A deep learning-based intelligent optimization method for grouting parameters in a coastal multi-coupling corrosion environment

Abstract

The application discloses a kind of based on deep learning's intelligent optimization method of grouting parameter under coastal multiple coupling corrosion environment, the method first obtains including corrosion medium concentration and the environmental characteristic data of soil void, subsequently constructs and trains deep neural network model;Establish the nonlinear mapping relationship between environmental characteristics and predicted expansion rate, predicted fluidity;On this basis, in combination with the void volume parameter of target grouting area and the maximum allowable stress threshold, using the particle swarm optimization algorithm with constraint, calculate and output the grouting control parameter including target micro-expansion time and target expansion rate value, finally it is output to grouting equipment control terminal to set material mix proportion and pumping parameter.The application realizes the accurate calculation and dynamic configuration of grouting parameter under multivariate coupling environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction and grouting control technology for underground engineering, and more specifically, to an intelligent optimization method for grouting parameters in a coastal multi-coupling corrosion environment based on deep learning. Background Technology

[0002] In the construction and maintenance of coastal underground spaces, the strata in the target area typically contain coastal swamp facies sediments. The groundwater and soil contain corrosive media such as chloride ions, sulfates, and organic matter at specific concentrations. When performing underground grouting operations in this environment with multiple coupled corrosion characteristics, the performance parameters of the grouting material, such as its expansion rate, fluidity, and micro-expansion time, are influenced by the combined effects of various corrosive media concentration variables and soil porosity characteristics, exhibiting a nonlinear mapping relationship.

[0003] Conventional grouting parameter settings mostly employ static lookup tables or univariate linear fitting models for estimation and configuration. However, when facing complex environments with multivariate coupling, these basic analytical methods have limitations in characterizing the nonlinear relationship between the environmental characteristic input matrix and the grouting material performance output matrix. This leads to some objective problems in practical engineering applications. When the system underestimates fluid resistance or overestimates expansion rate, the actual grouting volume will be smaller than the target void volume. Furthermore, when the expansion stress in the coupled environment is not accurately calculated, the expansion stress generated by the grouting material in the sealed soil risks exceeding the maximum allowable stress threshold of the surrounding structure. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based intelligent optimization method for grouting parameters in a multi-coupled corrosion environment in coastal areas. This method aims to solve the technical problem that existing conventional setting methods are unable to accurately characterize the nonlinear mapping relationship between environmental characteristics and grouting material performance in a multi-variable coupled environment, which leads to deviations in grouting volume and expansion stress exceeding the threshold of surrounding structures in actual engineering projects.

[0005] To achieve the above objectives, this invention proposes a deep learning-based intelligent optimization method for grouting parameters in a coastal multi-coupling corrosion environment, comprising: Environmental characteristic data of the target area is acquired, which is composed of corrosive medium concentration data and soil porosity data; wherein, the corrosive medium concentration data includes the concentration or mass fraction data of chloride ions, sulfates and organic matter; the soil porosity data includes porosity, effective porosity and permeability coefficient; A deep neural network model is constructed and trained, the deep neural network model including an input layer, at least one hidden layer and an output layer; the input layer nodes of the deep neural network model are configured to receive the environmental feature data, and the output layer nodes are configured to output the performance parameters of the grouting material, the performance parameters including the predicted expansion rate and the predicted flowability; the trained deep neural network model is used as a nonlinear mapping model between the environmental features and the performance parameters of the grouting material. The current environmental characteristic data is obtained and input into the nonlinear mapping model to obtain the corresponding predicted expansion rate and predicted mobility parameters. The void volume parameters of the target grouting area and the maximum allowable stress threshold of the surrounding existing structure are obtained. A constrained particle swarm optimization algorithm is used to calculate and output the corresponding grouting control parameters based on the predicted expansion rate and the predicted flowability parameters. The grouting control parameters include the target micro-expansion time and the target expansion rate value. In the constrained particle swarm optimization algorithm, the target micro-expansion time and the target expansion rate are used as particle positions, and the fitness function is defined as the algebraic sum of the basic deviation term and the penalty term. The basic deviation term includes a first deviation term and a second deviation term. The penalty term is triggered when the particle position violates the preset constraint conditions. The constraint conditions are configured as follows: the calculated initial pumped grouting volume coupled with the expansion rate is greater than or equal to the void volume parameter, and the grouting expansion stress calculated based on the target expansion rate and the formation mechanical parameters of the grouting area is less than the maximum allowable stress threshold. The calculated grouting control parameters are output to the control terminal of the grouting equipment to set the mixing ratio of the grouting material and the pumping parameters.

[0006] Preferably, before combining and generating the environmental feature data, the acquired data is preprocessed, including: A spatial proximity mapping rule based on inverse distance weighting is used to align and stitch the corrosive medium concentration data and the soil porosity data under the same three-dimensional spatial coordinate reference. Multidimensional inverse distance weighted interpolation is used to fill in the missing values ​​in the matrix caused by sampling omissions; The isolated forest algorithm is used to detect and remove abnormal noise data that deviates from the normal range. If a sample point is determined to be subject to strong environmental interference, the entire row vector containing the data of that sample point is removed from the dataset.

[0007] Preferably, in order to ensure that the model output satisfies the physical normalization boundary conditions, the output layer of the deep neural network model is configured with a specific activation function: For the node that outputs the predicted inflation rate, configure an exponential compression function to compress the real-domain feature tensor passed by the network hidden layer and map it to a continuous interval between 0 and 1. For nodes that output the predicted mobility, a non-negative truncated linear activation function is configured. When the received feature tensor scalar value is less than or equal to 0, it is forcibly truncated and output as 0. When it is greater than 0, an identity mapping is performed.

[0008] Preferably, in the process of constructing and training the deep neural network model: the weight matrix of the hidden layer is assigned values ​​using the Kemming normal distribution initialization rule; The loss function is defined as the weighted sum of mean squared error and weight decay regularization term, and random deactivation layers are configured between hidden layers to reduce the risk of overfitting.

[0009] Preferably, in the iterative process of the constrained particle swarm optimization algorithm: the penalty term includes a maximum penalty factor, which adopts a dynamic penalty mechanism. The calculation logic is to multiply the basic penalty coefficient by a specific exponent to ensure that there is an exponentially increasing penalty magnitude in the later stage of the iteration. After each particle position iteration update operation, the system configures a particle velocity boundary clamping logic unit, sets the physical boundary of the maximum absolute value of the velocity vector, and forces the assignment of values ​​to velocity scalars that are greater than or less than the negative physical boundary.

[0010] Preferably, obtaining the void volume parameters and maximum allowable stress threshold of the target grouting area specifically includes: Based on the set grouting influence radius and the effective porosity of the soil at the corresponding sampling point, the void volume parameter is calculated; wherein, the grouting influence radius is determined based on the rated maximum output pressure of the grouting pump, the formation permeability coefficient, the predicted plastic viscosity parameter, and the grouting diffusion model; Extract the tensile or compressive yield limit value of the most unfavorable section of the surrounding existing structure closest to the grouting hole in the target area, divide the limit value by the safety factor, and assign the resulting quotient value as the maximum allowable stress threshold.

[0011] Preferably, the control terminal uses the grouting control parameters to set the mix ratio, including: For the target micro-expansion time, the baseline setting time when no admixture is added is extracted. Based on the relationship and difference between the target time and the baseline time, the mass percentage of retarder or accelerator in the grouting dry material is adjusted through the preset setting time-admixture dosage correspondence. For the target expansion rate value, the mass percentage of the expansion agent component is obtained according to the preset expansion rate-expansion agent dosage correspondence.

[0012] Preferably, the control terminal also uses the predicted flowability parameter to dynamically adjust the grouting pump fluid, including: Based on the predicted flowability parameters output by the nonlinear mapping model, the initial yield shear stress and predicted plastic viscosity of the fluid are obtained by conversion calculation. Combining the pre-stored total length of the grouting pipeline, the inner diameter of the pipeline, the calculated initial yield shear stress of the fluid, and the predicted plastic viscosity, the fluid pumping resistance pressure loss variable is calculated by following the analytical logic of the rheological pipe flow resistance equation, and the pumping frequency parameter of the grouting pump is determined based on the pressure loss variable.

[0013] Preferably, the step of acquiring environmental feature data of the target area is configured as a dynamic monitoring process of cyclical collection based on a set time period; When it is detected that the change in any corrosive medium concentration data relative to the input value of the previous cycle exceeds the set percentage in the latest data record, or the change in any soil porosity parameter relative to the input value of the previous cycle exceeds the set percentage, it is determined that the trigger threshold change condition is met, a trigger calculation signal is generated, and the updated grouting control parameters are recalculated and output.

[0014] Preferably, the control terminal receives the updated grouting control parameters through a smooth transition filter set in the control logic to achieve seamless switching of material ratio and pumping pressure. The filter time constant is set according to the system response characteristics and is used to filter out high-frequency adjustment commands caused by instantaneous jitter of sensor signals.

[0015] The present invention has achieved the following beneficial effects: By acquiring environmental feature data from multiple dimensions and introducing a deep neural network model, a nonlinear mapping relationship was established between environmental features and grouting material performance parameters such as predicted expansion rate and predicted flowability, overcoming the limitation of the representation ability of a single linear model under multivariable coupled environments. Simultaneously, a constrained particle swarm optimization algorithm was used to optimize the physical boundary of the target grouting area and the mechanical thresholds of surrounding existing structures, restricting the calculated grouting parameters within safe boundaries. This allows the equipment terminal to set the material mix ratio and pumping status based on the optimized parameters. This method helps ensure that the grouting volume meets the actual void requirements, prevents the risk of expansion stress exceeding the safety threshold of surrounding structures, and improves the accuracy of grouting operation parameter configuration and the reliability of the overall project implementation in complex, multivariable coastal environments.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for intelligent optimization of grouting parameters in a coastal multi-coupling corrosion environment based on deep learning, according to an embodiment of the present invention. Figure 2 This is a flowchart of preprocessing acquired data according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the configuration of the output layer of a deep neural network model according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating how a control terminal, according to an embodiment of the present invention, uses the grouting control parameters to set the mix ratio; Figure 5 This is a flowchart illustrating how the control terminal, according to an embodiment of the present invention, also uses the predicted flowability parameter to dynamically adjust the grouting pump fluid. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] In the construction and maintenance of underground spaces at coastal vehicle depots, the target area typically contains coastal swamp facies sediments and is often adjacent to surrounding environments such as sewage treatment facilities. The groundwater and soil in the target area contain specific concentrations of corrosive media such as chloride ions, sulfates, and organic matter. When underground grouting is performed in such an environment with multiple coupled corrosion characteristics, the performance parameters of the grouting material, such as its expansion rate, fluidity, and micro-expansion time, are influenced by the combined effects of various corrosive media concentration variables and soil porosity characteristics, exhibiting a nonlinear mapping relationship.

[0021] Conventional grouting parameter settings typically employ static lookup tables or univariate linear fitting models. However, in multivariate coupled environments, these analytical methods struggle to accurately characterize the nonlinear mapping between the environmental feature input matrix and the grouting material performance output matrix. These technical limitations can lead to calculated grouting parameters deviating from actual physical requirements in practical engineering: underestimating fluid resistance or overestimating expansion rate results in grouting volume smaller than the target void volume; and failing to accurately calculate expansion stress in the coupled environment can cause the expansion stress generated by the grouting material in the enclosed soil to exceed the maximum allowable stress threshold of the surrounding structure. To address these technical problems, this application provides a deep learning-based intelligent optimization method for grouting parameters in a coastal multi-coupled corrosion environment. By introducing a deep neural network algorithm and physical boundary constraints, the method achieves the calculation and configuration of grouting parameters.

[0022] like Figure 1 As shown in the figure, this embodiment discloses an intelligent optimization method for grouting parameters in a coastal multi-coupling corrosion environment based on deep learning, which specifically includes the following steps: Step S1: Obtain environmental characteristic data of the target area. The environmental characteristic data is composed of corrosive medium concentration data and soil porosity data. The corrosive medium concentration data includes the concentration or mass fraction data of chloride ions, sulfates and organic matter. The soil porosity data includes the void ratio, effective porosity and permeability coefficient.

[0023] Specifically, in the grouting operation area, a grid of sampling points is set up to collect groundwater and soil samples from the target grouting area. For groundwater sampling, samples of groundwater fluid from aquifers at different depths are obtained through designated monitoring wells; for soil sampling, undisturbed soil samples are obtained using mechanical core drilling equipment. After sample collection, the samples are input into specific testing equipment to extract the concentration data of corrosive media. The corrosive media include chloride ions, sulfates, and organic matter.

[0024] Furthermore, the specific process for extracting the concentration data of corrosive media includes: using an ion chromatograph to perform ion separation and concentration determination on the groundwater sample and soil leachate to obtain the first concentration value of chloride ions (unit: mg / L) and the second concentration value of sulfate (unit: mg / L); using a total organic carbon analyzer to perform combustion oxidation and non-dispersive infrared detection on the soil sample to obtain the third mass fraction data of organic matter (unit: %). The above-mentioned first concentration value, second concentration value, and third mass fraction data are structured and stored according to the three-dimensional spatial coordinates of the sampling points to construct a chemical concentration data matrix. Specifically, the row dimension of this matrix strictly corresponds to the number of gridded sampling points, and the column dimension is fixed at 3 columns, corresponding one-to-one with chloride ion concentration, sulfate concentration, and organic matter mass fraction, respectively. The data type is uniformly converted to single-precision floating-point numbers and stored in the contiguous memory address of the electronic device.

[0025] While acquiring chemical parameters, soil porosity data for the target grouting area is obtained using surveying equipment, including an in-situ static cone penetrometer and indoor geotechnical testing equipment. The porosity, effective porosity, and permeability coefficient of the target stratum soil are acquired using these surveying equipment. The porosity, effective porosity, and permeability coefficient are aligned according to the same three-dimensional spatial coordinates of the sampling points to construct a physical porosity data matrix. Furthermore, to address the potential for non-complete overlap between chemical sampling points and physical survey points in spatial coordinates, the system employs a spatial proximity mapping rule based on inverse distance weighting. The specific mapping steps are as follows: using the three-dimensional spatial coordinates of any chemical sampling point as the origin, within a set spatial search radius, extract data from at most three adjacent effective physical survey points; subsequently, calculate the Euclidean distance from each physical survey point to the reference origin, and use the reciprocal of this distance as a weighting coefficient for weighted averaging to calculate the various physical parameters at the reference origin, thereby completing the alignment and stitching of the two types of data under the same three-dimensional spatial coordinate reference.

[0026] Understandably, due to the differences in dimensions and numerical ranges between the chemical concentration data matrix and the physical void data matrix, preprocessing of the acquired data is necessary before data combination. The specific data preprocessing process includes: First, using a linear interpolation algorithm to fill in missing values ​​in the matrix caused by sampling deficiencies; this linear interpolation is specifically set to a multidimensional inverse distance weighted interpolation method, where the system extracts data from valid sample points within a preset geometric range around the missing coordinate points, and performs weighted estimation and filling based on the inverse distance decay logic. Second, using an isolated forest algorithm to detect and remove abnormal noise data that significantly deviates from the normal range; the hyperparameters of the isolated forest algorithm are configured as follows: the abnormal data contamination rate parameter is set to 5%, and the number of trees in the random forest is fixed at 100; if the path length required for data isolation during random partitioning is less than the set path threshold, the sample point is determined to be subject to strong environmental interference, and the entire row vector containing the data of that sample point will be directly removed from the dataset matrix. The specific steps for deriving the set path threshold are as follows: extract the total number of samples in the current input dataset, calculate the difference between the total number of samples and 1, obtain the harmonic series of this difference, multiply the harmonic series by a constant 2, and then subtract the product of the constant 2 and the difference divided by the total number of samples to obtain the average path length baseline value. Multiply this average path length baseline value by a constant 0.5 to obtain the final set path threshold constant. The system uses this explicit algebraic calculation result as a hard comparison benchmark to eliminate subjective arbitrariness in the anomaly removal process. Finally, a standardization method is used to perform dimensionless processing on each feature data in the matrix. The execution steps of this standardization method are as follows: calculate the sample mean and sample standard deviation of each column of feature data, subtract the sample mean of the corresponding column from the original feature data, and then divide by the sample standard deviation to obtain the dimensionless feature value. After completing the standardization process, the corresponding concentration data vector and the gap data vector are concatenated to form a high-dimensional environmental feature dataset, where the total dimension constant of the input features is determined to be 6 dimensions. Furthermore, before generating the dataset, the system sets a hard threshold for the physical porosity data: if the obtained porosity ratio is less than 0.3 or greater than 2.0, the system will determine that the formation does not meet the conditions for conventional grouting, directly triggering a hardware warning interruption signal and terminating subsequent calculations. This environmental feature dataset is buffered and stored in the system memory as the input feature vector for subsequent deep neural networks.

[0027] Step S2: Construct and train a deep neural network model, the deep neural network model including an input layer, at least one hidden layer and an output layer; the input layer nodes of the deep neural network model are configured to receive the environmental feature data, and the output layer nodes are configured to output the performance parameters of the grouting material, the performance parameters including the predicted expansion rate and the predicted flowability; The current environmental characteristic data is obtained and input into the nonlinear mapping model to obtain the corresponding predicted expansion rate and predicted mobility parameters. Specifically, a deep neural network architecture is instantiated in the processor of an electronic device. A deep neural network comprising an input layer, multiple hidden layers, and an output layer is constructed. The initial configuration of the network structure is as follows: the number of nodes in the input layer of the deep neural network is configured to be 6, strictly corresponding to the feature dimension of the environmental feature dataset, and configured to receive the feature vectors of the environmental feature dataset; the number of hidden layers in the deep neural network is set to a range of 3 to 5 layers. The number of neurons in the hidden layers is set according to the input feature dimension and the complexity of the prediction task. During the network instantiation phase, all weight matrices are assigned values ​​using the Kemming normal distribution initialization rule to maintain the variance consistency of the input feature vector throughout the forward propagation process. The specific Kemming normal distribution initialization execution steps are as follows: for any hidden layer containing a specified number of input nodes, the constant of the number of input nodes in that hidden layer is extracted; the quotient of constant 2 divided by the constant of the number of input nodes is calculated; the square root of the quotient is taken, and the square root result is used as the standard deviation constant; then, with a mean of 0, a random number matrix following a normal distribution is generated using the obtained standard deviation constant, and this random number matrix is ​​overwritten into the memory address of the initial weight matrix of that layer. The activation function of the hidden layer uniformly adopts the modified linear unit with leakage, and its leakage parameter is configured to 0.01. The specific internal logic judgment conditions and data flow are as follows: a logical comparison branch is set at each neuron node. When the scalar data of the feature node input to the activation function is greater than 0, the scalar data itself is directly output and passed to the next layer; when the scalar data of the feature node input to the activation function is less than or equal to 0, the scalar data is multiplied by the constant 0.01 before output. This preserves the small gradient features of the negative half axis and avoids the phenomenon of the gradient completely returning to zero during the backpropagation of the neuron.

[0028] The output layer nodes of the deep neural network are configured to output performance parameters of the grouting material, including predicted expansion rate and predicted flowability. Therefore, the number of nodes in the output layer is configured to be 2. The activation function of the output layer adopts an exponential compression function or a non-negative truncated linear function. Specifically, for the predicted expansion rate output node in the output layer, since its physical expansion rate value has been normalized to a continuous interval of 0 to 1 in the preprocessing, this single node is configured with a nonlinear activation function, which is strictly defined as dividing a constant 1 by the sum of a constant 1 and a negative feature input value of the natural exponent. This nonlinear transformation compresses and maps the real-domain feature tensor passed by the hidden layer of the network to a continuous interval between 0 and 1, so as to ensure that the predicted expansion rate satisfies the normalization boundary conditions at the physical level. For the predicted flowability output node, the physical meaning of its output value is the initial flow expansion diameter of the grouting material, characterizing the macroscopic flow diffusion capacity of the grout. This type of parameter has a large numerical range and no strict upper limit; therefore, this node is configured with a non-negative truncated linear activation function. Its internal logic judgment condition is: when the received feature tensor scalar value is less than or equal to 0, it is forcibly truncated and output as 0; when the received feature tensor scalar value is greater than 0, an identity mapping is performed. The predicted flowability and rheological parameters have a preset quantitative mapping relationship: based on the reference water-cement ratio of the grouting material, the predicted flowability can be converted into the initial yield shear stress and predicted plastic viscosity of the fluid, thereby connecting the subsequent calculation logic of the pumping pipeline resistance. The two are a derived conversion relationship of "macroscopic performance index - microscopic rheological parameter," not the same physical quantity. Furthermore, before establishing the mapping model, the constructed deep neural network needs to be trained. Historical engineering grouting datasets are extracted from the engineering database. The basic size of this historical engineering grouting dataset is set to include multiple sets of mutually independent sample pairs. Each sample pair must include a complete and corresponding 6-dimensional environmental feature record, as well as the actual grouting material expansion rate measured 48 hours after grouting and the actual initial flowability index measured at the pump outlet. It should be noted that the extraction of the aforementioned specific corrosive media as input features is not arbitrary, but based on the specific multi-coupling chemical reaction mechanism of coastal grouting: chloride ions accelerate the early hydration of the cement matrix, while sulfates react with hydrated calcium aluminate to form ettringite, causing volume expansion. Simultaneously, high concentrations of organic matter create steric hindrance on the surface of the grouting material particles, competitively inhibiting the reaction rates of the two aforementioned media. This cross-competition and antagonism of the three media at different concentration gradients leads to highly nonlinear and abrupt changes in the expansion stress release curve and rheological curve of the grouting system. Conventional univariate linear regression or polynomial fitting cannot mathematically map this dynamic competition process of multiple chemical bonds; this is the core physical motivation for introducing a high-dimensional deep neural network to extract implicit mapping features. The aforementioned historical engineering grouting dataset is divided into training and test sets in an 80% to 20% ratio.

[0029] The deep neural network is iteratively trained using the training set. The training objective is to minimize the error between the predicted output value and the true label value. The loss function is defined as a weighted sum of mean squared error and weight decay regularization term. During training memory allocation, the number of samples in a single data batch is set according to computational resources; the regularization coefficient included in the loss function is used to control model complexity and reduce the risk of overfitting. The specific value of the regularization coefficient is set to 0.0001. To prevent overfitting of the deep neural network when fitting high-dimensional environmental feature data, a random deactivation layer is configured between the hidden layers of the network. During each forward propagation training phase, the random deactivation layer randomly sets the output value of some neurons to 0 according to a preset retention probability. The specific value of this retention probability is determined as follows: during the model cross-validation training phase, an exhaustive search is performed within the parameter grid interval of 0.5 to 0.9 with a step size of 0.05 to extract the retention probability value that minimizes the mean absolute error of the validation set and then solidifies it in the system.

[0030] An adaptive moment estimation optimization algorithm is used to backpropagate and update the network parameters. Specifically, the core hyperparameters of this optimization algorithm must be explicitly configured as follows: the exponential decay rate of the first-order moment estimation is set according to the training requirements, the exponential decay rate of the second-order moment estimation is fixed at 0.999, and the numerical stability constant, which is used to prevent the denominator from being 0, is set to 0.00000001. Before executing the first iteration, the system must initialize the moving average tensor of the first-order moment estimate and the moving average tensor of the gradient squared of the second-order moment estimate in memory, forcing all elements within these tensors to be 0, thus serving as the absolute zero starting point for dynamic adaptive adjustment. Subsequently, the internal data flow and derivation steps for parameter updates are as follows: First, calculate the moving average of the gradient of the target loss function and the moving average of the gradient squared. Specifically, extract the moving average from the previous time step, multiply it by the corresponding exponential decay rate, and add the difference between the current gradient value or the gradient squared value multiplied by a constant 1 and the corresponding exponential decay rate. Second, perform bias correction calculations at the beginning of the iteration, dividing the moving average by a constant 1 minus the difference between the exponential decay rate raised to the power of the current iteration step number. Third, calculate the network weight parameter update amount for this iteration. Specifically, multiply the current dynamic learning rate by the bias-corrected first-order moment estimate, and then divide by the sum of the square root of the bias-corrected second-order moment estimate and the numerical stability constant. Through these algebraic steps, the system transforms the loss gradient value into a precise correction amount for the network weights. In this backpropagation process, the initial learning rate parameter is set to 0.001, and a step-rate decay instruction is configured, meaning that the learning rate is reduced by 90% every 50 overall training cycles. The preset convergence condition is specifically quantified as follows: the absolute value of the decrease in the mean square error of the validation set is less than 0.00001 for 10 consecutive training cycles. The preset maximum number of iterations is set to a hardware limit of 2000. After training, the weights and bias parameters of the network model are solidified, completing the establishment of a nonlinear mapping model between environmental features and grouting material performance parameters.

[0031] Step S3: Obtain the void volume parameters and maximum allowable stress threshold of the target grouting area, and use a constrained particle swarm optimization algorithm to calculate and output the corresponding grouting control parameters based on the predicted expansion rate and the predicted flowability parameters. The grouting control parameters include the target micro-expansion time and the target expansion rate value. In the constrained particle swarm optimization algorithm, the target micro-expansion time and the target expansion rate are used as particle positions, and the fitness function is defined as the algebraic sum of the basic deviation term and the penalty term. The basic deviation term includes a first deviation term and a second deviation term. The penalty term is triggered when the particle position violates the preset constraint conditions. The constraint conditions are configured as follows: the calculated initial pumped grouting volume coupled with the expansion rate is greater than or equal to the void volume parameter, and the grouting expansion stress calculated based on the target expansion rate and the formation mechanical parameters of the grouting area is less than the maximum allowable stress threshold. Understandably, the current environmental characteristic data is acquired and input into the trained nonlinear mapping model. Through forward propagation calculations of the model, the system obtains the predicted expansion rate and predicted flowability parameters under the corresponding current environmental conditions. At this stage, the acquired predicted parameters serve as the theoretical physical response benchmark of the grouting material in the current coupled environment and under standard mix proportions.

[0032] Based on the set grouting influence radius and the effective porosity of the soil at the corresponding sampling point, the void volume parameter is calculated; wherein, the grouting influence radius is determined based on the rated maximum output pressure of the grouting pump, the formation permeability coefficient, the predicted plastic viscosity parameter, and the grouting diffusion model; The preset constraints are input into the parameter optimization algorithm. The goal of the optimization algorithm is to find the optimal combination of grouting control parameters within a given feasible region, including the target micro-expansion time and the target expansion rate. The constraints are configured as the following two simultaneous logical boundaries: First, a volume constraint, namely, the sum of the initial pumped grouting volume multiplied by a constant 1 and the target expansion rate must be greater than or equal to the void volume parameter. Second, a stress constraint, namely, the grouting expansion stress calculated based on the target expansion rate, the target formation elastic modulus, and the constraint stiffness must be less than the maximum allowable stress threshold.

[0033] Specifically, the analytical mathematical model for the grouting expansion stress calculated jointly based on the target expansion rate, the target formation elastic modulus, and the constraint stiffness is as follows: ; in, The calculated grouting expansion stress (unit: ); The target expansion rate value (dimensionless) represents the current iteration position of the particle. Represents the elastic modulus of the target formation (unit: This value was obtained by looking up the table from the in-situ static cone penetration data of the survey equipment; The equivalent constraint modulus of the surrounding existing structure on the grouting zone (unit: The value of this equation is obtained by extracting the product of the moment of inertia and elastic modulus of the surrounding structure sections, dividing it by the cube of the load span to obtain the structural linear stiffness, and then combining it with the equivalent load area and characteristic length to convert it into the equivalent constraint modulus. By introducing the above soil-structure co-deformation mechanical equation as a hard penalty trigger boundary, the shortcomings of traditional optimization algorithms that rely solely on purely numerical empirical thresholds and deviate from the actual geological mechanical characteristics are avoided.

[0034] Furthermore, based on the predicted expansion rate, predicted flowability parameters, and the aforementioned preset constraints, an iterative optimization process of the parameter optimization algorithm is executed. This embodiment employs a constrained particle swarm optimization algorithm for parameter calculation. The particle swarm is initialized and a fitness function is defined. To incorporate the constraints into the objective function, a penalty function method is introduced. The specific physical implementation details of initializing the particle swarm are as follows: an initial position matrix of 100 particles is randomly generated in system memory. The random value range of the target micro-expansion time dimension is rigidly limited to a closed interval of 40 to 120 minutes, and the random value range of the target expansion rate value dimension is rigidly limited to a closed interval of 0.05 to 0.15. Simultaneously, a corresponding initial velocity matrix is ​​generated, with the initial velocity value of each dimension set to 0.1 times the absolute value of the corresponding dimension's value range (a random floating-point number). This clearly defined initial value range ensures that the search space of the first-generation population of the algorithm strictly falls within the allowable range of the physical quantities of the actual grouting equipment. The fitness function is designed as an algebraic sum of a basic deviation term and a penalty term. The smaller the value of the fitness function, the better the overall performance of the grouting control parameters for the corresponding particle. The basic deviation term includes a first deviation term and a second deviation term, which are multiplied by a set weighting coefficient. The first deviation term is constructed based on the target micro-expansion time and the predicted flowability. The predicted flowability is the initial flow expansion diameter of the grouting material, characterizing the macroscopic flow diffusion capability of the grout. The specific calculation method is as follows: the upper limit of the micro-expansion time range and the upper limit of the flowability range are preset. The target micro-expansion time corresponding to the current particle is divided by the upper limit of the micro-expansion time range to obtain a dimensionless scalar of micro-expansion time in the range of 0 to 1. The predicted flowability is divided by the upper limit of the flowability range to obtain a dimensionless scalar of flowability in the range of 0 to 1. The mean square error between the two dimensionless scalars is calculated as the first deviation term. This deviation term is used to characterize the degree of matching between the setting and expansion timing of the grouting material and its pumping flow performance. The second deviation term is the mean square error between the target expansion rate and the predicted expansion rate, used to characterize the degree of fit between the expansion performance and the environmental adaptation benchmark. A penalty term is triggered when a particle position violates volume or stress constraints. The penalty term is obtained by multiplying the violation value by the maximum penalty factor. When handling particles that violate constraints, the maximum penalty factor employs a dynamic penalty mechanism, introducing a base penalty coefficient and a growth rate constant. The base penalty coefficient is specifically set to 10, and the growth rate constant is set to 2.0. The calculation logic of the dynamic penalty factor is to multiply the base penalty coefficient by a specific exponent of the constant 10, where the exponent is the product of the growth rate constant, the current iteration number, and the reciprocal of the maximum iteration number. This ensures that the exponentially increasing penalty magnitude in the later stages of iteration is sufficient to reject all infeasible solution particles falling outside the constraint boundaries.

[0035] In each iteration, the fitness value of each particle is calculated, and the individual historical best position of each particle and the global best position of the entire particle swarm are updated. The particle velocity update logic is as follows: multiply the inertia weight by the current velocity vector, add the product of the cognitive learning factor and a random number in the [0,1] interval, multiply by the difference between the individual historical best position and the current position, and finally add the product of the social learning factor and a random number in the [0,1] interval, multiply by the difference between the global best position and the current position. The inertia weight adopts a linear decreasing strategy, set to 0.9 in the early stage of the iteration and decreasing to 0.4 in the later stage of the iteration; the cognitive learning factor and the social learning factor are set according to the optimization requirements. Subsequently, the updated velocity vector is added to the current position vector to complete the position update. After each particle position iteration update operation, the system is configured with a particle velocity boundary clamping logic unit. This unit sets the maximum absolute value physical boundary of the velocity vector to 0.2 times the difference between the upper and lower limit intervals. The system iterates through the updated velocity vector and makes a judgment: when the velocity scalar of any dimension is greater than the physical boundary of the maximum absolute value, it is forcibly assigned the boundary value; when the velocity scalar of any dimension is less than the negative physical boundary of the maximum absolute value, it is forcibly assigned the negative boundary value.

[0036] The algorithm iterates continuously until the maximum number of iterations is reached or the fitness increment of the global optimal position is less than a set convergence threshold. The maximum number of iterations is set according to the optimization requirements. The specific logic of the set convergence threshold is as follows: during 50 consecutive position update iterations, the absolute difference increment of the fitness function of the global optimal position is always less than 0.001. After the iteration terminates, the coordinate values ​​corresponding to the global optimal position are output, which are the calculated grouting control parameters, including the target micro-expansion time and the target expansion rate.

[0037] Step S4: Output the calculated grouting control parameters to the control terminal of the grouting equipment to set the mixing ratio of the grouting material and the pumping parameters.

[0038] Specifically, after the processor completes the optimization calculations, it converts the calculated target micro-expansion time and target expansion rate values ​​into a standard data communication message format. The message data, including the grouting control parameters, is then output and transmitted to the control terminal of the automated grouting equipment on-site using the industrial Ethernet protocol.

[0039] After receiving the message data, the control terminal of the grouting equipment parses the corresponding values ​​and uses them to set the mix ratio and pumping parameters of the grouting material. The control terminal's internal memory contains a database of basic performance curves for a specific brand of grouting material. The microprocessor calls an interpolation program to query the theoretical addition amount of the expanding agent corresponding to the target expansion rate value from the basic performance curves. This database of basic performance curves specifically consists of several sets of discrete calibration data point coordinates pre-measured and entered into the control terminal. For the target value calculated and issued between the x-coordinates of two adjacent calibration data points, the control terminal strictly executes a piecewise linear interpolation subroutine, calculating a continuous and smooth addition amount setting value on the corresponding y-coordinate based on the geometric distance ratio between the target value and the x-coordinates of the adjacent known data points on the left and right sides.

[0040] Furthermore, the control terminal has a preset mix ratio conversion matrix. For the target micro-expansion time, the control terminal calculates and adjusts the percentage signal of the mass dosage of retarder or accelerator in the grouting dry material. The specific conversion rule is as follows: extract the baseline setting time when no admixture is added; when the calculated target micro-expansion time is greater than the baseline setting time, the controller triggers the retarding logic unit, and linearly converts the difference between the two target times according to the preset setting time-admixture dosage correspondence; conversely, if the target micro-expansion time is less than the baseline setting time, the accelerator logic unit is triggered, and the accelerator dosage is calculated according to the equivalent slope. For the target expansion rate value, the control terminal calculates and adjusts the percentage signal of the mass dosage of the expansive agent component. The specific conversion rule is as follows: based on the currently issued target expansion rate value, subtract the baseline expansion rate without admixtures, multiply the obtained expansion rate difference by a compensation conversion coefficient of 5.0, and the result is the percentage increase in the mass dosage of the expansive agent. To ensure safe feeding of the powder screw conveyor, the final physical value of the calculated admixture percentage is rigidly limited to a closed range of 5% to 15%. Subsequently, the control terminal converts the above-mentioned addition set value into the control variable of the specific actuator, calculates the screw conveyor operating frequency data to be set within a given feeding time period, encapsulates it, and sends it to the register address of the frequency converter, driving the motor to run at a specific speed to complete the precise addition of physical materials.

[0041] Simultaneously, the control terminal calculates the pressure loss variable of the target grouting pump pipeline based on the acquired predicted flowability parameters. The specific theoretical numerical solution steps strictly follow the analytical logic of the rheological pipe flow resistance equation: Here, based on the predicted flowability parameters output by the nonlinear mapping model, the exact logical mapping relationship between the initial yield shear stress and the predicted plastic viscosity of the fluid is derived as follows: the reference water-cement ratio parameter of the grouting material is extracted, the acquired predicted flowability parameter value is divided by the reference water-cement ratio constant, and the quotient is added to the constant 5.0 to obtain the initial yield shear stress value of the fluid, with the unit specified as Pascal; at the same time, the reciprocal of the predicted flowability parameter value is multiplied by the constant 0.02, and then added to the constant 0.1 to obtain the predicted plastic viscosity value, with the unit specified as Pascal-second. The first value is obtained by multiplying the pre-stored total length of the grouting pipeline in the control terminal by a constant of 4, dividing by the geometric value of the pipeline's inner diameter, and then multiplying by the initial yield shear stress of the fluid obtained from the predicted flowability parameters. The second value is obtained by multiplying the predicted plastic viscosity by a constant of 32, multiplying by the total pipeline length and the average linear velocity of the grout in the target pipeline, and finally dividing by the square of the geometric value of the pipeline's inner diameter. The first and second values ​​are then added together to obtain the macroscopic fluid pumping resistance pressure loss variable. Based on the calculated pressure loss value, the target frequency parameter of the grouting pump's inverter is set to dynamically adjust the grouting pump's flow rate and grouting output pressure.

[0042] During the grouting operation, flow meters and pressure sensors on the pipeline collect grouting status data in real time and feed it back to the control terminal. The control terminal uses a proportional-integral-derivative closed-loop control algorithm to dynamically fine-tune the operating frequency of the grouting pump. The core parameters of this closed-loop control algorithm are determined as follows: its control parameters are tuned according to the system characteristics. Within a single execution cycle of 20 milliseconds, the controller continuously calculates the discrete difference between the real-time electrical signals fed back by the pressure transmitter and flow meter and the target setpoint, and outputs an adjustment signal to the frequency converter based on the calculated adjustment increment.

[0043] Furthermore, in specific engineering implementation scenarios, the continuity and time-varying nature of environmental characteristic data in the target area require the system to have dynamic update capabilities. Due to groundwater flow and tidal effects in coastal swamp sediments, the process of acquiring environmental characteristic data of the target area is configured to perform periodic retests based on a set time period. When a new environmental characteristic data trigger threshold change is detected, the system will generate a trigger calculation signal. The specific judgment condition for the trigger threshold change is defined as follows: in the latest data record, the relative change of any corrosive medium concentration data compared to the input value of the previous calculation cycle exceeds the set threshold, or the change of any soil porosity parameter compared to the input value of the previous calculation cycle exceeds the set threshold. The system polls and scans the above comparison conditions within one machine cycle, and generates a trigger calculation signal once either condition is met. The data will be automatically transmitted to the model for update calculation, and the updated grouting control parameters will be re-output.

[0044] The controller in the control terminal receives updated parameter message data via an industrial Ethernet interface. A smooth transition filter is incorporated into the control logic to achieve seamless switching between material proportions and pumping pressure. The filter's time constant is set according to the system response characteristics, aiming to delay abrupt changes in the control output response and specifically filtering out high-frequency adjustment commands caused by momentary fluctuations in sensor signals.

[0045] The calculation and optimization processes of the aforementioned prediction parameters and target control parameters are all completed by the processor inside the electronic device by executing computer program instructions stored in the memory.

[0046] 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 deep learning-based intelligent optimization method for grouting parameters in a coastal multi-coupling corrosion environment, applied to grouting operations with a predetermined material system, characterized in that... Includes the following steps: Environmental characteristic data of the target area is acquired, which is composed of corrosive medium concentration data and soil porosity data; wherein, the corrosive medium concentration data includes the concentration or mass fraction data of chloride ions, sulfates and organic matter; the soil porosity data includes porosity, effective porosity and permeability coefficient; A deep neural network model is constructed and trained, the deep neural network model including an input layer, at least one hidden layer and an output layer; the input layer nodes of the deep neural network model are configured to receive the environmental feature data, and the output layer nodes are configured to output the performance parameters of the grouting material, the performance parameters including the predicted expansion rate and the predicted flowability; the trained deep neural network model is used as a nonlinear mapping model between the environmental features and the performance parameters of the grouting material. The current environmental characteristic data is obtained and input into the nonlinear mapping model to obtain the corresponding predicted expansion rate and predicted mobility parameters. The void volume parameters of the target grouting area and the maximum allowable stress threshold of the surrounding existing structure are obtained. A constrained particle swarm optimization algorithm is used to calculate and output the corresponding grouting control parameters based on the predicted expansion rate and the predicted flowability parameters. The grouting control parameters include the target micro-expansion time and the target expansion rate value. In the constrained particle swarm optimization algorithm, the target micro-expansion time and the target expansion rate are used as particle positions, and the fitness function is defined as the algebraic sum of the basic deviation term and the penalty term. The basic deviation term includes a first deviation term and a second deviation term. The penalty term is triggered when the particle position violates the preset constraint conditions. The constraint conditions are configured as follows: the calculated initial pumped grouting volume coupled with the expansion rate is greater than or equal to the void volume parameter, and the grouting expansion stress calculated based on the target expansion rate and the formation mechanical parameters of the grouting area is less than the maximum allowable stress threshold. The calculated grouting control parameters are output to the control terminal of the grouting equipment to set the mixing ratio of the grouting material and the pumping parameters.

2. The method according to claim 1, characterized in that, Before combining and generating the environmental feature data, the acquired data is preprocessed, including: A spatial proximity mapping rule based on inverse distance weighting is used to align and stitch the corrosive medium concentration data and the soil porosity data under the same three-dimensional spatial coordinate reference. Multidimensional inverse distance weighted interpolation is used to fill in the missing values ​​in the matrix caused by sampling omissions; The isolated forest algorithm is used to detect and remove abnormal noise data that deviates from the normal range. If a sample point is determined to be subject to strong environmental interference, the entire row vector containing the data of that sample point is removed from the dataset.

3. The method according to claim 1, characterized in that, To ensure that the model output satisfies the physical normalization boundary conditions, the output layer of the deep neural network model is configured with a specific activation function: For the node that outputs the predicted inflation rate, configure an exponential compression function to compress the real-domain feature tensor passed by the network hidden layer and map it to a continuous interval between 0 and 1. For nodes that output the predicted mobility, a non-negative truncated linear activation function is configured. When the received feature tensor scalar value is less than or equal to 0, it is forcibly truncated and output as 0. When it is greater than 0, an identity mapping is performed.

4. The method according to claim 1, characterized in that, In the process of constructing and training the deep neural network model: the weight matrix of the hidden layer is assigned values ​​using the Kemming normal distribution initialization rule; The loss function is defined as the weighted sum of mean squared error and weight decay regularization term, and random deactivation layers are configured between hidden layers to reduce the risk of overfitting.

5. The method according to claim 1, characterized in that, In the iterative process of the constrained particle swarm optimization algorithm: the penalty term includes a maximum penalty factor, which adopts a dynamic penalty mechanism. Its calculation logic is to multiply the basic penalty coefficient by a specific exponential power to ensure that the penalty magnitude grows exponentially in the later stage of the iteration. After each particle position iteration update operation, the system configures a particle velocity boundary clamping logic unit, sets the physical boundary of the maximum absolute value of the velocity vector, and forces the assignment of values ​​to velocity scalars that are greater than or less than the negative physical boundary.

6. The method according to claim 1, characterized in that, The acquisition of the void volume parameters of the target grouting area and the maximum allowable stress threshold of the surrounding existing structure specifically includes: Based on the set grouting influence radius and the effective porosity of the soil at the corresponding sampling point, the void volume parameter is calculated; wherein, the grouting influence radius is determined based on the rated maximum output pressure of the grouting pump, the formation permeability coefficient, the predicted plastic viscosity parameter, and the grouting diffusion model; Extract the tensile or compressive yield limit value of the most unfavorable section of the surrounding existing structure closest to the grouting hole in the target area, divide the limit value by the safety factor, and assign the resulting quotient value as the maximum allowable stress threshold.

7. The method according to claim 1, characterized in that, The control terminal uses the grouting control parameters to set the mix ratio, including: For the target micro-expansion time, the baseline setting time when no admixture is added is extracted. Based on the relationship and difference between the target time and the baseline time, the mass percentage of retarder or accelerator in the grouting dry material is adjusted through the preset setting time-admixture dosage correspondence. For the target expansion rate value, the mass percentage of the expansion agent component is obtained according to the preset expansion rate-expansion agent dosage correspondence.

8. The method according to claim 1, characterized in that, The control terminal also uses the predicted flowability parameters to assist in determining the grouting pumping parameters, including: Based on the predicted flowability parameters output by the nonlinear mapping model, the initial yield shear stress and predicted plastic viscosity of the fluid are obtained by conversion calculation. Combining the pre-stored total length of the grouting pipeline, the inner diameter of the pipeline, the calculated initial yield shear stress of the fluid, and the predicted plastic viscosity, the fluid pumping resistance pressure loss variable is calculated by following the analytical logic of the rheological pipe flow resistance equation, and the pumping frequency parameter of the grouting pump is determined based on the pressure loss variable.

9. The method according to claim 1, characterized in that, The step of acquiring environmental feature data of the target area is configured to perform periodic retesting based on a set time period; When it is detected that the change in any corrosive medium concentration data relative to the input value of the previous cycle exceeds the set percentage in the latest data record, or the change in any soil porosity parameter relative to the input value of the previous cycle exceeds the set percentage, it is determined that the trigger threshold change condition is met, a trigger calculation signal is generated, and the updated grouting control parameters are recalculated and output.

10. The method according to claim 9, characterized in that, The control terminal receives the updated grouting control parameters through a smooth transition filter set in the control logic, so as to achieve a smooth switching of material ratio and pumping pressure. The filter's filtering time constant is set according to the system response characteristics and is used to filter out high-frequency adjustment commands caused by instantaneous jitter in the sensor signal.

Images