A method for optimizing the parameters of a deep foundation pit casing jet grouting anchor cable in a dike riprap area

Abstract

The present application relates to a kind of dike riprap area deep foundation pit casing jet grouting anchor supporting parameter optimization method, specifically relates to civil engineering foundation treatment and computer aided design technical field, by real-time acquisition multivariate time series signal and extract high-dimensional feature vector set, combined with agent model and bayesian inference, three-dimensional stratum mechanics parameter probability field is inverted, then by the physical information neural network of fusing physical law optimization local process parameters, the precise adaptive matching to riprap stratum heterogeneity is realized, the safety and reliability of supporting structure are significantly improved, avoid the problem of under-reinforcement or over-reinforcement caused by traditional uniform parameter, reduce material waste and delay risk, while enhancing the robustness of construction process to complex geological conditions.

Description

Technical Field

[0001] This invention relates to the field of civil engineering foundation treatment and computer-aided design technology, and more specifically, to a method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in embankment riprap areas. Background Technology

[0002] Currently, in the construction of deep foundation pits in riprap embankment areas, the riprap is typically formed by artificial filling. Its particle size distribution is wide, gradation varies greatly, and its density and cementation exhibit high spatial randomness and heterogeneity. These unique engineering geological conditions lead to severe spatial variability in ground mechanical parameters such as internal friction angle, cohesion, and deformation modulus within the foundation pit area. Traditional engineering survey methods, including single-hole core drilling and in-situ standard penetration tests, can only obtain discrete point data, making it difficult to fully characterize the true distribution characteristics of ground mechanical parameters throughout the entire foundation pit area. When using casing jet grouting and anchor cable technology to traverse such diverse riprap strata for foundation pit support construction, practical engineering experience often encounters difficulties. Based on the average parameters or worst-case conditions provided in the geological survey report, and combined with the engineers' experience, a set of fixed and uniformly applicable construction process parameters are pre-calculated and determined through finite element numerical simulation, including jet grouting pressure, casing follow-up depth, and anchor cable prestressing tensioning sequence. However, this static parameter design method cannot adapt to the drastic spatial changes in riprap strata. In actual construction, jet grouting piles in some areas often suffer from insufficient pile diameter and low pile strength due to improper parameter matching, while in other areas, excessive diameter enlargement may occur due to overly conservative parameters. This not only causes serious waste of grouting materials but may also lead to environmental safety problems such as surface uplift or settlement of adjacent dike structures.

[0003] It is evident that the core flaw of existing construction methods based on static design parameters lies in their inability to dynamically adjust support parameters based on real-time geological feedback information acquired during construction. When construction machinery traverses rockfill formations of varying properties, drilling data such as rig torque, thrust, casing follow-up resistance, and pump pressure and flow fluctuations during jet grouting contain rich geological information. However, this valuable real-time data is currently not effectively collected, integrated, or used to guide parameter optimization. Therefore, there is an urgent need to develop a technical solution that can fully utilize multi-source sensor data from the construction process to invert the three-dimensional mechanical parameter field of rockfill formations in real time, and on this basis, achieve adaptive matching and dynamic optimization of casing, jet grouting, and anchor cable support parameters. This would fundamentally resolve the contradiction between traditional static design methods and dynamic, complex geological conditions, thereby improving the reliability and economy of the support structure. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method for optimizing the parameters of deep foundation pit casing jet grouting and anchor cable support in embankment riprap areas, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps: Step S1: In response to the construction start command, simultaneously collect multi-dimensional time-series signals from the drilling rig, jet grouting equipment, and anchor cable tensioning equipment, and preprocess and extract structured features from the multi-dimensional time-series signals to generate a high-dimensional feature vector set for the construction process of each anchor cable. Step S2: Based on the training dataset generated by finite element simulation, construct and train a deep neural network as a surrogate model. The surrogate model is used to quickly predict the construction response prediction value under any combination of geological parameters. Step S3: Map the high-dimensional feature vector set obtained in step S1 to the measured values ​​of the construction response, and use the surrogate model obtained in step S2 as the likelihood function calculation tool. Use the Bayesian inference method to inverse the posterior distribution of the stratum parameters at the location of each anchor cable, and then interpolate to construct the three-dimensional stratum mechanical parameter probability field of the entire foundation pit area. Step S4: Input the probability field of three-dimensional stratum mechanical parameters obtained in step S3 into a pre-trained physical information neural network. The physical information neural network integrates the fluid dynamics equations of the jet grouting process with the shear slip constitutive relationship of the anchor cable-soil interface. Through optimization, it outputs the optimal local process parameters for each anchor cable. The optimal local process parameters include the casing follow-up depth, jet grouting pressure and anchor cable prestressing tensioning sequence. In a preferred embodiment, in step S1, in response to the construction start command, the drilling rig's rotational torque, axial thrust, and drilling speed are synchronously acquired at a sampling frequency of not less than 200 Hz. The instantaneous values ​​of the high-pressure jet grouting pump outlet pressure and flow rate, and the casing follow-up resistance of the jet grouting equipment are also synchronously acquired. The three components of the drill rod vibration acceleration and the load-displacement curve during anchor tensioning are also synchronously acquired. All signals are synchronized by a hardware clock to generate a multi-dimensional time-series signal.

[0006] In a preferred embodiment, the preprocessing includes removing outliers using a robust thresholding method based on the absolute deviation of the median and filtering out high-frequency noise using a Butterworth low-pass filter. Structured feature extraction employs variational mode decomposition to adaptively decompose each preprocessed multivariate time-series signal into multiple intrinsic mode functions. Each intrinsic mode function corresponds to a center frequency and an instantaneous amplitude. The number of modes is adaptively determined by the center frequency separation criterion. Then, the ratio of the energy of each mode to the total energy is calculated to form a modal energy ratio vector. Then, the weighted spectral entropy is calculated. The calculation process of the weighted spectral entropy is as follows: for each mode, the ratio of the energy of each mode to the total energy is multiplied by the natural logarithm of the ratio, and then multiplied by the ratio of the modal center frequency to the maximum value of the center frequencies of all modes. Finally, the product of all modes is summed by taking the negative value. Finally, the impact energy index is calculated. The calculation process of the impact energy index is as follows: During a construction period, the second derivative of the processed multivariate time series signal is calculated, and the absolute value of the second derivative is taken. The absolute value of the second derivative is multiplied by the result of the calculation of the hyperbolic tangent function, where the independent variable of the hyperbolic tangent function is the absolute value of the first derivative of the multivariate time series signal minus 0.8 times the root mean square value of the drilling speed. Two adjustment parameters are introduced, which are taken as empirical values ​​of 0.5 and 0.8 respectively. Finally, the product result is integrated over the entire construction period and divided by the duration of the construction period.

[0007] In a preferred embodiment, the process of generating a high-dimensional feature vector set for each anchor cable construction process specifically involves: For the construction process of each anchor cable, features of all corresponding multivariate time series signals are extracted in the order of variational mode decomposition, modal energy ratio vector, weighted spectral entropy and impact energy index to obtain the modal energy ratio vector, weighted spectral entropy and impact energy index of each multivariate time series signal; Then, the modal energy ratio vector, weighted spectral entropy, and impact energy index of all multivariate time-series signals of the same anchor cable are concatenated into a high-dimensional feature vector. The dimension of the high-dimensional feature vector is equal to the sum of the total number of multivariate time-series signals multiplied by the number of modes and two. The total number of multivariate time-series signals includes ten signals in total: drilling rig rotation torque, axial thrust, drilling speed, casing follow-up resistance, instantaneous values ​​of high-pressure jet pump outlet pressure and flow rate, load-displacement curve during anchor cable tensioning, and drill pipe vibration acceleration. The high-dimensional feature vectors of all anchor cables together constitute a high-dimensional feature vector set.

[0008] In a preferred embodiment, the process of generating the training dataset in step S2 is as follows: The four formation parameters of the riprap strata are determined as follows: internal friction angle, cohesion, deformation modulus and Poisson's ratio. The range of values ​​for each parameter is set based on engineering survey data and uniform distribution is adopted as the a priori distribution. The Latin hypercube sampling method was used to generate 8,000 combinations of formation parameters from the four-dimensional parameter space. Each combination of formation parameters was input into the finite element simulation to simulate the entire process of casing follow-up and high-pressure jet grouting. The corresponding construction response simulation values ​​were output. The construction response simulation values ​​include five indicators: average drilling torque, maximum axial thrust, average jet grouting pump pressure, standard deviation of pump pressure, and pile diameter. In this way, a training dataset between formation parameters and construction response simulation values ​​was constructed. The calculation process for the standard deviation of pump pressure is as follows: For the outlet pressure of the high-pressure jet grouting pump during the construction period corresponding to each combination of formation parameters, calculate the variance of all sampling points, that is, sum the squares of the differences between the values ​​of each sampling point and the average value, divide by the total number of sampling points minus one, and then perform a square root operation on the variance.

[0009] In a preferred embodiment, the specific operation of constructing and training a deep neural network as a surrogate model is as follows: The proxy model is implemented using a deep neural network, which adopts a multilayer perceptron architecture, including one input layer, three hidden layers, and one output layer. The input layer has four neurons corresponding to the four stratum parameters, each hidden layer contains 128 neurons, and the output layer has five neurons, which output five construction response prediction values ​​respectively. The activation function of the hidden layer is the Swish function, and a batch normalization layer is added after each hidden layer, and Dropout regularization with a dropout rate of 0.2 is introduced. The training loss function of the proxy model adopts the adaptive weighted multi-task Huber loss. The specific calculation process of Huber loss is as follows: for the prediction residual between a single construction response prediction value and the corresponding construction response simulation value, the relationship between the absolute value of the prediction residual and 1 is judged. If the absolute value of the prediction residual is less than 1, then 0.5 multiplied by the square of the prediction residual is taken as the value of Huber loss. If the absolute value of the prediction residual is greater than or equal to 1, then the absolute value of the prediction residual minus 0.5 is taken as the value of Huber loss. The specific calculation process of the total loss function is as follows: For each of the five indicators of the construction response prediction value, firstly, obtain the value of the Huber loss corresponding to that indicator. Then, obtain the learnable noise parameter that corresponds to each indicator. Calculate the value of the Huber loss by dividing it by 2 and then by the square of the learnable noise parameter. Then, add the natural logarithm of the learnable noise parameter to this value. Sum all the values ​​calculated for the five indicators to form the value of the total loss function. During the training process, the Adam optimizer is used, with the initial learning rate set to 10 to the power of negative cube. Every twenty rounds, the learning rate is reduced to 0.9 times the original rate. The early termination condition is that the validation set loss no longer decreases. After training converges, the surrogate model is obtained.

[0010] In a preferred embodiment, the specific operation of mapping the high-dimensional feature vector set obtained in step S1 to the measured values ​​of the construction response in step S3 is as follows: A shallow regression network is pre-trained. The input of the shallow regression network is the high-dimensional feature vector of each anchor cable, and the output is the actual construction response value with the same dimension as the construction response prediction value in step S2. The actual construction response value includes five indicators: average drilling torque, maximum axial thrust, average jet grouting pump pressure, standard deviation of pump pressure, and pile diameter. The training data for the shallow regression network comes from a small number of field calibration experiments. In the calibration experiments, high-dimensional feature vectors and corresponding measured construction response values ​​are recorded simultaneously. After mapping, a set of measured construction response values ​​for all anchor cables is obtained.

[0011] In a preferred embodiment, the process of using Bayesian inference to inversely deduce the posterior distribution of formation parameters at each anchor cable location specifically involves: The surrogate model obtained in step S2 is used as the likelihood function calculation tool. The likelihood function is constructed as follows: the probability density of the five construction response prediction values ​​is calculated separately and then multiplied together to obtain the joint likelihood probability. The probability density function of each construction response prediction value adopts the Laplace distribution. The location parameter of the Laplace distribution is the construction response prediction value, and the scale parameter of the Laplace distribution is the estimated value obtained by calculating the median absolute deviation of the measured construction response value. The sampling of the posterior distribution adopts the No-U-TurnSampler algorithm, which extracts posterior sampling samples from the posterior distribution. After convergence diagnosis, the mean of the posterior sampling samples is taken as the maximum posterior estimate of the formation parameters at the anchor cable location. Then, based on the maximum a posteriori estimates and their spatial coordinates at all anchor locations, a three-dimensional probability field of geological mechanical parameters is constructed using Gaussian process regression. The mean function of the Gaussian process regression is a constant, and the covariance function uses the Matérn kernel. The covariance value of the Matérn kernel is determined by the signal variance, the length scale, and the Euclidean distance between the input points. Specifically, the Euclidean distance between any two spatial coordinates is calculated first, then divided by the length scale, multiplied by the square root of three, and then added by one. The result after adding one is multiplied by the negative square root of the natural constant e, the Euclidean distance divided by the length scale to the power of the natural constant e, and finally the product is multiplied by the signal variance to obtain the covariance value of the Matérn kernel. The signal variance and length scale of the Matérn kernel are obtained by maximizing the marginal likelihood estimation. The Gaussian process regression outputs the predicted mean and predicted variance at any location within the pit area, constituting a complete three-dimensional probability field of geological mechanical parameters.

[0012] In a preferred embodiment, the specific operation of constructing the physical information neural network in step S4 is as follows: Define the input and output parameters of the physical information neural network: The input parameters consist of two parts. The first part is the mean and variance of the formation parameters at each anchor cable location extracted from the three-dimensional formation mechanical parameter probability field obtained in step S3. The second part is the local process parameters to be optimized, which consist of the casing follow-up depth, jet grouting pressure, and anchor cable prestressing tensioning sequence. The output parameter is the predicted support effect vector, which specifically includes the predicted pile diameter and the predicted anchoring force. The loss function of the physical information neural network is composed of a weighted average of the data fitting term, the physical constraint term, and the process feasibility penalty term. The data fitting term measures the deviation between the predicted support effect vector and the preset target values ​​for pile diameter and anchoring force; Physical constraints are embedded in the fluid dynamics equations and shear slip constitutive relations of the anchor cable-soil interface in the jet grouting process. The fluid dynamics equations adopt a radial pressure attenuation model based on turbulent jet theory; the shear slip constitutive relations of the anchor cable-soil interface adopt a hyperbolic shear stress-displacement model. The calculation process for the total value of physical constraint terms is as follows: First, for each simulation sample, the residual square norm of the fluid dynamics equation and the residual square norm of the shear slip constitutive relation of the anchor cable-soil interface are calculated respectively. The two are added together to obtain the physical residual norm. Then, the arithmetic mean of the physical residual norms of all simulation samples is calculated. This arithmetic mean is used as the total value of physical constraint terms. The process feasibility penalty is used to limit local process parameters to within a preset upper limit range. The total value of the penalty is the sum of the squares of the portion of each parameter that exceeds the upper limit.

[0013] In a preferred embodiment, the process of optimizing and outputting the optimal local process parameters for each anchor cable specifically involves: The mean and variance of the formation parameters at each anchor cable location are used as fixed inputs to the physical information neural network, and the local process parameters to be optimized are used as trainable variables. The gradient descent method is used to minimize the loss function of the physical information neural network. In each iteration, the physical information neural network forward propagates to calculate the predicted support effect vector and physical residual norm, and backpropagates to update the local process parameters to be optimized. The local process parameters output after iterative convergence are the optimal casing follow-up depth, optimal jet grouting pressure, and optimal anchor cable prestressing tensioning sequence for the anchor cable. The gradient descent method uses the Adam optimizer, with a learning rate set to the negative square of 10 and an iteration count of 300 steps. In each iteration, the total loss function value of the current iteration step is calculated, and it is determined whether the absolute value of the difference between the total loss function value of the current iteration step and the total loss function value of the previous iteration step is less than 10 to the power of negative 4. If it is less, the iteration is determined to be converged and the optimization process is terminated; otherwise, the next iteration is continued. The total loss function is calculated as follows: First, the data fitting term, physical constraint term, and process feasibility penalty term are weighted and summed, where the weight of the data fitting term is 1, the weight of the physical constraint term is 0.1, and the weight of the process feasibility penalty term is 0.01. The specific method of weighted summation is as follows: multiply the value of the data fitting term by 1, multiply the value of the physical constraint term by 0.1, multiply the value of the process feasibility penalty term by 0.01, and then add the three weighted values ​​together to form the value of the total loss function.

[0014] The beneficial effects of this invention are as follows: by real-time acquisition of multivariate time-series signals and extraction of high-dimensional feature vector sets, combined with surrogate models and Bayesian inference to inversely derive the probability field of three-dimensional stratum mechanical parameters, and then by using a physical information neural network that integrates physical laws to optimize local process parameters, accurate adaptive matching of the heterogeneity of riprap strata is achieved, which significantly improves the safety and reliability of the support structure, avoids the under-reinforcement or over-reinforcement problems caused by traditional uniform parameters, reduces the risk of material waste and construction delays, and enhances the robustness of the construction process to complex geological conditions. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1: This example provides the following... Figure 1 The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes, as shown, specifically includes the following steps: Step S1: In response to the construction start command, simultaneously collect multi-dimensional time-series signals from the drilling rig, jet grouting equipment, and anchor cable tensioning equipment, and preprocess and extract structured features from the multi-dimensional time-series signals to generate a high-dimensional feature vector set for the construction process of each anchor cable. Step S2: Based on the training dataset generated by finite element simulation, construct and train a deep neural network as a surrogate model. The surrogate model is used to quickly predict the construction response prediction value under any combination of geological parameters. Step S3: Map the high-dimensional feature vector set obtained in step S1 to the measured values ​​of the construction response, and use the surrogate model obtained in step S2 as the likelihood function calculation tool. Use the Bayesian inference method to inverse the posterior distribution of the stratum parameters at the location of each anchor cable, and then interpolate to construct the three-dimensional stratum mechanical parameter probability field of the entire foundation pit area. Step S4: Input the probability field of three-dimensional stratum mechanical parameters obtained in step S3 into a pre-trained physical information neural network. The physical information neural network integrates the fluid dynamics equations of the jet grouting process with the shear slip constitutive relationship of the anchor cable-soil interface. Through optimization, it outputs the optimal local process parameters for each anchor cable. The optimal local process parameters include the casing follow-up depth, jet grouting pressure, and anchor cable prestressing tensioning sequence.

[0020] In this embodiment, it is specifically necessary to explain that in step S1, in response to the construction start command, the drilling rig's rotational torque, axial thrust, and drilling speed are synchronously collected at a sampling frequency of not less than 200 Hz. The instantaneous values ​​of the high-pressure jet grouting pump outlet pressure and flow rate, and the casing follow-up resistance of the jet grouting equipment are also synchronously collected. The three components of the drill rod vibration acceleration and the load-displacement curve during anchor tensioning are also synchronously collected. All signals are synchronized through a hardware clock to ensure time alignment and generate multi-dimensional time-series signals. The sampling frequency is set to no less than 200 Hz because the energy of the impact signal generated by the hard particles hitting the drill bit in the rock-filled strata is mainly concentrated in the 10 Hz to 150 Hz frequency band. Using a 200 Hz sampling frequency can ensure the requirements of the Nyquist sampling theorem and completely preserve the high-frequency components of the impact signal without aliasing. The hardware clock synchronization adopts a wired synchronization scheme based on the IEEE 1588 precision time protocol. Each sensor node is connected to the data acquisition master station through a dedicated synchronization cable. The synchronization accuracy is better than 100 microseconds, ensuring that the time alignment error between each component of the multivariate time sequence signal does not exceed one sampling interval, and providing an accurate time domain correspondence for subsequent variational mode decomposition. Preprocessing includes removing outliers using a robust thresholding method based on the absolute deviation of the median and filtering out high-frequency noise using a Butterworth low-pass filter; The Butterworth low-pass filter has a cutoff frequency of 50 Hz, which can effectively suppress mechanical noise such as motor rotation and hydraulic pump pulsation, while retaining low-frequency effective signals related to ground breaking and jet grouting. The specific implementation process of the robust thresholding method based on the absolute deviation of the median is as follows: calculate the median of each signal within the sliding window, then calculate the absolute deviation of each sampling point from the median, identify sampling points with an absolute deviation greater than three times the absolute deviation of the median as outliers, and replace the value of the outlier with the median. Structured feature extraction employs variational mode decomposition to adaptively decompose each preprocessed multivariate time series signal into multiple intrinsic mode functions (IMFs). Each IMF corresponds to a center frequency and instantaneous amplitude. The number of modes is adaptively determined by a center frequency separation criterion. The specific determination method of the center frequency separation criterion is as follows: In each iteration of variational mode decomposition, the center frequencies of all currently decomposed modes are calculated. When the difference between the center frequency of the newly added mode and the minimum center frequency of the existing modes is less than 5% of the average center frequency of all modes, the decomposition is stopped. At this time, the number of modes is the final determined value of K. K is a positive integer not less than 2 and not greater than 10, with a typical value range of 3 to 8. The specific value depends on the complexity of the multivariate time series signal and the degree of formation fragmentation. Subsequently, the ratio of the energy of each mode to the total energy is calculated to form a mode energy ratio vector, which reflects the energy distribution of the multivariate time series signal in different frequency bands. Then, the weighted spectral entropy is calculated. The calculation process of the weighted spectral entropy is as follows: for each mode, the ratio of the energy of each mode to the total energy is multiplied by the natural logarithm of the ratio, and then multiplied by the ratio of the mode center frequency to the maximum value of the center frequencies of all modes. Finally, the product of all modes is negative and summed. This is used to quantify the irregularity and complexity of the multivariate time series signal, so as to be more sensitive to the impact characteristics of hard particles hitting the drill bit in rock-filled formations. Finally, the impact energy index is calculated. The calculation process is as follows: During a construction period, the second derivative of the processed multivariate time series signal is calculated, and the absolute value of the second derivative is taken. This absolute value is then multiplied by the result of a hyperbolic tangent function, where the independent variable of the hyperbolic tangent function is the absolute value of the first derivative of the multivariate time series signal minus 0.8 times the root mean square value of the drilling speed. Two adjustment parameters are introduced, with empirical values ​​of 0.5 and 0.8 respectively. Finally, the product is integrated over the entire construction period and divided by the duration of the construction period to highlight the strong impact event. The root mean square value of the drilling speed refers to the root mean square value of the drilling speed signal during the construction period. Of the two adjustment parameters, 0.5 is used to control the steepness of the hyperbolic tangent function, and 0.8 is used to set the threshold for determining a sudden change in velocity. That is, when the absolute value of the first derivative of the multivariate time series signal exceeds 0.8 times the root mean square value of the drilling velocity, the output value of the hyperbolic tangent function approaches one, so that the impact energy index is given a near-full weight for this sudden change moment; conversely, when the absolute value of the first derivative is much lower than this threshold, the output value of the hyperbolic tangent function approaches zero, and the impact energy index contributes very little to this stable moment. The above parameter values ​​have been verified through a large number of field calibration tests and are suitable for typical working conditions in rockfill formations. The root mean square value of drilling speed is calculated as follows: the average value of all sampling points of drilling speed signal during the construction period is calculated by squaring them, and then the square root is taken to obtain the root mean square value of that period, which is used to measure the degree of normalization of speed change. The process of generating a high-dimensional feature vector set for each anchor cable construction process is as follows: For the construction process of each anchor cable, features are extracted from all corresponding multivariate time-series signals in the order of variational mode decomposition, modal energy ratio vector, weighted spectral entropy, and impact energy index, resulting in the modal energy ratio vector, weighted spectral entropy, and impact energy index for each multivariate time-series signal. The modal energy ratio vector consists of the ratio of the energy of each mode to the total energy in multiple dimensionless quantities, and the sum of the ratios of the energy of each mode to the total energy is one. The weighted spectral entropy is a dimensionless scalar with a value between zero and the natural logarithm of the number of modes. The impact energy index is a dimensionless scalar with a value greater than zero. The calculation of the impact energy index is based on the comparison between the absolute value of the first derivative of the multivariate time-series signal and the root mean square value of 0.8 times the drilling speed, and the weights are dynamically adjusted using a hyperbolic tangent function to assign greater weights at high-speed abrupt changes. Then, the modal energy ratio vector, weighted spectral entropy, and impact energy index of all multivariate time-series signals of the same anchor cable are concatenated into a high-dimensional feature vector. The dimension of the high-dimensional feature vector is equal to the sum of the total number of multivariate time-series signals multiplied by the number of modes and two. The total number of multivariate time-series signals includes ten signals in total: drilling rig rotation torque, axial thrust, drilling speed, casing follow-up resistance, instantaneous values ​​of high-pressure jet pump outlet pressure and flow rate, load-displacement curve during anchor cable tensioning, and drill pipe vibration acceleration. The high-dimensional feature vectors of all anchor cables together constitute a high-dimensional feature vector set, which is passed as the output of step S1 to step S3. In the formula for calculating the dimension of a high-dimensional feature vector, "two" represents the two scalar features obtained after feature extraction for each multivariate time series signal—the weighted spectral entropy and the impact energy index. K represents the number of modes adaptively determined by the center frequency separation criterion, and K is a positive integer not less than 2 and not greater than 10. Therefore, the total number of features for each multivariate time series signal is K plus two, and the total number of features for all ten multivariate time series signals is ten multiplied by K plus two, which constitutes the dimension of the high-dimensional feature vector. The specific operation of "concatenating the modal energy ratio vector, weighted spectral entropy, and impact energy index of all multivariate time-series signals of the same anchor cable into a high-dimensional feature vector" is as follows: following a fixed order of drilling rig rotation torque, axial thrust, drilling speed, casing follow-up resistance, high-pressure jet grouting pump outlet pressure, high-pressure jet grouting pump instantaneous flow rate, load-displacement curve during anchor cable tensioning, drill pipe vibration acceleration X component, drill pipe vibration acceleration Y component, and drill pipe vibration acceleration Z component, the modal energy ratio vector (length K), weighted spectral entropy (a scalar), and impact energy index (a scalar) of each signal are concatenated end to end to form a one-dimensional vector with a length of 10 times K plus 2. The modal energy ratio vector consists of the ratio of the energy of each of the K dimensionless modes to the total energy. The value of each ratio ranges from zero to one, and the sum of all K ratios is always equal to one. This modal energy ratio vector reflects the distribution ratio of signal energy in different frequency bands and is a key feature characterizing the changes in formation hardness. The weighted spectral entropy ranges between zero and the natural logarithm of K. Since K has a minimum of 2 and a maximum of 10, the weighted spectral entropy ranges between zero and ln(10)≈2.3026. The weighted spectral entropy reaches its maximum value when the signal energy is evenly distributed across all K modes, and approaches zero when the signal energy is concentrated in a single mode. This characteristic is particularly sensitive to broadband impact noise generated by hard particles in riprap strata. The impact energy index is a dimensionless scalar with a value greater than zero. Its magnitude directly reflects the intensity of encountering large boulders or hard interlayers during construction. The higher the value, the more severe the impact event. It can be used to identify high-risk boreholes that require special adjustment of jet grouting pressure.

[0021] In this embodiment, it is specifically necessary to explain the process of generating the training dataset in step S2 as follows: The four formation parameters for determining the riprap strata include internal friction angle, cohesion, deformation modulus, and Poisson's ratio. The range of values ​​for each parameter is set based on engineering survey data, and a uniform distribution is adopted as the a priori distribution. For example, the unit of internal friction angle is degrees, and the value range is between 15 and 45 degrees; the unit of cohesion is kilopascals, and the value range is between 5 and 50; the unit of deformation modulus is megapascals, and the value range is between 10 and 100; and Poisson's ratio is a dimensionless parameter, and the value range is between 0.2 and 0.4. Eight thousand combinations of formation parameters were generated from a four-dimensional parameter space using the Latin hypercube sampling method. Each combination was then input into a finite element simulation to simulate the entire process of casing advance and high-pressure jet grouting. The simulation output corresponding construction response values, which included five indicators: average drilling rig torque, maximum axial thrust, average jet grouting pump pressure, standard deviation of pump pressure, and pile diameter. This data was used to construct a training dataset between the formation parameters and the construction response simulation values. During the generation of the training dataset, for the standard deviation of pump pressure among the five indicators, a method was first used to calculate the standard deviation for each pressure... The standard deviation of pump pressure is obtained by taking the variance of the sampling points and then performing a square root operation. The specific calculation process of the standard deviation of pump pressure is as follows: For the outlet pressure of the high-pressure jet grouting pump during the construction period corresponding to each set of formation parameters, firstly, calculate the variance of all sampling points in the outlet pressure of the high-pressure jet grouting pump. That is, sum the squares of the differences between the values ​​of each sampling point in the outlet pressure of the high-pressure jet grouting pump and the average value of the outlet pressure of the high-pressure jet grouting pump. Then, divide the summation result by the total number of sampling points minus one. Finally, perform a square root operation on the obtained variance value. The result is the standard deviation of pump pressure corresponding to the set of formation parameters. Finite element simulation was performed using ABAQUS finite element analysis software. The formation model used eight-node hexahedral reduced integral elements (C3D8R), with the mesh size controlled between 0.1 and 0.3 meters. Local refinement was applied in the casing follow-up and jet grouting areas, with a mesh size of 0.05 meters in the refined areas. The formation constitutive model adopted the Mohr-Coulomb ideal elastoplastic model, and the contact interface between the riprap and the casing adopted a penalty function frictional contact with a friction coefficient of 0.3. The jet grouting process was simulated by applying a time-varying pressure load, and the pressure curve was obtained by fitting field measured data. The specific implementation of Latin hypercube sampling is as follows: using the LatinHypercube class in the Python scientific computing library scipy.stats.qmc, setting the seed value to 42 to ensure repeatability, generating 8,000 sets of four-dimensional samples, and then mapping the uniformly distributed samples to the specified value range of each parameter through the inverse transformation method; the unbiased estimate (divided by the total number of sampling points minus one) used in the calculation of pump pressure standard deviation is to eliminate the bias of sample variance, making the estimated value closer to the population variance, and thus more accurately reflecting the pressure fluctuation degree of the jet jet process; The specific steps for constructing and training a deep neural network as a surrogate model are as follows: The surrogate model is implemented using a deep neural network, which adopts a multilayer perceptron architecture, including one input layer, three hidden layers, and one output layer. The input layer has four neurons corresponding to the four stratum parameters, each hidden layer contains 128 neurons, and the output layer has five neurons, which output five construction response prediction values. The activation function of the hidden layers is the Swish function, and a batch normalization layer is added after each hidden layer. Dropout regularization with a dropout rate of 0.2 is introduced to prevent overfitting. The training loss function of the proxy model adopts the adaptive weighted multi-task Huber loss. The specific calculation process of Huber loss is as follows: for the prediction residual between a single construction response prediction value and the corresponding construction response simulation value, the relationship between the absolute value of the prediction residual and 1 is judged. If the absolute value of the prediction residual is less than 1, then 0.5 multiplied by the square of the prediction residual is taken as the value of Huber loss. If the absolute value of the prediction residual is greater than or equal to 1, then the absolute value of the prediction residual minus 0.5 is taken as the value of Huber loss. Based on this, the specific calculation process of the total loss function is as follows: For each of the five indicators of the construction response prediction value, firstly, obtain the value of the Huber loss corresponding to that indicator; then, obtain the learnable noise parameter corresponding to that indicator; calculate the value of the Huber loss divided by 2 and then divided by the square of the learnable noise parameter; then, add the natural logarithm of the learnable noise parameter to this value; sum all the values ​​calculated for the five indicators to form the value of the total loss function; the learnable noise parameter is automatically adjusted by the surrogate model during training, so that the surrogate model prioritizes fitting tasks with lower uncertainty; the Adam optimizer is used during training, with the initial learning rate set to 10 to the power of negative cube, decaying to 0.9 times the original rate every 20 rounds, and the early stopping termination condition is used when the validation set loss no longer decreases; after training convergence, the surrogate model is obtained, which can complete a positive prediction within milliseconds, and the prediction error is controlled within 5% of the construction response simulation value; The specific expression of the Swish activation function is: Swish(x) equals x multiplied by Sigmoid(x), where Sigmoid(x) equals one divided by one plus e raised to the power of negative x. The Swish function is non-monotonic on the negative half-axis, which helps alleviate the gradient vanishing problem and improves the model's expressive power. The specific operation of the batch normalization layer is: the output of each hidden layer is standardized to make its mean zero and variance one, and then restored by a learnable scaling factor and translation factor to accelerate training convergence and reduce internal covariate bias. Dropout regularization randomly drops hidden layer neurons with a probability of 0.2 during training, forcing the network to learn redundant features and enhancing generalization ability. The initial value of the learnable noise parameter is set to one, and during training, it is adjusted by the gradient. Automatic update descent, in its physical sense, refers to the logarithmic scale of uncertainty in each output dimension. The network automatically reduces the weights of high-noise (high-uncertainty) tasks, thereby improving overall prediction robustness. The early stopping criterion for validation set loss no longer decreasing is: if the validation set loss does not fall below the historical minimum value for ten consecutive rounds, training is stopped to prevent overfitting. The prediction error is measured as mean absolute percentage error, which is calculated as the percentage of the absolute difference between the predicted and simulated values ​​for each of the five output metrics relative to the simulated value, and then the average of the five metrics is taken. This average is controlled to be within 5%. The hyperparameters of the Adam optimizer are set as follows: the first-order moment decay coefficient beta1 equals 0.9, the second-order moment decay coefficient beta2 equals 0.999, and the numerical stability constant epsilon equals 10 to the power of negative 8.

[0022] In this embodiment, it is specifically necessary to explain the following steps in step S3: mapping the high-dimensional feature vector set obtained in step S1 to the measured values ​​of the construction response. A shallow regression network is pre-trained. The input of the shallow regression network is the high-dimensional feature vector of each anchor cable, and the output is the actual construction response value with the same dimension as the construction response prediction value in step S2. The actual construction response value includes five indicators: average drilling torque, maximum axial thrust, average jet grouting pump pressure, standard deviation of pump pressure, and pile diameter. The training data for the shallow regression network comes from a small number of field calibration experiments, in which high-dimensional feature vectors and corresponding measured construction response values ​​are recorded simultaneously. After mapping, a set of measured construction response values ​​for all anchor cables is obtained, which serves as the input for subsequent Bayesian inference. The shallow regression network adopts a three-layer fully connected architecture. The number of neurons in the input layer is consistent with the dimension of the high-dimensional feature vector, i.e., ten times K plus two. The hidden layer contains sixty-four neurons, and the activation function is a linear rectified function. The output layer contains five neurons, corresponding to the average torque of the drilling rig, the maximum axial thrust, and the rotational torque, respectively. Five measured values ​​of construction response were used: mean pump pressure, standard deviation of pump pressure, and pile diameter. A Dropout layer with a dropout rate of 0.1 was added after the hidden layer to prevent overfitting. The network training used the mean squared error loss function, the optimizer was Adam, the initial learning rate was set to 10 to the power of negative 4, the training epochs were 500, and the early stopping condition was that the validation set loss did not decrease for 20 consecutive epochs. The number of field calibration experiments was no less than 30 sets, each experiment was conducted under different geological conditions, and the high-dimensional feature vector and the corresponding measured values ​​of construction response were recorded to ensure the generalization ability of the mapping relationship. The process of retrieving the posterior distribution of formation parameters at each anchor cable location using Bayesian inference is as follows: Using the surrogate model obtained in step S2 as the likelihood function calculation tool, the likelihood function is constructed as follows: the probability density of each of the five construction response prediction values ​​is calculated and then multiplied together to obtain the joint likelihood probability. The probability density function of each construction response prediction value adopts a Laplace distribution. The location parameter of the Laplace distribution is the construction response prediction value, and the scale parameter of the Laplace distribution is the estimated value obtained by calculating the median absolute deviation of the measured construction response value. Specifically, for the m-th construction response prediction value of the i-th anchor cable, the calculation process of its likelihood probability density is as follows: first, obtain the absolute value of the difference between the construction response prediction value and the measured construction response value, multiply the absolute value by negative two to obtain the natural exponent, then divide by twice the scale parameter, and finally multiply the calculation result by the scale parameter to obtain the probability density value of the construction response prediction value; multiply the five probability density values ​​of the five construction response prediction values ​​together to construct the comprehensive likelihood function of the i-th anchor cable. The sampling of the posterior distribution adopts the No-U-TurnSampler algorithm, which extracts posterior sampling samples from the posterior distribution. After convergence diagnosis, the mean of the posterior sampling samples is taken as the maximum posterior estimate of the formation parameters at the anchor cable location. At the same time, the covariance matrix of the posterior sampling samples is calculated as an uncertainty measure. Then, based on the maximum a posteriori estimate and spatial coordinates of all anchor locations, a three-dimensional probability field of geological mechanical parameters is constructed using Gaussian process regression. The mean function of the Gaussian process regression is a constant, and the covariance function uses the Matérn kernel. The covariance value of the Matérn kernel is determined by the signal variance, the length scale, and the Euclidean distance between the input points. Specifically, the Euclidean distance between any two spatial coordinates is calculated first, then divided by the length scale, multiplied by the square root of three, and then one is added. The result after adding one is multiplied by the negative square root of the natural constant e, the Euclidean distance divided by the power of the length scale, and finally the product is multiplied by the signal variance to obtain the covariance value of the Matérn kernel. The signal variance and length scale of the Matérn kernel are obtained by maximizing the marginal likelihood estimation. The Gaussian process regression outputs the predicted mean and predicted variance at any location within the pit area, constituting a complete three-dimensional probability field of geological mechanical parameters. The specific calculation method for the scale parameter of the Laplace distribution is as follows: For the m-th predicted construction response value, firstly calculate the difference between the measured construction response values ​​of all anchor cables and the predicted values ​​of the surrogate model, take the absolute value of these differences, and then calculate the median of these absolute values. This median is used as the scale parameter of the Laplace distribution. The specific parameters of the No-U-TurnSampler algorithm are set as follows: the initial step size is 0.1, the target average acceptance rate is 0.8, the maximum tree depth is 10, the total number of iterations is 10,000, the first 2,000 iterations are discarded as a warm-up stage, and the remaining 8,000 iterations are used as posterior sampling samples. The convergence diagnosis uses the Gelman-Rubin statistic. When the Gelman-Rubin statistic of all formation parameters is less than 1.05, the sampling is considered to have converged. In Gaussian process regression, the smoothing parameter of the Matérn kernel is set to 3 / 2. Specifically, the Matérn kernel takes the form: the covariance equals the signal variance multiplied by 1 plus the square root of three Euclidean distance divided by the length scale, then multiplied by the negative square root of three Euclidean distance divided by the length scale power of the natural constant e. The optimization of the signal variance and length scale uses the L-BFGS-B algorithm, with a maximum of 200 iterations and a gradient tolerance set to 1 multiplied by 10 to the power of -5. Gaussian process regression independently models each formation parameter (internal friction angle, cohesion, deformation modulus, Poisson's ratio), and each model outputs the predicted mean and predicted variance of the corresponding parameter. The predicted variance is used to quantify interpolation uncertainty and serves as part of the physical information neural network input in subsequent step S4, used to adjust the confidence weight of the optimization objective.

[0023] In this embodiment, it is specifically necessary to explain the specific operation of constructing the physical information neural network in step S4: Define the input and output parameters of the physical information neural network: The input parameters consist of two parts. The first part is the mean and variance of the formation parameters at each anchor cable location extracted from the three-dimensional formation mechanical parameter probability field obtained in step S3. The second part is the local process parameters to be optimized, which consist of the casing follow-up depth, jet grouting pressure, and anchor cable prestressing tensioning sequence. The output parameter is the predicted support effect vector, which specifically includes the predicted pile diameter and the predicted anchoring force. The loss function of the physical information neural network is composed of a weighted average of the data fitting term, the physical constraint term, and the process feasibility penalty term. The data fitting term measures the deviation between the predicted support effect vector and the preset target values ​​for pile diameter and anchorage force. The specific calculation process is as follows: For the difference between the predicted pile diameter and the preset target value, and the difference between the predicted anchorage force and the preset target value, the difference is defined as a residual. The absolute value of the residual is compared to 5% of the preset design tolerance. If the absolute value of the residual is less than 5% of the preset design tolerance, then 0.5 multiplied by the square of the residual is taken as the Huber loss score for that indicator. The calculation process for the Huber loss component for a single indicator is as follows: First, determine whether the absolute value of the difference between the predicted pile diameter and the preset target pile diameter is less than 5% of the preset design tolerance. If it is less, calculate 0.5 multiplied by the square of the difference. If it is greater than or equal to, calculate the absolute value of the difference minus 0.5 times the preset design tolerance. Perform the same judgment on the difference between the predicted anchoring force and the preset target anchoring force, and sum the Huber loss components calculated from the two indicators to form a data fitting term. The physical constraints are embedded in the fluid dynamics equations and the constitutive relation of the anchor cable-soil interface shear slip in the jet grouting process. The fluid dynamics equations adopt a radial pressure attenuation model based on turbulent jet theory. Specifically, the product of the radial pressure change rate with distance, the grout density and the empirical coefficient, multiplied by the square of the flow rate, and divided by the square of the product of the radial distance and the casing radius is equal to zero. The fluid dynamics equations are used to constrain the attenuation law of the jet grouting pressure along the radial direction, ensuring that the predicted pile diameter is consistent with the pressure, flow rate, and follow-up depth. The constitutive relation of the anchor cable-soil interface shear slip adopts a hyperbolic shear stress-displacement model. Specifically, the product of the interface shear stress minus the product of the shear modulus and the relative displacement, divided by the product of the failure ratio and the normal stress, the interface friction angle tangent, and the shear modulus and the relative displacement divided by the ultimate shear stress, is equal to zero. The constitutive relation of the anchor cable-soil interface shear slip is used to constrain the relationship between the tensioning sequence and the anchoring force. The calculation process for the total value of physical constraint terms is as follows: First, for each simulation sample, the residual square norm of the fluid dynamics equation and the residual square norm of the shear slip constitutive relation of the anchor cable-soil interface are calculated respectively. The two are added together to obtain the physical residual norm. Then, the arithmetic mean of the physical residual norms of all simulation samples is calculated. This arithmetic mean is used as the total value of physical constraint terms. The process feasibility penalty item is used to limit local process parameters within a preset upper limit range. The calculation process of the total value of the process feasibility penalty item is as follows: For each local process parameter to be optimized, it is determined whether its current value exceeds the corresponding preset upper limit threshold. If the value exceeds the preset upper limit threshold, the difference between the value and the preset upper limit threshold is calculated, and the square of the difference is used as the penalty increment. If it does not exceed the threshold, the penalty increment is zero. The penalty increments corresponding to all local process parameters are accumulated to form the total value of the process feasibility penalty item. The upper limits of the local process parameters in the process feasibility penalty item are set as follows: the casing follow-up depth does not exceed 15 meters, the jet grouting pressure does not exceed 40 MPa, and the anchor cable prestressing tensioning time does not exceed 72 hours. The preset target pile diameter is 600 mm, the preset target anchoring force is 300 kN, and the preset design tolerance is 10% of the target pile diameter, i.e., 60 mm. Therefore, 5% of the preset design tolerance is 3 mm. The grout density is 1200 kg / m³, the empirical coefficient is 0.016, the flow rate is 80 liters / minute, and the casing radius is 0.07 m. The shear modulus is 5 MPa, the failure ratio is 0.9, the normal stress is calculated from the deformation modulus and Poisson's ratio in the formation parameters using the elasticity formula, the interface friction angle is 0.8 times the internal friction angle of the formation, and the ultimate shear stress is 200 kPa. In the calculation of the physical residual norm, the square norm of the fluid dynamics equation residual is obtained by calculating the sum of the squares of the differences between the left and right sides of the equation after discrete sampling of the radial distance. The square norm of the residuals of the soil interface shear slip constitutive relation is obtained by calculating the sum of the squares of the differences between the left and right sides of the equation after discrete sampling of the relative displacement; the physical information neural network adopts a fully connected architecture, including an input layer, three hidden layers and one output layer. The number of neurons in the input layer is equal to the sum of the dimensions of the mean and variance of the formation parameters plus the dimensions of the three local process parameters. Each hidden layer contains 128 neurons, and the activation function is the Swish function. The output layer contains two neurons corresponding to the predicted pile diameter and the predicted anchorage force, respectively; the network training uses the Adam optimizer, with a learning rate of 10 to the power of negative cube and 500 training rounds. The training data comes from the construction response simulation values ​​generated by the finite element simulation in step S2 and their corresponding combinations of formation parameters and process parameters; The process of optimizing and solving for the optimal local process parameters of each anchor cable is as follows: The mean and variance of the formation parameters at each anchor cable location are used as fixed inputs to the physical information neural network, and the local process parameters to be optimized are used as trainable variables. The gradient descent method is used to minimize the loss function of the physical information neural network. In each iteration, the physical information neural network forward propagates to calculate the predicted support effect vector and physical residual norm, and backpropagates to update the local process parameters to be optimized. The local process parameters output after iterative convergence are the optimal casing follow-up depth, optimal jet grouting pressure, and optimal anchor cable prestressing tensioning sequence for the anchor cable. The gradient descent method uses the Adam optimizer, with a learning rate set to the negative square of 10 and an iteration count of 300 steps. In each iteration, the total loss function value of the current iteration step is calculated, and it is determined whether the absolute value of the difference between the total loss function value of the current iteration step and the total loss function value of the previous iteration step is less than 10 to the power of negative 4. If it is less, the iteration is determined to be converged and the optimization process is terminated; otherwise, the next iteration is continued. The total loss function is calculated as follows: First, the data fitting term, physical constraint term, and process feasibility penalty term are weighted and summed, where the weight of the data fitting term is 1, the weight of the physical constraint term is 0.1, and the weight of the process feasibility penalty term is 0.01. The specific method of weighted summation is as follows: multiply the value of the data fitting term by 1, multiply the value of the physical constraint term by 0.1, multiply the value of the process feasibility penalty term by 0.01, and then add the three weighted values ​​together to form the value of the total loss function. In gradient descent, the hyperparameters of the Adam optimizer are set as follows: the first-order moment decay coefficient beta1 equals 0.9, the second-order moment decay coefficient beta2 equals 0.999, and the numerical stability constant epsilon equals 10 to the power of -8. The setting of 10 to the power of -4 in the iterative convergence condition is based on the assumption that the optimization has reached a stable state when the change in the total loss function value is less than one ten-thousandth of the initial loss function value. The physical residual norm is calculated by the physical constraint branch within the physical information neural network during each forward propagation. This branch does not participate in backpropagation to update network weights; it is only used to calculate the value of the physical constraint term. The weighted summation time... The reasons for setting the weight of the fitting term to 1, the weight of the physical constraint term to 0.1, and the weight of the process feasibility penalty term to 0.01 are as follows: the data fitting term directly reflects the degree of achievement of the design goal and should occupy a dominant position; the physical constraint term, as a soft constraint, should not have too high a weight to avoid suppressing the data fitting; the process feasibility penalty term is only used to prevent parameters from going out of bounds and has the smallest weight; during the optimization process, each anchor cable is optimized independently without affecting each other, and the final output of the optimal casing follow-up depth, optimal jet grouting pressure, and optimal anchor cable prestressing tensioning sequence is the best parameter combination that enables the anchor cable to achieve the design goal and satisfy physical laws and process constraints under the current geological conditions.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes, characterized in that, Specifically, the steps include the following: Step S1: In response to the construction start command, simultaneously collect multi-dimensional time-series signals from the drilling rig, jet grouting equipment, and anchor cable tensioning equipment, and preprocess and extract structured features from the multi-dimensional time-series signals to generate a high-dimensional feature vector set for the construction process of each anchor cable. Step S2: Based on the training dataset generated by finite element simulation, construct and train a deep neural network as a surrogate model. The surrogate model is used to quickly predict the construction response prediction value under any combination of geological parameters. Step S3: Map the high-dimensional feature vector set obtained in step S1 to the measured values ​​of the construction response, and use the surrogate model obtained in step S2 as the likelihood function calculation tool. Use the Bayesian inference method to inverse the posterior distribution of the stratum parameters at the location of each anchor cable, and then interpolate to construct the three-dimensional stratum mechanical parameter probability field of the entire foundation pit area. Step S4: Input the probability field of three-dimensional stratum mechanical parameters obtained in step S3 into a pre-trained physical information neural network. The physical information neural network integrates the fluid dynamics equations of the jet grouting process with the shear slip constitutive relationship of the anchor cable-soil interface. Through optimization, it outputs the optimal local process parameters for each anchor cable. The optimal local process parameters include the casing follow-up depth, jet grouting pressure, and anchor cable prestressing tensioning sequence.

2. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 1, characterized in that: In step S1, in response to the construction start command, the drilling rig's rotational torque, axial thrust, and drilling speed are synchronously acquired at a sampling frequency of not less than 200 Hz. The instantaneous values ​​of the high-pressure jet grouting pump outlet pressure and flow rate, and the casing follow-up resistance of the jet grouting equipment are also synchronously acquired. The three components of the drill rod vibration acceleration and the load-displacement curve during anchor tensioning are also synchronously acquired. All signals are synchronized through a hardware clock to generate a multi-dimensional time sequence signal.

3. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 2, characterized in that: The preprocessing includes removing outliers using a robust thresholding method based on the absolute deviation of the median and filtering out high-frequency noise using a Butterworth low-pass filter. Structured feature extraction employs variational mode decomposition to adaptively decompose each preprocessed multivariate time-series signal into multiple intrinsic mode functions. Each intrinsic mode function corresponds to a center frequency and an instantaneous amplitude. The number of modes is adaptively determined by the center frequency separation criterion. Then, the ratio of the energy of each mode to the total energy is calculated to form a modal energy ratio vector. Then, the weighted spectral entropy is calculated. The calculation process of the weighted spectral entropy is as follows: for each mode, the ratio of the energy of each mode to the total energy is multiplied by the natural logarithm of the ratio, and then multiplied by the ratio of the modal center frequency to the maximum value of the center frequencies of all modes. Finally, the product of all modes is summed by taking the negative value. Finally, the impact energy index is calculated. The calculation process of the impact energy index is as follows: During a construction period, the second derivative of the processed multivariate time series signal is calculated, and the absolute value of the second derivative is taken. The absolute value of the second derivative is multiplied by the result of the calculation of the hyperbolic tangent function, where the independent variable of the hyperbolic tangent function is the absolute value of the first derivative of the multivariate time series signal minus 0.8 times the root mean square value of the drilling speed. Two adjustment parameters are introduced, which are taken as empirical values ​​of 0.5 and 0.8 respectively. Finally, the product result is integrated over the entire construction period and divided by the duration of the construction period.

4. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 3, characterized in that: The process of generating a high-dimensional feature vector set for each anchor cable construction process is as follows: For the construction process of each anchor cable, features of all corresponding multivariate time series signals are extracted in the order of variational mode decomposition, modal energy ratio vector, weighted spectral entropy and impact energy index to obtain the modal energy ratio vector, weighted spectral entropy and impact energy index of each multivariate time series signal; Then, the modal energy ratio vector, weighted spectral entropy, and impact energy index of all multivariate time-series signals of the same anchor cable are concatenated into a high-dimensional feature vector. The dimension of the high-dimensional feature vector is equal to the sum of the total number of multivariate time-series signals multiplied by the number of modes and two. The total number of multivariate time-series signals includes ten signals in total: drilling rig rotation torque, axial thrust, drilling speed, casing follow-up resistance, instantaneous values ​​of high-pressure jet pump outlet pressure and flow rate, load-displacement curve during anchor cable tensioning, and drill pipe vibration acceleration. The high-dimensional feature vectors of all anchor cables together constitute a high-dimensional feature vector set.

5. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 4, characterized in that: In step S2, the process of generating the training dataset is as follows: The four formation parameters of the riprap strata are determined as follows: internal friction angle, cohesion, deformation modulus and Poisson's ratio. The range of values ​​for each parameter is set based on engineering survey data and uniform distribution is adopted as the a priori distribution. The Latin hypercube sampling method was used to generate 8,000 combinations of formation parameters from the four-dimensional parameter space. Each combination of formation parameters was input into the finite element simulation to simulate the entire process of casing follow-up and high-pressure jet grouting. The corresponding construction response simulation values ​​were output. The construction response simulation values ​​include five indicators: average drilling torque, maximum axial thrust, average jet grouting pump pressure, standard deviation of pump pressure, and pile diameter. In this way, a training dataset between formation parameters and construction response simulation values ​​was constructed. The calculation process for the standard deviation of pump pressure is as follows: For the outlet pressure of the high-pressure jet grouting pump during the construction period corresponding to each combination of formation parameters, calculate the variance of all sampling points, that is, sum the squares of the differences between the values ​​of each sampling point and the average value, divide by the total number of sampling points minus one, and then perform a square root operation on the variance.

6. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 5, characterized in that: The specific steps for constructing and training a deep neural network as a proxy model are as follows: The proxy model is implemented using a deep neural network, which adopts a multilayer perceptron architecture, including one input layer, three hidden layers, and one output layer. The input layer has four neurons corresponding to the four stratum parameters, each hidden layer contains 128 neurons, and the output layer has five neurons, which output five construction response prediction values ​​respectively. The activation function of the hidden layer is the Swish function, and a batch normalization layer is added after each hidden layer, and Dropout regularization with a dropout rate of 0.2 is introduced. The training loss function of the proxy model adopts the adaptive weighted multi-task Huber loss. The specific calculation process of Huber loss is as follows: for the prediction residual between a single construction response prediction value and the corresponding construction response simulation value, the relationship between the absolute value of the prediction residual and 1 is judged. If the absolute value of the prediction residual is less than 1, then 0.5 multiplied by the square of the prediction residual is taken as the value of Huber loss. If the absolute value of the prediction residual is greater than or equal to 1, then the absolute value of the prediction residual minus 0.5 is taken as the value of Huber loss. The specific calculation process of the total loss function is as follows: For each of the five indicators of the construction response prediction value, firstly obtain the value of the Huber loss corresponding to the indicator, then obtain the learnable noise parameter that corresponds to the indicator, calculate the value of the Huber loss divided by 2 and then divided by the square of the learnable noise parameter, then add the natural logarithm of the learnable noise parameter to the value, and sum all the values ​​calculated for the five indicators to form the value of the total loss function. The Adam optimizer was used during training, with an initial learning rate of 10 to the power of negative cube. The learning rate was reduced to 0.9 times every 20 rounds, and early stopping was used as the termination condition when the loss on the validation set no longer decreased. The surrogate model was obtained after training converged.

7. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 6, characterized in that: In step S3, the specific operation of mapping the high-dimensional feature vector set obtained in step S1 to the measured values ​​of the construction response is as follows: A shallow regression network is pre-trained. The input of the shallow regression network is the high-dimensional feature vector of each anchor cable, and the output is the actual construction response value with the same dimension as the construction response prediction value in step S2. The actual construction response value includes five indicators: average drilling torque, maximum axial thrust, average jet grouting pump pressure, standard deviation of pump pressure, and pile diameter. The training data for the shallow regression network comes from a small number of field calibration experiments. In the calibration experiments, high-dimensional feature vectors and corresponding measured construction response values ​​are recorded simultaneously. After mapping, a set of measured construction response values ​​for all anchor cables is obtained.

8. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 7, characterized in that: The process of using Bayesian inference to inversely derive the posterior distribution of formation parameters at each anchor cable location is as follows: The surrogate model obtained in step S2 is used as the likelihood function calculation tool. The likelihood function is constructed as follows: the probability density of the five construction response prediction values ​​is calculated separately and then multiplied together to obtain the joint likelihood probability. The probability density function of each construction response prediction value adopts the Laplace distribution. The location parameter of the Laplace distribution is the construction response prediction value, and the scale parameter of the Laplace distribution is the estimated value obtained by calculating the median absolute deviation of the measured construction response value. The sampling of the posterior distribution adopts the No-U-TurnSampler algorithm, which extracts posterior sampling samples from the posterior distribution. After convergence diagnosis, the mean of the posterior sampling samples is taken as the maximum posterior estimate of the formation parameters at the anchor cable location. Then, based on the maximum a posteriori estimates and their spatial coordinates at all anchor locations, a three-dimensional probability field of geological mechanical parameters is constructed using Gaussian process regression. The mean function of the Gaussian process regression is a constant, and the covariance function uses the Matérn kernel. The covariance value of the Matérn kernel is determined by the signal variance, the length scale, and the Euclidean distance between the input points. Specifically, the Euclidean distance between any two spatial coordinates is calculated first, then divided by the length scale, multiplied by the square root of three, and then added by one. The result after adding one is multiplied by the negative square root of the natural constant e, the Euclidean distance divided by the length scale to the power of the natural constant e, and finally the product is multiplied by the signal variance to obtain the covariance value of the Matérn kernel. The signal variance and length scale of the Matérn kernel are obtained by maximizing the marginal likelihood estimation. The Gaussian process regression outputs the predicted mean and predicted variance at any location within the pit area, constituting a complete three-dimensional probability field of geological mechanical parameters.

9. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 8, characterized in that: In step S4, the specific operation of constructing the physical information neural network is as follows: Define the input and output parameters of the physical information neural network: The input parameters consist of two parts. The first part is the mean and variance of the formation parameters at each anchor cable location extracted from the three-dimensional formation mechanical parameter probability field obtained in step S3. The second part is the local process parameters to be optimized, which consist of the casing follow-up depth, jet grouting pressure, and anchor cable prestressing tensioning sequence. The output parameter is the predicted support effect vector, which specifically includes the predicted pile diameter and the predicted anchoring force. The loss function of the physical information neural network is composed of a weighted average of the data fitting term, the physical constraint term, and the process feasibility penalty term. The data fitting term measures the deviation between the predicted support effect vector and the preset target values ​​for pile diameter and anchoring force; Physical constraints are embedded in the fluid dynamics equations and shear slip constitutive relations of the anchor cable-soil interface in the jet grouting process. The fluid dynamics equations adopt a radial pressure attenuation model based on turbulent jet theory; the shear slip constitutive relations of the anchor cable-soil interface adopt a hyperbolic shear stress-displacement model. The calculation process for the total value of physical constraint terms is as follows: First, for each simulation sample, the residual square norm of the fluid dynamics equation and the residual square norm of the shear slip constitutive relation of the anchor cable-soil interface are calculated respectively. The two are added together to obtain the physical residual norm. Then, the arithmetic mean of the physical residual norms of all simulation samples is calculated. This arithmetic mean is used as the total value of physical constraint terms. The process feasibility penalty is used to limit local process parameters to within a preset upper limit range. The total value of the penalty is the sum of the squares of the portion of each parameter that exceeds the upper limit.

10. The method for optimizing parameters of deep foundation pit casing jet grouting and anchor cable support in riprap areas of dikes according to claim 9, characterized in that: The process of optimizing and outputting the optimal local process parameters for each anchor cable is as follows: The mean and variance of the formation parameters at each anchor cable location are used as fixed inputs to the physical information neural network, and the local process parameters to be optimized are used as trainable variables. The gradient descent method is used to minimize the loss function of the physical information neural network. In each iteration, the physical information neural network forward propagates to calculate the predicted support effect vector and physical residual norm, and backpropagates to update the local process parameters to be optimized. The local process parameters output after iterative convergence are the optimal casing follow-up depth, optimal jet grouting pressure, and optimal anchor cable prestressing tensioning sequence for the anchor cable. The gradient descent method uses the Adam optimizer, with a learning rate set to the negative square of 10 and an iteration count of 300 steps. In each iteration, the total loss function value of the current iteration step is calculated, and it is determined whether the absolute value of the difference between the total loss function value of the current iteration step and the total loss function value of the previous iteration step is less than 10 to the power of negative 4. If it is less, the iteration is determined to be converged and the optimization process is terminated; otherwise, the next iteration is continued. The total loss function is calculated as follows: First, the data fitting term, physical constraint term, and process feasibility penalty term are weighted and summed, where the weight of the data fitting term is 1, the weight of the physical constraint term is 0.1, and the weight of the process feasibility penalty term is 0.

01. The specific method of weighted summation is as follows: multiply the value of the data fitting term by 1, multiply the value of the physical constraint term by 0.1, multiply the value of the process feasibility penalty term by 0.01, and then add the three weighted values ​​together to form the value of the total loss function.

Images