Foundation pit supporting structure deformation prediction and early warning method based on multi-source data fusion

Abstract

The invention provides a foundation pit supporting structure deformation prediction and early warning method based on multi-source data fusion, and the method comprises the steps: deploying a differentiated sensor network at a key part of a supporting structure, and collecting multi-source time sequence data; performing space-time alignment and layered noise suppression on the multi-source time series data, and constructing a data quality evaluation matrix to screen effective data; time features, spatial features and mechanical features of the effective data are extracted, and a fusion feature data set is constructed; establishing a physical mechanism model and a data driving model, and generating a dual-core prediction model through a weighted fusion mechanism; deformation index rolling prediction is carried out based on a dual-core prediction model, and third-level early warning management is executed according to a prediction result; and performing online performance monitoring on the dual-core prediction model, and starting incremental learning to optimize model parameters when a prediction error is greater than a set threshold value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of foundation pit construction, and in particular to a foundation pit support structure deformation prediction and early warning method based on multi-source sensor data fusion. BACKGROUND

[0002] Foundation pit engineering refers to a series of support, reinforcement, dewatering, earthwork excavation and backfilling engineering measures taken to ensure the safety of foundation pit construction and protect the surrounding environment when constructing the underground part of a building or structure. Under the background of urban construction developing towards underground space, the scale and complexity of foundation pit engineering are constantly improving, and its safety problems are increasingly prominent. Accurate prediction and timely early warning of foundation pit support structure deformation are of great significance to ensure engineering safety and prevent accidents.

[0003] However, the traditional foundation pit deformation monitoring and prediction method has the following shortcomings. First, in the data collection layer, the sensor network layout often lacks pertinence and system optimization, and it is difficult to differentiate deployment according to the stress and deformation characteristics of key parts of the support structure. The frequency and range of data collection are usually fixed and cannot be dynamically adjusted according to the risk level and deformation rate at different stages of foundation pit construction. At the same time, the synchronous access and integration of external environmental data and construction condition data are not comprehensive and timely enough, resulting in limitations in data sources and difficulty in fully reflecting the true state of the foundation pit.

[0004] Secondly, in the data processing aspect, the alignment accuracy of multi-source data in time and space dimensions is often insufficient, and the noise suppression methods for different types of data lack hierarchical pertinence. For example, GNSS displacement data, optical fiber strain data, and pore water pressure data have their own unique noise characteristics, and a unified denoising method cannot achieve the best results. In addition, the data quality evaluation system is not systematic and perfect, and the method for completing missing data is relatively simple, which will affect the reliability of subsequent analysis.

[0005] Thirdly, in the model construction layer, existing methods mostly rely on single physical mechanism model or data-driven model. Although the physical mechanism model can reflect the physical laws and constraints of foundation pit deformation, it has limited ability to describe nonlinear deformation relationships under complex conditions. Data-driven models have strong nonlinear fitting capabilities, but lack physical constraints and are prone to prediction errors when data is insufficient or conditions change greatly. Both types of models have their own advantages and disadvantages, and it is difficult to fully utilize their respective advantages when used alone. At the same time, there is a lack of cross-modal feature extraction and fusion, making it difficult to accurately capture the complex rules of foundation pit deformation in time, space, and mechanics.

[0006] Finally, the prediction model generally lacks online learning and full-cycle performance optimization mechanism. With the advancement of foundation pit engineering, geological conditions, construction conditions and other factors are constantly changing, and fixed model parameters cannot adapt to these dynamic changes in time, resulting in a decline in prediction accuracy over time. The accuracy and timeliness of the early warning system are difficult to guarantee, and the decision support capability of risk tracing and simulation deduction is relatively lacking, which cannot provide effective emergency decision basis for engineering management personnel.

[0007] Therefore, it is urgent to develop a foundation pit support structure deformation prediction and early warning method which can realize comprehensive collection and accurate fusion of multi-source data, has high model prediction accuracy, timely and effective early warning, decision support capability and sustainable optimization of model, so as to meet the needs of modern foundation pit engineering safety management. SUMMARY

[0008] The application provides a foundation pit support structure deformation prediction and early warning method based on multi-source sensor data fusion to solve the technical problems of incomplete data collection, low processing accuracy, inaccurate model prediction and untimely early warning in the prior art.

[0009] To solve the above technical problems, the application provides a foundation pit support structure deformation prediction and early warning method based on multi-source sensor data fusion, which comprises the following steps: Step S1: deploying a differentiated sensor network at key positions of the support structure to collect multi-source time series data; Step S2: performing space-time alignment and layered noise suppression on the multi-source time series data to construct a data quality evaluation matrix to screen effective data; Step S3: extracting time features, space features and mechanical features of the effective data to construct a fusion feature data set; Step S4: establishing a physical mechanism model and a data-driven model to generate a dual-core prediction model through a weighted fusion mechanism; Step S5: performing rolling prediction of deformation indicators based on the dual-core prediction model, and executing three-level early warning management according to the prediction results; Step S6: performing online performance monitoring on the dual-core prediction model, and starting incremental learning to optimize model parameters when the prediction error is greater than a set threshold.

[0010] Preferably, in step S1, the differentiated sensor network comprises a GNSS receiver, a fiber Bragg grating sensor, an inclinometer, a soil pressure cell and a pore water pressure gauge; the GNSS receiver is deployed at the top of the pile for monitoring displacement, the fiber Bragg grating sensor is deployed at the anchor rod point for monitoring strain, the inclinometer is deployed on the surface of the support structure for monitoring the inclination angle, and the soil pressure cell and the pore water pressure gauge are deployed in the rock-soil environment for monitoring environmental pressure.

[0011] Preferably, the multi-source time-series data acquisition in step S1 adopts a dynamic sampling strategy, which automatically adjusts the sampling frequency according to the construction stage; the sampling frequency during the excavation and support stages is set to A minutes per time, and the sampling frequency during the stabilization stage is set to B minutes per time, where B is greater than A.

[0012] Preferably, the spatiotemporal alignment method in step S2 includes: using a unified timestamp to mark the multi-source time-series data collected by all sensors, establishing a three-dimensional spatial coordinate system integrating BIM and GIS, mapping the multi-source time-series data to the corresponding spatial position in the three-dimensional spatial coordinate system, and realizing the matching of the multi-source time-series data in the time dimension and spatial dimension.

[0013] Preferably, the hierarchical noise suppression in step S2 employs corresponding denoising methods for different types of data: Kalman filtering is used to denoise the displacement data collected by the GNSS receiver; wavelet thresholding is used to denoise the strain data collected by the fiber Bragg grating sensor; and moving average filtering is used to denoise the water pressure data collected by the pore water pressure gauge. The data quality assessment matrix evaluates data quality from three dimensions: completeness, accuracy, and consistency, and K-nearest neighbor spatiotemporal interpolation is used to complete missing data.

[0014] Preferably, in step S3, the temporal features are deformation trend features and periodic features extracted using an LSTM neural network; the spatial features are spatial correlation features of adjacent measurement points extracted using a CNN convolutional neural network; the mechanical features are obtained by deriving the stress-deformation relationship through finite element analysis; and the importance of the temporal features, spatial features, and mechanical features is evaluated using a random forest algorithm, assigning higher weights to features with high contributions.

[0015] Preferably, the physical mechanism model in step S4 is established using the finite element method, simulating the working condition response based on geological parameters and support structure design parameters; the data-driven model includes an LSTM neural network model and an XGBoost gradient boosting model; the weight calculation method for the weighted fusion mechanism includes: (1) Calculate the historical prediction errors of the physical mechanism model, the LSTM neural network model and the XGBoost gradient boosting model on the validation dataset. The historical prediction errors are represented by the root mean square error (RMSE). (2) Evaluate the overall performance of each model according to the Bayesian Information Criterion (BIC), whereby the overall performance considers both prediction accuracy and model complexity. (3) Based on the historical prediction error and the overall performance, the weighting coefficients are calculated using an exponential decay function. The formula for calculating the weighting coefficients is as follows: ; in For the firsti The weight coefficients of each model, and To adjust the parameters, For the first i The root mean square error of each model For the first i Bayesian information criterion values ​​for each model; (4) Recalculate the weight coefficients based on the latest prediction performance to achieve dynamic updating of the weight coefficients.

[0016] Preferably, the rolling prediction of deformation indicators in step S5 adopts a sliding time window mechanism, the length of which is 72 hours. Based on the dual-core prediction model, the rolling prediction of deformation indicators for the next 12 hours, 24 hours, and 48 hours is performed. The three-level early warning management includes blue, yellow, and red warnings. The blue warning corresponds to the normal state and maintains the regular monitoring frequency. The yellow warning is triggered when the predicted value approaches the limit and the monitoring frequency is increased to once every 2 minutes. The red warning is triggered when the predicted value exceeds the limit and the audible and visual alarm device is activated.

[0017] Preferably, the warning threshold of the three-level early warning management adopts a dynamic update mechanism. The dynamic update mechanism is based on the Bayesian update algorithm and updates the warning threshold every 7 days by combining historical engineering data and real-time monitoring information. Under soft soil geological conditions or when the surrounding load increases, the warning threshold is reduced.

[0018] Preferably, the online performance monitoring method in step S6 includes: periodically comparing the prediction results of the dual-core prediction model with the actual monitoring data to calculate the prediction error; when the prediction error is greater than the initial error for several consecutive prediction cycles, it is determined that the prediction error is greater than a set threshold, and incremental learning is performed.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention, through differentiated sensor network deployment and dynamic sampling strategies, combined with the synchronous access of external environment and construction condition data, can acquire foundation pit-related data in a comprehensive, multi-dimensional, and timely manner, providing a sufficient and accurate data source for subsequent analysis.

[0020] This invention achieves accurate spatiotemporal alignment of multi-source data based on a unified timestamp and BIM+GIS spatial coordinate system. It employs a hierarchical noise suppression strategy to denoise different types of data and constructs a data quality assessment matrix system to filter data, effectively improving the accuracy, completeness, and consistency of the data and ensuring the reliability of subsequent analysis.

[0021] The application realizes intuitive visualization of deformation, internal force and other data by constructing a foundation pit dynamic digital twin, integrating BIM and GIS technologies to establish a high-precision model, and dynamically mapping multi-source real-time data to the twin, thereby providing an efficient monitoring and analysis platform for engineering managers and making the foundation pit state clear at a glance.

[0022] The application constructs a physical mechanism and data-driven dual-core prediction model, organically combines the advantages of the two types of models through a weighted fusion mechanism, utilizes the physical constraints of the physical mechanism model, and plays the nonlinear fitting capability of the data-driven model, compared with a single model method, the prediction accuracy is significantly improved, and a reliable prediction basis is provided for early warning.

[0023] In summary, the foundation pit supporting structure deformation prediction and early warning method based on multi-source sensor data fusion provided by the application optimizes and improves each link from data acquisition, processing, feature extraction, model prediction, early warning management to decision support, forms a complete technical system, can effectively solve many problems existing in the prior art, significantly improves the safety management level of foundation pit engineering, and has important engineering application value and broad popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The method flowchart of the embodiment of the application is shown in the figure; Figure 2 The multi-source sensor network optimization layout and dynamic data acquisition flowchart of the application is shown in the figure; Figure 3 The space-time alignment and layered data cleaning flowchart of the application is shown in the figure; Figure 4 The foundation pit dynamic digital twin construction flowchart of the application is shown in the figure; Figure 5 The cross-modal feature extraction and fusion data set construction flowchart of the application is shown in the figure; Figure 6 The physical mechanism + data-driven dual-core prediction model construction flowchart of the application is shown in the figure; Figure 7 The real-time rolling prediction and dynamic early warning management flowchart of the application is shown in the figure; Figure 8 The risk tracing and simulation deduction decision support flowchart of the application is shown in the figure; Figure 9 The model online learning and full-cycle performance optimization flowchart of the application is shown in the figure. DETAILED DESCRIPTION

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0026] like Figures 1 to 9 As shown, this embodiment of the invention provides a method for predicting and warning the deformation of foundation pit support structures based on multi-source data fusion. Its core lies in establishing a complete technical chain from data acquisition, processing, feature extraction, model prediction to early warning decision-making, including the following steps.

[0027] Step S1: Optimization and Deployment of Multi-Source Sensor Network and Dynamic Data Acquisition The purpose of this step is to construct a comprehensive, multi-dimensional foundation pit perception network to provide sufficient and accurate data sources for subsequent analysis.

[0028] First, differentiated sensor deployments are implemented based on key components of the support structure, such as pile tops, anchor points, and the surrounding ground surface. Specifically, GNSS receivers are deployed at pile tops to monitor displacement changes. Utilizing the high-precision positioning capabilities of the Global Navigation Satellite System, GNSS receivers can achieve real-time three-dimensional displacement monitoring with millimeter-level accuracy. Fiber Bragg grating sensors are deployed at anchor points to monitor strain. Based on the optical modulation principle of optical fibers, these sensors can accurately acquire deformation information when strain occurs in the support structure, achieving a resolution down to the micro-strain level. Inclinometers are deployed on the surface of the support structure to monitor tilt angles. Employing MEMS technology, these inclinometers achieve a measurement accuracy of 0.01°, enabling them to keenly detect tilting and instability trends. Earth pressure cells and pore water pressure gauges are deployed in the soil and rock environment to measure the pressure exerted by the soil and rock on the support structure and monitor changes in pore water pressure, respectively, providing crucial data support for analyzing the effects of the soil and rock environment on the support structure. This differentiated deployment forms a three-dimensional sensing network covering space, mechanics, and the environment, enabling comprehensive perception of the state changes of the foundation pit support structure and its surrounding environment.

[0029] Secondly, a dynamic sampling strategy was designed to adapt to the monitoring needs of different construction stages. During the excavation and support phases, the foundation pit deformation is relatively active and the risk level is high. Therefore, the sampling frequency was increased to 5 minutes per sampling to capture deformation details and key change points in a timely manner. During the stable phase, the foundation pit deformation tends to level off, and the sampling frequency can be reduced to 30 minutes per sampling. This meets the monitoring needs while reducing data redundancy and saving system resources. This strategy of dynamically adjusting the sampling frequency according to the risk level and deformation rate of the construction stage achieves an optimal balance between data acquisition efficiency and cost.

[0030] During the excavation period, the soil body of the foundation pit is continuously excavated, and the soil pressure and other load conditions of the supporting structure change dramatically, with a fast deformation rate. If the sampling frequency is too low, the key deformation nodes may be missed, leading to deviations in the subsequent prediction model due to data missing. During the supporting period, the supporting structure is being installed or adjusted, and the stress and deformation of the structure are also in a dynamic change process, requiring a high sampling frequency to track the deformation in real time. During the stable period, the main construction of the foundation pit has been completed, and the stress and deformation of the supporting structure gradually stabilize. At this time, a high sampling frequency will generate a large amount of repeated data with low value for subsequent analysis, increasing the cost of data storage and processing. Through this dynamic sampling strategy based on the risk level and deformation rate of the construction stage, the timeliness and detail of data collection can be ensured during key construction stages, meeting the data density requirements of deformation prediction. At the same time, the sampling frequency is reasonably reduced during the stable period, achieving a balance between data collection efficiency and cost. The collected data accurately reflects the real state of the foundation pit at different stages, avoiding unnecessary resource waste, and providing an efficient and high-quality data source for subsequent data processing and model construction.

[0031] Thirdly, synchronize external environmental data and construction condition data to ensure the comprehensiveness and timeliness of the data source. External environmental data includes precipitation, ground vibration, and other factors that may affect the stability of the foundation pit. Construction condition data includes excavation depth, support state, load changes, and other real-time construction information. Integrating these data with sensor monitoring data can more comprehensively reflect the dynamic environment of the foundation pit, providing a more complete information base for subsequent deformation prediction and early warning.

[0032] Step S2: Spatiotemporal alignment and layered data cleaning This step aims to solve the fusion problem of multi-source heterogeneous data, improving the accuracy, completeness, and consistency of the data.

[0033] Firstly, spatiotemporal alignment of multi-source data is achieved. In the time dimension, a unified timestamp technology is adopted, with the help of high-precision clock synchronization protocols such as NTP (Network Time Protocol), to ensure that the time references of different sensors, external environment monitoring devices, and construction condition collection devices are consistent, with an error within milliseconds, ensuring strict synchronization of various data in the time dimension. In the spatial dimension, precise correspondence is achieved based on the BIM+GIS spatial coordinate system. The BIM model provides accurate three-dimensional geometric information and attribute information of each component of the supporting structure, and GIS provides real geographic spatial references including terrain, geological layers, etc. Through coordinate conversion and matching algorithms, the sensor installation location, external environmental impact area, and construction condition occurrence location are accurately associated to a unified three-dimensional spatial coordinate system, achieving accurate correspondence of multi-source data in space, eliminating the misalignment problem in the time and space dimensions.

[0034] Secondly, a hierarchical noise suppression strategy is adopted to denoise different types of data according to their noise characteristics. For GNSS displacement data, which is susceptible to satellite signal blockage, multipath effects and other factors, Kalman filter is used for denoising. Kalman filter is based on state space model, combining the observation equation and motion equation of GNSS, which can make optimal estimation on displacement data containing noise, effectively filtering out random noise while preserving the true trend of displacement. For strain data collected by fiber Bragg grating sensors, wavelet threshold denoising method is used. First, wavelet decomposition is performed on strain data to obtain wavelet coefficients at different scales. Then, according to the preset threshold rule, the coefficients containing noise are suppressed or set to zero. Finally, through wavelet reconstruction, the pure strain signal is recovered. This method can well preserve the key features such as mutation in strain data. For water pressure data collected by pore water pressure gauges, moving average filter is used. By selecting an appropriate moving window size, the water pressure data within the window is averaged to smooth the random noise in the data, making the overall trend of water pressure change clearer.

[0035] Thirdly, a data quality evaluation matrix is constructed to systematically evaluate the data from three dimensions of integrity, accuracy and consistency. Integrity evaluation checks for missing data, accuracy evaluation judges whether the data is within a reasonable range and conforms to physical laws, and consistency evaluation verifies whether there are contradictions between data from different sources. For missing data selected, K- nearest neighbor spatiotemporal interpolation method is used for completion. This method considers the information of neighboring data points in time and space, estimates the missing value through weighted average, etc., effectively ensuring the integrity of the data.

[0036] Through the hierarchical noise suppression strategy, the quality of various types of data is greatly improved by using specialized denoising methods for different noise characteristics of data, reducing the interference of noise on subsequent feature extraction and model training, and making the features extracted from the data more truly reflect the actual state of the foundation pit, providing high-quality data support for building high-precision prediction models.

[0037] Step S3: Constructing dynamic digital twin of foundation pit This step constructs a virtual mapping of the foundation pit, providing a platform for visual monitoring and simulation of deformation.

[0038] First, the BIM and GIS technologies are integrated to establish high-precision three-dimensional geological and supporting structure models as the basis for digital twinning. Using BIM technology, based on three-dimensional modeling software, the three-dimensional geometric models of supporting piles, anchor rods, and supports are accurately created according to design drawings, and each component is given material properties, mechanical parameters, and other information, enabling the model to accurately reflect the actual structure and mechanical properties of the supporting structure. Using GIS technology, a three-dimensional geological model containing information such as stratum distribution, geotechnical mechanical properties, and groundwater distribution is constructed on the GIS platform using geological exploration data, and a spatial correlation between the geological model and the BIM supporting structure model is established, enabling seamless integration of the two in space and accurately reflecting the geological and supporting structure conditions of the foundation pit.

[0039] Second, the cleaned multi-source real-time data is mapped to the twin body through a streaming data processing platform such as Kafka, enabling dynamic visualization of deformation, internal forces, and other data. Cloud maps, vector arrows, and other visualization formats are used to visually display the deformation size and direction of each part of the supporting structure, internal force distribution, and changes in the surrounding geological environment. This dynamic visualization enables project managers to monitor the overall operation of the foundation pit in real time and comprehensively, and promptly identify any abnormalities.

[0040] Step S4: Cross-modal feature extraction and fusion dataset construction This step comprehensively mines information about foundation pit deformation from multiple dimensions.

[0041] In terms of time feature extraction, a long short-term memory (LSTM) neural network is used to process time series data. LSTM, with its unique memory cell structure, can learn the trends of time series data such as foundation pit deformation over long time spans, such as changes in deformation rates at different construction stages, while also capturing periodic patterns in the data, such as periodic deformations that may be caused by factors such as daily temperature differences and tides, extracting trend and periodic features.

[0042] In terms of spatial feature extraction, a convolutional neural network (CNN) is used to process spatial distribution data. CNN uses convolution kernels to slide across the space, enabling the extraction of correlation patterns between adjacent monitoring points, such as the synchronization or transmission of displacement changes at adjacent pile tops, reflecting the spatial distribution of foundation pit deformation.

[0043] In terms of mechanical feature extraction, the stress-deformation relationship is derived through finite element analysis. Based on the mechanical model of the foundation pit supporting structure, geological parameters, supporting structure design parameters, and other information are input, and the stress distribution of the supporting structure under different working conditions is calculated through finite element analysis, and the quantitative relationship between stress and deformation is obtained. These mechanical features reflect the physical mechanisms of foundation pit deformation.

[0044] The importance of the extracted temporal, spatial, and mechanical features is evaluated using a random forest algorithm. Random forest can quantify the contribution of each feature to the prediction of foundation deformation by constructing multiple decision trees and integrating their results. High-contribution features are given higher weights, allowing the subsequent model training to focus more on key features and improve prediction accuracy.

[0045] Finally, the extracted multi-source features are spliced and fused to construct a unified feature and label dataset for model training. This dataset contains rich information about foundation deformation in multiple dimensions such as time, space, and mechanics, laying a solid foundation for subsequent construction of high-precision prediction models.

[0046] Through multi-dimensional feature extraction, the information about foundation support structure deformation contained in multi-source sensor data is comprehensively and deeply mined, providing rich and targeted input for subsequent construction of "feature-label" dataset and training prediction model. This allows the model to learn and understand the rules of foundation deformation from multiple dimensions, improving the accuracy of foundation deformation prediction.

[0047] Step S5: Construction of physical mechanism and data-driven dual-core prediction model This step combines physical mechanism model and data-driven model to leverage their respective advantages.

[0048] First, establish the physical mechanism model. Using the finite element method, based on the geological parameters of the foundation (such as soil distribution, soil mechanical parameters, etc.) and support structure design parameters (such as support pile size, anchor prestress, etc.), simulate the stress and deformation response of the support structure under different construction conditions. The physical mechanism model can provide physical constraints that conform to the laws of mechanics, ensuring that the prediction results are physically reasonable. Especially when there is insufficient data or the working conditions change greatly, it can provide a reasonable prediction trend based on known physical laws.

[0049] Second, establish the data-driven model. Construct LSTM neural network model and XGBoost gradient boosting model. LSTM model is good at processing time series data and can learn the non-linear mapping relationship between multi-source features and deformation sequence, capturing the complex evolution law of deformation over time. XGBoost model integrates multiple weak learners through gradient boosting, has strong non-linear fitting ability and feature interaction mining ability, and can learn the complex relationship between features and deformation from another perspective.

[0050] Secondly, a weighted fusion mechanism is used to dynamically allocate weights based on the historical prediction errors of each model. Specifically, the historical prediction errors of the physical mechanism model, LSTM model, and XGBoost model on the validation dataset are first calculated, using the root mean square error (RMSE) as the error metric. Then, the overall performance of each model is evaluated according to the Bayesian Information Criterion (BIC), which considers both prediction accuracy and model complexity, thus avoiding overfitting. Based on the historical prediction errors and overall performance, the weight coefficients are calculated using an exponential decay function. The formula for calculating the weight coefficients is as follows: ; in For the first i The weight coefficients of each model, and To adjust the parameters, For the first i The root mean square error of each model For the first i The Bayesian information criterion value of each model.

[0051] Models with smaller errors and better performance will receive higher weights. For example, under a certain working condition, the LSTM model might be assigned 60% weight, the finite element physical model 25%, and the XGBoost model 15%. These weights are recalculated based on the latest predictive performance, achieving dynamic updates to ensure that the weight allocation always matches the model's current predictive capability. This weighted fusion mechanism fully combines the physical mechanism model's ability to accurately characterize the physical laws of foundation pit deformation with the data-driven model's ability to fit complex nonlinear deformation relationships, significantly improving overall prediction accuracy.

[0052] By using a weighted voting fusion mechanism, the physical mechanism model's ability to accurately depict the physical laws of foundation pit deformation and the data-driven model's ability to fit complex nonlinear deformation relationships are fully combined. This allows the fused model to leverage the advantages of each sub-model and compensate for the shortcomings of a single model, thereby significantly improving the overall prediction accuracy of foundation pit support structure deformation and providing more reliable prediction data for subsequent early warning management.

[0053] Step S6: Real-time rolling forecast and dynamic early warning management This step enables the prediction of future deformation of the foundation pit and timely early warning of risks.

[0054] Firstly, based on the latest fusion data and the dual-core prediction model, a sliding time window mechanism is used for rolling prediction. The length of the sliding time window is set to 72 hours, and the model uses the historical data within the window to predict the deformation indicators such as displacement size and strain change for the next 12 hours, 24 hours and 48 hours. This rolling prediction method can update the prediction results in real time as new data is continuously collected, ensuring the timeliness and accuracy of the prediction.

[0055] Secondly, a dynamic early warning threshold is constructed. Using the Bayesian update algorithm, historical engineering data is used as prior knowledge and real-time monitoring information as new evidence. The Bayesian formula is used to continuously update the probability distribution estimate of the foundation pit deformation risk, and then adjust the early warning threshold. Combined with historical engineering data and real-time monitoring information, the threshold is updated every 7 days, so that the threshold can adapt to the dynamic changes of the foundation pit working conditions. In special working conditions, such as soft soil conditions or increased surrounding load, the threshold will be adjusted downward to improve the sensitivity of the early warning system and capture potential risks in a timely manner.

[0056] Thirdly, a three-level early warning management is implemented, including blue, yellow and red early warnings. Blue early warning corresponds to the normal state, at which the regular monitoring frequency is maintained, and data is collected and analyzed regularly. Yellow early warning is triggered when the predicted value approaches the limit, indicating that the foundation pit deformation has a risk trend, and the monitoring frequency needs to be increased to 2 minutes per time to strengthen the tracking of deformation, and the geological review is started to check whether there are potential problems such as changes in geological conditions. Red early warning is triggered when the predicted value exceeds the limit, indicating that the foundation pit is in a dangerous state, at which the sound and light alarm device is triggered to send an emergency alarm to relevant personnel, and emergency suggestions such as stopping construction immediately, supporting and reinforcing, etc. are pushed, and supporting and reinforcing measures are started to maximize the safety of the project.

[0057] By constructing a dynamic early warning threshold based on the Bayesian update algorithm, the early warning threshold can be dynamically adjusted with the changes of the foundation pit construction process, external environment, geological conditions and other factors, rather than being fixed, thereby improving the accuracy and timeliness of the early warning, and enabling the early warning system to more accurately identify abnormal conditions of the foundation pit supporting structure deformation, providing more effective protection for the safety of the project.

[0058] Step S7: Risk Traceability and Simulation Deduction Decision Support This step provides risk analysis and decision-making basis for engineering management.

[0059] Firstly, the key contributing features of the early warning events are analyzed using explainable AI techniques such as SHAP (SHapley Additive exPlanations), enabling risk tracing. SHAP is based on the Shapley value concept in game theory, which can quantify the contribution of each feature to the prediction result, thereby identifying the key factors leading to early warning, such as sudden displacement of a monitoring point, abnormal increase in soil pressure, or external environmental factors (e.g., heavy rainfall). Through risk tracing, engineering managers can clearly identify the root cause of the problem and take targeted measures.

[0060] Secondly, the effects of various intervention measures are simulated in the digital twin, providing quantitative support for decision-making. By leveraging the simulation capabilities of the digital twin, the impact of different response plans, such as reinforcing the support structure, implementing dewatering measures, or adjusting the support state, on the deformation of the foundation pit is simulated, and the effects of each plan are evaluated and quantified. For example, the simulation shows how much the deformation of the support structure can be reduced by adding anchor rods at a certain location, and how the deformation rate changes. By comparing the simulation results of different plans, engineering managers can choose the optimal response plan and make scientific decisions.

[0061] Through the model performance monitoring mechanism, the changes in the model's prediction ability for the deformation of the foundation pit support structure can be monitored in real time. When the model's adaptability decreases, online incremental learning or fine-tuning is initiated in a timely manner, enabling the model to continuously absorb the information about changes in the foundation pit working conditions contained in new data, continuously optimizing its prediction performance, and ensuring that the model maintains high prediction accuracy throughout the entire foundation pit construction cycle, providing reliable model support for subsequent deformation warning and other work.

[0062] Step S8: Online learning of the model and optimization of performance throughout the cycle This step ensures that the model maintains good prediction performance throughout the entire foundation pit engineering cycle.

[0063] Firstly, a model performance monitoring mechanism is established. The prediction results of the dual-core prediction model are compared with the actual monitoring data on a regular basis, and the prediction error, such as mean squared error and mean absolute error, is calculated. When the prediction error of consecutive periods, such as 3 consecutive periods, is greater than the error threshold in the initial training phase, it is determined that the prediction error is continuously increasing, indicating that the model's adaptability to the current foundation pit working conditions has decreased. At this time, online incremental learning or fine-tuning mechanisms are initiated. Online incremental learning updates the model's parameters gradually without retraining the entire model, while fine-tuning adjusts some key parameters slightly based on the original model for new working condition data. Through this approach, the model can learn the working condition change information contained in new data in a timely manner, maintaining accurate prediction ability for the current foundation pit state.

[0064] Secondly, after completing each foundation pit project, the data of the project and the predicted effect are fed back to the database, and the model parameters are optimized through transfer learning. Transfer learning can transfer the experience and knowledge accumulated in the completed projects to new projects, improve the universality and adaptability of the model in different foundation pit projects, and make the method better applied to various foundation pit projects.

[0065] Finally, the system performance is continuously iterated and optimized through field test comparison and verification. Performance indicators such as prediction error less than 3%, early warning time more than 24 hours, system response time less than 30 minutes, etc. are set to verify in actual engineering. According to the verification results, the algorithm parameters and model structure are continuously adjusted and optimized to ensure that the system achieves the expected effect and continuously improves in actual application.

[0066] Through the model performance monitoring mechanism, the change of the model's prediction ability for the deformation of the foundation pit supporting structure can be mastered in real time. When the adaptability of the model decreases, online incremental learning or fine tuning is started in time, so that the model can continuously absorb the information of the change of the foundation pit working condition contained in the new data, continuously optimize its prediction performance, and ensure that the model can maintain high prediction accuracy throughout the whole cycle of the foundation pit construction, providing reliable model support for subsequent work such as deformation warning.

[0067] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, only the preferred embodiments of the present application are expressed, and the description is more specific and detailed, but it should not be understood as limiting the scope of the present application. As long as the combination of these technical features does not exist contradictory, it should be considered as the scope of the present application.

[0068] It should be noted that for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for predicting and warning the deformation of foundation pit support structures based on multi-source data fusion, characterized in that: Includes the following steps: Step S1: Deploy a differentiated sensor network at key parts of the support structure to collect multi-source time-series data; Step S2: Perform spatiotemporal alignment and hierarchical noise suppression on the multi-source time series data, and construct a data quality assessment matrix to filter valid data; Step S3: Extract the temporal, spatial, and mechanical features of the effective data to construct a fused feature dataset; Step S4: Establish a physical mechanism model and a data-driven model, and generate a dual-core prediction model through a weighted fusion mechanism; Step S5: Perform rolling prediction of deformation indicators based on the dual-core prediction model, and implement three-level early warning management based on the prediction results; Step S6: Perform online performance monitoring on the dual-core prediction model. When the prediction error exceeds a set threshold, start incremental learning to optimize the model parameters.

2. The method for predicting and warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 1, characterized in that: In step S1, the differentiated sensor network includes a GNSS receiver, a fiber Bragg grating sensor, an inclinometer, an earth pressure cell, and a pore water pressure gauge. The GNSS receiver is deployed at the top of the pile to monitor displacement, the fiber Bragg grating sensor is deployed at the anchor point to monitor strain, the inclinometer is deployed on the surface of the support structure to monitor the tilt angle, and the earth pressure cell and the pore water pressure gauge are deployed in the soil and rock environment to monitor environmental pressure.

3. The method for predicting and early warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 1, characterized in that: The multi-source time-series data collection in step S1 adopts a dynamic sampling strategy, which automatically adjusts the sampling frequency according to the construction stage; the sampling frequency during the excavation and support stages is set to A minutes per time, and the sampling frequency during the stabilization stage is set to B minutes per time, where B is greater than A.

4. The method for predicting and warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 1, characterized in that: The spatiotemporal alignment method described in step S2 includes: using a unified timestamp to mark the multi-source time-series data collected by all sensors, establishing a three-dimensional spatial coordinate system integrating BIM and GIS, mapping the multi-source time-series data to the corresponding spatial position in the three-dimensional spatial coordinate system, and realizing the matching of the multi-source time-series data in the time dimension and spatial dimension.

5. The method for predicting and early warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 1, characterized in that: The layered noise suppression described in step S2 employs corresponding denoising methods for different types of data: Kalman filtering is used for displacement data acquired by GNSS receivers; wavelet thresholding is used for strain data acquired by fiber Bragg grating sensors. The water pressure data collected by the pore water pressure gauge is filtered and denoised using a moving average method. The data quality assessment matrix evaluates the data quality from three dimensions: completeness, accuracy, and consistency. Missing data is filled using the K-nearest neighbor spatiotemporal interpolation method.

6. The method for predicting and early warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 1, characterized in that: In step S3, the temporal features are the deformation trend features and periodic features extracted using an LSTM neural network; the spatial features are the spatial correlation features of adjacent measurement points extracted using a CNN convolutional neural network. The mechanical characteristics were obtained by deriving the stress-deformation relationship through finite element analysis; the importance of the temporal characteristics, spatial characteristics and mechanical characteristics was evaluated using the random forest algorithm, and higher weights were assigned to characteristics with high contributions.

7. The method for predicting and early warning of deformation of foundation pit support structure based on multi-source data fusion according to claim 1, characterized in that: The physical mechanism model described in step S4 is established using the finite element method, simulating the working condition response based on geological parameters and support structure design parameters; the data-driven model includes an LSTM neural network model and an XGBoost gradient boosting model; the weight calculation method for the weighted fusion mechanism includes: (1) Calculate the historical prediction errors of the physical mechanism model, the LSTM neural network model and the XGBoost gradient boosting model on the validation dataset. The historical prediction errors are represented by the root mean square error (RMSE). (2) Evaluate the overall performance of each model according to the Bayesian Information Criterion (BIC), whereby the overall performance considers both prediction accuracy and model complexity. (3) Based on the historical prediction error and the overall performance, the weighting coefficients are calculated using an exponential decay function. The formula for calculating the weighting coefficients is as follows: ； in For the first i The weight coefficients of each model, and To adjust the parameters, For the first i The root mean square error of each model For the first i Bayesian information criterion values ​​for each model; (4) Recalculate the weight coefficients based on the latest prediction performance to achieve dynamic updating of the weight coefficients.

8. The method for predicting and early warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 1, characterized in that: The deformation index rolling prediction in step S5 adopts a sliding time window mechanism with a sliding time window length of 72 hours. Based on the dual-core prediction model, the deformation index is predicted in the rolling prediction of the next 12 hours, 24 hours, and 48 hours. The three-level early warning management includes blue warning, yellow warning, and red warning. The blue warning corresponds to the normal state and maintains the regular monitoring frequency. The yellow warning is triggered when the predicted value is close to the limit and the monitoring frequency is increased to once every 2 minutes. The red warning is triggered when the predicted value exceeds the limit and the audible and visual alarm device is activated.

9. The method for predicting and warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 8, characterized in that: The warning thresholds of the three-level early warning management adopt a dynamic update mechanism. The dynamic update mechanism is based on the Bayesian update algorithm and updates the warning thresholds every 7 days by combining historical engineering data and real-time monitoring information. Under soft soil geological conditions or when the surrounding load increases, the warning thresholds are reduced.

10. The method for predicting and early warning the deformation of foundation pit support structures based on multi-source data fusion according to claim 1, characterized in that: The online performance monitoring method in step S6 includes: periodically comparing the prediction results of the dual-core prediction model with the actual monitoring data to calculate the prediction error; when the prediction error is greater than the initial error for several consecutive prediction periods, it is determined that the prediction error is greater than a set threshold, and incremental learning is performed.

Images