0.82 and 0.76 to the corresponding time steps. The weights for the remaining stable settlement observation steps are distributed only in the 0.1 to 0.3 range. No manual labeling of abnormal time intervals is required; the computational adaptive locking of local windows with settlement abrupt changes precisely amplifies the feature contribution of non-uniform settlement segments and weakens the interference from indistinguishable samples in stable settlement.

[0077] The trend features, after attention weighting, are fed into the hidden layer of a multi-level standard LSTM for iterative fitting. The branch gating update uses the general neural computation formula described earlier. Settlement caused by soil consolidation and underground tectonic movements does not exhibit a fixed linear rate of change, but rather shows a nonlinear evolution pattern of rapid initial settlement followed by slower settlement. The strata monitored at Yuquan Building are alluvial-diluvial loose silt layers. During the early monitoring period (2018-2019), pore water was continuously discharged, resulting in a larger average settlement amplitude per period. From 2020 to 2021, the soil gradually compacted, and the settlement rate continuously slowed down. The multi-level LSTM automatically captures this gradual curve through multiple rounds of gradient descent iterations, avoiding the constraints of a single linear fitting method. Linear models would use a fixed rate of estimation throughout, resulting in a significant deviation from the measured rapid-then-slow settlement curve. The nonlinear fitting output curve of this branch closely matches the measured low-frequency trend.

[0078] After the model training iterations reach the preset convergence threshold, the branch output contains only pure long-term trend deformation sequences induced by geological factors and soil consolidation, without temperature cycles or random noise mixing in the values. The output time series dimension corresponds one-to-one with the nine observation time steps of the input window. The input window time series length remains fixed, and the output sequence length remains synchronously consistent, without any time series compression or stretching deformation. The long-term trend output components corresponding to each group of 34 window samples are stored independently in a separate cache area, without being mixed with the seasonal deformation sequences output by the physically induced branch. All trend components are uniformly transmitted to the downstream component weighted fusion unit, serving as dedicated material for long-term settlement in the entire calculation process of total deformation reconstruction, residual difference calculation, and medium- and long-term settlement extrapolation.

[0079] S106, weighted fusion of the two outputs to complete the separation of deformation components, and the use of 3σ residual threshold to identify sudden structural deformation anomalies; input the predicted meteorological sequence to extrapolate the medium and long-term deformation curve, and generate advanced quantitative assessment information on building settlement risk.

[0080] In one implementation, a deformation decoupling and prediction integrated computation link is established based on the seasonal deformation component and long-term trend deformation component output by the dual-branch network. This establishes a complete four-level computational logic: component weighted fusion, residual term solving, 3σ anomaly detection, and time series forward extrapolation. The input material for the weighted fusion computation comes from the time series curves independently output by the two branches of the dual-branch parallel LSTM network. The two types of components have completely independent physical meanings and there is no signal crossover or mixing. Specifically, the time series output by the physically induced seasonal LSTM branch only represents the building's thermoelastic expansion and contraction deformation caused by annual temperature and precipitation cycles, while the time series output by the adaptive attention trend LSTM branch only represents the long-term settlement deformation caused by soil consolidation and tectonic movement. The two sets of output time series are strictly synchronized and matched with the previous time series sliding window segmentation rules. Each window segment is uniformly set to correspond to 9 satellite observation time steps. The total number of samples, single window time series length, and time index of the two component sequences are completely one-to-one correspondences. Each observation moment has a set of seasonal deformation values ​​and a set of long-term subsidence values, ensuring that the time-by-time weighted summation operation can be performed synchronously without time series misalignment or length mismatch issues. The entire operation relies on the total deformation conservation relationship for constraint verification. The conservation calculation formula is: In the formula Represents the total form variable of the complete LOS. For long-term trend deformation components, For temperature-driven seasonal deformation components, Representing the residual displacement term, the weighted superposition operation is essentially the numerical restoration of the trend and seasonal terms within the formula.

[0081] A uniform weighted summation rule is used to directly sum the seasonal deformation and long-term subsidence values ​​at the same observation time step, with all weight coefficients set to 1. No differential scaling factor needs to be set, and the calculation only applies to values ​​at a single moment, without modifying the time series arrangement or the logical sequence of times within the window. A single sliding window sample from the Yuquan Building monitoring dataset is used as a calculation example. This window contains nine consecutive periods of COSMO-SkyMed satellite observation records, and the two component values ​​are read synchronously for summation at each observation step. The calculation is performed at the fifth observation time within the window: the seasonal deformation component value is 0.82 mm, and the long-term subsidence component value is -1.26 mm. Following the uniform weighted summation rule, the reconstructed total deformation is calculated as 0.82 + (-1.26) = -0.44 mm. The same summation rule is used for the remaining observation steps 1 to 4 and 6 to 9 within the window, generating a complete nine-step reconstructed total deformation time series curve. The entire weighted operation only performs the basic numerical addition operation, without time series truncation, smoothing, or reordering. All time series values ​​and their order of the original seasonal component and the original long-term settlement component are completely preserved. The reconstructed total deformation sequence generated by superposition is only used as derived auxiliary data and does not cover the original two types of physical components.

[0082] After weighted calculation, the data is stored in three independent cache areas. The first area retains the original, unmodified seasonal deformation time series. The second area stores the original long-term settlement time series separately. The third area stores the reconstructed total deformation sequence generated by superposition. The three types of data are archived separately, and any downstream process can retrieve a single component without secondary splitting of the reconstruction curve. This achieves permanent decoupling of the two physical causes, thermoelastic deformation and soil settlement, and avoids component contamination. The three types of cached data are synchronously transferred to the downstream residual calculation process as standard input material for residual term calculation. Subsequent residual statistics, sudden deformation identification, and medium- and long-term time series extrapolation can retrieve seasonal components, trend components, or reconstructed total deformation curves as needed. In the residual calculation process, the reconstructed total deformation is compared with the original LOS deformation measured by satellite at each time step. In the extrapolation process, two types of components can be retrieved separately to independently construct future deformation evolution curves. Multi-dimensional analysis is carried out throughout the process based on the separated components retained by this weighted fusion.

[0083] Weighted fusion operation rules were configured for two types of independent deformation components. Weighted fusion operations were performed on seasonal deformation components and long-term trend deformation components to achieve independent decoupling and separation of the two physical components: thermoelastic periodicity and soil consolidation. Complete time-series data were obtained from 42 consecutive interferometric observation records from the COSMO-SkyMed satellite from January 31, 2018 to October 6, 2021. Residual difference calculations were performed independently at all observation time points. Before the operation, the two types of components were reconstructed by superimposing them with equal weights to generate the total deformation. The superposition rule was to directly sum the seasonal component values ​​and the trend component values ​​at the same observation step, without any differential scaling factor. Taking the satellite observation time corresponding to the Yuquan Building monitoring sample in July 2021 as an example, the total measured LOS deformation obtained directly from the satellite during this period was -4.81 mm; the seasonal deformation value output by the physical induced branch was 0.82 mm, and the long-term settlement value output by the adaptive branch was -1.26 mm. The reconstructed total deformation value after weighted fusion of the two was 0.82 + (-1.26) = -0.44 mm. Substituting the two sets of values ​​into the residual solution formula, -1.65 mm was generated. This -1.65 mm residual value represents the displacement disturbance at the observation point in July 2021 that cannot be explained by the slow consolidation of the strata and the annual temperature expansion and contraction law. The entire set of 42 time series is calculated cyclically according to this time-by-time difference logic to generate a continuous and complete residual time series curve. Each set of satellite observation timestamps is matched with a unique residual value, and there is no misalignment or missing time index.

[0084] The complete long residual time series is segmented according to the unified cutting standard of the previous time series sliding window. The window length is fixed and corresponds to a 3-month satellite observation cycle. Each window contains 9 consecutive observation time steps, and each group of 9 consecutive residual values ​​is packaged separately to form a residual segment. All residual segments are arranged strictly in the order of satellite observations, without disrupting the causal arrangement of time. All segmented and packaged residual segments are uniformly sent to the three-fold standard difference anomaly discrimination module. This module uses the residual sequence as the only input material for judgment. The conventional deformation signals caused by ground subsidence and temperature cycles have been separated by dual-branch components. The residual sequence only retains information on sudden and abnormal displacements such as construction excavation, local structural damage to buildings, and underground pipeline disturbances, which can serve as an independent benchmark indicator for identifying abnormal deformation of buildings.

[0085] The time-series residuals are solved by interpolating the measured total deformation and the model reconstructed deformation. A three-standard-deviation threshold discrimination mechanism is introduced to identify sudden deformation anomalies induced by construction disturbances and building structural damage on a time-series basis. First, interpolation is performed on the measured total deformation and the double-branch weighted reconstructed total deformation at each observation time to obtain a complete full-time residual sequence. Then, global statistical operations are performed on the entire residual sequence to calculate the overall mean and standard deviation of the residuals. The upper and lower anomaly judgment boundaries are defined by the three-standard-deviation value. The residual value at each satellite observation time step is compared with the boundary interval to identify sudden deformation samples that exceed the interval.

[0086] Using the residual data from 42 satellite observations of Yuquan Building as a statistical example, the mean residual value for the entire time period was 0 mm, and the standard deviation of the residual σ was 0.872 mm. After a three-fold conversion, the upper and lower boundaries for anomaly judgment were +2.615 mm and -2.615 mm, respectively. Any residual value calculated at any observation time step that was higher than the upper limit or lower than the lower limit was considered a point where sudden building deformation existed. In July 2021, the residual value calculated from one set of observation steps was -1.65 mm, which fell within the range of -2.615 mm to +2.615 mm and was only classified as normal fluctuation. At the same time period, the residual value of another monitoring point was -3.01 mm, which was less than the negative three-standard-deviation judgment boundary and was directly marked as a sample of sudden structural anomaly.

[0087] All locations identified as anomaly will have three types of supporting information simultaneously stored: the corresponding observation time window, the unique PS number for that location, and the magnitude of the residual exceeding the threshold. All anomaly-related information will be archived and stored separately as the foundational data for building safety risk screening. The entire threshold classification is automatically generated based on the statistical characteristics of all residual samples, without the need for manual pre-setting of fixed anomaly cutoff values. Residuals within the normal range will be stored separately as random noise and will not be included in the analysis materials related to building safety early warning.

[0088] Future predicted meteorological sequences and time-series time codes are simultaneously integrated into a fully converged, dual-branch LSTM network to establish a medium- to long-term deformation pre-extrapolation computational pipeline. The input materials required for the computation are divided into two categories: dynamic materials—medium- to long-term predicted meteorological sequences and time-series time codes—and are simultaneously coupled with static features such as building height, building type, and construction interval. All materials undergo normalization processing according to standards consistent with the model training process. Numerical normalization is uniformly performed using Z-score standardization, with the corresponding calculation formula: In the formula, x represents the predicted temperature, the 30-day moving cumulative precipitation, and the original time-series encoded value. , The overall mean and standard deviation are fixed and taken from the historical satellite monitoring samples of the 42nd phase of COSMO-SkyMed at Yuquan Building. The mean and variance are not recalculated using data from the prediction period to avoid distortion caused by scale mismatch. Taking the 12-month extrapolation task at Yuquan Building as an example, the extrapolation period covers the complete climate cycle of the year following the end of satellite observations. Daily average temperature and daily precipitation forecasts are generated based on the average meteorological monitoring data for the same period in previous years in the region. A 30-day sliding cumulative precipitation sequence is calculated based on the daily precipitation. Corresponding extrapolation time series time codes are generated according to the fixed satellite observation intervals. All meteorological and coded values ​​are substituted into the above formulas to complete the standardization conversion. The rules for processing static building features are consistent with those in the training phase: the original building height of 96m and the construction interval of 38m are substituted into the same standardized formula for conversion; the commercial hotel use adopts a three-class one-hot encoding to generate a one-dimensional vector; the standardized height, the encoded vector, and the standardized construction interval are concatenated into a one-dimensional static feature vector, which is then broadcast and copied across the entire domain according to the observation duration of the 9 satellites corresponding to the extrapolation window. The same set of static feature values ​​is reused in each time step within the window to achieve matching between static features and the duration of time series extrapolation samples.

[0089] The standardized forecast meteorological sequence and time-series encoding, combined with the complete input of the static building features after global broadcasting, are used to input the features into the dual-branch network feature input layer. The splitting layer uses the fixed channel division rules from the training phase to split the features and distribute them to the two modeling branches.

[0090] Physically induced seasonal LSTM branch allocation materials: standardized predicted daily temperature, 30-day sliding cumulative precipitation sequence, time-series encoding, and a complete set of broadcast building static features; the branch internal control unit and building height correction weights maintain the fixed values ​​obtained from training, and rely on future meteorological data to replicate the lag deformation law of concrete thermal expansion and contraction caused by annual hot and cold cycles.

[0091] The adaptive attention-based trend LSTM branch allocates data only using time-series encoding and a full set of broadcast building static features, completely isolating it from periodic environmental data such as predicted temperature and precipitation, thus preventing seasonal signals from interfering with long-term settlement extrapolation results. The branch incorporates a scaling dot-product attention module to maintain the Q, K, and V mapping weights after training convergence, continuously identifying acceleration and deceleration intervals in the nonlinear evolution of soil consolidation. The two branches execute forward extrapolation calculations synchronously and in parallel, only calling fixed network parameters during the calculation process, without gradient backpropagation or weight correction operations.

[0092] Using the 12-month long-term extrapolation task of Yuquan Building as a calculation example, the entire extrapolation process uniformly maintains a single window corresponding to the observation duration of 9 satellites. The sliding truncation rules are completely consistent with the historical sample cutting standards. Multiple extrapolation windows are generated in batches, and the feature output of the two branches is completed simultaneously. The physical-induced seasonal LSTM branch outputs the extrapolation of seasonal deformation curves. The curve fluctuation period matches the annual temperature change characteristics of the region. During the high-temperature period in summer, the extrapolated deformation value increases positively, and during the low-temperature period in winter, the value decreases downward, completely replicating the annual thermoelastic expansion and contraction evolution law of the building. The adaptive attention trend LSTM branch outputs the long-term settlement extrapolation curve. Relying on fixed attention weights, it continuously continues the historical soil consolidation evolution characteristics of rapid initial settlement followed by slow settlement. The settlement change amplitude is relatively large in the early stage of the extrapolation, and the soil gradually compacts in the later stage of the extrapolation, with the settlement rate continuously and slowly decreasing. It does not use a fixed linear rate to estimate the settlement value. The time series lengths of both types of inference output curves strictly correspond to the nine observation steps of a single window. All seasonal deformation sequences and long-term subsidence sequences generated in all windows are cached and stored separately and independently, without being mixed with each other. Any component can be retrieved separately for risk quantification and assessment, or the overall inference total deformation curve can be generated by superimposing according to the equal weighting rule.

[0093] Based on the long-term settlement trend curve and seasonal thermoelastic deformation evolution curve output by the network forward extrapolation, an advanced quantitative assessment of the overall building safety risk is completed using the dual-component prediction results. After forward extrapolation operation by a convergent dual-branch LSTM network, two independent time-series results are obtained: the medium-to-long-term stratum long-term settlement trend curve and the medium-to-long-term seasonal thermoelastic deformation evolution curve. The time-series length and observation time index of the two sets of extrapolation curves are completely matched, following the window cutting standard consistent with historical samples, with each window uniformly retaining 9 satellite observation time steps. The two types of extrapolation components completely retain their independent physical evolution characteristics, and no numerical aliasing occurs. Subsequent fusion only performs synchronous numerical superposition, without modifying the internal time-series arrangement or single-time deformation value of any component.

[0094] The fusion operation employs an equal-weighted superposition rule. Long-term settlement and seasonal deformation values ​​at the same extrapolation observation time are directly summed to synthesize the total deformation sequence. The weighting coefficient is uniformly set to 1, and no differential scaling parameters are introduced throughout the process. Taking a set of observation times within the 12-month extrapolation window of Yuquan Building as an example, the extrapolated long-term settlement component for this period is -1.31 mm, and the concurrent seasonal thermoelastic deformation component is 0.78 mm. After substituting into the superposition logic, the overall predicted total deformation at this time is -0.53 mm. All extrapolation windows perform the same summation operation step-by-step. The synthesized overall predicted total deformation sequence serves only as a comprehensive deformation reference curve. The original settlement curve and the original seasonal curve are stored separately. These two independent components can be retrieved independently throughout the process for multi-dimensional risk assessment.

[0095] Two independent quantitative assessment indicators were extracted from the settlement trend curve and the seasonal deformation curve, respectively. These indicators correspond to the physical driving factors of the two types of deformation, allowing for the separate identification of soil consolidation risk and building temperature expansion and contraction fatigue risk. For the soil settlement dimension indicator: the average settlement rate within a single time period is calculated for each projection window. The difference in deformation between the beginning and end of the window is divided by the observation duration of the window to obtain the settlement rate for a single window, used to determine the acceleration or deceleration trend of soil consolidation. The entire rate calculation relies on the original values ​​of the projection trend curve, requiring no additional conversion processing. Using the historical benchmark rate of Yuquan Building as a comparison reference, the overall average settlement rate for 42 historical time periods is -0.21 mm / single observation period. For the seasonal deformation dimension indicator: the maximum and minimum deformation values ​​within a single complete annual projection time period are extracted, and the difference between the two is calculated to obtain the annual high and low temperature deformation peak difference. The magnitude of the difference directly reflects the building's annual thermal expansion and contraction amplitude; a larger difference indicates a higher degree of repeated expansion and contraction fatigue in the concrete structure. The historical seasonal peak difference of 1.05 mm for the same period in this region is used as the comparison standard.

[0096] By comparing the deviations of settlement rates and annual peak deformation differences between the various windows during the extrapolation period and historical baseline values, differentiated safety risk levels are established. If either indicator shows a sustained deterioration, the corresponding extrapolation window number and the quantitative deviation value are recorded simultaneously, preserving the complete timeframe and deformation magnitude of the risk occurrence. A case study using the 12-month extrapolation results of the Yuquan Building as an example: the average settlement rate of the four consecutive sliding windows in the latter half of the extrapolation was -0.29 mm / single observation period, compared to the historical average of -0.21 mm. The rate increase was calculated as (0.29-0.21)÷0.21≈38%, nearly a 40% increase. Simultaneously, the extrapolated summer peak deformation reached 1.83 mm, and the winter trough was -0.42 mm, with an annual peak difference of 2.25 mm, significantly higher than the historical baseline difference of 1.05 mm for the same period. The simultaneous sustained deterioration of both indicators quantitatively indicates an increase in the building's medium- to long-term safety risk level.

[0097] If only the settlement rate continues to increase, the risk factor is classified as uneven consolidation of the soil and slow movement of underground structures. If only the annual peak deformation difference increases significantly, the risk factor is classified as long-term thermal fatigue damage to high-rise concrete structures. If both types of indicators deteriorate simultaneously, it is judged as a compound risk, with foundation settlement and structural thermal fatigue problems existing simultaneously. All judgment conclusions are accompanied by corresponding quantitative values ​​such as window time intervals, rate increase ratios, and seasonal peak differences, forming standardized and traceable quantitative assessment outputs for building safety.

[0098] like Figure 2 As shown, Figure 2This is a spatial distribution map of the study area and the selected PS points. The base map is an optical satellite image of a certain location, overlaid with building vector boundary polygons, and includes a 7.5-meter scale bar and a north-facing label. The legend indicates the annual deformation rate range of the PS points, graded and colored from less than -4.5 mm / yr to greater than 4.5 mm / yr.

[0099] Only permanent scattering points (PS points) within a 5-meter buffer zone outside the building vector frame are retained in the image. Irrelevant scattering points on the ground are removed. All PS points are concentrated in the building's exterior walls and roof structure areas, which intuitively shows the spatial distribution of effective monitoring points on a single building and demonstrates the preprocessing effect of spatial scale screening based on the building vector frame.

[0100] like Figure 3 As shown, Figure 3 This is a verification graph of the 2×4 subgraph matrix for STL mode decomposition. Two representative PS points (PS#33 and PS#6) were selected by sampling with a fixed random seed. The subgraph matrix is ​​arranged in 2 rows and 4 columns. Each row corresponds to a monitoring point. The four columns from left to right are the standardized original deformation curve, the long-term trend component curve extracted by STL, the seasonal periodic component curve, and the random residual component curve.

[0101] The first column shows the original time series: the standardized original LOS displacement curve, which clearly demonstrates the high coupling between geological subsidence and thermoelastic deformation, resulting in significant non-stationary fluctuations in the signal. The second column shows the trend component: a smooth blue curve depicting the long-term, slow subsidence evolution caused by stratum consolidation and tectonic movement. The third column shows the seasonal component: a green periodic curve representing the quasi-periodic fluctuations in building thermal expansion and contraction driven by annual temperature. The fourth column shows the residual component: a red irregular oscillating curve representing the remaining noise and sudden deformation signals after removing the trend and seasonal terms. The entire set of charts demonstrates that STL decomposition can effectively isolate coupled deformation signals, separating three types of time series components with clear physical meanings.

[0102] like Figure 4 As shown, Figure 4 This chart compares the decoupling effects of the two-branch LSTM method. The horizontal axis represents time, and the vertical axis represents deformation (mm), with three core time-series curves. The original measured deformation curve reflects the total deformation observed at the Yuquan Building PS point using InSAR, with long-term subsidence superimposed on annual periodic fluctuations. The model-decoupled long-term trend deformation curve is a smooth, monotonically descending curve, accurately reproducing the continuous subsidence caused by the uneven consolidation of alluvial strata. The model-decoupled seasonal thermoelastic deformation curve is a standard annual periodic oscillation curve, perfectly matching the seasonal changes in regional temperature, reflecting the elastic displacement of building concrete due to thermal expansion and contraction. The comparison of these three curves visually demonstrates that the two-branch LSTM method can achieve accurate physical decoupling between the geological trend component and the environmental seasonal deformation component, with clear separation of the boundaries between the two types of deformation characteristics.

[0103] like Figure 5 As shown, Figure 5 This chart compares the original deformation with the model reconstruction prediction. The horizontal axis represents the monitoring period from January 2018 to July 2021, and the vertical axis represents the LOS deformation (mm). The baseline is a gray dashed line at 0mm. Dark solid line with scattered dots: InSAR measured original total deformation sequence. Red dashed line: Model reconstruction prediction curve obtained by overlaying the trend and seasonal components of the dual-branch LSTM. Light red filled area: Overestimation range where model predictions are higher than measured values; light blue filled area: Underestimation range where model predictions are lower than measured values.

[0104] The quantitative accuracy indicators marked above the figure include the maximum reconstruction error, the mean absolute error of prediction (MAE), and the root mean square error (RMSE). The values ​​are extremely low, and the visualization verifies that the overall reconstruction fitting accuracy of the model is excellent. The predicted curve closely matches the actual observed deformation, and the difference between the two is the residual term used for anomaly detection.

[0105] like Figure 6 As shown, Figure 6 This chart shows the residual anomaly detection and 3σ risk warning curves. The horizontal axis represents the complete monitoring period, and the vertical axis represents the residual displacement (mm) after decoupling. A red warning threshold of ±3 standard deviations (3σ) (±2.615mm) is set with 0mm as the baseline. The curve represents the time-series residual ε output by the model. For most periods, the residual oscillates randomly within the threshold range. Around July 2021, some PS points experienced significant residual deviations, exceeding the 3σ warning red line and being marked as anomalies. The chart includes statistics on the maximum absolute value of the global residual, the absolute deviation of the median residual (MAD), and the number of anomalies, visually demonstrating the sudden deformation identification logic based on the three-standard-deviation threshold for residuals. This logic can filter out seasonal interferences such as temperature and precipitation, accurately identifying non-periodic sudden displacement risks such as construction disturbances and localized building damage.

[0106] like Figure 7 As shown, Figure 7 This chart compares the extrapolated deformation trends for the next year. The horizontal axis represents the 12-month prediction period from October 2021 to October 2022, and the vertical axis represents the deformation (mm), including multiple comparative curves. The total deformation prediction curve of this invention's dual-branch model integrates long-term settlement trends and seasonal thermoelastic components. During the high-temperature summer period, thermal expansion and rebound occur, while in winter, low-temperature settlement overlaps with ground subsidence, completely reproducing the annual cyclical deformation pattern. The traditional linear rate prediction curve is a single, flat, monotonically decreasing line, completely ignoring the seasonal fluctuations of building thermal expansion and contraction, resulting in significant prediction errors during winter and summer. The component output curves separately display the decoupled long-term trend prediction curve and the seasonal deformation prediction curve. The chart comparison demonstrates the extrapolation advantages of this method: it can adaptively capture temperature-driven periodic elastic deformation, with no significant drift in long-term extrapolation errors, significantly reducing the probability of missed or misjudged building safety risks compared to traditional linear models.

[0107] In one embodiment, this application also provides a building deformation feature decoupling system based on PS-InSAR time series data, comprising: The multi-source heterogeneous data acquisition module is used to read and parse the PS-InSAR vector file of the target building area to obtain LOS time-series displacement data, simultaneously collect temperature and cumulative precipitation meteorological monitoring data, and extract various spatial attribute parameters such as building height, building type, and construction year. The feature cleaning and standardization module is used to filter effective PS monitoring points within a single building based on the building vector boundary; numerical features are standardized using Z-score, building category features are one-hot encoded, and the construction year is converted into the observation interval time series before standardization to uniformly eliminate the differences in the dimensions of each feature. The temporal mode decomposition module is used to pre-decompose the standardized deformation sequence by selecting STL local weighted regression or VMD variational mode algorithm, stripping the initial trend and seasonal features, weakening the non-stationary characteristics of deformation signal coupling, reducing the risk of overfitting during network training, and improving the convergence speed of the model. The 3D tensor construction module is used to set a fixed window and sliding step size to cut deformation and generate causal samples from long-term meteorological data. The static attributes of buildings are flattened and broadcast in the time dimension and spliced ​​with the time-series sample channels to generate a standardized 3D input tensor adapted to the LSTM network. The dual-branch deformation feature extraction module is used to input three-dimensional tensors into the dual-branch LSTM in parallel; the physical-induced seasonal branch learns the effect of temperature lag, corrects the thermoelastic response of high-rise buildings based on building height and coherence, and superimposes the total deformation conservation loss constraint; the adaptive trend branch accesses the decomposed trend features, uses scaling dot product attention to locate the settlement change interval, and fits the soil and structure-induced long-term settlement. The component fusion risk assessment module is used to weightedly fuse the dual-branch outputs to decouple two types of deformation components, identify sudden building deformation anomalies based on the 3σ residual threshold, import predicted meteorological time series to complete medium- and long-term deformation extrapolation, and output the advanced quantitative risk assessment results of building settlement.

[0108] The computer-readable storage medium provided in the above embodiments of this application and the building deformation feature decoupling method based on PS-InSAR time series data provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0109] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the decoupling method, system, electronic device, and readable storage medium based on PS-InSAR time-series data are basically similar to the embodiments of the decoupling method for building deformation features based on PS-InSAR time-series data described above, and are therefore described relatively simply. Relevant parts can be referred to in the descriptions of the embodiments of the decoupling method for building deformation features based on PS-InSAR time-series data described above.

Claims

1. A method for decoupling building deformation features based on PS-InSAR time series data, characterized in that, include: The PS-InSAR vector file of the target building area is obtained and parsed to obtain PS-InSAR time-series displacement data. Simultaneously, temperature and cumulative precipitation meteorological data are collected, and spatial attributes such as building height, type, and construction year are extracted. Based on the building vector frame, the spatial scale constraint screening of single building PS points is completed. The numerical features are standardized by Z-score to eliminate dimensions. The building use category features are processed by one-hot coding. The building construction year is converted into the observation start interval time series feature and then standardized. STL or VMD decomposition is used to remove trend and seasonal preliminary estimation features from time-series displacement data, reduce non-stationary coupling of deformation signals, alleviate network overfitting, and accelerate convergence. Set a fixed window length and sliding step size to cut deformation and long-term meteorological data, generate continuous samples with temporal causal relationships, flatten and broadcast the static spatial attributes of buildings in the time dimension, and stitch them with the time-series sample channel to construct a three-dimensional standard data tensor. The constructed data tensor is input into the dual-branch LSTM model to extract deformation features in parallel. The physical-induced seasonal LSTM branch learns the hysteresis effect of temperature on building action, and introduces building height and coherence into the input layer to correct the thermoelastic response of high-rise buildings. The physical constraint of total deformation conservation is added to the loss function. The adaptive attention trend LSTM branch receives the trend estimate features obtained from pre-decomposition, and introduces a scaling dot product attention mechanism to adaptively locate the settlement acceleration and step change time window, and fit the nonlinear long-term settlement trend components caused by soil consolidation and tectonic movement. The two outputs are weighted and fused to separate the deformation components. The 3σ residual threshold is used to identify sudden structural deformation anomalies. The input of the predicted meteorological sequence is used to extrapolate the medium- and long-term deformation curves and generate advanced quantitative assessment information on building settlement risk.

2. The method as described in claim 1, characterized in that, Using STL or VMD decomposition to remove trend and seasonal preliminary features from time-series displacement data reduces the non-stationary coupling of deformation signals, alleviates network overfitting, and accelerates convergence, including: One of the algorithms, STL Local Weighted Regression or VMD Variational Mode, was selected to perform pre-decomposition processing on the standardized time-series displacement data. When using STL decomposition, the time series is split by relying on a local weighted regression smoother to output three types of preliminary estimated components: long-term geological trend, annual and seasonal cycle, and random residual. When using VMD decomposition, the variational model is solved iteratively to extract the low-frequency zero center frequency IMF as the initial trend feature and the periodic high-frequency IMF as the initial seasonal feature. By pre-decomposing and removing the two core physical features coupled within the deformation signal in advance, the non-stationarity of the original time series is reduced, the coupling interference during network training is weakened, the overfitting phenomenon of deep learning models is alleviated, and the convergence speed of the models is accelerated.

3. The method as described in claim 1, characterized in that, By setting a fixed window length and sliding step size to cut deformation and long-term meteorological data, continuous samples with temporal causal relationships are generated. The static spatial attributes of buildings are flattened and broadcast in the time dimension, and then stitched together with the time-series sample channels to construct a three-dimensional standard data tensor, including: Set a fixed observation sub-interval length W and an independent sliding step size S, and simultaneously segment the trend estimate features, seasonal estimate features, and long-term meteorological features obtained by decomposing the standardized deformation sequence to generate continuous sample segments with temporal causal relationships. For static spatial attributes including building height, building type, and construction interval, perform dimension flattening and time axis full-domain broadcast copying operations in sequence to match the duration of all time series samples; The segmented trend features, seasonal features, and meteorological dynamic samples are combined with the broadcasted static building features through channel-dimensional splicing and fusion to generate a standardized three-dimensional data tensor with dimensions of sample number, window length, feature dimension, and adapted to an LSTM temporal neural network.

4. The method as described in claim 1, characterized in that, The constructed data tensor is input into a dual-branch LSTM model to extract deformation features in parallel. The physically induced seasonal LSTM branch learns the hysteresis effect of temperature on building action, introducing building height and coherence into the input layer to correct the thermoelastic response of high-rise buildings, and adding a physical constraint of total deformation conservation to the loss function. The adaptive attention trend LSTM branch receives the initial trend estimate features obtained from pre-decomposition, introduces a scaling dot product attention mechanism to adaptively locate the settlement acceleration and step change time windows, and fits the nonlinear long-term settlement trend components induced by soil consolidation and tectonic movement, including: A dual-branch parallel LSTM feature extraction computing architecture is built based on standardized three-dimensional data tensors, and the independent modeling and calculation of seasonal and long-term deformation components are completed simultaneously in two branches. To explore the thermal elastic periodic deformation of the environment, a physical-induced LSTM branch input parsing logic is configured. The network gating unit is used to learn the hysteresis response law of temperature transmission to buildings. The building height and coherence auxiliary weights are loaded into the input layer to correct the thermal elastic response characteristics of high-rise buildings. The model loss function is embedded with the physical constraint formula of total deformation conservation. The constraint output fits the periodic deformation law of temperature and captures the temperature-driven seasonal deformation component. An adaptive LSTM branch operation logic is configured for long-term settlement excavation. The branch is connected to the trend estimation feature tensor output by STL / VMD pre-decomposition. The scaling dot product attention calculation formula is introduced to adaptively identify three types of local time windows: settlement acceleration, deceleration, and step change. The nonlinear long-term settlement trend components caused by soil consolidation and tectonic movement are accurately fitted.

5. The method as described in claim 4, characterized in that, The two outputs are weighted and fused to separate the deformation components, and the 3σ residual threshold is used to identify sudden structural deformation anomalies. Input the predicted meteorological sequence to extrapolate the medium- and long-term deformation curves, and generate advanced quantitative assessment information on building settlement risk, including: Based on the seasonal deformation components and long-term trend deformation components output by the dual-branch network, an integrated computational link for deformation decoupling and prediction is established, and a complete four-level computational logic is established, including component weighted fusion, residual term solution, 3σ anomaly discrimination, and time series forward extrapolation. We configure weighted fusion operation rules for two types of independent deformation components, perform weighted fusion operation on seasonal deformation components and long-term trend deformation components, and complete the independent decoupling and separation of two types of physical components: thermoelastic period and soil consolidation. The time-series residual terms are solved by difference calculation using the measured total deformation and the model reconstruction deformation. A three-standard-deviation threshold discrimination mechanism is introduced to identify sudden deformation anomaly samples induced by construction disturbance and building structure damage in time sequence. The future predicted meteorological sequence and time series encoding are synchronously connected to the complete dual-branch LSTM network after training convergence to build a medium- and long-term deformation forward extrapolation and inference operation link. Based on the network forward extrapolation output of long-term settlement trend curves and seasonal thermoelastic deformation evolution curves, and relying on the dual-component prediction results, an advanced quantitative assessment of the overall building safety risk is completed.

6. A decoupling system for building deformation features based on PS-InSAR time series data, characterized in that, The system is used to execute executable instructions to perform the building deformation feature decoupling method based on PS-InSAR time series data as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, include: First processor; The processor also includes a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the building deformation feature decoupling method based on PS-InSAR time series data as described in any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the building deformation feature decoupling method based on PS-InSAR time series data as described in any one of claims 1 to 5.

Images