Karst cave pile foundation grouting real-time control system based on optical fiber sensing and AI dynamic optimization

Abstract

The invention discloses a karst cave pile foundation grouting real-time control system based on optical fiber sensing and AI dynamic optimization, which relates to the field of pile foundation grouting real-time control and comprises a data acquisition module, an intelligent analysis module, a dynamic optimization module and a feedback control module. Normalized optical fiber input vectors and geological input vectors are obtained by analyzing karst cave optical fiber sensing data and geological parameter data, fusion is performed through a double-branch convolutional network, a grouting diffusion prediction model is constructed through a convolutional neural network, the slurry diffusion radius and the filling saturation are output, the model is constructed based on an isolated forest algorithm, and the grouting diffusion prediction model is constructed. The method comprises the following steps: calculating abnormal score early warning, establishing a mapping model, and dynamically adjusting the weight to obtain a Pareto optimal solution set adjusted in each construction stage, and a closed-loop control module can perform timely early warning adjustment on an accident according to the Pareto optimal solution set adjusted by real-time progress and rescreening parameters, can sense data in real time to form closed-loop feedback, and meets the requirements of engineering economy.

Description

Technical Field

[0001] The present invention relates to the field of real-time control of pile foundation grouting, and specifically to a real-time control system for karst cave pile foundation grouting based on fiber optic sensing and AI dynamic optimization. Background Art

[0002] Traditional karst cave grouting construction relies on point sensors and manual experience for monitoring, and can only obtain local point data, unable to perceive the diffusion radius and filling saturation of karst cave slurry in real time. Although Patent 202021352717 realizes karst cave identification and positioning through distributed optical fibers, its technical solution focuses on karst cave detection in the static load test stage and is not linked to the grouting process, unable to dynamically adjust grouting parameters according to the karst cave morphology, resulting in rapid loss of slurry in high-permeability fracture areas or insufficient local filling. In addition, dynamic geological changes such as hidden fractures and groundwater seepage are difficult to be captured in time, easily leading to accidents such as hole wall collapse or slurry loss, and lacking early warning and adjustment of accidents.

[0003] At the same time, existing grouting control systems mostly use preset fixed parameters for construction and lack the ability to dynamically respond to geological conditions. At present, although some technologies attempt to introduce AI algorithms, their model training relies on historical data and does not form a closed-loop feedback with real-time sensing data, resulting in prediction lag. In addition, existing technologies mostly focus on single-objective optimization and ignore the coordination of multiple objectives such as material cost and construction period, making it difficult to meet the engineering economic requirements.

[0004] To solve the above defects, a technical solution is provided now. Summary of the Invention

[0005] To solve the technical problems raised in the above background art, the present invention is proposed. An embodiment of the present invention provides a real-time control system for karst cave pile foundation grouting based on fiber optic sensing and AI dynamic optimization.

[0006] The object of the present invention can be achieved by the following technical solutions: A real-time control system for karst cave pile foundation grouting based on fiber optic sensing and AI dynamic optimization includes a data acquisition module, an intelligent analysis module, a dynamic optimization module, and a feedback control module.

[0007] The data acquisition module includes a distributed fiber optic sensing network, arranges fiber Bragg grating arrays along the axial and radial directions of the pile hole, combines with flexible nano sensors to form a three-dimensional monitoring network covering the inside of the karst cave, uses optical frequency domain reflectometry technology to achieve high-precision continuous monitoring, synchronously collects strain, temperature, and acoustic vibration signals, collects geological parameter data such as karst cave volume, fracture density, and permeability, and transmits them to the intelligent analysis module;

[0008] The intelligent analysis module is used to receive the fiber optic sensing data and geological parameters of the data acquisition module, predict the grouting diffusion path through a convolutional neural network, and detect construction anomalies to trigger hierarchical early warnings;

[0009] A dynamic optimization module, which is used to generate a Pareto optimal solution set for grouting parameter adjustment by using a multi-objective genetic algorithm and transmit it to the feedback control module;

[0010] A feedback control module, which realizes dynamic regulation of grouting parameters based on the adjusted Pareto optimal solution set.

[0011] Furthermore, the following steps are also included:

[0012] S110: Data acquisition, obtaining karst cave optical fiber sensing data and geological parameters collected by the data acquisition module;

[0013] S120: Data fusion, the karst cave optical fiber sensing data and geological parameter data are respectively arranged and analyzed according to the time dimension to obtain a normalized optical fiber input vector and a geological input vector, and the optical fiber input vector and the geological input vector are fused through a double-branch convolutional network to obtain the fused final feature;

[0014] S130: Modeling and analysis, constructing a grouting diffusion prediction model through a convolutional neural network, inputting the fused final feature into a fully connected layer, calculating the total loss in combination with the physical constraint loss, and backpropagating to optimize the parameters until the training is terminated to obtain the grouting diffusion prediction model, and outputting the slurry diffusion radius and filling saturation;

[0015] S140: Early warning and adjustment, constructing a model based on the isolation forest algorithm with the slurry diffusion radius and filling saturation as features, calculating the anomaly score, and giving three-level early warnings according to the threshold to guide the construction operation;

[0016] S150: Dynamic optimization, establishing a mapping model based on a multi-layer perceptron, randomly generating an initial population, and dynamically adjusting the weights through operations such as screening, crossover, and mutation to obtain the Pareto optimal solution set adjusted for each construction stage;

[0017] S160: Closed-loop control, the closed-loop control module screens the parameters according to the Pareto optimal solution set adjusted in real time, compares the diffusion radius and interval output by the model to adjust the pressure, re-screens the parameters, and loops until stable to accurately control the diffusion path.

[0018] Furthermore, the analysis steps of the fused final feature are as follows:

[0019] S121: Organize the strain, temperature, and acoustic wave vibration of the optical fiber sensing data into a two-dimensional matrix X of the optical fiber input vector according to the time series f , where X f ∈R T×Df , T is the time step, data at T moments are collected in chronological order, and R T×Df represents that the matrix is in the real number space, D fis the fiber optic feature dimension, and the fiber optic sensing data at each moment are specifically strain, temperature, and acoustic wave vibration data; the geological data, including cave volume, fracture density, and permeability, are constructed into a geological input vector X according to the time series g , where X g ∈R T ×Dg , T is consistent with the time step of the fiber optic data to ensure time alignment, and D g is the geological feature dimension;

[0020] Through the normalization formulas X f norm =(X f -u f ) / σ f and X g norm =(X c -u g ) / σ g , the normalized fiber optic input vector X f norm and the geological input vector X g norm are obtained, where u and σ are the mean and standard deviation respectively;

[0021] S122: Perform one-dimensional convolution operation through a convolutional neural network, combined with dilated convolution, to extract high-frequency dynamic features. Through the fiber optic branch formula F t =Conv1D(X f norm ), F t ∈R T×df , the fiber optic feature F t at time t is obtained, where Conv1D is the one-dimensional convolution operation, which extracts features from the input data through the convolutional kernel filter, combines dilated convolution to expand the receptive field, captures the high-frequency dynamic changes of the fiber optic data, df is the set fiber optic branch feature dimension, captures the steady-state geological features through causal convolution, and through the geological branch formula G t =CausalConv(X g norm ), G t ∈R T×dg , the geological feature G t at time t is obtained, and CausalConv is the causal convolution;

[0022] S123: Calculate the fusion weight s t =Sigmoid(Wf×F t +Wg×G t +b), s tDenote the fusion weight of the fiber optic feature and the geological feature at time t. Sigmoid is the activation function, and Wf and Wg are learnable weight matrices. By training to optimize the parameters, through the feature fusion formula, fuse the fiber optic and geological features, specifically Ffin t = M fuse (s t ⊙ DWConv(F t ) + (1 - s t ) ⊙ GAP(G t-k：t )) to obtain the final fused feature Ffin at time t t , where ⊙ is element-wise multiplication, DWConv(F t ) is to apply depthwise separable convolution to the fiber optic feature F t , GAP(G t-k：t ) performs global average pooling on the geological feature sliding window G t-k：t with window size k, and M fuse is an additional fusion operation.

[0023] Furthermore, the analysis steps of the slurry diffusion radius and filling saturation are as follows:

[0024] Introduce a regularization loss constraint weight variance formula to perform dynamic weight correction on the model when fusing features; construct a grouting diffusion prediction model through a convolutional neural network, and input the final fused feature Ffin t into the fully connected layer. Through the linear transformation and activation function of the fully connected layer, output the slurry diffusion radius R and filling saturation β. The formula is [R, β] = FC(Ffin t ) = σ out (W fh ×Ffin t + b fh ), where σ out is the activation function of the output layer, and W fh and b fh are the learnable weights and biases of the fully connected layer. Calculate the physical constraint loss L ph , calculate the total loss L of the model through the formula tot , and use backpropagation to find the gradient. The optimizer updates the model parameters according to the gradient and iterates continuously until the preset number of training times is reached, and the training terminates. Obtain the grouting diffusion prediction model, input the karst cave fiber optic sensing data and geological parameters collected by the data acquisition module, and output the slurry diffusion radius and filling saturation.

[0025] Furthermore, the analysis steps of the warning adjustment are as follows:

[0026] Based on the Isolation Forest algorithm, by taking the output grouting diffusion radius R and filling saturation β as the two-dimensional feature vector Y = [R, β], organizing them in time series and normalizing the formula standard, an Isolation Forest model is constructed. The model generates multiple isolation trees by randomly splitting the feature space, calculates the sample path length h(Y), obtains the average path length E(h(Y)) through the formula, and obtains the anomaly score k(Y) through the formula. If the anomaly score of the real-time data is greater than the threshold TG1, a first-level warning signal is issued accordingly, and corresponding measure one is taken. If the anomaly score of the real-time data is less than or equal to the threshold TG1 and greater than the threshold TG2, a second-level warning signal is issued accordingly, and corresponding measure two is taken. If the anomaly score of the real-time data is less than or equal to the threshold TG2, a third-level warning signal is issued accordingly.

[0027] Further, the closed-loop control analysis steps are as follows:

[0028] The closed-loop control module realizes the dynamic regulation of grouting parameters based on the Pareto optimal solution set, obtains the Pareto optimal solution set adjusted in real time according to the real-time construction progress, and screens out the initial parameter combination (P0*, Q0*, r0* that adapts to the current stage goal from the solution set ) , and outputs the grout diffusion radius Rpred through the grouting diffusion model and compares it with the set interval [Rmin, Rmax]. If the grout diffusion radius Rpred is within the interval range, the current parameters are maintained; if the grout diffusion radius Rpred is greater than the maximum value Rmax of the interval, the adjusted pressure value Pnew is obtained according to the formula, and with this adjusted pressure value Pnew as the core, a parameter combination is re-screened from the Pareto optimal solution set; if the grout diffusion radius Rpred is less than the minimum value Rmin of the interval, the adjusted pressure value Pnew is obtained according to the formula, ρ is the sensitivity coefficient related to the construction stage, and with this adjusted pressure value Pnew as the core, a parameter combination is re-screened from the Pareto optimal solution set;

[0029] According to the re-matched parameter combination (P1*, Q1*, r1* ) Control the grouting equipment, repeat the process of "model prediction Rpred → deviation discrimination → parameter adjustment → solution set screening" until Rpred is stable within the interval [Rmin, Rmax], and dynamically adjust the parameters of the grouting equipment through the closed-loop control module to achieve precise control of the grout diffusion path.

[0030] Further, the analysis steps of the adjusted Pareto optimal solution set are as follows:

[0031] Construct the grouting parameter vector Q is the grouting flow rate, r1, r2, and r3 are the proportions of cement, water, and admixture, and the construction target vector δ is the filling rate, C is the cost, T is the construction period, through the function Establish a mapping. The f mapping function is obtained based on a neural network. Collect historical data of grouting parameter ratios and corresponding construction targets, perform normalization processing, and use a multi-layer perceptron to establish a non-linear mapping model. The input layer of the model is 3-dimensional, specifically the grouting pressure, grouting flow rate, and slurry ratio. The ReLU activation function is used in the hidden layer, and the output layer directly predicts the filling rate, cost, and construction period. During the training process, the Adam optimizer is used to minimize the mean square error loss, and L2 regularization and early stopping mechanism are combined to prevent overfitting. Finally, the f mapping function is obtained.

[0032] Randomly generate H combinations of grouting parameters to form an initial population. For each parameter combination, calculate the target through the f mapping function. Let i be the serial number of the parameter combination. Stratify according to the dominance relationship. Those with no individual dominance are in the first layer of the Pareto front, and divide layer by layer. Calculate the crowding distance of each individual in the target space to measure the surrounding density. Based on the non-dominated rank and crowding degree, select individuals through the roulette wheel method. Perform crossover such as simulated binary crossover and mutation such as polynomial mutation on the selected individuals to generate offspring. Combine the parent generation and offspring, and reselect to maintain the population size H. Repeat until the maximum number of iterations. All non-dominated solutions in the final population, that is, the individuals in the first layer of the Pareto front, constitute the Pareto optimal solution set. And optimize and adjust the weights for different construction stages, dynamically adjust the weights of the filling rate δ, cost C, and construction period T, determine the stage of construction and match the corresponding weight value range, and obtain the adjusted Pareto optimal solution set for each construction stage.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] 1. The present invention collects the karst cave optical fiber sensing data and geological parameters through the data acquisition module. The karst cave optical fiber sensing data and geological parameter data are respectively arranged and analyzed according to the time dimension to obtain the normalized optical fiber input vector and geological input vector. The optical fiber input vector and geological input vector are fused through a double-branch convolutional network to obtain the final fused feature. A grouting diffusion prediction model is constructed through a convolutional neural network. The final fused feature is input into the fully connected layer, and the total loss is calculated by combining the physical constraint loss. The parameters are optimized by backpropagation until the training terminates to obtain the grouting diffusion prediction model, and the slurry diffusion radius and filling saturation are output. Based on the isolated forest algorithm, a model is constructed with the slurry diffusion radius and filling saturation as features, the anomaly score is calculated, and three-level early warnings are given according to the threshold to guide the construction operation. It can real-time sense the slurry diffusion radius and filling saturation of the karst cave, timely capture dynamic geological changes such as hidden fractures and groundwater seepage, and timely give early warning and adjustment for accidents such as hole wall collapse or slurry loss.

[0035] 2. The present invention establishes a mapping model based on a multi-layer perceptron, randomly generates an initial population, and through operations such as screening, crossover, and mutation, dynamically adjusts the weights to obtain the Pareto optimal solution set adjusted for each construction stage. The closed-loop control module screens parameters according to the Pareto optimal solution set adjusted for the real-time progress, compares the diffusion radius output by the model with the interval to adjust the pressure and re-screen the parameters, and loops until stable, precisely controlling the diffusion path. It has the ability to dynamically respond to geological conditions, can form a closed-loop feedback with real-time sensing data, and can also coordinate multiple objectives such as material cost and construction period, meeting the engineering economic requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. The following drawings are not deliberately drawn to scale in actual size, and the focus is on showing the gist of the present invention.

[0037] Figure 1 is the system block diagram of the present invention;

[0038] Figure 2 is the method flow chart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts also belong to the scope of protection of the present invention.

[0040] As Figure 1 shown, the real-time control system for karst cave pile foundation grouting based on fiber optic sensing and AI dynamic optimization includes a data acquisition module, an intelligent analysis module, a dynamic optimization module, and a feedback control module.

[0041] The data acquisition module includes a distributed fiber optic sensing network, which arranges a fiber optic grating array along the axial and radial directions of the pile hole, combines with a flexible nano sensor to form a three-dimensional monitoring network covering the inside of the karst cave. The specific fiber layout is dynamically adjusted according to the geometric shape of the karst cave to ensure the monitoring coverage of key areas such as fracture development zones. The optical frequency domain reflectometry technology is used to achieve high-precision continuous monitoring, synchronously collect strain, temperature, and acoustic vibration signals, and collect geological parameter data specifically including the volume of the karst cave, fracture density, and permeability, and transmit them to the intelligent analysis module;

[0042] The intelligent analysis module is used to receive the fiber optic sensing data and geological parameters of the data acquisition module, predict the grouting diffusion path through a convolutional neural network, and detect construction anomalies to trigger hierarchical warnings;

[0043] The dynamic optimization module is used to generate a Pareto optimal solution set for grouting parameter adjustment by using a multi-objective genetic algorithm and transmit it to the feedback control module;

[0044] The feedback control module realizes dynamic regulation of grouting parameters based on the adjusted Pareto optimal solution set.

[0045] As Figure 2 shown, the present invention further includes the following steps:

[0046] S110: Data acquisition, obtaining the karst cave optical fiber sensing data and geological parameters collected by the data acquisition module;

[0047] S120: Data fusion, the karst cave optical fiber sensing data and geological parameter data are respectively arranged and analyzed according to the time dimension to obtain a normalized optical fiber input vector and a geological input vector, and the optical fiber input vector and the geological input vector are fused through a double-branch convolutional network to obtain the final fused feature;

[0048] S130: Modeling and analysis, constructing a grouting diffusion prediction model through a convolutional neural network, inputting the final fused feature into a fully connected layer, calculating the total loss in combination with the physical constraint loss, backpropagating to optimize the parameters until the training is terminated, obtaining the grouting diffusion prediction model, and outputting the slurry diffusion radius and filling saturation degree;

[0049] S140: Early warning and adjustment, constructing a model based on the isolation forest algorithm with the slurry diffusion radius and filling saturation degree as features, calculating the anomaly score, and giving three-level early warnings according to the threshold to guide the construction operation;

[0050] S150: Dynamic optimization, establishing a mapping model based on a multi-layer perceptron, randomly generating an initial population, and dynamically adjusting the weights through operations such as screening, crossover, and mutation to obtain a Pareto optimal solution set adjusted for each construction stage;

[0051] S160: Closed-loop control, the closed-loop control module screens the parameters according to the Pareto optimal solution set adjusted in real time, compares the diffusion radius and interval output by the model to adjust the pressure, re-screens the parameters, and loops until stable to accurately control the diffusion path.

[0052] Specifically, the analysis steps of step S120 data fusion are as follows:

[0053] S121: Organize the optical fiber sensing data strain, temperature, and acoustic wave vibration into a two-dimensional matrix X of the optical fiber input vector according to the time series f , where X f ∈R T×Df , T is the time step, data at T moments are collected in time order, and R T×Df represents that the matrix is in the real number space, and D fis the fiber optic feature dimension. The fiber optic sensing data at each moment is specifically strain, temperature, and acoustic wave vibration data, including D f features. Specifically, 3 features of strain, temperature, and acoustic wave vibration are collected at each time. When T = 100, then X f is a 100×3 matrix; the geological data of cave volume, fracture density, and permeability are constructed into a geological input vector X g , where X g ∈R T×Dg , T is consistent with the time step of the fiber optic data to ensure time alignment, and D g is the geological feature dimension;

[0054] Through the normalization formulas X f norm =(X f -u f ) / σ f and X g norm =(X c -u g ) / σ g , the normalized fiber optic input vector X f norm and the geological input vector X g norm are obtained, where u and σ are the mean and standard deviation respectively;

[0055] S122: Perform one-dimensional convolution operation through a convolutional neural network, combine dilated convolution, extract high-frequency dynamic features, and through the fiber optic branch formula F t =Conv1D(X f norm ), F t ∈R T×df , the fiber optic feature F t at time t is obtained, where Conv1D is the one-dimensional convolution operation. Specifically, the feature extraction of the input data is performed through a convolutional kernel filter size = 5, and the receptive field is expanded by combining dilated convolution with a dilation rate = 2 to capture the high-frequency dynamic changes of the fiber optic data. df is the set fiber optic branch feature dimension. Through causal convolution, the steady-state geological features are captured. Through the geological branch formula G t =CausalConv(X g norm ), G t ∈R T×dg , the geological feature G t at time t is obtained. CausalConv is causal convolution, which specifically calculates only using the current time and historical data before it. Through the convolution operation with a filter size = 3 and a stride = 1, the features of the geological data are extracted;

[0056] S123: Calculate the fusion weight s t = Sigmoid(Wf × F t + Wg × G t + b), s t represents the fusion weight of the fiber optic feature and the geological feature at time t. Sigmoid is the activation function. Wf and Wg are learnable weight matrices responsible for automatically adjusting the importance of the fiber optic feature F t and the geological feature G t . Through training to optimize the parameters, and through the feature fusion formula, fuse the fiber optic and geological features. Specifically, Ffin t = M fuse (s t ⊙ DWConv(F t )) + (1 - s t ) ⊙ GAP(G t-k：t )) to obtain the final fused feature Ffin at time t t , where ⊙ is element-wise multiplication, DWConv(F t ) is to apply depthwise separable convolution to the fiber optic feature F t , GAP(G t-k：t ) performs global average pooling on the geological feature sliding window G t-k：t with window size k, and M fuse is an additional fusion operation, specifically a fully connected layer or normalization.

[0057] The specific analysis steps of step S130 modeling analysis are as follows:

[0058] Introduce a regularization loss constraint weight variance formula to perform dynamic weight correction on the model during feature fusion L re is the regularization loss, which constrains the stability of the fusion weight s t . η is a hyperparameter, and E[s] is the mean of the fusion weight s t in the time dimension, to avoid the model relying too much on a single modality;

[0059] Construct a grouting diffusion prediction model through a convolutional neural network. Input the final fused feature Ffin t into the fully connected layer. Through the linear transformation and activation function of the fully connected layer, output the slurry diffusion radius R and the filling saturation β. The formula is [R, β] = FC(Ffin t ) = σ out (W fh × Ffin t + b fh ), where σ out is the activation function of the output layer, and W fh and b fhare the learnable weights and biases of the fully connected layer, calculating the physical constraint loss L ph , specifically Φ is a hyperparameter used to adjust the weight of the physical constraint loss. is the slurry diffusion velocity, P inj is the grouting pressure. is the slurry viscosity, χ is the proportionality coefficient related to the change rate of grouting pressure, slurry viscosity and diffusion radius. Through the formula L tot = η × L re + Φ × L ph , the total loss L of the model is calculated tot , and the gradient is obtained using backpropagation. The optimizer updates the model parameters according to the gradient and iterates continuously until the preset number of training times is reached. The training terminates, and a grouting diffusion prediction model is obtained. Inputting the karst cave optical fiber sensing data and geological parameters collected by the data acquisition module, the slurry diffusion radius and filling saturation are output.

[0060] Specifically, the analysis steps of the early warning adjustment in step S140 are as follows:

[0061] Based on the isolation forest algorithm, by taking the output grouting diffusion radius R and filling saturation β as the two-dimensional feature vector Y = [R, β], organizing them in time series and normalizing the formula standard, an isolation forest model is constructed. The model generates multiple isolation trees by randomly splitting the feature space, calculates the sample path length h(Y), and through the formula the average path length E(h(Y)) is obtained, j is the serial number of the isolation tree, taking positive integers, and the maximum value is M. Through the formula k(Y) = 2 - E( h(Y)) / c (N) , the anomaly score k(Y) is obtained, where c(N) = 2H(N - 1) - 2(N - 1) / N, N is the number of samples used for training the isolation forest, H(N - 1) is the harmonic number. If the anomaly score of the real-time data is greater than the threshold TG1, a first-level early warning signal is sent correspondingly, and it is reminded to suspend construction and start the emergency plan. If the anomaly score of the real-time data is less than or equal to the threshold TG1 and greater than the threshold TG2, a second-level early warning signal is sent correspondingly, and it is reminded that the grouting parameters need to be adjusted. If the anomaly score of the real-time data is less than or equal to the threshold TG2, a third-level early warning signal is sent correspondingly, and there is no corresponding operation.

[0062] Specifically, the analysis steps of the dynamic optimization in step S150 are as follows:

[0063] Construct the grouting parameter vector Q is the grouting flow rate, r1, r2, and r3 are the proportions of cement, water, and admixture, and the construction target vector δ is the filling rate, C is the cost, T is the construction period, through the function Establish a mapping. The f mapping function is obtained based on a neural network. Collect historical data of grouting parameter ratios and corresponding construction targets, perform normalization processing, and use a multi-layer perceptron to establish a non-linear mapping model. The input layer of the model is 3-dimensional, specifically the grouting pressure, grouting flow rate, and slurry ratio. The ReLU activation function is used in the hidden layer. The output layer directly predicts the filling rate, cost, and construction period. During the training process, the Adam optimizer is used to minimize the mean squared error loss, and L2 regularization and early stopping mechanism are combined to prevent overfitting. Finally, the f mapping function is obtained.

[0064] Randomly generate H combinations of grouting parameters to form an initial population. For each parameter combination, calculate the target through the f mapping function. Let i be the serial number of the parameter combination. Stratify according to the dominance relationship. Those without individual dominance are in the first layer of the Pareto front, and divide layer by layer. Calculate the crowding distance of each individual in the target space to measure the surrounding density. Based on the non-dominated rank and crowding degree, select individuals through the roulette wheel method. Perform crossover (such as simulated binary crossover) and mutation (such as polynomial mutation) on the selected individuals to generate offspring. Combine the parent generation and offspring, and re-select to maintain the population size H. Repeat until the maximum number of iterations. All non-dominated solutions in the final population, that is, the individuals in the first layer of the Pareto front, constitute the Pareto optimal solution set, and perform weight optimization and adjustment for different construction stages. Dynamically adjust the weights of the filling rate δ, cost C, and construction period T, which are w1, w2, and w3 respectively, and w1 + w2 + w3 = 1. If the construction progress is between 0 - 30%, it is judged as the initial construction stage. If the construction progress is between 30 - 70%, it is judged as the middle construction stage. If the construction progress is between 70 - 100%, it is judged as the late construction stage. Corresponding to the initial construction stage, the value ranges of w1, w2, and w3 are 0.5 - 0.7, 0.6 - 0.5, and 0.2 - 0.4 respectively. For the middle construction stage, the value ranges of w1, w2, and w3 are 0.1 - 0.3, 0.2 - 0.4, and 0.4 - 0.6 respectively. For the late construction stage, the value ranges of w1, w2, and w3 are 0.1 - 0.3, 0.3 - 0.4, and 0.5 - 0.6 respectively, to obtain the adjusted Pareto optimal solution set for each construction stage.

[0065] Specifically, the analysis steps of the closed-loop control in step S160 are as follows:

[0066] The closed-loop control module realizes the dynamic regulation of grouting parameters based on the Pareto optimal solution set, obtains the Pareto optimal solution set adjusted in real time according to the real-time construction progress, and screens out the initial parameter combination (P0 * , Q0 * , r0 *) The grout diffusion radius Rpred is output through the grouting diffusion model and compared with the set interval [Rmin, Rmax]. If the grout diffusion radius Rpred is within the interval, the current parameters are maintained; if the grout diffusion radius Rpred is greater than the maximum value Rmax of the interval, according to the formula Pnew = Rpred × [1 - γ × (Rpred - Rmax) / Rpred], the adjusted pressure value Pnew is obtained, where γ is the sensitivity coefficient related to the construction stage, and the value ranges in the initial stage, middle stage, and later stage of construction are 0.3 - 0.5, 0.5 - 0.7, and 0.7 - 0.8 respectively. Taking this adjusted pressure value Pnew as the core, a parameter combination is re-screened from the Pareto optimal solution set; if the grout diffusion radius Rpred is less than the minimum value Rmin of the interval, according to the formula Pnew = Rpred × [1 + ρ × (Rmin - Rpred) / Rpred], the adjusted pressure value Pnew is obtained, where ρ is the sensitivity coefficient related to the construction stage, and the value ranges in the initial stage, middle stage, and later stage of construction are 0.7 - 0.8, 0.5 - 0.7, and 0.3 - 0.5 respectively. Taking this adjusted pressure value Pnew as the core, a parameter combination is re-screened from the Pareto optimal solution set;

[0067] According to the re-matched parameter combination (P1*, Q1*, r1* ) Control the grouting equipment and repeat the process of "model prediction Rpred → deviation discrimination → parameter adjustment → solution set screening" until Rpred is stable within the interval [Rmin, Rmax]. Dynamically adjust the parameters of the grouting equipment through the closed-loop control module to achieve precise control of the grout diffusion path.

[0068] The above is the description of the present invention and should not be regarded as a limitation thereof. Although several exemplary embodiments of the present invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the present invention. Therefore, all such modifications are intended to be included within the scope of the present invention as defined by the claims. It should be understood that the above is the description of the present invention and should not be considered limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present invention is defined by the claims and their equivalents.

Claims

1. A real-time control system for grouting of karst cave pile foundation based on fiber optic sensing and AI dynamic optimization, including a data acquisition module, an intelligent analysis module, a dynamic optimization module and a feedback control module, characterized in that: The data acquisition module includes a distributed fiber optic sensing network, arranging fiber Bragg grating arrays along the axial and radial directions of the pile hole, combining with flexible nano-sensors to form a three-dimensional monitoring network covering the inside of the karst cave, and adopting optical frequency domain reflectometry technology to achieve high-precision continuous monitoring, synchronously collecting strain, temperature, acoustic vibration signals and geological parameter data, and transmitting them to the intelligent analysis module; The intelligent analysis module is used to receive the fiber optic sensing data and geological parameters of the data acquisition module, predict the grouting diffusion path through a convolutional neural network and detect construction anomalies to trigger hierarchical early warnings; The dynamic optimization module is used to generate a Pareto optimal solution set for adjusting grouting parameters by using a multi-objective genetic algorithm and transmit it to the feedback control module; The feedback control module realizes dynamic regulation of grouting parameters based on the Pareto optimal solution set.

2. The real-time control system for karst cave pile foundation grouting based on optical fiber sensing and AI dynamic optimization according to claim 1, wherein It also includes: Obtaining the karst cave fiber optic sensing data and geological parameters collected by the data acquisition module; The karst cave fiber optic sensing data and geological parameter data are respectively arranged and analyzed according to the time dimension to obtain the normalized fiber input vector and geological input vector, and the fiber input vector and geological input vector are fused through a double-branch convolutional network to obtain the final fused feature; Construct a grouting diffusion prediction model through a convolutional neural network, input the final fused feature into the fully connected layer, calculate the total loss by combining the physical constraint loss, and backpropagate to optimize the parameters until the training ends to obtain the grouting diffusion prediction model, and output the slurry diffusion radius and filling saturation; Based on the isolation forest algorithm, build a model with the slurry diffusion radius and filling saturation as features, calculate the anomaly score, and issue early warnings in three levels according to the threshold to guide construction operations; Establish a mapping model based on a multi-layer perceptron, randomly generate an initial population, and through screening, crossover, and mutation operations, dynamically adjust the weights to obtain the Pareto optimal solution set adjusted for each construction stage; The closed-loop control module selects parameters according to the Pareto optimal solution set adjusted according to the real-time progress, compares the output diffusion radius and interval of the model to adjust the pressure, re-selects the parameters, and loops until stable.

3. The real-time control system for karst cave pile foundation grouting based on optical fiber sensing and AI dynamic optimization according to claim 2, wherein The analysis of the final fused feature is as follows: Organize the fiber optic sensing data of strain, temperature, and acoustic wave vibration into a two-dimensional matrix X of fiber optic input vectors according to the time series f , where X f ∈R T×Df , T is the time step, and data at T moments are collected in chronological order. R T×Df represents that the matrix is in the real number space. D f is the fiber optic feature dimension. The fiber optic sensing data at each moment are specifically strain, temperature, and acoustic wave vibration data; construct the geological data of karst cave volume, fracture density, and permeability into a geological input vector X g , where X g ∈R T×Dg , T is the same as the time step of the fiber optic data, and D g is the geological feature dimension; Through the normalization formula X f norm =(X f -u f ) / σ f and X g norm =(X c -u g ) / σ g , the normalized optical fiber input vector X f norm and the geological input vector X g norm are obtained, where u and σ are the mean and standard deviation respectively; Perform one-dimensional convolution operation through a convolutional neural network, combine dilated convolution to extract high-frequency dynamic features, and use the optical fiber branch formula F t = Conv1D(X f norm ), F t ∈R T×df , to obtain the optical fiber feature Ft at time t. Here, Conv1D is the one-dimensional convolution operation, which extracts features from the input data through the convolutional kernel filter, combines dilated convolution to expand the receptive field, captures the high-frequency dynamic changes of the optical fiber data, df is the set optical fiber branch feature dimension, captures the steady-state geological features through causal convolution, and uses the geological branch formula G t = CausalConv(X t g norm ), G t ∈R T×dg , to obtain the geological feature Gt at time t. CausalConv is the causal convolution;​ Calculate the fusion weight s t = Sigmoid(Wf × F t + Wg × G t + b), s t represents the fusion weight of the fiber optic feature and the geological feature at time t. Sigmoid is the activation function. Wf and Wg are learnable weight matrices. By training to optimize the parameters, through the feature fusion formula, the fiber optic and geological features are fused. Specifically, Ffin t = M fuse (s t ⊙ DWConv(F t + (1 - s t ) ⊙ GAP(G t-k：t )) to obtain the final fused feature Ffin at time t t , where ⊙ is element-wise multiplication, DWConv(F t ) is to apply depthwise separable convolution to the fiber optic feature F t , GAP(G t-k：t ) performs global average pooling on the geological feature sliding window G t-k：t with window size k, and M fuse is an additional fusion operation.

4. The real-time control system for karst cave pile foundation grouting based on optical fiber sensing and AI dynamic optimization according to claim 2, wherein The analysis of the slurry diffusion radius and filling saturation is as follows: Introduce the regularization loss constraint weight variance formula to dynamically correct the weights when the model fuses features; construct a grouting diffusion prediction model through a convolutional neural network, and fuse the final feature Ffin t Input it into the fully connected layer. Through the linear transformation and activation function of the fully connected layer, output the grout diffusion radius R and the filling saturation β. The formula is [R, β] = FC(Ffin t ) = σ out (W fh ×Ffin t +b fh ), where σ out is the activation function of the output layer, W fh and b fh are the learnable weights and biases of the fully connected layer. Calculate the physical constraint loss L ph . Through the formula, calculate the total loss L tot of the model, and use backpropagation to calculate the gradient. The optimizer updates the model parameters according to the gradient and iterates continuously until the preset number of training times is reached and the training terminates, obtaining the grouting diffusion prediction model. Input the karst cave optical fiber sensing data and geological parameters collected by the data acquisition module, and output the grout diffusion radius and the filling saturation.

5. The real-time control system for karst cave pile foundation grouting based on optical fiber sensing and AI dynamic optimization according to claim 2, wherein The analysis of the early warning adjustment is as follows: Based on the isolation forest algorithm, taking the output grouting diffusion radius R and filling saturation β as the two-dimensional feature vector Y = [R, β], organizing and normalizing them according to the time series formula standard, constructing an isolation forest model, the model generates multiple isolation trees by randomly splitting the feature space, calculates the sample path length h(Y) and the average path length E(h(Y)), and then obtains the anomaly score k(Y). If the anomaly score of the real-time data is greater than the threshold TG1, a first-level early warning signal is sent correspondingly, and corresponding measure one is taken. If the anomaly score of the real-time data is less than or equal to the threshold TG1 and greater than the threshold TG2, a second-level early warning signal is sent correspondingly, and corresponding measure two is taken. If the anomaly score of the real-time data is less than or equal to the threshold TG2, a third-level early warning signal is sent correspondingly.

6. The real-time control system for karst cave pile foundation grouting based on optical fiber sensing and AI dynamic optimization according to claim 2, characterized in that, The analysis of the closed-loop control is as follows: The closed-loop control module realizes the dynamic regulation of grouting parameters based on the Pareto optimal solution set, obtains the real-time adjusted Pareto optimal solution set according to the real-time construction progress, and screens out the initial parameter combination (P0 * , Q0 * , r0 * ) that adapts to the current stage goal from the solution set. The grout diffusion radius Rpred is output through the grout diffusion model and compared with the set interval [Rmin, Rmax]. If the grout diffusion radius Rpred is within the interval, the current parameters are maintained; if the grout diffusion radius Rpred is greater than the maximum value Rmax of the interval, the adjusted pressure value Pnew is obtained according to the formula, and with this adjusted pressure value Pnew as the core, a parameter combination is re-screened from the Pareto optimal solution set; if the grout diffusion radius Rpred is less than the minimum value Rmin of the interval, the adjusted pressure value Pnew is obtained according to the formula, where ρ is the sensitivity coefficient related to the construction stage, and with this adjusted pressure value Pnew as the core, a parameter combination is re-screened from the Pareto optimal solution set; Control the grouting equipment according to the re-matched parameter combination (P1 * , Q1 * , r1 * ), repeat the process of "model prediction Rpred → deviation discrimination → parameter adjustment → solution set screening" until Rpred is stabilized within the interval [Rmin, Rmax], and dynamically adjust the parameters of the grouting equipment through the closed-loop control module to achieve precise control of the slurry diffusion path.

7. The real-time control system for karst cave pile foundation grouting based on optical fiber sensing and AI dynamic optimization according to claim 6, characterized in that, The analysis steps of the adjusted Pareto optimal solution set are as follows: Construct the grouting parameter vector Q is the grouting flow rate, r1, r2, and r3 are the proportions of cement, water, and admixture, and the construction target vector δ is the filling rate, C is the cost, T is the construction period, and through the function Establish a mapping. The f mapping function is obtained based on a neural network. Collect historical data on grouting parameter ratios and corresponding construction targets, perform normalization processing, and use a multi-layer perceptron to establish a non-linear mapping model The input layer of the model is 3-dimensional, specifically the grouting pressure, grouting flow rate, and slurry ratio. The ReLU activation function is used in the hidden layer, and the output layer directly predicts the filling rate, cost, and construction period. The training process minimizes the mean square error loss through the Adam optimizer, combines L2 regularization and early stopping mechanism, and finally obtains the f mapping function; Randomly generate H combinations of grouting parameters to form an initial population. For each parameter combination, calculate the target through the f mapping function. Let i be the serial number of the parameter combination. Stratify according to the dominance relationship. Those with no individual dominance form the first layer of the Pareto front, and divide layer by layer. Calculate the crowding distance of individuals in the target space for each layer to measure the surrounding density. Based on the non-dominated rank and crowding degree, select individuals through the roulette wheel method. Perform crossover (such as simulated binary crossover) and mutation on the selected individuals to generate offspring. Combine the parent generation and the offspring, and reselect to maintain the population size H. Repeat until the maximum number of iterations. All non-dominated solutions in the final population, that is, the individuals in the first layer of the Pareto front, constitute the Pareto optimal solution set. And optimize and adjust the weights for different construction stages, dynamically adjust the weights of the filling rate δ, cost C, and construction period T, determine the stage of construction and match the corresponding weight value range, and obtain the adjusted Pareto optimal solution set for each construction stage.

Images