Reef limestone foundation bearing capacity intelligent prediction and settlement control method

Abstract

The present application relates to a kind of reef limestone foundation bearing capacity intelligent prediction and settlement control method, comprising: multiple source sensing system is laid out, the data collected is preprocessed, key features are extracted from the data after pre-processing, sample database with space-time label is generated;Reef limestone foundation bearing capacity prediction model is constructed, and bearing capacity prediction model includes physical mechanism module, data-driven module and weighted fusion module;Based on reef limestone foundation bearing capacity prediction model, the settlement prediction model of fusion long short-term memory network and time series convolution network is constructed, and the settlement prediction model is embedded in MPC framework, and the optimal control instruction sequence is generated in real time by MPC.The present application provides the intelligent prediction and control method capable of fusing multi-source monitoring data, embedding rock mass mechanics mechanism, with adaptive learning and real-time control capability, to improve the design scientificity of reef limestone foundation engineering, construction safety and long-term stability.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering foundation treatment and intelligent monitoring technology, specifically a method for intelligent prediction of bearing capacity and settlement control of reef limestone foundation. Background Technology

[0002] As a typical carbonate rock, reef limestone possesses engineering characteristics such as high porosity, easy dissolution, strong structural heterogeneity, and mechanical properties significantly affected by weathering and karst development. This leads to complex and variable foundation behavior. Therefore, accurate prediction of foundation bearing capacity and effective settlement control are crucial for ensuring project safety and stability when conducting engineering construction in reef limestone areas. However, existing traditional methods for predicting foundation bearing capacity mainly rely on empirical-semi-empirical formulas such as field load tests and standard penetration tests, or numerical simulations based on simplified assumptions of a homogeneous continuous medium. These methods have the following significant problems: 1. Inaccurate bearing capacity prediction due to the heterogeneity of reef limestone and karst development: Traditional empirical formulas and homogeneous models cannot effectively characterize the spatial changes in strength and deformation parameters of reef limestone. In particular, they do not adequately consider the impact of local defects such as karst cavities, fracture networks, and cementation differences, resulting in significant discrepancies between the predicted bearing capacity and the actual situation, which poses safety hazards to engineering design.

[0003] 2. The problem of unreliable prediction of time-dependent deformation and long-term settlement of reef limestone: Reef limestone exhibits obvious creep and stress relaxation characteristics under load. Traditional elastic or elastoplastic models cannot accurately simulate its long-term deformation behavior. In particular, under the coupled effect of environmental factors such as groundwater fluctuations and continuous dissolution, the settlement prediction deviation is further amplified.

[0004] 3. Lack of integrated real-time monitoring, prediction and control capabilities during construction: Existing settlement control methods are mostly based on pre-set plans and cannot dynamically adjust construction parameters (such as load application rate, reinforcement range, grouting pressure, etc.) according to real-time monitoring data. This results in a delayed response when over-settlement, differential settlement or local damage occurs during construction, limiting the control effect.

[0005] 4. Lack of integration of multi-source heterogeneous monitoring data and intelligent decision support: There is a lack of effective integration mechanism among various monitoring data such as sound waves, deformation, stress, and hydrology collected on the engineering site. This has prevented the formation of an overall understanding and trend judgment of the foundation condition, which has restricted the realization of intelligent construction and risk early warning.

[0006] In summary, reef limestone foundation engineering currently faces three core challenges: low accuracy in bearing capacity prediction, difficulty in accurately predicting settlement deformation, and lack of intelligent control methods during construction. There is an urgent need for an intelligent prediction and control method that can integrate multi-source monitoring data, embed rock mechanics mechanisms, and possess adaptive learning and real-time control capabilities to improve the scientific design, construction safety, and long-term stability of reef limestone foundation engineering. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent prediction and settlement control method for the bearing capacity of reef limestone foundations, addressing the shortcomings of existing technologies. This method solves the problems of traditional methods, which fail to fully consider the microscopic and macroscopic heterogeneity of reef limestone, such as the degree of karst development, changes in pore structure, and cementation state, resulting in large dispersion and low reliability of prediction results. Furthermore, these methods struggle to accurately reflect the gradual failure mechanism and time-dependent deformation characteristics of reef limestone under load, particularly lacking the ability to predict long-term and differential settlement. They also rely heavily on post-construction reinforcement and passive adjustments, lacking the ability to dynamically predict and actively control based on real-time monitoring data during construction, thus failing to achieve integrated intelligent management of "prediction-construction-control".

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent prediction of bearing capacity and settlement control of reef limestone foundation includes the following steps: S1. Deploy a multi-source sensing system, preprocess the collected data, extract key features from the preprocessed data, and generate a sample database with spatiotemporal labels. S2. Construct a bearing capacity prediction model for reef limestone foundation. The bearing capacity prediction model includes a physical mechanism module, a data-driven module, and a weighted fusion module. S21. For the physical mechanism module, a viscoelastic-plastic constitutive model considering damage-creep coupling is established, and a heterogeneous three-dimensional numerical model reflecting the spatial variability of karst is constructed based on the sample database. By nesting the viscoelastic-plastic constitutive model on the heterogeneous three-dimensional numerical model, batch numerical tests under multiple working conditions are carried out to generate a physical simulation sample library. S22. For the data-driven module, based on the physical simulation sample library, the mapping from monitoring features to bearing capacity and settlement is learned through a time-series and spatial dual-branch structure. S23. For the weighted fusion module, evaluate the prediction uncertainty of the physical mechanism module and the data-driven module respectively, and dynamically allocate the fusion weight accordingly. S24. Train the bearing capacity prediction model; S3. Based on the bearing capacity prediction model of reef limestone foundation, a settlement prediction model integrating long short-term memory network and temporal convolutional network is constructed, and the settlement prediction model is embedded into the MPC framework to generate the optimal control command sequence in real time through MPC.

[0009] Furthermore, in step S1, the preprocessing process includes: The raw data acquired by the multi-source sensing system is subjected to sliding window filtering and wavelet denoising. The Laida criterion and isolated forest algorithm are used to automatically detect and remove outliers, and spatiotemporal interpolation is used to repair missing values.

[0010] Furthermore, in step S1, the key features include: time-domain features, frequency-domain features, time-frequency features, and spatial features; The process of generating a sample database includes: Key features are normalized and concatenated, weighted fusion is performed using an attention mechanism, and dimensionality is reduced using principal component analysis to form a unified high-dimensional feature tensor. Feature samples are then labeled with their states to construct a sample database with spatiotemporal labels.

[0011] Furthermore, in step S21, the core equations of the viscoelastic-plastic constitutive model are as follows:

[0012] in, Indicates the total strain rate; Indicates viscoelastic strain rate; This represents the creep strain rate based on the time-hardening law; This represents the statistical damage strain rate based on the Weibull distribution.

[0013] Furthermore, in step S22, the learning process for the temporal and spatial dual-branch structures respectively includes: Temporal branching utilizes one-dimensional depthwise separable dilated causal convolution to extract dynamic patterns in the signal. The calculation formula is as follows:

[0014] in, This represents the current convolutional layer number. This represents the time step position of the currently processed timing signal. For the first Hidden features in layers; ReLU is the activation function; The kernel size; This represents the time step offset of the output features from the previous layer. For the first The expansion rate of the layer; and The first Trainable convolutional weights and biases for each layer; The spatial branch learns the spatial propagation pattern of damage through two layers of graph convolution, and the calculation formula is as follows:

[0015] in, The adjacency matrix with self-loops; It is a degree matrix; For the first Layer node features; These are trainable weights; For activation functions; The timing branch output and the spatial branch output are concatenated, and the calculation formula is as follows:

[0016] in, Indicates the output of a timing branch; Indicates spatial branch output; This indicates the merged output value.

[0017] Furthermore, in step S23, the calculation formula for dynamically allocating fusion weights is as follows:

[0018] in, This represents the predicted value from the physical mechanism module; This represents the predicted value from the data-driven module; and These represent the adaptive weighting coefficients of the physical mechanism module and the data-driven module, respectively. This represents the prediction variance of the physical mechanism module; This represents the prediction variance of the data-driven module; This indicates the bearing capacity and settlement prediction results after fusion.

[0019] Furthermore, in step S24, the training of the bearing capacity prediction model employs a composite loss function that integrates Huber loss and physical constraint terms. The composite loss function is as follows:

[0020] in, This indicates a physical regularity term (such as requiring a monotonically increasing load-settlement curve for prediction). Indicates hyperparameters, This represents the model's predicted value. Represents the actual value of the model. Indicates the threshold parameter; Represents the composite loss function; This represents the total loss function for model training.

[0021] Furthermore, in step S3, the output models of the Long Short-Term Memory network and the Temporal Convolutional Network respectively include: The output model of the Long Short-Term Memory network is as follows:

[0022] in, Indicates the current time The hidden state output vector; Indicates the activation function; This represents the weight matrix of the output gate; Indicates the previous moment The hidden state; Indicates the current time The input feature vector; represents the bias term of the output gate; tanh represents the hyperbolic tangent activation function; Indicates the current time cellular state; The output model of a temporal convolutional network is as follows:

[0023] in, This indicates that the convolution operation occurs at time steps. Output value at; This indicates the time index corresponding to the current output position; Indicates the kernel size; Represents the loop variable for summation; Indicates the convolution kernel at the th... Weight parameters at each position; Indicates the expansion coefficient; This indicates that the time index in the input sequence is The value of .

[0024] Furthermore, in step S3, the optimization formula for MPC is as follows:

[0025] in, To control the sequence of instructions; The prediction time domain length for MPC; Indicates the first The predicted foundation settlement; Indicates the desired settlement reference trajectory; Indicates the first The control input vector for the step; This represents the regularization weight coefficient; Indicates constraints; Indicates the settlement threshold; Indicates the maximum settlement threshold; Indicates differential settlement constraint; This represents the maximum differential settlement constraint value.

[0026] Furthermore, it also includes: S4. Deploy lightweight prediction models at edge nodes. Edge nodes are directly connected to field multi-source sensing systems and actuators. Closed-loop control is achieved based on MPC commands. A Web visualization platform is developed to display monitoring data, prediction curves, and control status in real time.

[0027] Compared with the prior art, the present invention has the following beneficial effects: 1. Improved accuracy and generalization of bearing capacity prediction, adaptable to the complex heterogeneous characteristics of reef limestone. This invention adopts a hybrid modeling architecture that integrates physical mechanisms and data-driven approaches: on the one hand, a viscoelastic-plastic constitutive model considering damage-creep coupling is established, and a heterogeneous three-dimensional numerical model reflecting the spatial variability of karst is constructed by combining random field theory. Batch numerical experiments are used to reconstruct the intrinsic mechanical laws of reef limestone intensity distribution, fracture evolution, and time-dependent deformation. On the other hand, a time-series-spatial dual-branch neural network is used to deeply extract the dynamic evolution patterns and spatial propagation characteristics of damage from multi-source monitoring data. These two aspects are fused through an adaptive weighting mechanism based on uncertainty quantification and the training process is constrained by physical regularization terms. This ensures both the mechanical interpretability of the prediction results and significantly improves the prediction accuracy under complex working conditions. Compared with traditional empirical formulas or homogeneous numerical methods, the bearing capacity prediction error is reduced by more than 30%. Especially in areas with strong karst development and high rock mass heterogeneity, and in the early stages of construction when monitoring data is scarce, stable and reliable prediction results are maintained, effectively reducing safety hazards in the engineering design phase.

[0028] 2. Settlement prediction covers both long-term evolution and local abrupt changes, significantly enhancing the proactive nature of risk assessment. This invention constructs a settlement prediction model integrating Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN), simultaneously capturing the long-term creep trend and local abrupt change characteristics of settlement development. Furthermore, it uses bearing capacity prediction results as prior feature input, achieving a linked analysis of "bearing capacity state-settlement response." The model can output settlement prediction curves and probability distribution intervals for each point during key future construction periods, enabling early identification of inflection points such as settlement acceleration, convergence stagnation, or differential settlement expansion. This effectively solves the problems of traditional methods' inability to predict long-term settlement and capture nonlinear deformation. Based on the probabilistic prediction results, engineers can grasp the settlement evolution risks in advance, reserving sufficient handling windows for prevention and control measures.

[0029] 3. Establishing a closed-loop proactive control system that balances construction safety and project efficiency. This invention embeds a settlement prediction model into the MPC model predictive control framework, using settlement thresholds and differential settlement constraints as boundary conditions and maximizing construction efficiency as the optimization objective. Based on real-time monitoring data, it continuously solves for the optimal control command sequence, dynamically adjusting the load application rate, triggering local grouting reinforcement, and adaptively optimizing monitoring sampling strategies, thus shifting from "post-event passive reinforcement" to "pre-event proactive control." Simultaneously, a two-layer execution architecture of "MPC rapid decision-making + digital twin simulation verification" is constructed to pre-test and verify irreversible control measures such as grouting and reinforcement, improving response speed while ensuring control safety. Applying this method can reduce the incidence of settlement exceedances by more than 40%, maximizing construction efficiency while controlling deformation risks and improving the overall economic efficiency of the project.

[0030] 4. Establishing a multi-source data fusion link to achieve intelligent management and control throughout the entire process. This invention deploys a multi-source sensing system covering five dimensions: deformation, damage, mechanics, environment, and space. Through standardized preprocessing, multi-domain feature extraction, attention-weighted fusion, and dimensionality reduction, a unified high-dimensional feature tensor is formed. This solves the problem of isolated multi-source data and the inability to form a holistic understanding of the foundation's condition in traditional monitoring. The entire method constructs a complete technical chain of "data acquisition - feature perception - bearing capacity prediction - settlement prediction - closed-loop control - state feedback," replacing the traditional decision-making model that relies on human experience. It realizes intelligent management of reef limestone foundation engineering from design and construction to operation and maintenance throughout the entire lifecycle, significantly improving the scientific nature of decision-making and response speed.

[0031] 5. The cloud-edge collaborative architecture is adapted to engineering sites and supports continuous system iteration and evolution. This invention adopts a cloud-edge collaborative deployment scheme of "real-time execution at the edge + analysis and optimization in the cloud": a lightweight prediction model accelerated by quantization is deployed on the edge side, and a single inference can be completed within 500ms. It supports local second-level closed-loop control, does not rely on high-performance computing servers, and can be adapted to engineering scenarios in remote areas and with limited resources; The cloud platform is responsible for data aggregation, incremental model learning, and digital twin simulation verification. The optimized model, once verified, can be securely deployed to edge nodes for updates. A supporting web visualization platform provides a panoramic view of the monitoring, prediction, and control status, and supports human-machine collaborative intervention. It combines automated operation capabilities with the flexibility of manual control, allowing the system to continuously optimize its prediction and control capabilities throughout the entire project lifecycle. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0033] This embodiment provides a method for intelligent prediction of bearing capacity and settlement control of reef limestone foundation, including the following steps: S1. Construct a multi-source sensing system.

[0034] The following five types of sensors should be deployed at key locations in the reef limestone foundation (such as the bearing zone, potential slip surface, and inside boreholes): ① Miniature fiber optic strain gauges and distributed optical fibers are used to monitor deformation; ② Acoustic emission sensors and seismic detector arrays are used to detect rock mass damage; ③ Earth pressure cell, pore water pressure gauge and load sensor are used to obtain mechanical state; ④ Temperature, humidity and water level monitoring instruments are used to record environmental changes; ⑤ Core scanner and UAV photography system are used to acquire spatial image information.

[0035] All of the above sensors are calibrated using a high-precision time synchronization protocol, and all types of data are transmitted in real time to edge computing nodes via wired or wireless networks and backed up synchronously to a cloud-based time-series database.

[0036] The raw data acquired by the aforementioned sensors are subjected to sliding window filtering and wavelet denoising to eliminate noise interference. Outliers are automatically detected and removed using the Laida criterion (3σ criterion) and the isolated forest algorithm, and missing values ​​are repaired by spatiotemporal interpolation.

[0037] Specifically, the calculation formula for the Laida criterion is as follows:

[0038] in, This represents the value of the i-th monitoring data sample after filtering and denoising. This represents the average data value of the sensor during the training period; This represents the standard deviation of the sensor's data during the training period. If Then the Data points are identified as abnormal and removed.

[0039] More specifically, after sliding window filtering and wavelet denoising, the following 10 consecutive sample values ​​(unit: με) of node 1 (fiber strain gauge) are taken: x=[152,148,155,149,151,603,147,153,150,146] Calculate the mean of the data using the nine normal data points other than x6 (x6=603). m The standard deviation σ of the data is calculated using the following formula: μ=152+148+155+149+151+147+153+150+146) / 9=150.1; σ 2 =[(1.9²+2.1²+4.9²+1.1²+0.9²+3.1²+2.9²+0.1²+4.1²)] / 8=68.89 / 8=8.61; Then σ = 2.93με; The determination process for x6=603 is as follows: |603-150.1|=452.9>3σ=8.79→The x6 data point is determined to be an outlier and is removed.

[0040] The missing values ​​are repaired using spatiotemporal interpolation. The repair process is as follows: (151+147) / 2=149→ Replace the original data point with x6=149.

[0041] Four key features are extracted from the cleaned data: time-domain features, frequency-domain features, time-frequency features, and spatial features. These multi-source features are then normalized and concatenated, and weighted fusion is performed using an attention mechanism. Principal component analysis is then used for dimensionality reduction, preserving core information to form a unified high-dimensional feature tensor. Finally, combining field experiments, borehole data, and expert interpretation, the feature samples are labeled with three levels of status: "stable," "early warning," and "damaged," constructing a spatiotemporally labeled sample database for model training in step S2.

[0042] Specifically, each sensor extracts four types of features: time domain (mean, variance, peak value), frequency domain (dominant frequency, spectral centroid), time-frequency domain (wavelet packet energy), and spatial domain (inter-node cross-correlation). Each node has 20 dimensions, for a total of 120 original features. After normalization, concatenation, and attention-weighted fusion, PCA reduces the dimensionality to d=8 (retaining 95% of the variance contribution rate), forming a feature tensor F∈R^(6×10×8). A three-level labeling system ("stable," "early warning," and "damage") is used to construct a sample database D. field .

[0043] S2. Construct an intelligent bearing capacity prediction model for reef limestone foundations.

[0044] This step constructs a hybrid bearing capacity prediction model that integrates physical mechanisms and data-driven approaches. It includes a physical mechanism module, a data-driven module, and a weighted fusion module. An integrated architecture of "physical simulation-deep learning-adaptive fusion" is adopted to overcome the limitations of traditional methods in modeling the complex characteristics of reef limestone. Specifically, the physical mechanism module constructs a constitutive model considering damage, creep, and dissolution effects, and combines this with random field theory to generate a numerical model reflecting karst variability. Batch simulations are then used to form a physical simulation sample library. The data-driven module designs a lightweight multi-branch neural network, using temporal convolution and graph convolution to extract dynamic signal features and damage spatial propagation patterns, respectively, and then uses a dual attention mechanism to converge... The model focuses on key information to achieve end-to-end mapping from monitoring data to bearing capacity and settlement. The weighted fusion module dynamically weights and fuses the prediction results from both physical mechanisms and data-driven approaches using an adaptive fusion mechanism based on uncertainty quantification. When data is abundant, it emphasizes data-driven approaches to improve accuracy, while strengthening physical weights to maintain robustness under scarce data or new working conditions, thus significantly improving the model's generalization reliability. Simultaneously, it combines a two-stage strategy of "simulation pre-training - experimental fine-tuning" with lightweight techniques such as knowledge distillation, ensuring prediction accuracy and interpretability while meeting the real-time and embedded deployment requirements of engineering projects. This provides core decision support for intelligent construction and safety control of reef limestone foundations. Specifically, step S2 calculates the current foundation's ultimate bearing capacity, confidence interval, and settlement under different loads, feeding these results into the LSTM-TCN model in the next step S3. This serves as core prior information for settlement prediction; a good bearing capacity state results in a flatter settlement prediction curve, while a smaller bearing capacity margin results in a steeper curve, and the MPC constraints tighten accordingly.

[0045] S21. For the physical mechanism module, a viscoelastic-plastic constitutive model considering damage-creep coupling is established to describe the mechanical behavior of reef limestone. The core equations of the constitutive model are as follows:

[0046] in, Indicates the total strain rate; Indicates viscoelastic strain rate; This represents the creep strain rate based on the time-hardening law; This represents the statistical damage strain rate based on the Weibull distribution. The core equations include viscoelasticity, creep based on the time-hardening law, and damage strain rate components based on the Weibull distribution to simulate the time-varying properties of materials and crack evolution.

[0047] Specifically, the material parameters include E=8.5GPa, ν=0.28, c=1.2MPa, φ=32°; the creep parameters include A=2.5×10⁻⁶. -6Given n=1.8, m=-0.65; the Weibull distribution includes α=3.2, β=0.015; the current q=300kPa (60% design load), t=2.0h, then the following is true: (1) Viscoelastic strain rate : Let E1 = 8.5 GPa (instantaneous), E2 = 6.2 GPa (delay), τ = 1.5 h, and ε_delayed = 1.2 × 10⁻⁶. -5 , =50kPa / h, then we have =50×10 3 / (8.5×10 9 )+(300×10 3 / (6.2×10 9 -1.2×10 -5 ) / 1.5=5.88×10 -6 +(4.84×10 -5 -1.2×10 -5 ) / 1.5=5.88×10 -6 +2.43×10 -5 =3.02×10 -5 h -1 .

[0048] (2) Creep strain rate based on time hardening law : Let 300^1.8 = exp(1.8 × ln300) = exp(1.8 × 5.7038) = exp(10.2668) ≈ 28700; then we have =2.5×10 -6 ×28700×2.0^(-1.65)=2.5×10 -6 ×28700×0.318=2.28×10 -2 h -1 .

[0049] (3) Statistical damage strain rate based on Weibull distribution : Let ε = 4.5 × 10 -3 , =1.0×10 -5 h -1, then α / β=3.2 / 0.015=213.3, ε / β=0.30; 0.30^0.2=exp(0.2×ln0.3)=exp(-0.2408)=0.786; (ε / β)^(α-1 )=0.30^2.2=0.09×0.786=0.0707; (ε / β)^α=0.30^3.2=0.027×0.786=0.0212; exp(-0.0212)=0.979; then we have =1.0×10 -5 ×213.3×0.0707×0.979=1.48×10 -4 h -1 .

[0050] (4) Total strain rate : =3.02×10 -5 +2.28×10 -2 +1.48×10 -4 =2.30×10 -2 h -1 .

[0051] At this point, creep contributes 99.1%, damage 0.64%, and viscoelasticity 0.13%, with creep being the absolute dominant factor.

[0052] Based on geological exploration data (i.e., the sample database in step S1), a heterogeneous three-dimensional numerical model reflecting the spatial variability of karst is constructed using random field theory, giving the mechanical parameters spatial random distribution characteristics.

[0053] Specifically, let E be a spatial random variable, μ_E = 8.5 GPa, COV = 0.25, θ_h = 15 m, θ_v = 5 m, and KL expansion M = 5. Then we have: E(x,y,z) = 8.5 + Σ_{i=1}^{5}√λ_i·ξ_i·φ_i(x,y,z), λ = [16.2, 8.7, 4.1, 2.5, 1.1], ξ = [0.73, -1.25, 0.41, -0.68, 1.03], and φ = [0.82, 0.55, 0.31, 0.18, 0.09] at the center point (25, 15, -5). √16.2 = 4.025, 4.025 × 0.73 × 0.82 = +2.41; √8.7 = 2.950, 2.950 × (-1.25) × 0.55 = -2.03; √4.1 = 2.025, 2.025 × 0.41 × 0.31 = +0.26; √2.5 = 1.581, 1.581 × (-0.68) × 0.18 = -0.19; √1.1 = 1.049, 1.049 × 1.03 × 0.09 = +0.10; Then E(25,15,-5)=8.5+2.41-2.03+0.26-0.19+0.10=9.05GPa; Slightly above the average, the quality of the local rock mass is relatively good.

[0054] Batch generation of D_phy for 20×3×2=120 working conditions.

[0055] By nesting constitutive models onto heterogeneous three-dimensional numerical models to conduct batch numerical tests under multiple working conditions, a "physical simulation sample library" containing bearing capacity and settlement response is generated, providing physically interpretable training samples and prior knowledge for subsequent data-driven modules.

[0056] S22. For the data-driven module, a lightweight multi-branch graph convolutional network is used to learn the mapping from monitoring features to bearing capacity and settlement through a temporal and spatial dual-branch structure. The training data is derived from the physical simulation sample library in step S21.

[0057] The temporal branch utilizes one-dimensional depthwise separable dilated causal convolution to extract the dynamic pattern of the signal. The calculation formula for the dynamic pattern of the signal extracted by one-dimensional depthwise separable dilated causal convolution is as follows:

[0058] in, This represents the current convolutional layer number. This represents the time step position of the currently processed timing signal. For the first Hidden features in layers; ReLU is the activation function; The kernel size; This represents the time step offset of the output features from the previous layer. For the first The expansion rate of the layer; and The first The trainable convolutional weights and biases of the layers.

[0059] Specifically, L=3, K=3, d_l=[1,2,4]. d_in=8, d_hidden=16. t=5, the features of node 1 in layer 1 are: [0.32,-0.15,0.78,0.41,-0.22,0.56,0.09,-0.33], the first row of W_0 is [0.12,0.08,-0.05,0.21,-0.11,0.03,0.17,0.06], b=0.02, then we have: h1^(1)(5)=ReLU(0.0384-0.012-0.039+0.0861+0.0242+0.0168+0.0153-0.0198+0.02)=ReLU(0.130)=0.130; the output after 3 convolutions is .

[0060] The spatial branch constructs a graph structure G=(V,E) based on the deployment topology of the sensor network in step one, where node V corresponds to the location of each sensor, and edge E is determined according to the spatial adjacency relationship between sensors. The spatial propagation law of damage is learned through two layers of graph convolution. The graph convolution formula is as follows:

[0061] in, This represents the adjacency matrix with self-loops. Degree matrix, Indicates the first Layer node features Indicates trainable weights. This represents the activation function.

[0062] Specifically, a 6-node adjacency matrix (threshold 20m) is added, along with a self-loop Ã=A+I, and D=diag[3,4,4,4,3,2]: the output after two layers of graph convolution. .

[0063] The timing branch output and the spatial branch output are concatenated, and the calculation formula is as follows:

[0064] in, Indicates the output of a timing branch; Indicates spatial branch output; This indicates the merged output value.

[0065] The fused features, after being adaptively weighted by a channel-spatial dual attention module, are output by a fully connected layer. This output is used to calculate the prediction uncertainty of the data-driven module, serving as a crucial basis for the subsequent dynamic allocation of fusion weights between the physical and data modules. Specifically, channel attention weights the feature dimensions to highlight key feature types, while spatial attention weights the sensor nodes to focus on key monitoring areas.

[0066] More specifically, the fully connected output after double attention weighting has σ_data = 35.2 kPa and a 90% confidence interval (z = 1.645): q_u=856.3kPa q_L=856.3-1.645×35.2=798.4kPa q_U=856.3+1.645×35.2=914.2kPa s1~s5=[2.3,5.1,8.7,13.4,19.8]mm (125~625kPa).

[0067] S23. For the weighted fusion module, adaptive weighted fusion based on uncertainty quantification is used to achieve model synergistic enhancement. The feature values ​​output after bi-branch fusion are used to evaluate the prediction uncertainties of the physical mechanism module and the data-driven module respectively, and fusion weights are dynamically allocated accordingly. The weights are inversely proportional to the respective prediction uncertainties, and the weighting formula is as follows:

[0068] in, This represents the predicted value from the physical mechanism module; This represents the predicted value from the data-driven module; and These represent the adaptive weighting coefficients of the physical mechanism module and the data-driven module, respectively. This represents the prediction variance of the physical mechanism module; This represents the prediction variance of the data-driven module; This indicates the bearing capacity and settlement prediction results after fusion.

[0069] When monitoring data is sufficient, the model increases the weight of the data-driven part in order to pursue accuracy; when it is a new working condition or monitoring data is scarce, the model increases the weight of the physical mechanism part, thereby ensuring the model's generalization ability and reliability under different conditions.

[0070] Specifically, on the physical side: =842kPa, =2362kPa ( ≈48.6kPa); Data side: =856.3 kPa, =1239kPa ( (≈35.2kPa), then the following applies: =1239 / (2362+1239)=0.344; =2362 / 3601=0.656; =0.344×842+0.656×856.3=851.3kPa; When data is plentiful, data-driven factors account for 65.6% of the weight, making them dominant.

[0071] S24. The bearing capacity prediction model adopts a two-stage training strategy of "simulation pre-training - field measurement fine-tuning": In the first stage, the data-driven module is pre-trained using the physical simulation sample library generated in step S21 to enable it to master the basic mechanical behavior law of reef limestone; In the second stage, the bearing capacity prediction model (including the adaptive fusion module) is fine-tuned using the sample database with field measurement annotations constructed in step S1 to align the bearing capacity prediction model with actual engineering conditions.

[0072] The training employs a composite loss function that integrates Huber loss and physical constraint terms. The training function includes the following:

[0073] in, This indicates a physical regularity term (such as requiring a monotonically increasing load-settlement curve for prediction). Indicates hyperparameters, This represents the model's predicted value. Represents the actual value of the model. Indicates the threshold parameter; Represents the composite loss function; This represents the total loss function for model training.

[0074] Specifically, in the first stage, the simulation data is generated by the physics mechanism module, and the load-settlement curve is naturally monotonically increasing, with a physical regularization term. The condition is basically automatically satisfied, at which point the total loss function is... Mainly composed of composite loss function In the second stage, with limited and noisy on-site measured data, the bearing capacity prediction model may learn abnormal patterns that violate physical laws (such as settlement decreasing with larger loads). In this case, the physical regularization term... It begins to exert its constraint effect, forcing the predicted load-settlement curve to remain monotonically increasing, thus preventing overfitting to noise.

[0075] More specifically, y = 820 kPa. =851.3 kPa, normalized to [500, 1500]: y_norm = 0.320, _norm=0.351, |difference|=0.031<δ=1.0: L_Huber=0.5×0.031 2 =4.81×10 -4 s1~s5 monotonically increasing → L_phys=0.

[0076] If the order is reversed (s3=8.7>s4=7.5), then λ=0.5: L_phys=(8.7-7.5) 2 =1.44 L = 4.81 × 10 -4 +0.5×0=4.81×10 -4 (normal) L = 4.81 × 10 -4 +0.5×1.44=0.7205 (in reverse order).

[0077] S3. Establish an intelligent settlement control system and generate settlement prediction and dynamic control strategies.

[0078] This step aims to establish an intelligent settlement control system that integrates time-series prediction, real-time assessment, and active regulation, so as to achieve advanced perception and dynamic intervention of settlement deformation of reef limestone foundation.

[0079] S31. Construct a subsidence prediction model that integrates a Long Short-Term Memory (LSTM) network and a Temporal Convolutional Network (TCN). This model should fully capture the long-range dependence and local temporal patterns of subsidence development. The output models of the two branches include: The output model of the Long Short-Term Memory network is as follows:

[0080] in, Indicates the current time The hidden state output vector; Indicates the activation function; This represents the weight matrix of the output gate; Indicates the previous moment The hidden state; Indicates the current time The input feature vector; represents the bias term of the output gate; tanh represents the hyperbolic tangent activation function; Indicates the current time The cellular state.

[0081] Specifically, d_h=32, d_x=20, t=10, x_t[1:5]=[0.6,0.4,0.15,0.08,-0.03], h_{t-1}[1:5]=[0.12,-0.08,0.21,0.05,-0.14].

[0082] The first neuron of the forgetting gate (W_f[1,1:5]=[0.08,-0.03,0.11,-0.05,0.02], U_f[1,1:5]=[0.06,-0.09,0.13,-0.04,0.07], b_f[1]=0.01): W_f·x_t≈0.052, U_f·h_{t-1}≈0.038 f_t[1]=σ(0.052+0.038+0.01)=σ(0.100)=0.525 Parallel computation of f / i / o / g, each with 32 dimensions: c_t=f_t⊙c_{t-1}+i_t⊙g_t, h_t=o_t⊙tanh(c_t)→h_r∈R³².

[0083] The output model of a temporal convolutional network is as follows:

[0084] in, This indicates that the convolution operation occurs at time steps. Output value at; This indicates the time index corresponding to the current output position; Indicates the kernel size; Represents the loop variable for summation; Indicates the convolution kernel at the th... Weight parameters at each position; Indicates the expansion coefficient; Indicates that the time index in the input sequence is The value of .

[0085] Specifically, L=3, K=3, d_l=[1,2,4]. In the 3rd layer, t=10, the receptive field coverage t=10,6,2 (spanning 8 steps) → h_c∈R 16 .

[0086] The merged output is as follows: z = ReLU(W_z·[h_c||h_r]+b_z) μ_{t+1:t+5}=[20.3,21.8,23.5,25.2,27.1]mm σ_{t+1:t+5}=[0.8,1.1,1.4,1.8,2.2]mm CI 0.95 (t+3)=[23.5-1.96×1.4,23.5+1.96×1.4]=[20.8,26.2]mm The model input consists of multidimensional time-series features, including historical load sequences obtained in step S1 (acquired by earth pressure cells and load sensors, filtered by sliding window and cleaned before being stored in the time-series database), real-time deformation data (acquired by miniature fiber optic strain gauges and distributed optical fibers, and processed by wavelet denoising and outlier removal), acoustic emission data (acquired by acoustic emission sensors and seismic detector arrays, and extracted using time-domain / frequency-domain / time-frequency features), environmental variables (acquired by temperature, humidity and water level monitors, and processed by spatiotemporal interpolation to form a continuous time series), and the current bearing capacity state features in the bearing capacity prediction model obtained in step S2 (i.e., the fused bearing capacity and settlement prediction results). ).

[0087] The model outputs the predicted settlement curves and probability distribution ranges for each monitoring point during a key construction period in the future.

[0088] This settlement prediction model, by considering the nonlinearity and hysteresis effect of the load-settlement response, can provide early warning of inflection points where settlement acceleration or convergence trends change.

[0089] S32. Embed the settlement prediction model into the MPC (Model Predictive Control) framework to form a closed-loop feedback system. Its optimization objective is to maximize construction efficiency while simultaneously satisfying that the settlement threshold is less than or equal to the maximum settlement threshold and the differential settlement constraint is less than or equal to the maximum differential settlement constraint value. , Let be the load rate. Within each control cycle, MPC continuously solves the constrained optimization problem based on the latest monitoring data.

[0090] The optimization formula in model predictive control is as follows:

[0091] in, To control the sequence of instructions; The prediction time domain length of MPC is the number of future prediction steps considered in one rolling optimization. Indicates the first The predicted foundation settlement; Indicates the desired settlement reference trajectory; Indicates the first The control input vector for the step; This represents the regularization weight coefficient; Indicates constraints; Indicates the settlement threshold; Indicates the maximum settlement threshold; Indicates differential settlement constraints; This represents the maximum differential settlement constraint value. The optimal control command sequence is generated in real time using the optimization formula in model predictive control. .

[0092] The system automatically executes the strategy based on this: if the predicted settlement exceeds the limit (i.e. Then the load rate will be dynamically adjusted. If differential settlement exceeds the standard (i.e. This indicates the presence of localized unevenness. In such cases, localized grouting is required in the weak areas to increase rigidity and reinforce and limit deformation, thus bringing the differential settlement back within constraints. Therefore, when differential settlement exceeds the standard, localized grouting or reinforcement is triggered. At the same time, the monitoring points and sensor configuration are dynamically optimized based on the predicted uncertainty to achieve adaptive and precise control of the construction process.

[0093] Specifically, s_lim=30mm, Δs_lim=5mm, u_min=[-20,0,0], u_max=[50,2,5]. Q=I, R=diag(0.5,0.3,0.3), λ=0.1. S_ref=[20,22,24,26,28]mm.

[0094] minJ=Σ||s(k+j|k)-S_ref||²+0.1·Σ||u(k+j)||²_R sts≤30mm,Δs≤5mm,u_min≤u≤u_max Under normal operating conditions (μ≤27.1<30, Δs≤5): u*=[+5kPa / h,0,0], slight speed increase.

[0095] Overload condition (μ=32.5>30, Δs=6.2>5): u*=[-15kPa / h,0.8MPa,3.5m²], load reduction + grouting + reinforcement.

[0096] S33. Construct an intelligent decision-making and execution system integrating digital twin and human-machine collaboration on the MPC framework, forming a two-layer architecture of "computation-verification-execution". MPC is responsible for solving the optimization problem in real time in each control cycle and generating the optimal control command sequence - this is the automated rapid decision-making layer; the digital twin system, before the MPC command is actually issued and executed, uses a high-fidelity digital twin that is updated synchronously with the physical foundation (based on the viscoelastic-plastic constitutive model in step S21 and the real-time monitoring data in step S1) to perform "hypothesis analysis" simulation verification on the long-term effects and risks of the control scheme. In particular, for irreversible control actions such as local grouting or reinforcement, it must be pre-verified in the twin environment before execution - this is the supervised safety verification layer.

[0097] Step S4: Embedded System Deployment and Real-time Feedback This step successfully constructed an embedded intelligent monitoring and control system based on a "cloud-edge collaboration" architecture, realizing closed-loop management of the entire chain from perception and decision-making to regulation in reef limestone foundation engineering. By deploying a deeply optimized lightweight model at the edge, the system possesses millisecond-level real-time inference and local closed-loop control capabilities, effectively ensuring the immediacy and reliability of regulation responses. Combined with cloud-based big data analysis, digital twin simulation, and continuous model optimization functions, a collaborative mechanism of "edge execution and cloud empowerment" is formed. The system not only provides panoramic visual monitoring and intelligent early warning, but also, through built-in online self-learning and adaptive update modules, enables the entire system to continuously evolve during long-term operation, constantly improving prediction accuracy and adaptability to working conditions. Ultimately, it forms a highly reliable, adaptive, and evolvable intelligent engineering management closed loop, providing solid technical support for the construction and long-term safety of complex reef limestone foundations.

[0098] S41. Use industrial-grade embedded devices (such as the Jetson series) as edge nodes to deploy lightweight prediction models accelerated by TensorRT.

[0099] The lightweight prediction formula is as follows:

[0100] in, Indicates the original floating-point weight or activation value; Indicates the scaling factor; Indicates zero-point offset; This represents the quantized 8-bit integer value.

[0101] Edge nodes have a built-in real-time data pipeline to complete data acquisition, feature extraction and model inference, ensuring that a single prediction is completed within 500ms.

[0102] Specifically, INT8 quantization: W_out weight w=0.184 (FP32), layer range [-0.452, 0.621]: S_W=(0.621+0.452) / 255=0.00421,Z_W=round(0.452 / 0.00421)=107 w_int8=round(0.184 / 0.00421+107)=151 Inverse quantization: w_dequant=(151-107)×0.00421=0.1852 Error = |0.184 - 0.1852| = 0.0012 (relative 0.65%) S42. Edge nodes are directly connected to field sensors and actuators, and realize local second-level closed-loop control based on MPC instructions. At the same time, they synchronize data with the cloud platform through the 5G network, forming a collaborative architecture of "real-time execution at the edge and analysis and optimization in the cloud".

[0103] S43. Develop a web visualization platform to display monitoring data, forecast curves, and control status in real time.

[0104] The adaptive sampling rate adjustment formula is as follows:

[0105] in, This indicates the dynamically adjusted sensor sampling frequency; Indicates the basic sampling frequency; This represents the standard deviation of the current predicted output; This indicates a preset upper limit for uncertainty; This represents the adjustment factor.

[0106] Specifically, f_base = 1 time / h, σ_max = 5.0mm, α = 2.0.

[0107] f = f_base × (1 + α × σ / σ_max) Normal (σ=1.4mm): f=1×(1+2.0×1.4 / 5.0)=1.6 times / h (approximately 38 minutes / time).

[0108] High risk (σ=4.8mm): f=1×(1+2.0×4.8 / 5.0)=2.9≈3 times / h (approximately 20 minutes / time).

[0109] The system automatically pushes early warning information based on the risk level and provides a human-machine collaboration interface to support the confirmation or adjustment of automatic control commands.

[0110] S44. The system continuously collects on-site data and performs incremental model learning and periodic updates in the cloud. After verification through digital twins, the optimized model is securely deployed to the edge to improve system performance.

[0111] Although the present invention has been described using the above preferred embodiments, it is not intended to limit the scope of protection of the present invention. Any changes and modifications made by those skilled in the art to the above embodiments without departing from the spirit and scope of the present invention shall still fall within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of bearing capacity and settlement control of reef limestone foundation, characterized in that, Includes the following steps: S1. Deploy a multi-source sensing system, preprocess the collected data, extract key features from the preprocessed data, and generate a sample database with spatiotemporal labels. S2. Construct a bearing capacity prediction model for reef limestone foundation. The bearing capacity prediction model includes a physical mechanism module, a data-driven module, and a weighted fusion module. S21. For the physical mechanism module, a viscoelastic-plastic constitutive model considering damage-creep coupling is established, and a heterogeneous three-dimensional numerical model reflecting the spatial variability of karst is constructed based on the sample database. By nesting the viscoelastic-plastic constitutive model on the heterogeneous three-dimensional numerical model, batch numerical tests under multiple working conditions are carried out to generate a physical simulation sample library. S22. For the data-driven module, based on the physical simulation sample library, the mapping from monitoring features to bearing capacity and settlement is learned through a time-series and spatial dual-branch structure. S23. For the weighted fusion module, evaluate the prediction uncertainty of the physical mechanism module and the data-driven module respectively, and dynamically allocate the fusion weight accordingly. S24. Train the bearing capacity prediction model; S3. Based on the bearing capacity prediction model of reef limestone foundation, a settlement prediction model integrating long short-term memory network and temporal convolutional network is constructed, and the settlement prediction model is embedded into the MPC framework to generate the optimal control command sequence in real time through MPC.

2. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 1, characterized in that, In step S1, the preprocessing process includes: The raw data acquired by the multi-source sensing system is subjected to sliding window filtering and wavelet denoising. The Laida criterion and isolated forest algorithm are used to automatically detect and remove outliers, and spatiotemporal interpolation is used to repair missing values.

3. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 1, characterized in that, In step S1 Key features include: time-domain features, frequency-domain features, time-frequency features, and spatial features; The process of generating a sample database includes: Key features are normalized and concatenated, weighted fusion is performed using an attention mechanism, and dimensionality is reduced using principal component analysis to form a unified high-dimensional feature tensor. Feature samples are then labeled with their states to construct a sample database with spatiotemporal labels.

4. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 1, characterized in that, In step S21, the core equations of the viscoelastic-plastic constitutive model are as follows: ， in, Indicates the total strain rate; Indicates viscoelastic strain rate; This represents the creep strain rate based on the time-hardening law; This represents the statistical damage strain rate based on the Weibull distribution.

5. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 4, characterized in that, In step S22, the learning processes for the temporal and spatial dual-branch structures respectively include: Temporal branching utilizes one-dimensional depthwise separable dilated causal convolution to extract dynamic patterns in the signal. The calculation formula is as follows: ， in, This represents the current convolutional layer number. This represents the time step position of the currently processed timing signal. For the first Hidden features in layers; ReLU is the activation function; The kernel size; This represents the time step offset of the output features from the previous layer. For the first The expansion rate of the layer; and The first Trainable convolutional weights and biases for each layer; The spatial branch learns the spatial propagation pattern of damage through two layers of graph convolution, and the calculation formula is as follows: ， in, The adjacency matrix with self-loops; It is a degree matrix; For the first Layer node features; These are trainable weights; For activation functions; The timing branch output and the spatial branch output are concatenated, and the calculation formula is as follows: ， in, Indicates the output of a timing branch; Indicates spatial branch output; This indicates the merged output value.

6. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 5, characterized in that, In step S23, the calculation formula for dynamically allocating fusion weights is as follows: ， in, This represents the predicted value from the physical mechanism module; This represents the predicted value from the data-driven module; and These represent the adaptive weighting coefficients of the physical mechanism module and the data-driven module, respectively. This represents the prediction variance of the physical mechanism module; This represents the prediction variance of the data-driven module; This indicates the bearing capacity and settlement prediction results after fusion.

7. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 6, characterized in that, In step S24, the bearing capacity prediction model is trained using a composite loss function that integrates Huber loss and physical constraint terms. The composite loss function is as follows: , in, Represents the physical regularity term. Indicates hyperparameters, This represents the model's predicted value. Represents the actual value of the model. Indicates the threshold parameter; Represents the composite loss function; This represents the total loss function for model training.

8. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 1, characterized in that, In step S3, the output models of the Long Short-Term Memory network and the Temporal Convolutional Network respectively include: The output model of the Long Short-Term Memory network is as follows: ， in, Indicates the current time The hidden state output vector; Indicates the activation function; This represents the weight matrix of the output gate; Indicates the previous moment The hidden state; Indicates the current time The input feature vector; represents the bias term of the output gate; tanh represents the hyperbolic tangent activation function; Indicates the current time cellular state; The output model of a temporal convolutional network is as follows: ， in, This indicates that the convolution operation occurs at time steps. Output value at; This indicates the time index corresponding to the current output position; Indicates the kernel size; Represents the loop variable for summation; Indicates the convolution kernel at the th... Weight parameters at each position; Indicates the expansion coefficient; This indicates that the time index in the input sequence is The value of .

9. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 8, characterized in that, In step S3, the optimization formula for MPC is as follows: ， ， in, To control the sequence of instructions; The prediction time domain length for MPC; Indicates the first The predicted foundation settlement; Indicates the desired settlement reference trajectory; Indicates the first The control input vector for the step; This represents the regularization weight coefficient; st Indicates constraints; Indicates the settlement threshold; Indicates the maximum settlement threshold; Indicates differential settlement constraint; This represents the maximum differential settlement constraint value.

10. The intelligent prediction and settlement control method for bearing capacity of reef limestone foundation according to claim 1, characterized in that, Also includes: S4. Deploy lightweight prediction models at edge nodes. Edge nodes are directly connected to field multi-source sensing systems and actuators. Closed-loop control is achieved based on MPC commands. A Web visualization platform is developed to display monitoring data, prediction curves, and control status in real time.

Images