A method and system for predicting soft soil foundation deformation based on big data

Abstract

The present invention discloses a method and system for predicting deformation of soft soil foundation based on big data, which relates to the field of geotechnical monitoring technology. InSAR satellite data, Beidou GNSS displacement data, fiber optic strain data and meteorological and geological parameters are integrated through a multi-source sensor network. ERA5 reanalysis data is used to establish an atmospheric delay compensation function, and dynamic sliding window filtering and strain gradient constraints are combined to achieve data spatiotemporal alignment and anomaly cleaning. Based on the generalized Kelvin creep constitutive model, the environmental coupling function of temperature, humidity and pore water pressure is incorporated to dynamically correct the model parameters, thereby enhancing the creep response characterization capability under complex environments. Regionalized geological data are distributedly jointly trained through a federated learning framework, and differential privacy encryption and node credibility verification mechanisms are integrated to achieve secure aggregation and migration optimization of cross-regional data, generating geological partition adaptive deformation prediction results. This method effectively improves the reliability of soft soil foundation deformation prediction.

Description

Technical Field

[0001] The present invention relates to the field of rock and soil monitoring technology, and in particular to a soft soil foundation deformation prediction method and system based on big data. Background Art

[0002] Predicting soft soil foundation deformation is a core technical challenge in geotechnical engineering safety monitoring. Its accuracy is directly related to the long-term stability of major infrastructure and the effectiveness of disaster warning. Traditional prediction methods often rely on single monitoring methods, such as GNSS displacement measurements or geological survey data, which cannot fully capture the multi-dimensional coupled characteristics of soft soil creep. Especially in complex geological areas such as coastal areas and river networks, foundation deformation is dynamically affected by multiple environmental parameters such as temperature, humidity, and pore water pressure. Existing constitutive models often use static parameter corrections, resulting in significant lag bias in long-term predictions. In addition, while new monitoring technologies such as InSAR and fiber optic sensing can provide high-density spatiotemporal data, the lack of unified spatiotemporal benchmarks for multi-source heterogeneous data, atmospheric delay interference, and abnormal noise seriously restrict the accuracy of data fusion. At the model construction level, traditional centralized training models face conflicts such as the regional sensitivity of geological data, barriers to cross-regional sharing, and privacy protection, making it difficult to achieve distributed learning that balances data security and model generalization capabilities. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention provides a soft soil foundation deformation prediction method and system based on big data.

[0004] The present invention provides a soft soil foundation deformation prediction method based on big data, comprising:

[0005] S1. Acquire deformation monitoring data of soft soil foundation through a multi-source sensor network, including InSAR satellite phase data, BeiDou GNSS displacement time series data, distributed optical fiber strain data, meteorological data, and geological survey parameters;

[0006] S2. performing spatiotemporal alignment and outlier cleaning on the deformation monitoring data, constructing an atmospheric delay compensation function using ERA5 reanalysis data, and generating a preprocessed standardized data set;

[0007] S3. Construct a generalized Kelvin creep constitutive model based on the standardized data set and introduce an environmental coupling function Dynamic correction model parameters, where T is temperature, W is humidity, and P is pore water pressure;

[0008] S4. A federated learning framework is used to perform distributed joint training on the generalized Kelvin creep constitutive model, and the data characteristics of different geological partitions are optimized through transfer learning to generate regional adaptive deformation prediction results.

[0009] Optionally, the method for constructing the atmospheric delay compensation function in S2 includes:

[0010] The total atmospheric water vapor parameters and ionospheric total electron content parameters of the monitoring area were extracted based on ERA5 reanalysis data;

[0011] The atmospheric delay compensation function is constructed by combining the residual phase noise, the total water vapor parameter, the total electron content parameter and the corresponding regional atmospheric parameter compensation coefficient.

[0012] Optionally, outlier cleaning in S2 includes the following steps:

[0013] Perform wavelet multi-scale decomposition on BeiDou GNSS displacement time series data to separate low-frequency deformation signals and high-frequency noise components;

[0014] Dynamically adjust the sliding filter window length based on the ratio of the current window's signal-to-noise ratio to the benchmark signal-to-noise ratio;

[0015] The dynamically adjusted sliding filter window is used to perform adaptive threshold filtering on the high-frequency noise components, while retaining the low-frequency deformation signals.

[0016] Optionally, outlier cleaning in S2 also includes a real-time dynamic threshold adjustment mechanism:

[0017] Based on the spatial distribution characteristics of strain gradients of adjacent distributed optical fiber sensing nodes, the strain gradient safety constraint threshold is established in combination with the node spacing, soil shear strength and elastic modulus.

[0018] When it is detected that the strain change rate of an adjacent node exceeds the safety constraint threshold, the following processing is performed: the time coverage of the current sliding filter window is automatically extended, and the window expansion amplitude is dynamically adjusted based on the duration of the anomaly; the Beidou GNSS displacement data and InSAR deformation inversion results of the corresponding time period are synchronously integrated, and the strain measurement values ​​in the abnormal area are compensated through a multi-source data joint filtering algorithm.

[0019] Optionally, the environmental coupling function in S3 is defined as:

[0020] ;

[0021] Among them, f env (T, W, P) is the environmental coupling function, T represents temperature, W represents humidity, P represents pore water pressure, W represents sat is the saturated humidity, m1, m2, and m3 are the coupling coefficients dynamically updated by the LSTM network, and t represents time. represents the pore water pressure gradient.

[0022] Optionally, the implementation of the federated learning framework in S4 includes:

[0023] S41. Each edge node trains a sub-model of the generalized Kelvin model based on local data and uses a differential privacy encryption algorithm to perform noise perturbation on the model gradient parameters.

[0024] S42. The cloud aggregation node calculates the weighted average value of the global model parameters based on the encrypted gradient parameters uploaded by each node and the proportion of the monitoring data volume of each node to the total data volume as the weighting coefficient, thereby realizing the distributed joint update of the cross-regional model parameters.

[0025] Optionally, S42 also includes edge node credibility verification:

[0026] When uploading gradient parameters, each edge node simultaneously submits the mean, variance, and difference index of the local data set with the global data distribution;

[0027] The cloud selects a set of trusted nodes that meet the consistency constraints based on a threshold of the ratio of the difference index to the data dispersion, and a tolerance range of deviation between the local mean and the global mean distribution;

[0028] When weighted aggregation is performed on the gradient parameters of the trusted node set, an exponential decay correction is performed on the data amount weight based on the difference in data distribution of each node to generate a global model parameter update amount.

[0029] In a second aspect, the present invention further provides a soft soil foundation deformation prediction system based on big data, comprising:

[0030] The acquisition module is used to obtain deformation monitoring data of soft soil foundation through a multi-source sensor network, including InSAR satellite phase data, Beidou GNSS displacement time series data, distributed optical fiber strain data, meteorological data and geological survey parameters;

[0031] A generation module is used to perform spatiotemporal alignment and outlier cleaning on the deformation monitoring data, construct an atmospheric delay compensation function using ERA5 reanalysis data, and generate a preprocessed standardized data set;

[0032] A correction module is used to construct a generalized Kelvin creep constitutive model based on the standardized data set and introduce an environmental coupling function Dynamic correction model parameters, where T is temperature, W is humidity, and P is pore water pressure;

[0033] The prediction module is used to perform distributed joint training on the generalized Kelvin creep constitutive model using a federated learning framework, optimize the data characteristics of different geological partitions through transfer learning, and generate regional adaptive deformation prediction results.

[0034] The present invention has the following technical effects:

[0035] This invention significantly improves the spatial coverage density and temporal resolution of soft soil deformation monitoring through multi-dimensional data fusion from a multi-source sensor network, overcoming the perception limitations of traditional single-source monitoring methods. An atmospheric delay compensation function constructed based on ERA5 reanalysis data effectively eliminates tropospheric and ionospheric interference in InSAR satellite phase data. Adaptive noise reduction of Beidou GNSS displacement data is achieved in combination with a dynamic sliding window filtering algorithm, ensuring the spatiotemporal consistency of multi-source heterogeneous data. An environmental coupling function is introduced to dynamically modify the parameters of the generalized Kelvin creep constitutive model, enhancing the model's sensitivity to temperature and humidity changes and pore water pressure fluctuations and addressing the hysteresis problem of static parameter models in complex environments. A federated learning framework is employed to overcome barriers to cross-regional geological data sharing. While ensuring data security through differential privacy encryption, a node credibility verification mechanism is employed to select training nodes that meet global distribution characteristics, enhancing the generalization capability of the regional adaptive model. Furthermore, strain gradient constraint equations are used to implement anomaly mutual verification of distributed fiber optic data. Combined with the multi-source data compensation mechanism of the Kalman filter, the fault tolerance and robustness of the monitoring system are enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 A schematic flow chart of a method for predicting soft soil foundation deformation based on big data provided by an embodiment of the present invention;

[0038] Figure 2 A schematic structural diagram of a soft soil foundation deformation prediction system based on big data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0040] Figure 1 A schematic flow chart of a method for predicting soft soil foundation deformation based on big data provided by an embodiment of the present invention includes:

[0041] S1. Acquire deformation monitoring data of soft soil foundation through a multi-source sensor network, including InSAR satellite phase data, BeiDou GNSS displacement time series data, distributed optical fiber strain data, meteorological data, and geological survey parameters;

[0042] S2. Perform spatiotemporal alignment and outlier cleaning on the deformation monitoring data, construct an atmospheric delay compensation function using ERA5 reanalysis data, and generate a preprocessed standardized data set;

[0043] S3. Construct a generalized Kelvin creep constitutive model based on a standardized data set and introduce an environmental coupling function Dynamic correction model parameters, where T is temperature, W is humidity, and P is pore water pressure;

[0044] S4. A federated learning framework is used to perform distributed joint training on the generalized Kelvin creep constitutive model. The data characteristics of different geological partitions are optimized through transfer learning to generate regional adaptive deformation prediction results.

[0045] When implementing this method, a multi-source sensor network monitoring system covering the soft soil foundation area is first established. Beidou GNSS base stations and monitoring stations are deployed in the monitoring area according to the geological structure characteristics to form a satellite positioning observation array, which is synchronously connected to the InSAR phase data acquisition channel of the space-borne synthetic aperture radar. Distributed fiber optic sensor cables are laid along the key sections of the roadbed, and Brillouin optical time domain reflectometry technology is used to obtain the foundation strain distribution in real time. A meteorological monitoring station is installed to collect atmospheric temperature, precipitation and air pressure data in real time, and geological parameters such as soil porosity and permeability are extracted in combination with geological survey reports. Multi-source data is connected to the edge computing node through the Internet of Things gateway to form a four-dimensional monitoring matrix including deformation displacement, strain distribution, environmental status and geological characteristics.

[0046] During the data preprocessing phase, a unified spatiotemporal benchmark approach was used to address the heterogeneity of multiple sensors. To address the differences in temporal sampling intervals between InSAR phase data and Beidou GNSS displacement data, minute-level time alignment was achieved through cubic spline interpolation. The spatial coordinate system conversion module projected the geocoded coordinate system of InSAR data, the WGS84 coordinate system of Beidou GNSS, and the engineering coordinate system of fiber strain data onto the regional construction coordinate system. A multi-stage filtering strategy was employed in the outlier cleaning process, performing wavelet multiscale decomposition on the Beidou GNSS displacement time series data to separate low-frequency deformation trend terms from high-frequency vibration noise. Adaptive noise reduction was achieved by dynamically adjusting the sliding window length based on the signal energy ratio. Spatial continuity verification was used to eliminate local abnormal jump points in the distributed fiber strain data, and strain gradient constraints of adjacent sensor nodes were used to identify distorted data caused by equipment failure or external interference.

[0047] When establishing a generalized Kelvin creep constitutive model, an environmental parameter coupling term is added to the classical viscoelastic theory. The model input layer integrates preprocessed displacement and strain time series data, and the hidden layer introduces temperature, humidity, and pore water pressure as dynamic correction factors. A long-short-term memory neural network is used to learn the correlation between environmental parameters and creep rate online, adjusting the viscoelastic coefficient and delay time parameters in real time. The model training utilizes a transfer learning strategy, using training weights from typical geological regions as initialization parameters. Fine-tuning and optimization are performed based on local monitoring data to enhance the model's adaptability to regional geological characteristics.

[0048] When deploying the federated learning framework, a cloud-based aggregation server and multiple edge computing nodes are set up. Each edge node stores local monitoring data and trains a local creep model, using a homomorphic encryption algorithm to protect the privacy of the model's gradient parameters. The cloud server periodically initiates model aggregation instructions, receives the encrypted gradients uploaded by each node, performs a weighted average operation, generates global model parameters, and issues updates. During training, a data distribution feature comparison mechanism is established. Anomalous nodes are identified through statistical analysis of the KL divergence between node data and the global model, and the aggregation weight distribution strategy is dynamically adjusted. The resulting regional adaptive prediction model can simultaneously output the immediate deformation state and long-term creep trend of the soft soil foundation, providing a decision-making basis for engineering safety assessments.

[0049] In some embodiments, the method of constructing the atmospheric delay compensation function in S2 includes:

[0050] The total atmospheric water vapor parameters and ionospheric total electron content parameters of the monitoring area were extracted based on ERA5 reanalysis data;

[0051] The atmospheric delay compensation function is constructed by combining the residual phase noise, total water vapor parameter, total electron content parameter and the corresponding regional atmospheric parameter compensation coefficient.

[0052] Specifically, the tropospheric temperature and humidity profiles and ionospheric TEC maps can be extracted based on the ERA5 reanalysis data to construct a phase delay compensation function:

[0053] ;

[0054] in, is the phase delay, PWV is the total amount of atmospheric water vapor, TEC is the total electron content of the ionosphere, k1 and k2 are compensation coefficients, is the residual phase noise.

[0055] To implement atmospheric delay compensation, ERA5 reanalysis data are first obtained from the Weather Forecast Center to extract continuous atmospheric parameters covering the monitoring area in both time and space. The total atmospheric water vapor (PWV) value is calculated by integrating the vertical profiles of geopotential height and relative humidity in the reanalysis dataset along the satellite radar wave propagation path. Gridded ionospheric total electron content (TEC) data are simultaneously acquired and mapped to the InSAR satellite imaging geometry using a bilinear interpolation algorithm to eliminate the effects of the gradient of ionospheric delay with spatial position.

[0056] During the construction of the phase delay compensation function, a regionalized parameter calibration mechanism was established based on the physical correlation between the atmospheric delay component of the InSAR interferometric phase and meteorological factors. Through regression analysis of historical monitoring data, a multivariate linear fit was performed between the InSAR phase residuals for the rainy and dry seasons in the same region and the ERA5 meteorological parameters to obtain the regional atmospheric parameter compensation coefficients k1 and k2. The PWV component reflects the tropospheric wet delay effect, while the TEC component characterizes the ionospheric group delay. These two components are weighted together to form a deterministic model compensation for atmospheric delay.

[0057] Residual Phase Noise The processing method combines spatial filtering with time series analysis. Wavelet multiresolution decomposition is performed on the compensated phase residuals, extracting low-frequency phase components correlated with terrain height for terrain-related atmospheric correction. High-frequency residual components are suppressed by constructing an adaptive filter, whose parameters are dynamically adjusted based on the stability of the atmospheric state during the satellite revisit period. A TEC change rate threshold monitoring mechanism is established to address sudden ionospheric disturbances. When the TEC time derivative exceeds a preset warning value, a real-time update process for the compensation coefficient k2 is automatically triggered.

[0058] Dynamic optimization of regional atmospheric parameter compensation coefficients is achieved through an online learning module. Within the Beidou GNSS base station coverage area, the zenith tropospheric delay derived from GNSS is cross-validated with the PWV values ​​calculated using ERA5. The linear correlation coefficient between the two is calculated and used as the correction factor for k1. TEC observations at ionospheric penetration points from dual-frequency GNSS receivers are simultaneously analyzed for spatial matching with the TEC map provided by ERA5. The regional weighting for k2 is adjusted based on the matching error distribution. This mechanism ensures that the compensation model can adapt to seasonal and sudden changes in the atmospheric environment.

[0059] The application of the compensation model adopts a staged processing strategy. In the InSAR data processing front end, Δφ is used atmThe function performs an initial atmospheric correction on the raw interferometric phase, eliminating large-scale spatially correlated delay components. In the deformation inversion backend, the residual phase is iteratively optimized based on the deformation model predictions. The coupling parameters between the deformation rate and the residual atmospheric phase are solved through least-squares adjustment. The final output deformation monitoring data is verified for accuracy by comparing it with the displacement values ​​measured by Beidou GNSS base stations, forming a closed-loop quality control system for atmospheric delay compensation.

[0060] In some embodiments, outlier cleaning in S2 includes the following steps:

[0061] Perform wavelet multi-scale decomposition on BeiDou GNSS displacement time series data to separate low-frequency deformation signals and high-frequency noise components;

[0062] Dynamically adjust the sliding filter window length based on the ratio of the current window's signal-to-noise ratio to the benchmark signal-to-noise ratio;

[0063] The dynamically adjusted sliding filter window is used to perform adaptive threshold filtering on the high-frequency noise components, while retaining the low-frequency deformation signals.

[0064] Specifically, the outlier cleaning in S2 adopts a dynamic sliding window filtering algorithm:

[0065] The BeiDou GNSS displacement data is decomposed into low-frequency deformation signals and high-frequency noise through wavelet transform, and the sliding window length is dynamically adjusted according to the signal-to-noise ratio:

[0066] ;

[0067] Among them, L w represents the sliding window length, SNR current is the current signal-to-noise ratio, SNR base is the reference signal-to-noise ratio, L base is the initial window length.

[0068] When cleaning outliers in Beidou GNSS displacement data, the raw observation data is first preprocessed for multipath suppression and cycle slip repair. Dual-frequency carrier phase observations are used to construct a geometry-free combined observation equation. Cycle slip events are detected through statistical analysis of the carrier phase change rate within a sliding window, and phase discontinuities are repaired using the Kalman smoothing algorithm. After completing the data integrity check, the displacement time series is input into a wavelet transform module, where Daubechies wavelet basis functions are used for multiscale decomposition, decomposing the signal into a low-frequency component representing the creep trend of the foundation and a high-frequency component containing measurement noise and high-frequency vibrations.

[0069] The number of wavelet decomposition layers is determined by the relationship between the signal sampling frequency and the characteristic frequency of geological creep. For Beidou GNSS displacement data sampled at the minute level, a five-layer decomposition structure is used to divide the signal frequency band into a main creep frequency band below 0.5 Hz and a noise frequency band above 0.5 Hz. The current signal-to-noise ratio (SNR) is dynamically assessed by calculating the signal energy contribution of each frequency band. When the energy contribution of the high-frequency sub-band exceeds a preset threshold, the data is deemed to be significantly affected by transient interference noise.

[0070] The calculation of the dynamic sliding window length follows the formula The implementation logic of the initial window length L base According to the typical vibration period of the engineering site, the baseline signal-to-noise ratio SNR base The ratio of low-frequency component energy to high-frequency component energy of the historical clean data set is used for calibration. The instantaneous ratio of low-frequency component energy to high-frequency noise energy in the current sliding window is calculated in real time as SNR. current When the signal-to-noise ratio decreases, the window length is proportionally expanded to enhance the filtering smoothing effect, and when the signal-to-noise ratio rises, the window is contracted to retain effective high-frequency details.

[0071] The sliding window filter uses an improved Savitzky-Golay convolution algorithm, fitting a quadratic polynomial within the window to achieve trend smoothing. During window length adjustment, a boundary extension algorithm is used to avoid signal distortion at the beginning and end of the data sequence. For the extended window, a piecewise weighting strategy is used to reduce the weight of edge data points to suppress filter ringing caused by sudden window changes. High-frequency noise components are threshold filtered and then fused with low-frequency trend components using a wavelet reconstruction algorithm to generate denoised displacement time series data.

[0072] The filtering effect is verified by using a multi-sensor cross-comparison mechanism to perform a consistency check on the time domain integration operation of the processed Beidou GNSS displacement data and the distributed optical fiber strain data at the same point. When the difference in the cumulative deformation of the two within the sliding window exceeds the allowable range, the re-optimization process of the window length parameter is triggered, and L is dynamically corrected through the feedback adjustment mechanism. base At the same time, a noise energy spectrum feature library is established to perform pattern recognition and classification on high-frequency noise components, distinguishing different types of noise sources such as inherent equipment noise, vehicle vibration interference, and rainfall impact, providing environmental context information for window adjustment strategies.

[0073] In some embodiments, the outlier cleaning in S2 also includes a real-time dynamic threshold adjustment mechanism:

[0074] Based on the spatial distribution characteristics of strain gradients of adjacent distributed optical fiber sensing nodes, the strain gradient safety constraint threshold is established in combination with the node spacing, soil shear strength and elastic modulus.

[0075] When it is detected that the strain change rate of an adjacent node exceeds the safety constraint threshold, the following processing is performed: the time coverage of the current sliding filter window is automatically extended, and the window expansion amplitude is dynamically adjusted based on the duration of the anomaly; the Beidou GNSS displacement data and InSAR deformation inversion results of the corresponding period are synchronously integrated, and the strain measurement values ​​in the abnormal area are compensated through a multi-source data joint filtering algorithm.

[0076] Specifically, outlier cleaning in S2 also includes a real-time dynamic threshold adjustment mechanism:

[0077] Based on the spatial continuity characteristics of distributed optical fiber strain data, the strain gradient constraint equation of adjacent sensor nodes is constructed:

[0078] ;

[0079] in, represents the strain gradient, ε i , ε j is the strain value of the adjacent node, d i,j is the node spacing, σ ult is the ultimate shear strength of soil, E soil is the elastic modulus, α is the safety factor, represents the strain gradient threshold;

[0080] When the strain gradient exceeds the threshold, the following actions are triggered:

[0081] 1) Automatically extend the sliding window length to , where β is the time decay factor, Δt is the duration of the abnormality, Indicates the adjusted sliding window length;

[0082] 2) Aggregate BeiDou GNSS displacement data and InSAR phase data, and perform data compensation through Kalman filtering.

[0083] When implementing anomaly detection for distributed fiber-optic strain data, a strain gradient constraint network based on spatial continuity is first established. Sensing nodes are deployed at equal intervals along the fiber-optic sensing cable. The spacing between adjacent nodes is determined based on the uniformity of the engineering geology, and the node density is increased in areas of sudden soil changes. Each node collects axial strain values ​​in real time. Data exchange channels are established between adjacent nodes via internal communication links within the fiber-optic cable, forming a local topological structure for strain gradient calculation.

[0084] The physical meaning of the strain gradient constraint equation is to quantify the strain change rate of adjacent nodes. When calculating, the average strain of three consecutive sampling moments is taken as the current ε i , ε j Input value, node spacing d i,jThe precise value is obtained through the construction record of the optical cable laying. The threshold setting stage introduces a dynamic correction mechanism for soil mechanical parameters, and the ultimate shear strength of the soil obtained from geological survey is converted to ult and elastic modulus E soil The input calculation module adjusts the safety factor α based on real-time pore water pressure monitoring values. When the pore water pressure increases, causing the effective stress in the soil to decrease, the α value is automatically reduced to tighten the strain gradient threshold.

[0085] The anomaly triggering mechanism utilizes a multi-condition joint decision strategy. When the strain gradient of a single node and two or more adjacent nodes simultaneously exceeds the limit, the region is deemed to have experienced a true abnormal deformation. If the gradient of only one node exceeds the limit, a self-test program is initiated to verify the working status of the fiber optic interrogator. Once a data anomaly is confirmed, a sliding window length expansion operation is executed. The time decay factor β in the window increment ΔL = β·Δt is dynamically adjusted based on the duration of the anomaly. A larger β value is used for rapid window expansion during the initial anomaly phase, while the β value decays exponentially over time during persistent anomalies to prevent excessive window expansion.

[0086] Expanded sliding window A multi-source data fusion compensation mechanism is introduced into data filtering. Beidou GNSS displacement data and InSAR deformation inversion results are simultaneously retrieved from the same region during the anomaly period. A spatiotemporal registration module is used to unify these three data types to the same spatiotemporal reference. The Kalman filter's state equation is constructed as a deformation rate-acceleration model, and the observation equation integrates three observations: fiber strain-integrated displacement, GNSS displacement, and InSAR deformation. The filter gain matrix dynamically assigns weights to each data source based on its confidence level. During periods of fiber data anomalies, its weight is reduced, thereby enhancing the compensation effect of GNSS and InSAR data.

[0087] During the data compensation process, a residual convergence monitoring mechanism is established. When the modulus of the Kalman filter innovation vector exceeds the stable range for multiple consecutive cycles, the sensor cross-calibration process is triggered. The strain field in the abnormal area is reconstructed using redundant node data from distributed optical fibers. The optimal strain distribution is solved using the least squares fitting method. The reconstructed result is compared with the original observation data to generate a calibration compensation value. The calibrated strain data is re-entered into the strain gradient constraint equation for secondary verification until the system determines that the anomaly has been eliminated and the normal sliding window length is restored. The closed-loop control logic formed by this process effectively maintains the consistency of the multi-source data system and avoids the cascading error diffusion caused by local anomalies.

[0088] In some embodiments, the environmental coupling function in S3 is defined as:

[0089] ;

[0090] Among them, f env(T, W, P) is the environmental coupling function, T represents temperature, W represents humidity, P represents pore water pressure, W represents sat is the saturated humidity, m1, m2, and m3 are the coupling coefficients dynamically updated by the LSTM network, and t represents time. represents the pore water pressure gradient.

[0091] Environmental coupling function f env In the equation, the coefficient m1 (s - ¹)、m2(dimensionless)、m3(m -1 ) to achieve the unification of the dimensions of each component, and the final output unit is strain rate (s - ¹), which is compatible with the constitutive equations of the generalized Kelvin model.

[0092] When constructing the environmental coupling function, a multi-physics parameter acquisition system was first established. Temperature parameters were acquired using an array of thermistors embedded at varying depths in the soil. Vertical temperature measurement chains were deployed according to the layered structure of the soft soil foundation to simultaneously measure the surface atmospheric temperature and the underground temperature gradient. Moisture parameters were measured using a combination of dielectric constant sensors and matrix suction sensors. The former monitored the volumetric moisture content of the soil, while the latter measured the water potential energy in the unsaturated zone using a tensiometer. Pore water pressure was acquired using a network of vibrating-wire piezometers, forming a three-dimensional groundwater level observation system within the monitoring section. Spatial interpolation correction was performed using water content distribution maps inverted by ground-based radar.

[0093] The physical significance of the environmental coupling function is to quantify the time-varying effect of environmental excitation on soil creep. The five-point central difference format is used to suppress the interference of measurement noise through differential calculation to reflect the transient effect of thermal expansion and contraction on the viscoelasticity of soil. sat ) Saturated humidity W sat Through dynamic calculation of soil-water characteristic curve and nonlinear correction of porosity parameters monitored in real time, the regulatory effect of water content change on effective stress can be accurately characterized. Finite element interpolation method was used to calculate and construct the pressure field contour map based on the piezometer network data, and the maximum principal stress direction gradient value at the monitoring point was extracted.

[0094] A long short-term memory neural network is used to dynamically update the coupling coefficients m1, m2, and m3. The network input layer contains creep rate observations, a time series of environmental parameters, and soil constitutive parameters. The hidden layer is designed as a two-layer gated recurrent unit structure. The first layer extracts the short-term fluctuation correlation characteristics between environmental parameters and creep rate, while the second layer captures the long-term dependence of geological parameters on coupling effects. The output layer generates a normalized weight distribution of the coupling coefficients using a softmax function, which is then denormalized and mapped to the actual physical dimension range. A transfer learning strategy is used in the training phase, using the network weights pre-trained with laboratory triaxial creep test data as the initial values ​​and then fine-tuned online using field monitoring data.

[0095] The real-time update mechanism is implemented through a sliding time window. During each forecast period, the latest environmental monitoring data and creep observations are fed into the LSTM network to calculate coupling coefficient increments Δm1, Δm2, and Δm3. A momentum factor is introduced into the update process to suppress parameter abrupt changes. An exponentially weighted moving average algorithm is used to smooth the coefficient change curve and prevent model oscillation caused by localized anomalous data. A coefficient change tracking module is also established. When the m2 value is detected to deviate continuously from the historical mean by more than three standard deviations, the humidity sensor calibration process is automatically triggered to ensure dimensional consistency of the environmental coupling term.

[0096] The model parameter correction process is deeply coupled with the constitutive equation solver. At each iteration of the generalized Kelvin model, the calculated creep rate is compared with the output of the environmental coupling function, and the relaxation time parameters of the viscoelastic elements are adjusted using an adaptive step-size algorithm. When a sudden environmental change causes the output of the coupling function to jump in magnitude, the implicit Euler method is automatically used to solve the differential equations, enhancing the model's numerical stability in unsteady environments. The corrected model parameters are verified in a closed-loop manner by comparing the integrated results of the fiber strain data with the GNSS displacement monitoring values. Exceeding the residual limit triggers the retraining process of the LSTM network.

[0097] In some embodiments, the implementation of the federated learning framework in S4 includes:

[0098] S41. Each edge node trains a sub-model of the generalized Kelvin model based on local data and uses a differential privacy encryption algorithm to perform noise perturbation on the model gradient parameters.

[0099] S42. The cloud aggregation node calculates the weighted average value of the global model parameters based on the encrypted gradient parameters uploaded by each node and the proportion of the monitoring data volume of each node to the total data volume as the weighting coefficient, thereby realizing the distributed joint update of the cross-regional model parameters.

[0100] Specifically, the implementation of the federated learning framework in S4 includes:

[0101] S41. Each edge node trains a sub-model based on local data and uses differential privacy to encrypt gradient parameters.

[0102] S42. The cloud aggregation node updates the global model parameters by weighted average aggregation gradient:

[0103] ;

[0104] in, is the updated global parameter model, i represents the i-th node, N is the total number of nodes, is the local model gradient parameter, n i is the data volume of the i-th node, n total is the total data volume.

[0105] During the deployment phase of the federated learning framework, a distributed architecture consisting of a cloud-based coordination server and multiple edge computing nodes was established. Each edge node was deployed in the project area data center and equipped with an independent data storage unit and model training engine. It stored deformation monitoring data of the local soft soil foundation and a historical database of environmental parameters. At the start of the training cycle, the edge node downloaded the current global model parameters from the cloud as initialization weights and performed forward and backward propagation calculations using the local dataset to generate the gradient tensor matrix. The gradient parameters were differentially privacy-preserving before being transmitted. Privacy protection was achieved by adding noise perturbations that satisfied the Laplace distribution. The noise amplitude was dynamically adjusted based on the data sensitivity level to ensure that attackers could not reverse engineer the original monitoring data.

[0106] The cloud server sets the trigger conditions for model aggregation. When the number of online edge nodes reaches the minimum collaborative training threshold or the preset time window, a global aggregation instruction is initiated. Each node synchronously uploads the encrypted local gradient parameters and the data distribution feature vector. The feature vector contains the mean vector, covariance matrix, and KL divergence value of the local data set with the global model prediction result. The cloud verification module parses the feature vector and performs a two-level node screening: the first level calculates the ratio of the KL divergence value to the variance of each node to filter out abnormal nodes whose data distribution deviates significantly from the mainstream; the second level compares the node data mean with the range of three times the standard deviation of the global distribution to eliminate outlier nodes with systematic deviations.

[0107] When performing weighted averaging operations on the gradient parameters of a set of trusted nodes, an exponential decay adjustment strategy is used to optimize weight distribution. Initial weights are determined based on the data volume contribution of each node, and a KL divergence penalty factor is introduced to dynamically decay the weights. The weight coefficients in the aggregation formula are modified to the product of the original data volume weight and the divergence exponent, ensuring that nodes with high-quality and consistent data distribution dominate the model update. The aggregation process utilizes a tensor operation protocol under homomorphic encryption, completing weighted summation operations in ciphertext space to prevent the plaintext gradients from being intercepted and parsed during transmission.

[0108] After global model parameters are updated, they are distributed to all participating edge nodes through a version control mechanism. Upon receiving the new parameters, the nodes perform a local model hot update, retaining some historical parameters as the base model for ensemble learning. A voting mechanism is used to balance global consistency with regional characteristics. For edge nodes with unique geological conditions, regularization constraints are introduced during the local training phase to limit model updates to within the elastic bounds of the global parameters, preventing overfitting to regional noise. A model degradation monitoring mechanism is also established. When the accuracy of the local validation set continuously decreases, a global model rollback process is triggered, loading the historically optimized parameters to restore prediction stability.

[0109] During the training process, a secure multi-party computation channel is established. The cloud only captures gradient statistics, preventing edge nodes from observing the data distribution of other participants. Data feature vectors are compared using a secure multi-party computation protocol, and KL divergence values ​​are jointly calculated using an obfuscated circuit. Participants complete the screening of trusted nodes without disclosing local data details. To address abnormalities such as communication interruptions, a model parameter caching and breakpoint-resume mechanism are designed to ensure that the federated learning process is resilient to network fluctuations. The resulting regional adaptive deformation prediction model integrates multi-regional engineering experience while ensuring the privacy of monitoring data, significantly improving the generalization prediction accuracy for areas not participating in training.

[0110] In some implementations, S42 also includes edge node credibility verification:

[0111] When uploading gradient parameters, each edge node simultaneously submits the mean, variance, and difference index of the local data set with the global data distribution;

[0112] The cloud selects a set of trusted nodes that meet consistency constraints based on the ratio threshold of the difference index and the degree of data dispersion, as well as the deviation tolerance range between the local mean and the global mean distribution;

[0113] When performing weighted aggregation on the gradient parameters of the trusted node set, the data volume weight is exponentially attenuated based on the difference in data distribution of each node to generate the global model parameter update amount.

[0114] Specifically, S42 adds edge node credibility verification before gradient aggregation:

[0115] 1) Each node uploads gradient parameters When , the local data distribution feature vector is submitted synchronously ,in is the local data mean, is the local data variance, KL i is the KL divergence between the local data distribution and the global data distribution;

[0116] 2) The cloud filters the trusted node set through the following formula :

[0117] ;

[0118] Among them, δ is the threshold of the ratio of KL divergence to variance, is the tolerance multiple of mean deviation, μ g , σ g is the mean and standard deviation of the global data distribution;

[0119] 3) Aggregation The gradient of the midpoint, dynamically adjust the weight coefficient to , where λ is the attenuation coefficient, KL i is the KL discreteness, and the final global model parameter formula is .

[0120] When implementing edge node credibility verification, we first establish a standardized process for generating data distribution feature vectors. After the local training cycle is completed, each edge node extracts the time series statistics of displacement, strain and environmental parameters from the preprocessed monitoring data set. The data trend characteristics are retained by sliding window mean filtering calculation; variance The KL divergence value KL is calculated by using the detrended fluctuation analysis method to eliminate the influence of seasonal changes; i By comparing the local data distribution with the global distribution histogram sent from the cloud, the kernel density estimation method is used to generate the probability density function and calculate the relative entropy value.

[0121] Eigenvector The transmission adopts a lightweight encapsulation protocol, and an independent data segment is opened in the gradient parameter upload channel for carrying. After the cloud verification module receives the feature vector, it starts the multi-dimensional joint analysis process: first, the KL divergence value is normalized and divided by the local difference , eliminating the impact of data size differences on the discrimination threshold. The preset δ threshold is dynamically adjusted based on the 95th percentile of the KL / σ² ratio distribution of normal nodes in the historical training cycle, ensuring that the screening condition adapts to gradual changes in the data distribution.

[0122] The mean deviation check uses a sliding reference window mechanism. The cloud maintains the global data mean μ g and standard deviation σ g The rolling update queue retains the statistics of the latest N cycles after each aggregation and calculates the current reference value by exponential weighting. i With μ g The absolute difference exceeds times σ gWhen the deviation direction is consistent with the geological zoning law, it is determined to be a reasonable regional characteristic and the deviation is relaxed. If the deviation appears random, it is classified as a data anomaly.

[0123] Trusted node set The generation of the trustworthy nodes utilizes a combination of soft and hard decision strategies. Initial screening involves retaining candidate nodes through threshold comparison. Cluster analysis is then performed on the candidate node set to identify regional node clusters. Isolated nodes undergo secondary verification, testing the spatial correlation of their data distribution characteristics with adjacent geographic nodes to eliminate pseudo-anomalous nodes caused by communication failures. The final list of trusted nodes is stored on the blockchain, ensuring traceability of the screening process.

[0124] The dynamic weight adjustment link introduces the divergence attenuation factor, which reduces the original data weight n i and exp(-λ·KL i ) are multiplied together to achieve penalized downweighting. The attenuation coefficient λ is adaptively adjusted based on the global model convergence state: a larger λ value is used in the early stages of training to accelerate convergence, while a smaller λ value is used later to preserve regional characteristics. During weight normalization, weight compensation is applied to regions with fewer edge nodes to prevent data volume differences from biasing the model towards high-frequency monitoring areas.

[0125] The trusted node gradient aggregation process utilizes a fault-tolerant synchronization mechanism, setting a gradient reception timeout window and data verification retransmission requests. For nodes with missing gradients, a historical gradient interpolation algorithm is used to complete the parameter matrix to ensure aggregation integrity. A blacklist mechanism is also established. Nodes that fail verification for multiple consecutive cycles trigger an offline diagnostic process, requiring the operation and maintenance terminal to intervene for device calibration or data traceability. This mechanism forms a closed-loop quality control system, ensuring the efficiency of federated learning while maintaining the reliability of model updates.

[0126] In some implementations, to further improve data fusion reliability, the present invention introduces a multi-source data mutual verification method based on a variational autoencoder:

[0127] A deep generative model is constructed, taking BeiDou GNSS displacement time series, fiber strain spatial distribution, and InSAR deformation inversion results as joint inputs. After the encoder extracts the potential feature vectors, the decoder reconstructs the physical quantities of each data source. The reconstruction loss function is defined as:

[0128] ;

[0129] in, To reconstruct the loss function, x here i It can be x1, x2, x3, representing three types of data (Beidou GNSS displacement time series, optical fiber strain spatial distribution and InSAR deformation inversion results). To reconstruct the output data vector, is the original output data vector, is the dynamic weight coefficient, is the squared error term. When the reconstruction error of any data source exceeds 2 times the historical standard deviation, the data source is automatically downgraded and the redundant sensor data compensation process is started.

[0130] In some embodiments, a physics-data dual-driven correction of the creep model may also be introduced, specifically:

[0131] Physical Information Neural Network (PINN) embedded in the generalized Kelvin model:

[0132] The constitutive equation is incorporated into the network structure as a hard constraint to construct a hybrid loss function:

[0133] ;

[0134] in is the mixed loss function, is the mean square error between the predicted value and the measured value, is the residual term of the control equation, and λ1 and λ2 are coefficients. The ratio of λ1 and λ2 is dynamically adjusted using an adaptive weighting algorithm. When sudden changes in environmental parameters increase data noise, the physical constraint weight is increased to above 0.7 to ensure that the model evolution conforms to the basic laws of soil mechanics.

[0135] Figure 2 A schematic diagram of the structure of a soft soil foundation deformation prediction system based on big data provided by an embodiment of the present invention includes:

[0136] Acquisition module 100, for acquiring deformation monitoring data of soft soil foundation through a multi-source sensor network, including InSAR satellite phase data, BeiDou GNSS displacement time series data, distributed optical fiber strain data, meteorological data and geological survey parameters;

[0137] Generation module 200 is used to perform spatiotemporal alignment and outlier cleaning on the deformation monitoring data, construct an atmospheric delay compensation function using ERA5 reanalysis data, and generate a preprocessed standardized data set;

[0138] A correction module 300 is used to construct a generalized Kelvin creep constitutive model based on a standardized data set and introduce an environmental coupling function to dynamically correct model parameters, where T is temperature, W is humidity, and P is pore water pressure;

[0139] The prediction module 400 is used to perform distributed joint training on the generalized Kelvin creep constitutive model using a federated learning framework, optimize the data features of different geological partitions through transfer learning, and generate regional adaptive deformation prediction results.

[0140] The system provided by the embodiment of the present invention has the same technical features as the above method, and therefore can solve the same technical problems and achieve the same technical effects, which will not be described in detail here.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A soft soil foundation deformation prediction method based on big data, characterized in that: include: S1. Acquire deformation monitoring data of soft soil foundation through a multi-source sensor network, including InSAR satellite phase data, BeiDou GNSS displacement time series data, distributed optical fiber strain data, meteorological data, and geological survey parameters; S2. performing spatiotemporal alignment and outlier cleaning on the deformation monitoring data, constructing an atmospheric delay compensation function using ERA5 reanalysis data, and generating a preprocessed standardized data set; The method for constructing the atmospheric delay compensation function includes: The total atmospheric water vapor parameters and ionospheric total electron content parameters of the monitoring area were extracted based on ERA5 reanalysis data; The atmospheric delay compensation function is constructed by combining the residual phase noise, the total water vapor parameter, the total electron content parameter, and the corresponding regional atmospheric parameter compensation coefficient; S3. Construct a generalized Kelvin creep constitutive model based on the standardized data set and introduce an environmental coupling function Dynamic correction model parameters, where T is temperature, W is humidity, and P is pore water pressure; The environmental coupling function is defined as: ； Among them, f env (T, W, P) is the environmental coupling function, T represents temperature, W represents humidity, P represents pore water pressure, W represents sat is the saturation humidity, m1, m2, and m3 are the coupling coefficients dynamically updated through the LSTM network, t represents time, and ∇P represents the pore water pressure gradient; S4. A federated learning framework is used to perform distributed joint training on the generalized Kelvin creep constitutive model, and the data characteristics of different geological partitions are optimized through transfer learning to generate regional adaptive deformation prediction results.

2. The method according to claim 1, characterized in that Outlier cleaning in S2 includes the following steps: Perform wavelet multi-scale decomposition on BeiDou GNSS displacement time series data to separate low-frequency deformation signals and high-frequency noise components; Dynamically adjust the sliding filter window length based on the ratio of the current window's signal-to-noise ratio to the benchmark signal-to-noise ratio; The dynamically adjusted sliding filter window is used to perform adaptive threshold filtering on the high-frequency noise components, while retaining the low-frequency deformation signals.

3. The method according to claim 2, characterized in that Outlier cleaning in S2 also includes a real-time dynamic threshold adjustment mechanism: Based on the spatial distribution characteristics of strain gradients of adjacent distributed optical fiber sensing nodes, the strain gradient safety constraint threshold is established in combination with the node spacing, soil shear strength and elastic modulus. When it is detected that the strain change rate of an adjacent node exceeds the safety constraint threshold, the following processing is performed: the time coverage of the current sliding filter window is automatically extended, and the window expansion amplitude is dynamically adjusted based on the duration of the anomaly; the Beidou GNSS displacement data and InSAR deformation inversion results of the corresponding time period are synchronously integrated, and the strain measurement values ​​in the abnormal area are compensated through a multi-source data joint filtering algorithm.

4. The method according to claim 1, wherein The implementation of the federated learning framework described in S4 includes: S41. Each edge node trains a sub-model of the generalized Kelvin model based on local data and uses a differential privacy encryption algorithm to perform noise perturbation on the model gradient parameters. S42. The cloud aggregation node calculates the weighted average value of the global model parameters based on the encrypted gradient parameters uploaded by each node and the proportion of the monitoring data volume of each node to the total data volume as the weighting coefficient, thereby realizing the distributed joint update of the cross-regional model parameters.

5. The method according to claim 4, characterized in that S42 also includes edge node credibility verification: When uploading gradient parameters, each edge node simultaneously submits the mean, variance, and difference index of the local data set with the global data distribution; The cloud selects a set of trusted nodes that meet the consistency constraints based on a threshold of the ratio of the difference index to the data dispersion, and a tolerance range of deviation between the local mean and the global mean distribution; When weighted aggregation is performed on the gradient parameters of the trusted node set, an exponential decay correction is performed on the data amount weight based on the difference in data distribution of each node to generate a global model parameter update amount.

6. A soft soil foundation deformation prediction system based on big data, characterized in that: include: The acquisition module is used to obtain deformation monitoring data of soft soil foundation through a multi-source sensor network, including InSAR satellite phase data, Beidou GNSS displacement time series data, distributed optical fiber strain data, meteorological data and geological survey parameters; A generation module is used to perform spatiotemporal alignment and outlier cleaning on the deformation monitoring data, construct an atmospheric delay compensation function using ERA5 reanalysis data, and generate a preprocessed standardized data set; The method for constructing an atmospheric delay compensation function includes: extracting a total atmospheric water vapor parameter and a total ionospheric electron content parameter of a monitoring area based on ERA5 reanalysis data; and constructing the atmospheric delay compensation function by combining residual phase noise, the total atmospheric vapor parameter, the total electron content parameter, and corresponding regional atmospheric parameter compensation coefficients. A correction module is used to construct a generalized Kelvin creep constitutive model based on the standardized data set, and introduce an environmental coupling function to dynamically correct model parameters, where T is temperature, W is humidity, and P is pore water pressure; the environmental coupling function is defined as: ; Among them, f env (T, W, P) is the environmental coupling function, T represents temperature, W represents humidity, P represents pore water pressure, W represents sat is the saturation humidity, m1, m2, and m3 are the coupling coefficients dynamically updated through the LSTM network, t represents time, and ∇P represents the pore water pressure gradient; The prediction module is used to perform distributed joint training on the generalized Kelvin creep constitutive model using a federated learning framework, optimize the data characteristics of different geological partitions through transfer learning, and generate regional adaptive deformation prediction results.

Images