Numerical simulation coupling-based foundation pit deformation prediction and early warning method

Abstract

The application discloses a kind of based on numerical simulation coupling foundation pit deformation prediction and early warning method, to solve the noise of coupling numerical simulation prediction and field monitoring data, missing and the change of working condition lead to early warning threshold static, false alarm and miss report and lack of credibility quantization problem, the initial coupling numerical model is established by obtaining the monitoring data with time mark and construction record in the application, data quality vector is generated by executing self-supervision anomaly detection to monitoring data and carries out abnormal point suppression and missing completion;According to data quality vector, the working condition change point detection of multiple sensor weight and consistency gate is carried out, and the working condition state machine is updated in combination with hysteresis rule;According to this, the working condition and boundary condition of coupling numerical model are updated and are corrected on line with quality weighted residual, and the reference threshold is generated by rolling prediction, and the threshold is self-adaptively modified by data quality and working condition, and the credibility of monitoring, working condition identification, model consistency and numerical stability is decomposed and synthesized, early warning and review shunt disposal are carried out according to overall credibility, and the technical effects of dynamic threshold early warning, reducing false alarm and miss report and quantifying early warning credibility are realized.

Description

Technical Field

[0001] This invention relates to the field of underground engineering safety monitoring and early warning, and in particular to a method for predicting and warning of foundation pit deformation based on numerical simulation coupling. Background Technology

[0002] During the construction of foundation pit projects, factors such as excavation and unloading, support installation and removal, dewatering initiation and shutdown, and the surrounding environment affect monitoring indicators such as retaining structure displacement, surface settlement, support axial force, and groundwater level, resulting in significant phased and time-varying characteristics. Current engineering practices commonly employ automated monitoring systems to acquire multi-source monitoring data and issue alarms for exceeding limits based on standard limits, empirical thresholds, or statistical thresholds. Simultaneously, numerical simulation methods such as the finite element method are used for foundation pit deformation prediction and scheme comparison. Some technologies further incorporate parameter inversion, data assimilation, or rolling prediction to achieve matching and updating between the model and the actual site conditions. In recent years, data-driven methods such as anomaly identification, missing data completion, and deformation trend analysis based on monitoring data have also been gradually applied to foundation pit safety management.

[0003] Existing technologies still have shortcomings in engineering applications:

[0004] 1. Monitoring data is often affected by sensor noise, drift, jumps, missing measurements and communication interruptions. Existing methods mainly rely on simple filtering, manual removal or single-point rule processing, lacking online quantification of data quality and quality-based multi-source fusion mechanisms, resulting in insufficient reliability of early warning and model correction.

[0005] 2. Construction conditions and boundary conditions change frequently with each stage. Existing methods mostly rely on manual input of construction nodes or post-event interpretation, lacking automatic change point identification based on multi-sensor consistency and a working condition confirmation mechanism with hysteresis. This can easily lead to delays or misjudgments in working condition switching, affecting the stage updates and prediction accuracy of the numerical model.

[0006] 3. Most early warning systems still use static thresholds or single threshold strategies, which are difficult to adapt to changes in operating conditions and data quality. Furthermore, they lack reliable quantification and triage mechanisms for monitoring, operating condition identification, model consistency, and numerical calculation stability, which can easily lead to false alarms, missed alarms, or difficulty in guiding subsequent review and handling.

[0007] Therefore, a method for predicting and warning of foundation pit deformation that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a method for predicting and warning of foundation pit deformation based on numerical simulation coupling. Addressing the problems in existing technologies, such as poor data reliability due to noise, drift, jumps, and missing data in monitoring data; reliance on manual interpretation of changes in construction conditions and boundary conditions leading to delayed model updates; and the prevalence of static warning thresholds lacking reliable quantification, resulting in frequent false alarms and missed alarms, this invention proposes a technical solution based on "self-supervised quality control of monitoring data - detection of changes in working conditions - automatic switching and online correction of coupled numerical models - dynamic threshold adaptation - reliability-driven warning triage." This solution generates a data quality vector through self-supervised anomaly detection and achieves anomaly suppression and missing data completion; detects changes in working conditions based on quality-weighted multi-sensor consistency gating and updates the working condition state machine using hysteresis rules; smoothly updates the working conditions and boundary conditions of the coupled numerical model according to the working condition status; and uses quality-weighted residual iterative correction of model parameters to generate a baseline threshold through rolling prediction. The threshold is then adjusted based on data quality and working conditions. Simultaneously, monitoring reliability, working condition identification reliability, model consistency reliability, and numerical stability reliability are decomposed and synthesized to achieve triage of warning and review handling. This invention has the technical effects of dynamic threshold adaptation, improved model prediction accuracy, quantifiable early warning reliability, and reduced false alarms and missed alarms.

[0009] This invention provides a method for predicting and warning of foundation pit deformation based on numerical simulation coupling, comprising:

[0010] S1. Obtain time-stamped monitoring data and construction records for the foundation pit project, and establish an initial coupled numerical model based on the survey data, design parameters, and construction records; S2. Perform self-supervised anomaly detection on the monitoring data, generate a data quality vector corresponding to the monitoring data, suppress outliers and complete missing data in the monitoring data to obtain quality control monitoring data; S3. Set multi-sensor weights for the quality control monitoring data based on the data quality vector, and under its constraints, perform condition change point detection based on multi-sensor consistency on the quality control monitoring data, generate candidate points for condition changes, and calculate the condition... S4. Confirm candidate points for change points according to preset hysteresis judgment rules, update the working condition state machine and generate the current working condition state, and generate a coupled numerical model update instruction; S5. Perform a smooth transition update of the working condition and boundary conditions on the initial coupled numerical model according to the coupled numerical model update instruction, compare the calculation results of the quality control monitoring data with the updated coupled numerical model, weight the comparison residuals according to the data quality vector to obtain the weighted residuals, use them as the target to iteratively update the model parameters, obtain the coupled numerical model after online correction, perform rolling prediction, obtain the prediction results, and generate a numerical stability index characterizing the stability of the rolling prediction calculation; S6. For the preset early warning index, calculate the benchmark threshold according to the prediction results. The benchmark threshold is the prediction envelope threshold or prediction quantile threshold of the preset early warning index in the prediction time domain; S7. Correct the benchmark threshold according to the data quality vector and the current working condition state to obtain the corrected threshold, calculate the monitoring confidence according to the data quality vector, calculate the working condition identification confidence according to the working condition change point confidence, calculate the model consistency confidence according to the quality control monitoring data and the prediction results, and calculate the numerical stability index according to the numerical stability index. Calculate the reliability of numerical stability and combine the reliability of monitoring, reliability of working condition identification, reliability of model consistency and reliability of numerical stability into the overall reliability; S8, compare the quality control monitoring data with the correction threshold, calculate the degree and rate of exceeding the preset warning indicators, and carry out warning diversion and handling according to the overall reliability. When the overall reliability is not less than the preset reliability threshold and the degree or rate of exceeding the limit meets the preset limit criterion, output the warning result. When the overall reliability is less than the preset reliability threshold and the degree or rate of exceeding the limit meets the preset limit criterion, output the review and handling result.

[0011] Optionally, S1 includes:

[0012] Acquire time-stamped monitoring data for the foundation pit project, wherein the monitoring data includes two or more of the following: retaining structure displacement data, surface settlement data, support axial force data, and groundwater level data;

[0013] The monitoring data is time-aligned to form a unified time reference.

[0014] Obtain construction records aligned with the unified time reference, the construction records including excavation depth, support installation information, support removal information, and dewatering start and stop information;

[0015] The calculation domain and stratigraphic parameters are determined based on the survey data, the support structure parameters and support parameters are determined based on the design parameters, and the construction stage is determined based on the construction records.

[0016] A coupled model of soil and support structure or a coupled model of soil, support structure and groundwater is established in the computational domain. The initial coupled numerical model includes the excavation unloading settings, support installation and dismantling settings and dewatering start and stop settings corresponding to the construction stage, and includes the initial stress conditions and the initial groundwater level conditions.

[0017] Optionally, S2 includes:

[0018] The monitoring data are grouped according to monitoring type and monitoring location, and multivariate time series are formed under a unified time base.

[0019] A self-supervised anomaly detection model is constructed based on the multivariate time series. The self-supervised anomaly detection model obtains the reconstruction result by partially masking the multivariate time series and reconstructing the masked data.

[0020] Based on the reconstruction residual between the reconstruction result and the monitoring data, a data quality vector is generated for each time step. The data quality vector includes noise level, drift probability, jump probability, and missing measurement confidence. The noise level is determined by the statistics of the reconstruction residual. The drift probability is determined by the low-frequency trend change of the reconstruction residual. The jump probability is determined by the degree of deviation between the adjacent time step difference of the monitoring data and the adjacent time step difference of the reconstruction result. The missing measurement confidence is determined by the missing measurement marker of the monitoring data and the stability of the reconstruction result.

[0021] Based on the data quality vector, outlier suppression and missing data completion processing are performed on the monitoring data. Specifically, monitoring data with a jump probability exceeding a preset threshold are replaced or smoothed using the reconstruction result, monitoring data with a drift probability exceeding a preset threshold are drift corrected, and monitoring data with missing data are completed using the reconstruction result. This process yields quality control monitoring data, and the data quality vector and the quality control monitoring data are output.

[0022] Optionally, S3 includes:

[0023] Based on the data quality vector, multi-sensor weights are determined for each monitoring type and each monitoring point in the quality control monitoring data. The multi-sensor weights are negatively correlated with the noise level in the data quality vector and positively correlated with the confidence level of missing data in the data quality vector.

[0024] Under the constraint of the multi-sensor weights, a multi-monitoring type consistency index is calculated for the quality control monitoring data. The multi-monitoring type consistency index includes one or more of the following: consistency between the change in the displacement of the retaining structure and the change in the surface settlement, consistency between the change in the groundwater level and the change in the surface settlement, and consistency between the change in the axial force of the support and the change in the displacement of the retaining structure.

[0025] Based on the consistency index of the multiple monitoring types, identify the abrupt change position of the consistency index along the time axis to generate candidate points of operating condition change.

[0026] The confidence level of the operating condition change point is calculated for each candidate point. The confidence level is determined by the abrupt change of the consistency index at the candidate point, the weight of the multi-sensor system, and the stability of the consistency index within the time window before and after the candidate point. The candidate point and the confidence level of the operating condition change point are then output.

[0027] Optionally, S4 includes:

[0028] Arrange the candidate points of the operating condition change point in chronological order, and read the corresponding confidence level of the operating condition change point for each candidate point;

[0029] The candidate points of the working condition change point are confirmed one by one according to the preset hysteresis determination rule. The preset hysteresis determination rule includes a confidence threshold condition and a duration condition. The confidence threshold condition is that the confidence of the working condition change point is not less than the preset confidence threshold. The duration condition is that the confidence of the working condition change point within the preset duration range before and after the candidate point of the working condition change point continuously meets the confidence threshold condition.

[0030] When the candidate point of the working condition change point is confirmed as a working condition change point, the working condition state machine is updated according to the transition relationship of the working condition change point in the working condition state machine and the current working condition state is generated.

[0031] Based on the current working condition, a coupled numerical model update instruction is generated. The coupled numerical model update instruction includes construction stage update information corresponding to the current working condition and boundary condition update information corresponding to the construction stage update information. The boundary condition update information includes one or more of the following: excavation unloading settings, support installation and removal settings, and dewatering start and stop settings.

[0032] Optionally, S5 includes:

[0033] According to the coupled numerical model update instructions, the initial coupled numerical model is updated for the construction stage and boundary conditions. A step-by-step loading method is used to smoothly transition the updates for excavation unloading, support installation and removal, and dewatering start / stop, resulting in an updated coupled numerical model. A residual is established between the calculation results of the updated coupled numerical model and the quality control monitoring data, where the residual is the difference between the calculation results of the quality control monitoring data and the updated coupled numerical model at the same time and monitoring location. Weights are generated for the residuals based on the data quality vector, and the residuals are weighted to obtain weighted residuals. Using the reduction of the objective function value of the weighted residuals as the iterative update criterion, the model parameters of the updated coupled numerical model are iteratively updated to complete online correction, resulting in an online corrected coupled numerical model. Based on the online corrected coupled numerical model, rolling predictions are performed for subsequent time periods according to a preset prediction step size to obtain prediction results. The prediction results include one of the following: predicted retaining structure displacement, predicted surface settlement, predicted support axial force, and predicted groundwater level. The number of nonlinear iterations or the residual convergence ratio is recorded during each rolling prediction process to generate a numerical stability index.

[0034] Optionally, S6 includes:

[0035] For the preset early warning indicators, the corresponding prediction sequence is extracted within the prediction time domain based on the prediction results;

[0036] When the preset early warning indicator is a displacement indicator or a settlement indicator, the prediction envelope threshold is determined by the maximum or minimum value of the prediction sequence in the prediction time domain as the benchmark threshold.

[0037] When the preset early warning indicator is a rate-type indicator, the prediction sequence is differentially divided between adjacent time points to obtain a prediction rate sequence, and the prediction envelope threshold is determined as the benchmark threshold based on the maximum or minimum value of the prediction rate sequence in the prediction time domain.

[0038] When a predicted quantile threshold is used as a baseline threshold, a threshold corresponding to a preset quantile is calculated for the predicted sequence or the predicted rate sequence as a baseline threshold, wherein the preset quantile is a quantile greater than 50% and less than 100%.

[0039] Optionally, the S7 includes:

[0040] A quality correction coefficient is generated based on the data quality vector, wherein the quality correction coefficient increases with the increase of noise level in the data quality vector, increases with the increase of drift probability in the data quality vector, increases with the increase of jump probability in the data quality vector, and increases with the decrease of missing measurement confidence in the data quality vector; a working condition correction coefficient is generated based on the current working condition state, wherein the working condition correction coefficient is determined by the construction stage corresponding to the current working condition state; a correction threshold is obtained by performing correction operations on the benchmark threshold with the quality correction coefficient and the working condition correction coefficient respectively; monitoring confidence is calculated based on the data quality vector; working condition identification confidence is calculated based on the working condition change point confidence; model consistency residual is calculated based on the difference between quality control monitoring data and prediction results at the same time, and model consistency confidence is calculated based on the statistics of the model consistency residual; numerical stability confidence is calculated based on the numerical stability index; the monitoring confidence, working condition identification confidence, model consistency confidence, and numerical stability confidence are combined according to preset synthesis weights to obtain the overall confidence, wherein the preset synthesis weights are non-negative and the sum of the weights is 1.

[0041] Optionally, S8 includes:

[0042] The quality control monitoring data is used to calculate the measured value of the early warning indicator according to the preset early warning indicator, and the measured value of the early warning indicator is compared with the correction threshold at each time step.

[0043] When the measured value of the warning indicator exceeds the correction threshold, the degree of exceeding the limit is calculated according to the difference or ratio between the measured value of the warning indicator and the correction threshold, and the rate of exceeding the limit is calculated by the difference between adjacent time moments of the measured value of the warning indicator.

[0044] Whether a preset over-limit criterion is met is determined based on the degree of over-limit and the rate of over-limit, wherein the preset over-limit criterion includes one of an over-limit degree threshold condition or an over-limit rate threshold condition;

[0045] When the preset over-limit criterion is met, a warning and diversion process is performed based on the overall credibility. When the overall credibility is not less than the preset credibility threshold, a warning result is output. When the overall credibility is less than the preset credibility threshold, a review process result is output. The review process result includes one of triggering monitoring review, triggering encrypted monitoring, and triggering backtracking recalculation.

[0046] The beneficial effects of this invention are:

[0047] 1. Achieve online quantification and quality-driven integration of monitoring data quality: Generate a data quality vector containing noise level, drift probability, jump probability and missing measurement confidence through self-supervised anomaly detection, suppress anomalies and complete missing measurements, and set multi-sensor weights accordingly, so that working condition identification, model correction and early warning judgment are constrained by data quality, reducing misjudgment and false alarm caused by monitoring anomalies;

[0048] 2. Achieve automatic identification of working conditions and adaptive updating of coupled numerical models: Generate candidate points of working condition change points and calculate confidence scores through quality-weighted multi-sensor consistency gating change point detection. Combine hysteresis rules to update the working condition state machine, drive the smooth transition and switching of working conditions and boundary conditions of coupled numerical models, and iteratively correct parameters online to improve the continuity and accuracy of prediction under changes in the construction stage.

[0049] 3. Implement dynamic threshold early warning and credibility-based triage: Based on rolling prediction, a baseline threshold for early warning indicators is constructed, and dynamic thresholds are obtained by adjusting the thresholds according to data quality and current operating conditions. At the same time, the credibility of monitoring, operating condition identification, model consistency and numerical stability is decomposed and synthesized to form an overall credibility. When the limit criterion is met, an early warning or review and handling result is output, thereby suppressing low credibility alarms and improving the executability of handling while ensuring sensitivity. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of a method for predicting and warning the deformation of a foundation pit based on numerical simulation coupling proposed in this invention. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0053] refer to Figure 1 A method for predicting and warning of foundation pit deformation based on numerical simulation coupling, comprising:

[0054] S1. Obtain time-stamped monitoring data and construction records for the foundation pit project, and establish an initial coupled numerical model based on the survey data, design parameters, and construction records; S2. Perform self-supervised anomaly detection on the monitoring data, generate a data quality vector corresponding to the monitoring data, suppress outliers and complete missing data in the monitoring data to obtain quality control monitoring data; S3. Set multi-sensor weights for the quality control monitoring data based on the data quality vector, and under its constraints, perform condition change point detection based on multi-sensor consistency on the quality control monitoring data, generate candidate points for condition changes, and calculate the condition... S4. Confirm candidate points for change points according to preset hysteresis judgment rules, update the working condition state machine and generate the current working condition state, and generate a coupled numerical model update instruction; S5. Perform a smooth transition update of the working condition and boundary conditions on the initial coupled numerical model according to the coupled numerical model update instruction, compare the calculation results of the quality control monitoring data with the updated coupled numerical model, weight the comparison residuals according to the data quality vector to obtain the weighted residuals, use them as the target to iteratively update the model parameters, obtain the coupled numerical model after online correction, perform rolling prediction, obtain the prediction results, and generate a numerical stability index characterizing the stability of the rolling prediction calculation; S6. For the preset early warning index, calculate the benchmark threshold according to the prediction results. The benchmark threshold is the prediction envelope threshold or prediction quantile threshold of the preset early warning index in the prediction time domain; S7. Correct the benchmark threshold according to the data quality vector and the current working condition state to obtain the corrected threshold, calculate the monitoring confidence according to the data quality vector, calculate the working condition identification confidence according to the working condition change point confidence, calculate the model consistency confidence according to the quality control monitoring data and the prediction results, and calculate the numerical stability index according to the numerical stability index. Calculate the reliability of numerical stability and combine the reliability of monitoring, reliability of working condition identification, reliability of model consistency and reliability of numerical stability into the overall reliability; S8, compare the quality control monitoring data with the correction threshold, calculate the degree and rate of exceeding the preset warning indicators, and carry out warning diversion and handling according to the overall reliability. When the overall reliability is not less than the preset reliability threshold and the degree or rate of exceeding the limit meets the preset limit criterion, output the warning result. When the overall reliability is less than the preset reliability threshold and the degree or rate of exceeding the limit meets the preset limit criterion, output the review and handling result.

[0055] In this specific embodiment, S1 includes:

[0056] Read the time-stamped monitoring data from the automated monitoring system for the foundation pit project and establish a monitoring data structure. The monitoring data includes retaining structure displacement data, surface settlement data, support axial force data, and groundwater level data. The retaining structure displacement data is obtained from horizontal displacement monitoring points deployed at different depths of the retaining structure and is uniformly symbolized with the direction of the foundation pit's outer normal as the positive direction. The surface settlement data is obtained from settlement monitoring points deployed around the foundation pit and is uniformly symbolized with the downward direction as the positive direction. The support axial force data is obtained from axial force gauges deployed on each support component and is uniformly symbolized with the direction of compression as the positive direction. The groundwater level data is obtained from water level gauges deployed in observation wells inside and outside the foundation pit and is expressed as the water head height relative to the unified engineering elevation benchmark. For each type of monitoring data, establish a mapping relationship between monitoring point number and spatial coordinates and complete unit unification and abnormal timestamp removal. Abnormal timestamp removal is performed according to the rule of "retaining the last written record when multiple records appear at the same monitoring point at the same timestamp".

[0057] The above monitoring data are time-aligned to form a unified time base, and the construction records are aligned accordingly. The unified time base is constructed using the following formula:

[0058] ;

[0059] in Indicates the first A unified time reference point This indicates the start time of the unified time base and takes the earliest timestamp from all monitoring data and construction records. The time step number represents a non-negative integer. This indicates a uniform sampling time step of 1 hour. Then, for each monitoring point, linear interpolation is used between two adjacent original timestamp data points to obtain its position in time. The monitored values ​​are retained, and the original values ​​at the endpoints of the interpolation interval are preserved. For values ​​exceeding the original time range of the monitoring point... The missing data is marked and kept null for subsequent processing. Simultaneously, construction records aligned with a unified time base are read and converted into a "stage event sequence." This sequence includes excavation depth, support installation information, support removal information, and dewatering start / stop information, with each event carrying its corresponding data. The markings include the excavation depth as the cumulative excavation depth relative to the design ground level of the foundation pit, the support installation and removal information including the support number, installation or removal time and design prestress value, and the dewatering start and stop information including the dewatering well group number, start and stop time and target control water level.

[0060] Based on the exploration data, the calculation domain and stratigraphic parameters are determined. The support structure and bearing parameters are determined based on the design parameters. The calculation domain is established using a three-dimensional rectangular coordinate system, extending outwards from the excavation plane to the area affected by the excavation. The area affected is defined as "three times the excavation depth," and the bottom depth of the calculation domain is taken as "twice the excavation depth." The stratigraphic parameters are determined by the bottom elevation of each soil layer given in the exploration data and are laid out horizontally within the calculation domain. For the same soil layer, when different physical and mechanical properties are given at different exploration borehole locations, the arithmetic mean is taken as the parameter value for that soil layer. The soil constitutive model adopts the Mohr-Coulomb model, and an elastic modulus is assigned to each soil layer. Poisson's ratio Cohesion internal friction angle and severe The groundwater seepage parameters are based on the permeability coefficients of each soil layer. In the 3D model, values ​​are assigned isotropically, and the support structure is built into structural units of continuous wall or piles according to the design drawings, and the elastic modulus of concrete is assigned. Poisson's ratio ,thickness With the moment of inertia of the cross section The supporting components are constructed as beam elements according to the design drawings, and the steel or concrete is assigned an elastic modulus. Cross-sectional area With linear elastic axial stiffness At the same time, the prestress value in the construction record is used as the initial axial force of the supporting member and applied to the corresponding beam element;

[0061] A coupled numerical model of soil, support structure, and groundwater was established within the aforementioned computational domain, and the construction stages were determined based on construction records. The coupled numerical model adopted a consolidation analysis framework of displacement-pore pressure coupling, and simultaneously set displacement and pore pressure degrees of freedom in the soil elements. Displacement degrees of freedom were set in the support structure and supporting member elements, and the interaction between the soil and the support structure was realized through node sharing or contact constraints. The initial stress conditions were generated based on the self-weight stress field, with the ground surface as the zero vertical stress reference, and the unit weight was accumulated layer by layer. The initial geostress distribution was obtained. The initial groundwater level was determined based on the stable water level before the start and stop of dewatering in the exploration data and construction records, and set as the initial head boundary in the pore pressure field. The model boundary conditions were set as follows: the three-dimensional displacement of the bottom surface of the computational domain was fixed, the lateral normal displacement was fixed, and the tangential displacement was free. A constant head boundary consistent with the initial groundwater level was set on the groundwater boundary. The construction stage generated a stage list based on the stage event sequence, and each stage corresponded to a unique excavation unloading setting, support installation and removal setting, and dewatering start and stop setting. Excavation unloading was achieved by deactivating the soil element of the excavation area and releasing its self-weight and pore pressure in the corresponding stage. Support installation was achieved by activating the support beam element and applying the initial axial force in the corresponding stage. Support removal was achieved by deactivating the corresponding support beam element in the corresponding stage. Dewatering start and stop was achieved by switching the water level value of the constant head boundary in the corresponding stage. Thus, an initial coupled numerical model containing construction stage information, initial stress conditions, and initial groundwater level conditions was obtained.

[0062] In this specific embodiment, S2 includes:

[0063] The time-aligned monitoring data were grouped by monitoring type and monitoring location to form a multivariate time series. Each group corresponds to a monitoring type, and its multivariate dimension consists of all monitoring locations under that type. The unified time reference is [missing information]. and The time step number represents the time step number for each group at time [time]. The observation vector is obtained by splicing the observation values ​​of all points in the group according to the fixed point number order and forming a data matrix arranged in time. At the same time, normalization is performed on the monitoring point sequence of each dimension to eliminate the influence of dimensions. The normalization adopts a robust standardization with the median of the historical observation values ​​of the point as the center and the interquartile range as the scale. The center and scale parameters are fixed and saved for subsequent inverse normalization in the online inference stage.

[0064] A self-supervised anomaly detection model is constructed based on the multivariate time series. This model employs a masked reconstruction temporal convolutional autoencoder structure and trains a set of model parameters for each monitoring type. The model input is a sliding window sequence with a length of 48 time steps, containing multivariate data from all monitoring points of that monitoring type. The model encoder consists of three concatenated one-dimensional dilated convolutional residual blocks, with each layer having a kernel size of 3, 64 channels, and a dilation rate of [missing information]. Layer normalization and linear rectified activation are performed after each residual block. The model decoder uses a three-layer one-dimensional convolutional residual block symmetrical to the encoder and outputs a reconstructed sequence with the same dimension as the input. During the training phase, a masking matrix is ​​generated within the input window according to the rule of "random point masking and continuous segment masking superposition". The input value at the masked position is set to zero while retaining the missing test mark. The loss function only calculates the reconstruction error at the masked position and uses the mean absolute error as the optimization objective. The optimizer is Adam with a learning rate of 0.001, a batch size of 32, and an initial training round of 50. The training data is taken from the historical window samples of the previous 336 time steps under a unified time base. During the online operation phase, the model is updated once a day with sliding window samples of the most recent 336 time steps in an incremental manner to adapt to the distribution drift caused by changes in working conditions.

[0065] Based on the reconstruction residuals between the reconstruction results and the monitoring data, a data quality vector is generated for each time step. The reconstruction results are then analyzed for each time step. The most recent 48 time step windows are extracted as model inputs, and forward inference with different masking matrices is performed 5 times. The average value of the reconstructed output is then calculated to obtain a stable reconstructed estimate for the same monitoring point at the same time. The reconstructed residual is used to characterize the degree of anomaly at that point at that time. The reconstructed residual is calculated using the following formula: ;

[0066] in Indicates time The Reconstructed residuals of each monitoring point Indicates time The The measured values ​​from each monitoring point are taken from the original monitoring data after time alignment, and the units and symbols are standardized. Indicates time The The reconstructed estimates of each monitoring point are taken from the reconstructed output of the self-supervised anomaly detection model and inversely normalized. Indicates the time step number under a unified time base. This indicates the point index arranged in fixed point number order within this monitoring type group, based on... For each moment Each monitoring point outputs a data quality vector, which consists of four components: noise level, drift probability, jump probability, and missing data confidence. The noise level is calculated from the data from the most recent 24 time steps. The robustness of the dispersion is determined and the noise scale threshold is set for this monitoring type. After normalization, the cutoff point is... The drift probability is determined by the last 168 time steps. Low-frequency bias and monotonicity are jointly determined, and the absolute value of the bias exceeds the drift threshold for this monitoring type. If the sign consistency is not less than 0.9, take 1; otherwise, take the absolute value of the bias. The ratio is cut off to The transition probability is determined by the difference between adjacent time steps. and The degree of deviation is determined and the threshold is jumped to this monitoring type. After normalization, the cutoff point is... The confidence level of missing tests is jointly determined by the missing test markers and the reconstruction stability, and when To determine the results of 5 inferences when there are missing tests The variance and the stability threshold of this monitoring type Compare and ensure the variance is no greater than 1. If the value is 1, then the value is 0; otherwise, the value is 0. When there is no missing data, the confidence level for missing data is set to 1, where displacement and settlement monitoring types... Pick Pick Pick Pick Axial force monitoring type Pick Pick Pick Pick Groundwater level monitoring type Pick Pick Pick Pick ;

[0067] Based on the data quality vector, outlier suppression and missing data completion are performed on the monitoring data to obtain quality control monitoring data. Specifically, when the probability of a jump is not less than 0.8, the data at that point at that moment is... Replace directly with A three-point moving average was applied to the replaced sequence to suppress residual spikes. When the drift probability was not less than 0.7, the median of the reconstructed residuals from the most recent 168 time steps was used as the drift bias estimate, and the subsequent observations at that point from the current time were biased to complete the drift correction. When a missing data marker existed, the missing data at that point at that time was used... The missing data is filled in and the confidence level of the missing data decreases linearly with the completion time. When the continuous missing data time reaches 24 time steps, the confidence level of the missing data is fixed at 0.2 to reflect the uncertainty of long-term extrapolation. The quality control monitoring data and the corresponding data quality vector are output time by time according to a unified time base and written into the data bus used in subsequent steps.

[0068] In this specific embodiment, S3 includes:

[0069] Read the quality control monitoring data and its corresponding data quality vector, and establish multi-sensor weights according to monitoring type and monitoring point location. The monitoring types include retaining structure displacement, ground surface settlement, support axial force, and groundwater level, and the set of monitoring points under each monitoring type remains constant. The data quality vector changes at each time step. Each monitoring point includes two items: noise level and missing data confidence level, which are used to construct the weights. The noise level is used to characterize the point at time [time value missing]. The short-term fluctuation intensity and the numerical range are Furthermore, a larger value indicates stronger noise; the missing measurement confidence level is used to characterize the location at time [time]. The availability and completion reliability and the numerical range are Furthermore, a larger value indicates a higher availability of that location at that moment;

[0070] For each monitoring point within the same monitoring type at each time point Calculate the multi-sensor weights using the following formula and ensure that the sum of the weights is 1:

[0071] ;

[0072] in Represents a unified time reference time. Next The multi-sensor weights for each monitoring point have a value range of [value range missing]. Indicates time Next The reliability of missing data at individual monitoring points Indicates time Next The noise level at each monitoring point This indicates the total number of monitoring points participating in the fusion under this monitoring type. This represents the summation index of the monitoring points under this monitoring type. The time step number representing the unified time base. This represents the point index arranged in a fixed point number order under this monitoring type. Through the above construction, the higher the noise level, the smaller the weight, and the higher the confidence level of missing data, the larger the weight, and normalization is completed within the same monitoring type.

[0073] After obtaining the weights, change sequences for consistency analysis were extracted from the quality control monitoring data. These change sequences were calculated based on "differences between adjacent time points" and residual spikes were suppressed using three-point median filtering. Specifically, the change in retaining structure displacement was represented by the weighted average of the differences between the displacement points of each retaining structure; the change in surface settlement was represented by the weighted average of the differences between the surface settlement points of each surface settlement point; the change in support axial force was represented by the weighted average of the differences between the support axial force points of each support axial force; and the change in groundwater level was represented by the weighted average of the differences between the groundwater level points of each groundwater level. To eliminate the influence of different physical quantities' dimensions, normalization was performed on these four representative change sequences. The normalization scale was the median of the absolute values ​​of the differences over the most recent 168 time steps of the sequence. When this scale was less than the minimum scale threshold for the corresponding monitoring type, the scale was fixed to the minimum scale threshold. Specifically, the minimum scale threshold for retaining structure displacement and surface settlement was 0.5 mm, and the minimum scale threshold for support axial force was 10 kN. The minimum scale threshold for groundwater level is taken as 0.05 m, and the normalized change represents the sequence restricted to... Within the scope and used for cross-type consistency calculations;

[0074] The normalized change represents the sequence at each time step. Consistency indices for multiple monitoring types were calculated. These indices included consistency between changes in retaining structure displacement and changes in surface settlement, consistency between changes in groundwater level and changes in surface settlement, and consistency between changes in support axial force and changes in retaining structure displacement. All three indices were generated using the criterion that "the smaller the difference, the higher the consistency." The consistency score of the interval is set to 0 when the difference reaches 2 and 1 when the difference is 0. At the same time, the average weight of the monitoring type for each consistency indicator is calculated as the weight constraint of that consistency indicator and used for subsequent gating.

[0075] For each consistency index sequence along the time axis, mutation identification is performed to generate candidate points of change in operating conditions. The mutation identification is performed using a fixed sliding window method with a window length of 6 time steps, and at each potential boundary moment... Calculate the mean and variance of the consistency index for the front and back windows respectively. A window is considered valid only if the absolute value of the difference between the means of the consistency index for the front and back windows is not less than 0.25, the variance of the consistency index for both the front and back windows is not greater than 0.04, and the weight constraint of this consistency index is not less than 0.5. This is recorded as a candidate point for the operating condition change point corresponding to this consistency index;

[0076] Candidate points generated by different consistency indicators at the same time are merged to form a global working condition change point candidate point set. The merging rule is that when multiple candidate points appear at the same time, only one record is retained, and the consistency indicator with the largest mean change amplitude among the consistency indicators at that time is retained as the dominant indicator.

[0077] For each candidate point of the operating condition change, calculate and output the confidence score of the operating condition change. The confidence score of the operating condition change is determined by three parts, and the product of the three parts is truncated to a certain value. The first part is the mutation amplitude factor, which is obtained by dividing the absolute value of the difference between the mean of the consistency index of the dominant index in the back window and the front window by 0.5, and is set to 1 when it exceeds 1. The second part is the weight factor, which is the weight constraint of the dominant index itself to reflect the constraint of multi-sensor quality on the judgment of change points. The third part is the stability factor, which is obtained by subtracting the sum of the variances of the consistency index of the front window and the back window from 1 and dividing by 0.08, and is set to 0 when it is less than 0. Finally, candidate points with an interval of less than 3 time steps between adjacent time steps are regarded as the same change point cluster, and only the candidate point with the highest confidence is retained as the output, so as to obtain the working condition change point candidate points and their corresponding working condition change point confidence scores sorted by time.

[0078] In this specific embodiment, S4 includes:

[0079] Read the time-sorted candidate point sequence of operating condition changes and read it against the unified time base. Aligned working condition change point confidence sequence ,in Indicates the first A unified time reference point and The sequence of candidate points for a change in operating condition is represented by its time step number. and Candidate point number, confidence sequence of operating condition change points The range of values ​​is Furthermore, the larger the value, the stronger the evidence that a change in operating conditions occurred at that moment;

[0080] For each candidate point Confirmation is performed according to a preset hysteresis determination rule. The hysteresis determination rule includes both a confidence threshold condition and a duration condition, and in this embodiment, the confidence threshold is set to... Duration Each time step and candidate points In front and behind The time window consisting of 1 time step is denoted as Candidate points are selected if and only if the confidence level of the operating condition change points within the time window continuously meets the confidence threshold condition. The condition is confirmed to be a change point, and its continuous satisfaction relationship is given by the following formula:

[0081] ;

[0082] in Indicates time step The confidence level of the working condition variable point, Represents the integer offset relative to the candidate point within the time window. Indicates the number of time steps corresponding to the duration and takes... Indicates the confidence threshold and takes This indicates the operation that takes the minimum value within the time window;

[0083] To avoid fluctuations in operating conditions caused by repeated switching within a short period, a minimum switching interval constraint is applied and fixed at the time of confirming a change in operating condition. If the time step interval between the current candidate point and the most recently confirmed change point is less than [number] time steps, then [the following is true]. Then discard the candidate point and continue processing subsequent candidate points;

[0084] Regarding the update of the working condition state machine, the stage list generated in step S1 based on the construction records is solidified into a state set of the working condition state machine, and each state is associated with a unique construction stage, with the construction stage numbered sequentially. Indicates and ,in Indicates the initial unexcavated stage and The final stage number is determined by the length of the stage list. The state transition relationship is fixed as a directed transition that increments by the stage number and skips stages is not allowed.

[0085] During system initialization, the operating state machine is set to... When candidate points When a change in operating condition is confirmed and the minimum switching interval constraint is met, the operating condition state machine will switch from the current stage. Transfer to the next stage Output the current operating status as ,when Already equal to Staying at No further transfers;

[0086] Regarding the generation of update instructions for the coupled numerical model, an update instruction is generated for each state machine transition and written into the update instruction queue. The update instruction contains both construction stage update information and boundary condition update information, where the construction stage update information includes the stage number. Phase Effective Time With a unified time base interval for the duration of each stage, boundary condition update information is generated from the stage list and stage number. The bound construction records are parsed and include three parts: excavation unloading settings, support installation and removal settings, and dewatering start / stop settings. The excavation unloading settings include the target cumulative excavation depth. The spatial extent of the corresponding excavation area is identified and used to drive the phased unloading of the corresponding soil elements in the subsequent numerical model. The support installation and removal settings include a set of support component numbers. and And respectively represent the stages The support set that needs to be activated for installation and the support set that needs to be deactivated for removal, and includes... The corresponding design prestress value is used as the initial axial force input, and the precipitation start / stop setting includes precipitation status indicators. With the target control water level or head value It is also used to drive the start and stop of constant head boundary and water level switching in subsequent numerical models;

[0087] The current operating condition and the coupled numerical model update command are used as inputs to step S5 to trigger the update of the operating condition and boundary conditions of the coupled numerical model.

[0088] In this specific embodiment, S5 includes:

[0089] Read the coupled numerical model update instructions and at the unified time base time This triggers the construction phase update and boundary condition update of the initial coupled numerical model. The construction phase update is based on the phase number. To index the excavation unloading settings, support installation and removal settings, and dewatering start / stop settings, write them into the numerical solver's stage control table and lock their effective range. Boundary condition updates employ a step-by-step loading method to achieve a smooth transition, with a fixed number of steps. Each sub-step maintains a constant time step, and excavation and unloading occur at each sub-step. By applying an equivalent unloading ratio to the target excavation area Gradually release the soil's self-weight and pore pressure and in the first After each sub-step is completed, the corresponding soil element is deactivated to eliminate nonlinear oscillations caused by abrupt stiffness changes. Supports are installed in the sub-step. This is achieved by linearly increasing the axial stiffness of the supporting beam element from 0 to the design axial stiffness and simultaneously linearly increasing the design prestress from 0 to the target value. Support removal is then performed in the sub-step. By linearly reducing the axial stiffness of the target support beam element from the design value to 0, and at the first... Inactivation occurs after each sub-step, and precipitation starts and stops at the sub-step. By linearly transitioning the boundary head value from the previous stage to the target head of the current stage, an updated coupled numerical model is obtained. This model is then solved using an implicit backward Euler scheme with Newton iterations at each time step, a maximum nonlinear iteration count of 20, and a residual convergence threshold of [value missing]. ;

[0090] The quality control monitoring data and the updated coupled numerical model calculation results are aligned at the same time and at the same monitoring location. The correspondence between the monitoring location and the finite element mesh has been solidified in step S1 as a mapping from "monitoring point to element interpolation". After solving, the simulated values ​​of the retaining structure displacement, surface settlement, support axial force and groundwater level at the monitoring location are obtained from the element node displacement and pore pressure through shape function interpolation.

[0091] After alignment, residuals are established, and weights are generated based on the data quality vector to weight the residuals. The weighted residuals then drive online correction. The parameter vector for online correction is denoted as... Furthermore, it is composed of and fixed by the elastic modulus of each soil layer and the interface stiffness. ,in Indicates the first The elastic modulus of the soil layer and This represents the total number of soil layers, with its initial value taken from the survey data and upper and lower bounds applied. Indicates the first Initial value of the elastic modulus of the soil layer. This represents the normal contact stiffness between the soil and the supporting structure, with the initial value taken as... And apply upper and lower bound constraints as follows This represents the tangential contact stiffness between the soil and the supporting structure, with the initial value taken as... And apply upper and lower bound constraints as follows Online correction with the latest The correction time window, consisting of n time steps, is used as input, and the parameter iterative update is completed by minimizing the objective function, which is defined as:

[0092] ;

[0093] in Represents the parameter vector The weighted residual objective function value, Represents the set of time steps within the correction time window and takes , This represents the set of monitoring sequence indices participating in the calibration, containing all monitoring point indices for retaining structure displacement, surface settlement, support axial force, and groundwater level, arranged in order of fixed point number. Indicates time Next The weights of each monitoring sequence are calculated as the confidence level for missing data using the data quality vector. With noise suppression, drift suppression, and jump suppression factors The product is normalized within the same monitoring type at the same time to ensure comparability of weights. Indicates time Next Quality control monitoring values ​​for each monitoring sequence This indicates that the parameter vector is The updated coupled numerical model at time 10:00 With the Simulated values ​​at each monitoring location Indicates the time step number under a unified time base. Indicates the monitoring sequence index;

[0094] The parameter iterative update uses the damped Gauss-Newton method with a fixed maximum number of iterations of 8, and the objective function decreases by less than 8 between two adjacent iterations. The iteration is terminated and the iteration results that satisfy the upper and lower bound constraints are written back to the coupled numerical model as the online corrected parameter vector, thus obtaining the online corrected coupled numerical model;

[0095] Based on the online-calibrated coupled numerical model, rolling predictions are performed for subsequent time steps according to a preset prediction step size, where the prediction step size is consistent with the unified time step size and is taken as... The prediction time domain length is fixed at 24 time steps, and the current time is changed each time step is rolled over. The model state is used as the initial state and then advanced to... It outputs the predicted displacement of the retaining structure, the predicted surface settlement, the predicted axial force of the support, and the predicted groundwater level as the prediction results;

[0096] In each rolling prediction process, the maximum number of Newton iterations at each time step is recorded, and the residual convergence ratio is also recorded. The two are combined to form a numerical stability index, which is output. The number of Newton iterations is used to characterize the difficulty of nonlinear solution, and the residual convergence ratio is used to characterize the convergence speed and stability.

[0097] In this specific embodiment, S6 includes:

[0098] The rolling prediction results are read, and a prediction sequence is generated for a pre-fixed set of early warning indicators. This set includes indicators for retaining structure displacement, ground subsidence, retaining structure displacement rate, and ground subsidence rate, all using the convention that "increased values ​​represent increased risk." For each prediction time, the retaining structure displacement indicator takes the maximum predicted displacement value at all monitored locations of the retaining structure at that time. Similarly, the ground subsidence indicator takes the maximum predicted settlement value at all monitored locations of the ground subsidence at that time. The retaining structure displacement rate indicator is calculated by dividing the difference between the retaining structure displacement indicator values ​​at two adjacent prediction times by the prediction step size. The maximum value among all adjacent differences is obtained and taken as the index value at that moment. The surface subsidence rate index is calculated at each prediction moment by dividing the difference between the surface subsidence index values ​​of two adjacent prediction moments by the prediction step size. The maximum value among all adjacent differences is obtained and taken as the index value at that time, where the prediction step size is... Consistent with a unified time step of 1 hour, the prediction time domain length is fixed at [value missing]. Each time step and the predicted sequence covers from the current time step. After to ;

[0099] For each early warning indicator, the prediction envelope threshold of its predicted sequence in the prediction time domain is used as the baseline threshold and calculated according to the following formula: ;

[0100] in Indicates the current unified time base time. The corresponding baseline threshold for this early warning indicator, This indicates that the early warning indicator is in the prediction time. The predicted index value is determined by the prediction result of step S5 according to the "maximum point aggregation and adjacent difference" rule. This represents the prediction step number within the prediction time domain, and takes values ​​from 1 to... Indicates the prediction time domain length and takes This represents the operation that takes the maximum value within the prediction time domain;

[0101] Each early warning indicator received Its corresponding warning indicator and current time Predicted time domain length and prediction step size Write them into the threshold result table.

[0102] In this specific embodiment, S7 includes:

[0103] Read the baseline threshold And data quality vector and read the current operating status. Compared with the confidence score record of the change point used to confirm the working condition status, the To unify time standards Time step number;

[0104] For the set of monitoring points covered by the current early warning indicators Summarize the data quality vector components and obtain the summarization noise level. Summary drift probability Summary of jump probability Reliability of aggregated missing data For each component, the multi-sensor weights within the same monitoring type are applied as in step S3. A weighted average was used to ensure that the summarized data quality results were consistent with the effectiveness of the sensors.

[0105] According to the above Generate quality correction factor And the set of values ​​is fixed. ,when Time to take ,when and or Time to take ,when and and and Time to take In other cases, take To ensure that the threshold relaxation range is increased when the noise level increases, the drift probability increases, the jump probability increases, or the confidence of missing measurements decreases;

[0106] Based on the current operating conditions The corresponding construction phase update command generates the working condition correction coefficient. And the set of values ​​is fixed. When the stage update instruction includes changes to excavation unloading settings, support removal settings, or "Enable Dewatering" in the dewatering start / stop settings, then... When the phase update instruction does not meet the aforementioned conditions and includes support installation settings, take... In other cases, take This ensures that the threshold can be relaxed more drastically when the risk is higher during the construction phase;

[0107] Baseline threshold sequentially with the quality correction factor and operating condition correction factor Performing a multiplication correction operation yields the correction threshold. And and Bind storage to support subsequent tracing;

[0108] Calculate monitoring reliability based on data quality vector And the set of values ​​is fixed. ,when Time to take ,when and or Time to take ,when and and and Time to take In other cases, take This is to ensure that the reliability of monitoring decreases as noise, drift, jumps, or missing data become more severe.

[0109] Calculate the reliability of working condition identification based on the confidence level of the working condition change points. The confidence level corresponding to the most recently confirmed change point in the operating condition during system maintenance. And when step S4 completes the state machine transition at the current time, Update to the confidence level of the variable point used in this confirmation and let Maintain state when no state machine transition occurs at the current time. Unchanged and let This ensures that the credibility of operational condition identification is continuous between adjacent stages and is constrained by the strength of evidence from the most recent change point.

[0110] Calculate the model consistency residuals and the model consistency reliability based on the quality control monitoring data and prediction results. The model consistency residual, defined at the current early warning index scale, is "the measured value of the early warning index at the current moment minus the simulated value of the early warning index at the current moment." The measured value of the early warning index is calculated from the quality control monitoring data according to the same aggregation rule in step S6, and the simulated value of the early warning index is calculated by coupling the numerical model after online correction in step S5 at the current moment. The calculation results were obtained according to the same aggregation rule and in the most recent The root mean square is calculated on the residual sequence at each time step as a residual statistic to eliminate the influence of single-point random fluctuations.

[0111] When the early warning indicator is a displacement or settlement type, the threshold for the residual statistic is fixed at [value]. and And according to "residual statistics not greater than hour , between and hour greater than hour "Confirmed, when the early warning indicator is a rate-type indicator, the threshold for the residual statistic will be fixed at..." and And the same rule is used to determine the consistency of the model, and the reliability decreases as the deviation between the model and the monitoring increases;

[0112] Calculate the reliability of numerical stability based on numerical stability index The numerical stability metric includes the maximum number of Newton iterations recorded during the current rolling forecast process. With maximum residual convergence ratio ,in To predict the maximum number of Newton iterations at each time step in the time domain and To predict the maximum value of the ratio of convergent residual to initial residual at each time step in the time domain, when and Time to take ,when and And if the aforementioned conditions are not met, take In other cases, take The reliability of numerical stability decreases as nonlinear solutions become more difficult and convergence deteriorates.

[0113] Will monitor credibility Reliability of working condition identification Model consistency credibility With numerical stability confidence The overall credibility is calculated by combining the pre-set composite weights. The composite weight is fixed at 1. All values ​​are non-negative and the sum of their weights is 1. The overall credibility is calculated using the following formula:

[0114] ;

[0115] in Represents a unified time reference time. The following is an assessment of the overall reliability of the current early warning indicators, with a value range of [missing information]. Indicates the composite weight of monitoring credibility. Indicates the reliability of the monitoring. This indicates the weighting of the confidence level for identifying operating conditions. Indicates the reliability of operating condition identification. The composite weights represent the consistency and credibility of the model. Indicates the consistency and reliability of the model. The composite weight represents the reliability of numerical stability. Indicates the reliability of numerical stability. The time step number representing the unified time base is ultimately output as the correction threshold. Overall credibility .

[0116] In this specific embodiment, S8 includes:

[0117] Read the quality control monitoring data and calculate the measured value of the early warning indicator according to the aggregation rule consistent with the current early warning indicator in step S6. The displacement index of the enclosure structure is taken at time 1. The maximum value of quality control displacement at all monitoring locations of the enclosure structure displacement, and the time of taking the surface settlement index. The quality control settlement value at all surface settlement monitoring locations is determined by dividing the difference in the retaining structure displacement rate between adjacent time points by a uniform time step. The surface subsidence rate index is calculated by dividing the difference in surface subsidence rates at adjacent time points by a uniform time step. and will With the corrected threshold At the same time Next-step time-by-time comparison;

[0118] If and only if The degree and rate of exceeding the limit are calculated and used for the determination of exceeding the limit. With the over-limit rate Calculate using the following formula:

[0119] ;

[0120] in Represents a unified time reference time. The degree of exceeding the limit, This represents the operation of finding the maximum value. Represents a unified time reference time. The warning indicators are based on measured values ​​and aggregated calculations derived from quality control monitoring data. Represents a unified time reference time. The correction threshold below, Represents a unified time reference time. The rate of exceeding the limit, Represents a unified time reference time. Actual measured values ​​of the warning indicators below. Indicates a unified time step and takes Indicates the time step number of the unified time base and when season make ;

[0121] Based on the threshold conditions for the degree of exceedance and the threshold conditions for the rate of exceedance, a preset exceedance criterion is executed, and the two are combined into a trigger condition by logical OR. When the warning indicator is a displacement indicator or a settlement indicator, the threshold condition for the degree of exceedance is fixed. And the over-limit rate threshold is fixed at 1. When the warning indicator is a rate-related indicator, the threshold for exceeding the limit will be fixed at [value]. And the over-limit rate threshold is fixed at 1. If and only if or The preset over-limit criterion is then satisfied.

[0122] Read the overall credibility when the preset over-limit criterion is met. and with a preset confidence threshold Implement early warning and triage measures when Output the early warning results in real time and write the early warning indicator identifier and time into the early warning results. Measured values ​​of early warning indicators Correction threshold Degree of exceeding limits Exceeding the limit rate Overall credibility To support on-site response, when The system outputs the review and handling results in real time, and writes the same fields as the warning results into the review and handling results, as well as the review action type. The review action type is determined by the minimum value of the sub-item confidence level and is based on the monitoring confidence level. Reliability of working condition identification Model consistency credibility With numerical stability confidence Choose the action corresponding to the smallest value, where the smallest value is... The system triggers monitoring review and generates a list of monitoring points requiring review, with the smallest one being [the smallest value]. When the encrypted monitoring is triggered, the unified time step will be changed from Adjusted to And continue for 12 time steps, until the smallest is or When the time is triggered, backtracking and recalculation are performed, and the backtracking starting point is set to the most recent confirmed change point in the operating condition to re-execute online correction and rolling prediction, forming an early warning and review diversion output consistent with the overall credibility.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0124] This invention constructs a closed-loop chain encompassing online quality control of monitoring data, operating condition identification, coupled numerical model update and correction, rolling prediction, dynamic threshold early warning, and credibility-driven handling. This ensures that each link forms a synergistic constraint around the same technical problem and produces technical effects. First, self-supervised anomaly detection suppresses outliers and fills in missing data in multi-source monitoring data, outputting a quantifiable data quality vector. This reduces the disturbances caused by noise, drift, jumps, and missing data to subsequent identification and prediction from the source. Second, under data quality constraints, operating condition change point detection and hysteresis confirmation based on multi-sensor consistency are performed to obtain a stable and reliable operating condition state and generate model update instructions, thereby driving a smooth transition update of the operating condition and boundary conditions in the coupled numerical model. Subsequently, the monitoring and simulation residuals are weighted using the data quality vector, and the model parameters are iteratively updated to complete online correction. Rolling prediction is then performed to obtain prediction results consistent with the current operating condition, and a baseline threshold for the early warning indicator is generated accordingly. Finally, based on the baseline threshold, a modified threshold is obtained by introducing data quality and operating condition corrections. The threshold is then determined by combining the degree and rate of exceeding the limit, and the handling process is diverted based on the overall reliability. This improves the adaptability to changes in operating conditions and data fluctuations, reduces false alarms and missed alarms, and enhances the executability of the early warning results.

[0125] To address the aforementioned technical issues, this invention improves the algorithm structure for engineering scenarios. First, it proposes a data quality vector generated from the output of a self-supervised model, uniformly characterizing noise level, drift probability, jump probability, and missing measurement reliability. This vector is then used throughout sensor weight setting, condition change detection gating, residual weighted correction, and threshold correction, achieving a quantifiable and transferable constraint mechanism for data quality. Second, it constructs a quality-weighted multi-sensor consistency change detection and a condition state machine with hysteresis judgment rules to reduce false triggers caused by single-point anomalies and short-term fluctuations. Step-by-step loading enables smooth transition updates of boundary conditions and conditions, thereby improving the continuity and stability of coupled numerical calculations. Third, it employs a two-layer dynamic threshold mechanism and decomposes and synthesizes monitoring reliability, condition identification reliability, model consistency reliability, and numerical stability reliability to form an overall reliability. This allows the system to distinguish between high-reliability risks and low-reliability anomalies when exceeding limits and output different handling results, thus more effectively achieving a closed loop of dynamic threshold early warning and risk handling.

Claims

1. A method for predicting and warning of foundation pit deformation based on numerical simulation coupling, characterized in that, include: S1. Obtain time-stamped monitoring data and construction records for the foundation pit project, and establish an initial coupled numerical model based on the survey data, design parameters, and construction records; S2. Perform self-supervised anomaly detection on the monitoring data, generate a data quality vector corresponding to the monitoring data, suppress outliers and complete missing data in the monitoring data to obtain quality control monitoring data; S3. Set multi-sensor weights for the quality control monitoring data based on the data quality vector, and perform condition change point detection based on multi-sensor consistency on the quality control monitoring data under its constraints, generate condition change point candidate points, and calculate the condition change point confidence; S4. Confirm the condition change point candidate points according to the preset hysteresis judgment rules, update the condition state machine and generate the current condition state, and generate the coupled numerical model update instruction; S5. Based on the coupled numerical model update command, perform a smooth transition update of the working conditions and boundary conditions on the initial coupled numerical model. Compare the quality control monitoring data with the calculation results of the updated coupled numerical model. Based on the data quality vector, weight the comparison residuals to obtain the weighted residuals. Use the weighted residuals as the target to iteratively update the model parameters to obtain the coupled numerical model after online correction. Perform rolling prediction to obtain the prediction results and generate a numerical stability index that characterizes the stability of the rolling prediction calculation. S6. For the preset early warning indicators, calculate the benchmark threshold based on the prediction results. The benchmark threshold is the prediction envelope threshold or prediction quantile threshold of the preset early warning indicators in the prediction time domain. S7. Based on the data quality vector and the current working condition, the baseline threshold is corrected to obtain the corrected threshold. The monitoring credibility is calculated based on the data quality vector. The working condition identification credibility is calculated based on the working condition change point confidence. The model consistency credibility is calculated based on the quality control monitoring data and the prediction results. The numerical stability credibility is calculated based on the numerical stability index. The monitoring credibility, working condition identification credibility, model consistency credibility, and numerical stability credibility are combined into the overall credibility. S8. Compare the quality control monitoring data with the correction threshold, calculate the degree and rate of exceeding the preset warning indicators, and carry out warning diversion and handling based on the overall credibility. When the overall credibility is not less than the preset credibility threshold and the degree or rate of exceeding the limit meets the preset exceeding criterion, output the warning result. When the overall credibility is less than the preset credibility threshold and the degree or rate of exceeding the limit meets the preset exceeding criterion, output the review and handling result.

2. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S1 includes: Acquire time-stamped monitoring data for the foundation pit project, wherein the monitoring data includes two or more of the following: retaining structure displacement data, surface settlement data, support axial force data, and groundwater level data; The monitoring data is time-aligned to form a unified time reference. Obtain construction records aligned with the unified time reference, the construction records including excavation depth, support installation information, support removal information, and dewatering start and stop information; The calculation domain and stratigraphic parameters are determined based on the survey data, the support structure parameters and support parameters are determined based on the design parameters, and the construction stage is determined based on the construction records. A coupled model of soil and support structure or a coupled model of soil, support structure and groundwater is established in the computational domain. The initial coupled numerical model includes the excavation unloading settings, support installation and dismantling settings and dewatering start and stop settings corresponding to the construction stage, and includes the initial stress conditions and the initial groundwater level conditions.

3. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S2 include: The monitoring data are grouped according to monitoring type and monitoring location, and multivariate time series are formed under a unified time base. A self-supervised anomaly detection model is constructed based on the multivariate time series. The self-supervised anomaly detection model obtains the reconstruction result by partially masking the multivariate time series and reconstructing the masked data. Based on the reconstruction residual between the reconstruction result and the monitoring data, a data quality vector is generated for each time step. The data quality vector includes noise level, drift probability, jump probability, and missing measurement confidence. The noise level is determined by the statistics of the reconstruction residual. The drift probability is determined by the low-frequency trend change of the reconstruction residual. The jump probability is determined by the degree of deviation between the adjacent time step difference of the monitoring data and the adjacent time step difference of the reconstruction result. The missing measurement confidence is determined by the missing measurement marker of the monitoring data and the stability of the reconstruction result. Based on the data quality vector, outlier suppression and missing data completion processing are performed on the monitoring data. Specifically, monitoring data with a jump probability exceeding a preset threshold are replaced or smoothed using the reconstruction result, monitoring data with a drift probability exceeding a preset threshold are drift corrected, and monitoring data with missing data are completed using the reconstruction result. This process yields quality control monitoring data, and the data quality vector and the quality control monitoring data are output.

4. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S3 includes: Based on the data quality vector, multi-sensor weights are determined for each monitoring type and each monitoring point in the quality control monitoring data. The multi-sensor weights are negatively correlated with the noise level in the data quality vector and positively correlated with the confidence level of missing data in the data quality vector. Under the constraint of the multi-sensor weights, a multi-monitoring type consistency index is calculated for the quality control monitoring data. The multi-monitoring type consistency index includes one or more of the following: consistency between the change in the displacement of the retaining structure and the change in the surface settlement, consistency between the change in the groundwater level and the change in the surface settlement, and consistency between the change in the axial force of the support and the change in the displacement of the retaining structure. Based on the consistency index of the multiple monitoring types, identify the abrupt change position of the consistency index along the time axis to generate candidate points of operating condition change. The confidence level of the operating condition change point is calculated for each candidate point. The confidence level is determined by the abrupt change of the consistency index at the candidate point, the weight of the multi-sensor system, and the stability of the consistency index within the time window before and after the candidate point. The candidate point and the confidence level of the operating condition change point are then output.

5. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S4 include: Arrange the candidate points of the operating condition change point in chronological order, and read the corresponding confidence level of the operating condition change point for each candidate point; The candidate points of the working condition change point are confirmed one by one according to the preset hysteresis determination rule. The preset hysteresis determination rule includes a confidence threshold condition and a duration condition. The confidence threshold condition is that the confidence of the working condition change point is not less than the preset confidence threshold. The duration condition is that the confidence of the working condition change point within the preset duration range before and after the candidate point of the working condition change point continuously meets the confidence threshold condition. When the candidate point of the working condition change point is confirmed as a working condition change point, the working condition state machine is updated according to the transition relationship of the working condition change point in the working condition state machine and the current working condition state is generated. Based on the current working condition, a coupled numerical model update instruction is generated. The coupled numerical model update instruction includes construction stage update information corresponding to the current working condition and boundary condition update information corresponding to the construction stage update information. The boundary condition update information includes one or more of the following: excavation unloading settings, support installation and removal settings, and dewatering start and stop settings.

6. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S5 includes: According to the coupled numerical model update instructions, the initial coupled numerical model is updated for the construction stage and boundary conditions. A step-by-step loading method is used to smoothly transition the updates for excavation unloading, support installation and removal, and dewatering start / stop, resulting in an updated coupled numerical model. A residual is established between the calculation results of the updated coupled numerical model and the quality control monitoring data, where the residual is the difference between the calculation results of the quality control monitoring data and the updated coupled numerical model at the same time and monitoring location. Weights are generated for the residuals based on the data quality vector, and the residuals are weighted to obtain weighted residuals. Using the reduction of the objective function value of the weighted residuals as the iterative update criterion, the model parameters of the updated coupled numerical model are iteratively updated to complete online correction, resulting in an online corrected coupled numerical model. Based on the online corrected coupled numerical model, rolling predictions are performed for subsequent time periods according to a preset prediction step size to obtain prediction results. The prediction results include one of the following: predicted retaining structure displacement, predicted surface settlement, predicted support axial force, and predicted groundwater level. The number of nonlinear iterations or the residual convergence ratio is recorded during each rolling prediction process to generate a numerical stability index.

7. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S6 include: For the preset early warning indicators, the corresponding prediction sequence is extracted within the prediction time domain based on the prediction results; When the preset early warning indicator is a displacement indicator or a settlement indicator, the prediction envelope threshold is determined by the maximum or minimum value of the prediction sequence in the prediction time domain as the benchmark threshold. When the preset early warning indicator is a rate-type indicator, the prediction sequence is differentially divided between adjacent time points to obtain a prediction rate sequence, and the prediction envelope threshold is determined as the benchmark threshold based on the maximum or minimum value of the prediction rate sequence in the prediction time domain. When a predicted quantile threshold is used as a baseline threshold, a threshold corresponding to a preset quantile is calculated for the predicted sequence or the predicted rate sequence as a baseline threshold, wherein the preset quantile is a quantile greater than 50% and less than 100%.

8. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S7 includes: A quality correction coefficient is generated based on the data quality vector, wherein the quality correction coefficient increases with the increase of noise level in the data quality vector, increases with the increase of drift probability in the data quality vector, increases with the increase of jump probability in the data quality vector, and increases with the decrease of missing measurement confidence in the data quality vector; a working condition correction coefficient is generated based on the current working condition state, wherein the working condition correction coefficient is determined by the construction stage corresponding to the current working condition state; a correction threshold is obtained by performing correction operations on the benchmark threshold with the quality correction coefficient and the working condition correction coefficient respectively; monitoring confidence is calculated based on the data quality vector; working condition identification confidence is calculated based on the working condition change point confidence; model consistency residual is calculated based on the difference between quality control monitoring data and prediction results at the same time, and model consistency confidence is calculated based on the statistics of the model consistency residual; numerical stability confidence is calculated based on the numerical stability index. The overall credibility is obtained by combining the monitoring credibility, the working condition identification credibility, the model consistency credibility, and the numerical stability credibility according to a preset composite weight, wherein the preset composite weight is non-negative and the sum of the weights is 1.

9. The method for predicting and warning of foundation pit deformation based on numerical simulation coupling according to claim 1, characterized in that, S8 includes: The quality control monitoring data is used to calculate the measured value of the early warning indicator according to the preset early warning indicator, and the measured value of the early warning indicator is compared with the correction threshold at each time step. When the measured value of the warning indicator exceeds the correction threshold, the degree of exceeding the limit is calculated according to the difference or ratio between the measured value of the warning indicator and the correction threshold, and the rate of exceeding the limit is calculated by the difference between adjacent time moments of the measured value of the warning indicator. Whether a preset over-limit criterion is met is determined based on the degree of over-limit and the rate of over-limit, wherein the preset over-limit criterion includes one of an over-limit degree threshold condition or an over-limit rate threshold condition; When the preset over-limit criterion is met, a warning and diversion process is performed based on the overall credibility. When the overall credibility is not less than the preset credibility threshold, a warning result is output. When the overall credibility is less than the preset credibility threshold, a review process result is output. The review process result includes one of triggering monitoring review, triggering encrypted monitoring, and triggering backtracking recalculation.

Images