A grouting construction quality analysis method and device, electronic equipment and storage medium

Abstract

This application discloses a method, device, electronic equipment, and storage medium for analyzing the quality of grouting construction. It extracts core features through spatiotemporal mapping of multi-source parameters and a triple screening mechanism, and actively integrates engineering interactive features that reflect the physical evolution law to construct an optimized feature matrix. Combined with hyperparameter optimization and time-series dynamic weighting operations, it synergistically fuses multiple heterogeneous basic models, thereby effectively eliminating measurement noise and dimensional interference. It overcomes the bottleneck of prediction distortion of single traditional models in complex and variable construction periods, and realizes high-precision, real-time inversion and deduction of the permeability of underground rock mass after grouting. It accurately generates cement grouting construction quality analysis results for guiding on-site quality control.

Description

Technical Field

[0001] This invention relates to the field of grouting construction quality analysis technology, and in particular to a grouting construction quality analysis method, a grouting construction quality analysis device, an electronic device, and a readable storage medium. Background Technology

[0002] Cement grouting is a core method for seepage prevention and reinforcement in water conservancy and hydropower projects such as dams. The post-grouting permeability is a key indicator for evaluating the construction quality of concealed works. However, the grouting construction site environment is complex, and the development of underground rock fissures exhibits strong nonlinearity and uncertainty. The foundation construction data, dynamic geological parameters, and construction environment parameters collected during construction often suffer from severe spatiotemporal misalignment, dimensional differences, and measurement noise. Related technologies often rely on qualitative analysis based on human experience or simple predictions using single machine learning models. These approaches generally suffer from insufficient multi-source data fusion, simplistic preprocessing strategies, lack of engineering physical mechanism support for feature extraction, and simplistic model structures with weak generalization capabilities. Consequently, the prediction accuracy and stability of the post-grouting permeability are low, making it difficult to meet the real-time, accurate quality analysis and control requirements under complex geological conditions. Summary of the Invention

[0003] The present invention provides a method, apparatus, electronic device, and readable storage medium for analyzing the quality of grouting construction, in order to overcome or at least partially solve the above-mentioned problems.

[0004] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for analyzing the quality of grouting construction, including: Acquire basic construction data, dynamic geological parameters, and construction environment parameters; Based on the basic construction data, the dynamic geological parameters, and the construction environment parameters, core features that are strongly correlated with post-irrigation permeability are obtained. Determine the engineering mechanism information and calculate the engineering interaction characteristics based on the engineering mechanism information; An optimized feature matrix is ​​constructed by combining the associated core features and the engineering interaction features; The optimized feature matrix is ​​used to train multiple initial base model groups composed of heterogeneous base models; The initial base model group is trained by hyperparameter optimization and temporal dynamic weighting operations to construct a fusion model; The fusion model is controlled to output predicted values, and quality analysis results of cement grouting construction are generated based on the predicted values.

[0005] Optionally, before the step of obtaining the core features strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters, the method further includes: Spatiotemporal correlation mapping is performed on the basic construction data, the dynamic geological parameters, and the construction environment parameters to obtain a multi-source fusion dataset; The multi-source fusion dataset is filtered and sorted according to the data integrity and validity standards. After removing invalid data, a sorted dataset with hole segment identifiers is generated according to the hole order. Based on the borehole segment identifier, the grouting process data in the time series dataset is normalized in two dimensions to obtain the normalized time series dataset.

[0006] Optionally, the step of obtaining the core features strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters includes: By utilizing physical constraints based on engineering principles and statistical detection based on box plots and an improved three-standard-deviation criterion, outliers in the normalized time-series dataset are identified and processed to obtain an outlier-corrected dataset. Based on the parameter type, the missing values ​​of the outlier correction dataset are filled using differentiated strategies such as filling with the median under the same conditions, interpolating adjacent wells and correcting for geological stratification, and filling with the mean of the environment during the same period, to obtain a complete filled dataset. Improved Z-score normalization is performed on the target construction parameters of the complete filled dataset, maximum and minimum value normalization is applied to the target geological parameters of the complete filled dataset, and global standardization normalization is applied to the target construction environment parameters of the complete filled dataset to eliminate the dimensional differences and data noise of the target construction parameters, target geological parameters, and target construction environment parameters, thus obtaining an adaptive normalized dataset. Through a triple screening mechanism of Pearson correlation analysis, mutual information entropy calculation, and expert engineering significance verification, core features strongly correlated with post-irrigation permeability are extracted from the adaptive normalized dataset.

[0007] Optionally, the engineering mechanism information includes the amount of ash injected per unit time, the pressure-flow synergy coefficient, and the crack development-ash injection volume adaptation coefficient. The step of determining the engineering mechanism information and calculating the engineering interaction characteristics based on the engineering mechanism information includes: The core features are used to calculate engineering interaction features including the amount of ash injected per unit time, the synergy coefficient of pressure and flow rate, and the matching coefficient between crack development and ash injection amount. The core features include input parameters and output parameters, and the step of constructing an optimized feature matrix through the associated core features and the engineering interaction features includes: An optimized feature matrix is ​​constructed by concatenating the input parameters, the output parameters, and the engineering interaction features.

[0008] Optionally, the step of training an initial base model group consisting of multiple heterogeneous base models using the optimized feature matrix includes: The optimized feature matrix is ​​divided into a training set and a test set. The training set is used to train and generate an initial base model set including a support vector regression model, an improved long short-term memory network model, and a lightweight gradient boosting tree model. With the objective function of minimizing the average absolute percentage error of the fusion model composed of the initial basic model group, the improved Grey Wolf optimization algorithm is used to optimize the core hyperparameters of each model in the initial basic model group, and the optimal parameter combination of each model in the initial basic model group is output. The optimal parameter combination is configured into each model in the initial basic model group, and the fusion weight is calculated based on the average absolute percentage error of each model in the initial basic model group on the test set and the time-series prediction error for different construction periods. Based on the fusion weights and the preset error correction terms, a multi-model dynamic weighted fusion model is constructed.

[0009] Optionally, the step of controlling the output prediction value of the fusion model includes: The construction unit data to be predicted in the adaptive normalized dataset is input into the multi-model dynamic weighted fusion model to obtain the final post-irrigation permeability prediction value.

[0010] Optionally, the step of generating quality analysis results for cement grouting construction based on the predicted values ​​includes: The performance of the predicted post-irrigation permeability is evaluated using mean square error, mean absolute error, mean absolute percentage error, and prediction stability coefficient to obtain model performance indicators. When it is determined that the model performance index has not reached the preset accuracy or the engineering geological environment has changed, the multi-model dynamic weighted fusion model is triggered to enter an adaptive update process including incremental training and transfer learning, so as to freeze the core parameters of the multi-model dynamic weighted fusion model and adjust the target network layer weights of the multi-model dynamic weighted fusion model to obtain an adaptively optimized fusion model. The adaptively optimized fusion model, the optimized feature matrix, and the post-grouting permeability prediction value are integrated and input into an intelligent visualization system developed based on a graphical user interface component and a dynamic chart engine to generate a visualized cement grouting construction quality control and decision-making suggestion report.

[0011] Secondly, embodiments of this application provide a grouting construction quality analysis device, comprising: The data acquisition module is used to acquire basic construction data, dynamic geological parameters, and construction environment parameters. The core feature acquisition module is used to acquire core features that are strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters. The engineering interaction feature calculation module is used to determine engineering mechanism information and calculate engineering interaction features based on the engineering mechanism information. An optimized feature matrix construction module is used to construct an optimized feature matrix through the associated core features and the engineering interaction features; The initial base model group training module is used to train multiple initial base model groups composed of heterogeneous base models using the optimized feature matrix; The fusion model construction module is used to train the initial basic model group through hyperparameter optimization and temporal dynamic weighting operations to construct the fusion model; The prediction output module is used to control the fusion model to output prediction values ​​and generate quality analysis results of cement grouting construction based on the prediction values.

[0012] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0013] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0014] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the method as described in the first aspect.

[0015] The embodiments of the present invention have the following advantages: This invention extracts core features through spatiotemporal mapping of multi-source parameters and a triple screening mechanism, and actively integrates engineering interaction features that reflect the physical evolution law to construct an optimized feature matrix. Combined with hyperparameter optimization and time-series dynamic weighting operations, multiple heterogeneous basic models are synergistically fused, thereby effectively eliminating measurement noise and dimensional interference, overcoming the bottleneck of prediction distortion in complex and variable construction periods by a single traditional model, achieving high-precision, real-time inversion and deduction of the permeability of underground rock mass after grouting, and accurately generating cement grouting construction quality analysis results for guiding on-site quality control. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of a grouting construction quality analysis method provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a grouting construction quality analysis device provided in an embodiment of the present invention; Figure 3 This is a hardware structure block diagram of an electronic device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0017] 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, 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.

[0018] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Reference Figure 1 The diagram illustrates a flowchart of a grouting construction quality analysis method provided in an embodiment of the present invention, which may specifically include the following steps: Step 101: Obtain basic construction data, dynamic geological parameters, and construction environment parameters; Step 102: Based on the basic construction data, the dynamic geological parameters, and the construction environment parameters, obtain the core features that are strongly correlated with the post-irrigation permeability. Step 103: Determine the engineering mechanism information and calculate the engineering interaction features based on the engineering mechanism information; Step 104: Construct an optimized feature matrix using the associated core features and the engineering interaction features; Step 105: Use the optimized feature matrix to train multiple initial base model groups composed of heterogeneous base models; Step 106: Train the initial basic model group through hyperparameter optimization and temporal dynamic weighting operations to construct the fusion model; Step 107: Control the fusion model to output predicted values, and generate quality analysis results of cement grouting construction based on the predicted values.

[0021] The embodiments of the present invention can acquire basic construction data, dynamic geological parameters, and construction environment parameters to comprehensively collect the underlying original physical quantities of the multi-source heterogeneous dam construction site, providing complete and fundamental data raw material support for subsequent feature extraction, physical mechanism fusion, and AI model training.

[0022] Basic construction data refers to the real-time core process parameters directly collected by sensors and monitoring equipment at the grouting construction site, such as real-time grouting pressure, grout flow rate, cumulative grout injection volume, and grouting duration.

[0023] Dynamic geological parameters refer to the rock mass geological characteristics that change dynamically with the location of the borehole section or with the progress of grouting construction, such as the original permeability of the construction section, rock mass quality indicators, and groundwater level.

[0024] Construction environment parameters refer to the physical state parameters of the external environment or equipment during grouting construction, including but not limited to the ambient temperature of the construction site, the grout mixing temperature, and the ambient humidity.

[0025] In this embodiment of the invention, based on the basic construction data, the dynamic geological parameters, and the construction environment parameters, core features strongly correlated with post-irrigation permeability can be obtained. This allows for the precise removal of redundant and interfering information unrelated to quality from the messy, high-dimensional raw dataset using mathematical statistics and significant correlation analysis, thereby selecting the core input parameters that are most expressive and explanatory to the final quality indicators.

[0026] Post-grouting permeability refers to the seepage prevention performance index of underground rock mass measured by water pressure test after the cement grouting construction is completed and the grout has solidified. It is the ultimate target label for evaluating whether the grouting quality is qualified.

[0027] The core feature refers to the combination of high-value basic physical variables that are extracted from the original multi-source parameters and have a strong causal relationship or high correlation with the final post-irrigation permeability.

[0028] The embodiments of the present invention can determine engineering mechanism information and calculate engineering interaction features based on the engineering mechanism information, so as to actively inject the engineering physical laws and expert experience of cement grouting into the pure mathematical features, break the black box limitation of traditional AI models that rely purely on data correlation, and improve the physical credibility of the prediction model.

[0029] Engineering mechanism information refers to the objective physical and mechanical laws governing the flow, diffusion, and sealing of cement grout in rock fissures, as well as the engineering experience of industry experts.

[0030] Engineering interaction features refer to composite feature indicators generated by combining two or more single basic core features through cross-mathematical operations based on engineering mechanism information. Examples include pressure and flow synergy coefficients that reflect the fracture plugging state, and fracture development and ash injection volume matching coefficients that reflect formation adaptability.

[0031] In this embodiment of the invention, an optimized feature matrix can be constructed by the associated core features and the engineering interaction features to deeply fuse and format-align the selected high-quality basic physical features with the artificially constructed engineering physical mechanism features, and encapsulate them into a high-information-density data entity that can be directly fed to various heterogeneous basic models for efficient training.

[0032] The core features are the core basic features extracted above that are strongly correlated with post-irrigation permeability, which serve as the basic blocks for subsequent matrix splicing in the data flow.

[0033] Optimized feature matrix refers to the standardized, high-density structured dataset generated by concatenating and fusing associated core features, engineering interaction features, and corresponding output labels into a matrix, which is used to directly support the training of AI algorithm groups.

[0034] In this embodiment of the invention, the optimized feature matrix can be used to train multiple initial base model groups composed of heterogeneous base models. This allows multiple machine learning algorithms with different mathematical mechanisms and learning preferences to jointly learn the optimized feature matrix, thereby comprehensively capturing the multi-dimensional predictive advantages of small samples, high nonlinearity, and time-series evolution during dam grouting.

[0035] Heterogeneous underlying models refer to machine learning models with completely different underlying mathematical principles and algorithms, such as support vector regression models that are good at handling nonlinear regression, long short-term memory network models that are good at capturing causal relationships over time, and lightweight gradient boosting tree models that are good at handling structured tabular data.

[0036] The initial basic model group refers to a collection of multiple heterogeneous basic models that have not yet undergone parameter optimization and weight combination and are in an independent and initial running state.

[0037] In this embodiment of the invention, the initial basic model group can be trained through hyperparameter optimization and time-series dynamic weighting operations to construct a fusion model. First, the algorithm automatically optimizes to avoid the empirical errors introduced by manual parameter tuning, ensuring that each independent expert reaches the best prediction state. Then, the speaking weights are dynamically allocated according to the physical evolution characteristics of different grouting construction periods to achieve complementary advantages of multiple models.

[0038] Hyperparameter optimization refers to using optimization algorithms to automatically find the optimal configuration (such as learning rate, number of network layers, etc., which control the training performance of the model) in the solution space of the model, so as to avoid the model getting trapped in local optima or prediction distortion.

[0039] The time-series dynamic weighted operation refers to breaking away from the traditional one-size-fits-all fixed weight combination and dynamically adjusting the weight ratio of each basic model in the overall prediction result according to the different time stages and physical characteristics of dam grouting from the early to the later stages.

[0040] Fusion models refer to an integrated prediction architecture that combines multiple base models that have undergone hyperparameter optimization with a time-series dynamic weighting strategy to form an integrated prediction architecture with higher stability and generalization ability.

[0041] In this embodiment of the invention, the output predicted value of the fusion model can be controlled, and the quality analysis results of cement grouting construction can be generated based on the predicted value. This allows the trained and ready fusion model to perform real-time inversion and deduction of unknown units, and directly transforms the complex digital predicted values ​​into intuitive analytical conclusions that can guide on-site engineers to evaluate and control the grouting quality.

[0042] The predicted value refers to the future or unknown post-irrigation permeability value derived by the fusion model based on the current input construction condition data and through internal algorithms.

[0043] Quality analysis results refer to intelligent quality control and evaluation reports generated based on model predictions, used to quantitatively assess the current underground rock mass seepage prevention quality status and guide construction decisions.

[0044] This invention extracts core features through spatiotemporal mapping of multi-source parameters and a triple screening mechanism, and actively integrates engineering interaction features that reflect the physical evolution law to construct an optimized feature matrix. Combined with hyperparameter optimization and time-series dynamic weighting operations, multiple heterogeneous basic models are synergistically fused, thereby effectively eliminating measurement noise and dimensional interference, overcoming the bottleneck of prediction distortion in complex and variable construction periods by a single traditional model, achieving high-precision, real-time inversion and deduction of the permeability of underground rock mass after grouting, and accurately generating cement grouting construction quality analysis results for guiding on-site quality control.

[0045] Based on the above embodiments, modified embodiments of the above embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in the modified embodiments.

[0046] In an optional embodiment of the present invention, before the step of obtaining the core features strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters, the method further includes: Spatiotemporal correlation mapping is performed on the basic construction data, the dynamic geological parameters, and the construction environment parameters to obtain a multi-source fusion dataset; The multi-source fusion dataset is filtered and sorted according to the data integrity and validity standards. After removing invalid data, a sorted dataset with hole segment identifiers is generated according to the hole order. Based on the borehole segment identifier, the grouting process data in the time series dataset is normalized in two dimensions to obtain the normalized time series dataset.

[0047] In this embodiment of the invention, the basic construction data, the dynamic geological parameters, and the construction environment parameters can be spatiotemporally correlated and mapped to obtain a multi-source fusion dataset. This solves the inconsistency and lack of coordination between the original field data in terms of spatial acquisition points and temporal sampling frequencies. The three types of parameters—basic construction, dynamic geology, and construction environment—that were originally isolated from each other are precisely aligned in a unified four-dimensional time and space coordinate system, establishing a unified correlation link between the multi-source heterogeneous underlying data.

[0048] Basic construction data refers to physical quantities that reflect the characteristics of the grouting process, dynamically recorded by on-site monitoring instruments, such as grouting pressure, grout flow rate, and cumulative grout volume at various times.

[0049] Dynamic geological parameters refer to the geomechanical state parameters of underground rock masses that change in real time depending on the spatial location of the grouting hole section or as construction progresses, such as the original permeability and groundwater flow velocity.

[0050] Construction environment parameters refer to the external boundary environmental characteristics that affect the overall physical state of the construction site, such as mixing temperature, atmospheric temperature, and on-site relative humidity.

[0051] Spatiotemporal correlation mapping refers to a data alignment method that projects heterogeneous data collected at different sampling frequencies and geographical locations onto the same four-dimensional spatiotemporal reference by establishing a unified timestamp and three-dimensional spatial coordinate (X, Y, Z axis) indexing mechanism.

[0052] A multi-source fusion dataset refers to a mixed-source dataset that, after spatiotemporal alignment, combines data from three different sources—construction, geology, and environment—according to a unified spatiotemporal coordinate system, and is thus formed by merging and encapsulating the data with preliminary spatial and temporal correlations.

[0053] In this embodiment of the invention, the multi-source fusion dataset can be screened and sorted according to data integrity and validity standards. After removing invalid data, a sorted dataset with hole segment identifiers is generated according to the hole sequence. This allows for preliminary quality control of the fusion data at a macro level, filtering out completely missing and illogical garbage data caused by power outages, temporary sensor malfunctions, or misoperations. Furthermore, the structure is strictly divided according to the "hole segment" engineering logic of dam grouting, transforming the disordered mixed data stream into an ordered engineering entity dataset.

[0054] Data integrity and validity standards refer to the compliance judgment rules that are pre-set based on the boundaries of on-site construction technology and statistical characteristics of data, and are used to judge whether a certain row or a certain time period of data is usable (for example, a row of key parameters that are completely empty is judged as incomplete, and a value that exceeds the physical range of the equipment is judged as invalid).

[0055] Sequence segmentation refers to the process of logically classifying and grouping continuous data streams based on the technological logic of different construction sequences (such as first-sequence holes, second-sequence holes, and third-sequence holes) and different construction stages in grouting projects.

[0056] The sequence of grouting holes refers to the order in which they are arranged during on-site construction. It is a key engineering logic sequence that determines how the rock mass gradually approaches the design requirements for seepage prevention.

[0057] Hole segment identifiers are digital coded labels specifically used to uniquely identify and mark a particular grouting hole at a specific depth segment (such as the second segment of the third hole).

[0058] Sequential datasets refer to modular, structured datasets that have been cleaned of junk data and where each record clearly identifies the hole segment and its construction sequence.

[0059] In this embodiment of the invention, based on the borehole segment identifier, the grouting process data in the sequence dataset can be subjected to two-dimensional structural regularization to obtain a regularized time-series dataset. This addresses the problem of inconsistent data row lengths and time-series dimensions caused by intermittent work stoppages and sudden changes in grouting speed during cement grouting. Through high-level structural regularization, the dynamic grouting process records of varying borehole segment lengths are recorded, trimmed, filled, and transformed into a standard fixed-length two-dimensional time-series matrix that can be directly read by the algorithm model, ensuring the input stability for subsequent deep feature extraction.

[0060] Grouting process data refers to the time-series data stream recorded at high frequency by sensors within a specific borehole section, from the start of grouting to the end of the sealing process, throughout the entire dynamic continuous construction cycle.

[0061] Two-dimensional structure regularization refers to the process of transforming an originally irregular, non-fixed-length multi-hole data stream into a standard two-dimensional data matrix with consistent row sequence numbers and uniform column feature numbers, without destroying the original temporal causal relationship, through abstract processing methods such as fixed-length pruning of data rows and lossless regularization of missing positions.

[0062] The normalized time series dataset refers to a standardized dataset that is output after being normalized by a two-dimensional structure. Each data row of each construction unit within it has a standard fixed dimension and is arranged strictly in chronological order, and can be directly read by machine learning algorithms.

[0063] This invention effectively solves the problems of data silos and format incompatibility caused by spatiotemporal correlation mapping of original construction, geological and environmental parameters, and after sorting and invalid removal based on engineering logic, it uses borehole segment identification to perform two-dimensional structure regularization on non-fixed length grouting process data. This solves the problems of data silos and format incompatibility caused by spatiotemporal misalignment, dimensional chaos and measurement anomalies in the underlying original data. It realizes the lossless transformation of unstructured complex and messy data on site into a time series dataset with high completeness and high standard dimensions, laying a solid and orderly data matrix foundation for the accurate inversion of core features in the future.

[0064] In an optional embodiment of the present invention, the step of obtaining the core features strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters includes: By utilizing physical constraints based on engineering principles and statistical detection based on box plots and an improved three-standard-deviation criterion, outliers in the normalized time-series dataset are identified and processed to obtain an outlier-corrected dataset. Based on the parameter type, the missing values ​​of the outlier correction dataset are filled using differentiated strategies such as filling with the median under the same conditions, interpolating adjacent wells and correcting for geological stratification, and filling with the mean of the environment during the same period, to obtain a complete filled dataset. Improved Z-score normalization is performed on the target construction parameters of the complete filled dataset, maximum and minimum value normalization is applied to the target geological parameters of the complete filled dataset, and global standardization normalization is applied to the target construction environment parameters of the complete filled dataset to eliminate the dimensional differences and data noise of the target construction parameters, target geological parameters, and target construction environment parameters, thus obtaining an adaptive normalized dataset. Through a triple screening mechanism of Pearson correlation analysis, mutual information entropy calculation, and expert engineering significance verification, core features strongly correlated with post-irrigation permeability are extracted from the adaptive normalized dataset.

[0065] In this embodiment of the invention, physical constraints based on engineering principles and statistical detection based on box plots and an improved three-standard-deviation criterion can be used to identify and process outliers in the normalized time-series dataset, resulting in an outlier correction dataset. This constructs a dual defense barrier of "physical laws" and "mathematical statistics," accurately identifying and removing "dirty data" noise in the time-series dataset caused by instrument electromagnetic interference, sensor failures, or occasional field jumps, preventing outliers from damaging the training accuracy and fitting stability of subsequent algorithm models.

[0066] The normalized time series dataset refers to the original construction time series dataset output by the preceding steps, which is aligned and normalized in space and time and has a standard fixed-length two-dimensional matrix format.

[0067] The physical constraint range based on engineering principles refers to the upper and lower limits of rigid physical values ​​that are pre-set according to the rated maximum range of the equipment at the grouting site, the hydraulic limit of the pipeline, and the maximum pressure bearing boundary of the dam design (for example, the output pressure of the grouting pump cannot be negative; if it is, it is directly judged as a physical anomaly).

[0068] Box plot: A classic statistical test tool that describes the distribution characteristics of data based on statistical quartiles (i.e., the 25th, 50th, and 75th percentiles of data), used to identify deviations at both ends of a distribution sequence without relying on the assumption of a normal distribution.

[0069] The improved three-standard-deviation criterion, also known as the improved Laida criterion, is a statistical outlier that is determined by calculating the mean and standard deviation of data and identifying data that exceeds the range of "mean pm three standard deviations". This "improved" design is made by introducing the median or robust standard deviation to prevent extreme outliers from causing distortion or contamination of the standard deviation itself.

[0070] Outliers refer to erroneous or distorted measurement values ​​in a dataset that significantly deviate from the normal physical measurement range or violate statistical distribution rules.

[0071] Outlier correction datasets refer to clean datasets generated by removing identified outliers (usually by setting them to null values ​​to be filled in the next step).

[0072] In this embodiment of the invention, the missing values ​​of the outlier correction dataset can be filled using differentiated strategies based on parameter type, such as filling with the median under the same conditions, interpolation of adjacent boreholes with geological stratification correction, and filling with the mean of the environment during the same period. This results in a complete filled dataset, which avoids the simple one-size-fits-all approach of filling with the mean. Instead, it deeply integrates engineering physical mechanisms to perform personalized filling for parameters with three distinct physical properties: construction, geology, and environment. This perfectly repairs data gaps without disrupting the continuity of time sequence, restoring the true overall picture of the construction.

[0073] Median filling under the same conditions: This strategy for construction parameters refers to filling in the core median of such process parameters under the same construction stage, the same slurry ratio and other process conditions, so as to ensure that the filling value conforms to the normal operation of the actual process on site.

[0074] Adjacent borehole interpolation plus geological stratification correction: This strategy for dynamic geological parameters first uses data from the nearest known grouting borehole for spatial interpolation, and then combines the physical characteristics of the actual rock geological stratification (such as hard rock or fracture zone) of the current borehole segment with weighted coefficient correction to make the filling depth conform to the abrupt change law of the underground geological structure.

[0075] Simultaneous environmental mean filling: This strategy for construction environmental parameters refers to filling in the average measurement value of the macro environment within the same time period (because environmental temperature and humidity have global consistency within a short period of time and within the same work area).

[0076] The differentiation strategy refers to a refined control mechanism that uses completely different proprietary mathematical and engineering mechanisms to fill in the algorithm for parameters with different physical properties and different evolution laws.

[0077] Missing values ​​refer to blank data positions left over due to missing data in the original data collection or after outlier removal in the previous step.

[0078] A fully filled dataset refers to a dataset that, after undergoing the aforementioned differential fine-grained filling process, eliminates all data gaps and voids, achieving a highly complete dataset with fully closed temporal and spatial sequences and continuous, undistorted information.

[0079] In this embodiment of the invention, improved Z-score normalization is performed on the target construction parameters of the complete filled dataset, maximum-minimum normalization is used on the target geological parameters of the complete filled dataset, and global standardization normalization is used on the target construction environment parameters of the complete filled dataset. This eliminates the dimensional differences and data noise of the target construction parameters, target geological parameters, and target construction environment parameters, resulting in an adaptively normalized dataset. This addresses the "large number eats small number" and gradient explosion problems in model training caused by different physical dimensions (e.g., pressure in megapascals and flow rate in liters per minute, with values ​​differing by hundreds or thousands of times). Through adaptive multi-strategy normalization, all parameters are losslessly mapped to a unified mathematical magnitude range, while further smoothing out residual data noise.

[0080] Target construction parameters / target geological parameters / target construction environmental parameters refer to a subset of specific indicators that are about to undergo dimension conversion, which are classified and allocated from the complete incomplete dataset according to their respective physical attributes.

[0081] Improved Z-score normalization is a dimensionless data transformation method based on mean and standard deviation (Z-score). This method improves upon Z-score by introducing a more robust statistic for outliers, transforming frequently fluctuating construction parameters into dimensionless scores that conform to a standard distribution while preserving their high-frequency dynamic fluctuation characteristics.

[0082] Minimum-maximum normalization, also known as deviation normalization (Min-Max), is a method that linearly and proportionally maps geological parameter data to the range [0, 1]. It is extremely suitable for geological indicators with clear boundaries that require strict preservation of the original spatial distribution.

[0083] Global standardization and normalization refers to breaking the limitations of local samples and, from the global perspective of the entire construction cycle, standardizing and scaling the global mean and variance of environmental parameters with long-term stable characteristics.

[0084] Dimensional differences refer to the phenomenon that, due to differences in physical units and numerical ranges, it is impossible to make a direct and fair mathematical comparison between various characteristics.

[0085] An adaptive normalized dataset refers to a pure numerical normalized dataset that outputs a completely uniform global feature scale and removes the interference of large dimensional numbers after adaptively pairing a dedicated dimensionless normalization algorithm according to the distribution characteristics of the parameter data.

[0086] In this embodiment of the invention, a triple screening mechanism of Pearson correlation analysis, mutual information entropy calculation, and expert engineering significance verification is used to extract core features that are strongly correlated with post-irrigation permeability from the adaptive normalized dataset. This constructs a three-dimensional funnel screening mechanism of "linear correlation + nonlinear mapping + physical engineering mechanism". From the standardized full-dimensional parameter class, redundant features that do not contribute to quality prediction are precisely filtered out like gold, preventing feature overload from causing "curse of dimensionality" and overfitting in the AI ​​model.

[0087] Pearson correlation analysis is a classic mathematical algorithm used to quantitatively measure the degree of linear correlation between two variables. Its value is between [-1, 1] and it is used to quickly capture features that have a direct linear or inverse relationship with post-irrigation permeability.

[0088] Mutual information entropy calculation is a variable correlation measurement method based on information theory. It can effectively capture the deep information dependence between variables that are complex, nonlinear, and even multi-valued mappings, thus making up for the shortcomings of linear correlation analysis.

[0089] Expert engineering significance verification refers to introducing objective engineering experience and mechanical common sense from the dam grouting industry to conduct industry compliance review on the features selected by the mathematical algorithms in the first two steps (for example, some parameters may be mathematically correlated but have no physical causality, so they are eliminated through expert verification).

[0090] The triple screening mechanism refers to a closed-loop feature selection architecture that integrates Pearson (linear), mutual information entropy (nonlinear), and expert verification (physical mechanism) in a progressive manner.

[0091] The core features refer to a set of carefully selected feature variables that have a decisive impact on the permeability of underground rock mass after grouting and have extremely high information density, and which have passed the triple assessment.

[0092] This invention employs a dual mechanism of physical constraints and statistical detection to accurately eliminate abnormal noise. Furthermore, it adaptively matches a "differentiated filling strategy" and a "targeted normalization method" to address the physical heterogeneity of construction, geological, and environmental parameters. Finally, it combines linear, nonlinear, and engineering mechanism triple screening mechanisms to extract core features. This effectively eliminates negative data interference caused by dimensional differences, measurement mutation noise, and data gaps from multiple sources, deeply preserving the physical evolution of on-site time-series information. It achieves extremely high feature information density and clear physical reliability, providing the most perfect and accurate input representation for subsequent multi-model dynamic fusion prediction.

[0093] In an optional embodiment of the present invention, the engineering mechanism information includes the amount of ash injected per unit time, the pressure-flow synergy coefficient, and the crack development-ash injection volume adaptation coefficient. The step of determining the engineering mechanism information and calculating the engineering interaction characteristics based on the engineering mechanism information includes: The core features are used to calculate engineering interaction features including the amount of ash injected per unit time, the synergy coefficient of pressure and flow rate, and the matching coefficient between crack development and ash injection amount. The core features include input parameters and output parameters, and the step of constructing an optimized feature matrix through the associated core features and the engineering interaction features includes: An optimized feature matrix is ​​constructed by concatenating the input parameters, the output parameters, and the engineering interaction features.

[0094] In this embodiment of the invention, engineering interaction features including the amount of grout injected per unit time, the synergy coefficient of pressure and flow rate, and the adaptation coefficient of fracture development and grout injection volume can be calculated through the core features. This transforms the abstract, purely numerical core features into "high-dimensional features" that conform to the physical laws of dam grouting. By actively calculating composite indicators that reflect the evolution of underground rock fractures and the hydraulic diffusion state of grout, the algorithm model is endowed with profound industry physical common sense, solving the drawback of traditional data-driven models that lack physical boundaries and whose prediction results violate engineering common sense.

[0095] The core features refer to the set of basic physical variables with the highest information density extracted by the preceding steps through a triple screening mechanism (such as time, single-step ash injection volume, real-time pressure, real-time flow, etc.).

[0096] The amount of cement mortar injected per unit time refers to the actual mass of dry cement mortar injected into the underground rock mass per unit time during a specific construction period. It is a core dynamic physical quantity that directly reflects the rate of grout accumulation and bonding in the fissures.

[0097] The pressure-flow synergy coefficient is a composite hydraulic index obtained by cross-coupling grouting pressure and grout flow rate. It is used to quantitatively characterize the dynamic synergy state of "pressure surge - flow drop" or "stable pressure and flow" in pipelines and formations. It is the key characteristic for judging whether the formation is about to be saturated.

[0098] The fissure development and grouting volume matching coefficient refers to the evolution coefficient calculated by comparing the fissure opening degree and permeability characteristics of the original geological rock mass in the current site with the current actual grouting rate. It is used to quantitatively assess whether the strength of the current grouting process matches the complex underground geological structure.

[0099] Engineering interactive features refer to high-order derived features with clear engineering physical meanings that break through the limitations of single original features, deeply integrate the engineering mechanism of dam grouting, and cross-combine multiple original core features through mathematical means such as multiplication, division, integration or coupling functions.

[0100] In this embodiment of the invention, the input parameters, the output parameters, and the engineering interaction features can be matrix-concatenated to construct an optimized feature matrix. This allows for standardized data encapsulation, ensuring that the "basic input" used as the training initiation, the "interaction features" representing human engineering experience, and the "prediction labels" representing the ultimate goal are rigorously aligned and horizontally stitched together in a two-dimensional space. This creates a high-value feature matrix with complete information and a unified format, providing a one-stop data foundation for the subsequent collaborative training of heterogeneous model groups.

[0101] Input parameters refer to the subset of basic construction, geological and environmental indicators (such as normalized pressure, flow rate, original permeability, etc.) separated from the core features and used as independent variables of the prediction model.

[0102] Output parameters refer to the target label of the dependent variable that the model training and prediction are benchmarked against. In this case, it specifically refers to the "post-irrigation permeability" that can qualitatively characterize the seepage prevention quality of the dam.

[0103] Matrix splicing refers to a structured assembly operation that merges the input parameter matrix, output parameter sequence, and engineering interaction feature matrix horizontally along the column dimension (feature axis) while ensuring strict alignment of each construction unit (data row). This can be done by a mathematical concat operation.

[0104] The optimized feature matrix refers to the final generated intelligent full-dimensional dataset that perfectly integrates "basic physical features (input) + industry engineering mechanisms (interaction) + ultimate quality indicators (output)" and has extremely high information richness and a standard two-dimensional structure.

[0105] This invention, through deep inversion of core features, calculates a series of engineering interaction features with clear hydraulic diffusion mechanisms, such as "cement injection volume per unit time," "pressure and flow rate coordination coefficient," and "crack development and cement injection volume matching coefficient." The input parameters, output parameters, and engineering interaction features are then structurally and horizontally matrix-wise concatenated. This successfully eliminates the black-box prediction distortion problem caused by the lack of physical constraints in traditional pure data-driven models. It deeply injects the physical evolution of cement grouting into the digital matrix, providing logically rigorous and informationally complete feature matrix support for subsequent multi-heterogeneous model groups to capture microscopic physical and mechanical variations and improve high-precision quality extrapolation across work areas.

[0106] In an optional embodiment of the present invention, the step of training an initial base model group composed of multiple heterogeneous base models using the optimized feature matrix includes: The optimized feature matrix is ​​divided into a training set and a test set. The training set is used to train and generate an initial base model set including a support vector regression model, an improved long short-term memory network model, and a lightweight gradient boosting tree model. With the objective function of minimizing the average absolute percentage error of the fusion model composed of the initial basic model group, the improved Grey Wolf optimization algorithm is used to optimize the core hyperparameters of each model in the initial basic model group, and the optimal parameter combination of each model in the initial basic model group is output. The optimal parameter combination is configured into each model in the initial basic model group, and the fusion weight is calculated based on the average absolute percentage error of each model in the initial basic model group on the test set and the time-series prediction error for different construction periods. Based on the fusion weights and the preset error correction terms, a multi-model dynamic weighted fusion model is constructed.

[0107] In this embodiment of the invention, the optimized feature matrix can be divided into a training set and a test set. The training set is used to train and generate an initial basic model group that includes a support vector regression model, an improved long short-term memory network model, and a lightweight gradient boosting tree model. This allows for the use of heterogeneous machine learning algorithms with completely different mathematical mechanisms, learning preferences, and structural characteristics to perform multi-dimensional synchronous fitting on the feature matrix infused with engineering mechanisms. This fully explores and utilizes their complementary predictive advantages in dealing with small samples, nonlinearity, and time series evolution in the dam grouting process.

[0108] The optimized feature matrix refers to the standard two-dimensional matrix output from the preceding steps, which horizontally incorporates the basic physical input, engineering mechanism interaction, and post-irrigation permeability output labels.

[0109] The training set and test set refer to the existing historical feature matrix being logically divided into two parts according to a certain ratio (such as 80% and 20%). The "training set" is used to feed the AI ​​model to learn the rules, while the "test set" does not participate in the training at all and is reserved for subsequent blind testing to verify the model's prediction accuracy and generalization ability.

[0110] Support Vector Regression (SVR) is a classic machine learning algorithm based on the principle of minimizing structural risk. It excels at finding the global optimum through high-dimensional kernel function mapping under small sample conditions, and here it mainly leverages its strong generalization advantage on small sample data.

[0111] The improved Long Short-Term Memory (LSTM) network model is a deep learning recurrent neural network specifically designed for processing time series. By introducing a gating mechanism, it has long-term memory capabilities. The "improvement" here aims to deeply capture the temporal trend characteristics of the grouting process as it evolves backward along the construction time axis, with close causal connections between the preceding and following phases. In other words, it is an LSTM suitable for grouting time series.

[0112] Lightweight Gradient Boosting Tree Model (LightGBM) is a highly efficient distributed gradient boosting framework based on decision trees. It runs extremely fast, consumes very little memory, and has a natural and powerful ability to fit multidimensional table features. Here, it mainly leverages its advantages in rapidly scanning and nonlinear segmentation of mixed static geological and dynamic construction parameters.

[0113] The initial basic model group refers to the collection of three heterogeneous machine learning algorithm models—SVR, LSTM, and LightGBM—that have not yet undergone parameter optimization and weight integration and are in an independent, unrefined state.

[0114] In this embodiment of the invention, the objective function is to minimize the average absolute percentage error of the fusion model composed of the initial basic model group. The improved Grey Wolf optimization algorithm is used to optimize the core hyperparameters of each model in the initial basic model group and output the optimal parameter combination of each model in the initial basic model group. This method rejects the traditional parameter tuning method that relies on trial and error based on human experience. It unifies the internal parameter configurations of three independent AI experts into a single optimization problem. By simulating the collective intelligence evolution of wolf packs in nature, it automatically equips each heterogeneous model with the golden core parameters that can unleash its highest predictive potential, thus avoiding overfitting or prediction distortion.

[0115] A fusion model refers to a composite large model architecture that integrates multiple basic heterogeneous algorithms to jointly determine the final output result.

[0116] Mean absolute percentage error (MAPE) is a regression evaluation index used to quantitatively measure the relative degree to which predicted values ​​deviate from the true values. The smaller the value, the higher the prediction accuracy. Here, it serves as the ultimate indicator for controlling the overall model error.

[0117] The objective function refers to the core mathematical expression that needs to be minimized or maximized during the mathematical optimization process (here, it means "controlling the MAPE of the global fusion model to reach the theoretical minimum value").

[0118] The Improved Grey Wolf Optimization Algorithm (IGWO) is a swarm intelligence global optimization algorithm that simulates the predation behavior and strict social hierarchy of grey wolves (Alpha, Beta, Delta wolves). By introducing specific improvement mechanisms (such as adaptive weights and nonlinear convergence factors), the algorithm is prevented from getting stuck in local optima when optimizing the hyperparameter space.

[0119] Core hyperparameters refer to the key control parameters that must be set manually before algorithm training to determine the model's learning framework and evolution speed (such as the penalty factor and kernel radius of SVR, the number of hidden layer nodes and learning rate of LSTM, and the number of leaf nodes and tree depth of LightGBM).

[0120] The optimal parameter combination refers to the set of hyperparameter values ​​that, after improving the Grey Wolf algorithm through brute force and fine-grained search in the global space, can minimize the overall prediction error of the fusion model.

[0121] In this embodiment of the invention, the optimal parameter combination can be configured into each model in the initial basic model group. Based on the average absolute percentage error of each model in the initial basic model group on the test set and the temporal prediction error of different construction periods, the fusion weight is calculated to break the rigid mechanism of fixed weights in traditional ensemble learning. By evaluating the "big test score" (static error) of each model on the overall static test set and the "real-world temporal performance" (dynamic error) at different grouting stages (such as the early stage of high flow in cracks and the later stage of gradual sealing), the contribution of each model at different time periods is adaptively quantified, so that the model group can deeply adapt to the dynamic evolution of grouting physical characteristics.

[0122] Configuration refers to the process of back-injecting and writing the optimal hyperparameter values ​​obtained through optimization into the initialization code of each heterogeneous model, so that the algorithm can restart in the best possible state.

[0123] The time-series prediction error for different construction periods refers to the dynamic inversion error exhibited by the model at these different time sections and process stages, which fluctuate with time due to the large amount of grout absorbed in the early stage and the gradual saturation and sealing in the later stage, resulting in drastically different physical characteristics before and after the grouting project.

[0124] Fusion weight refers to the weight, contribution, or percentage coefficient of each independent heterogeneous model in the final integrated output.

[0125] In this embodiment of the invention, a multi-model dynamic weighted fusion model can be constructed based on the fusion weights and preset error correction terms to execute the final "collaborative team battle architecture" encapsulation. The calculated dynamic weight coefficients are mathematically combined with the safety valve (error correction term) used to resist occasional fluctuations on site, and a multi-model fusion body for intelligent simulation of dam grouting with ultra-high fitting accuracy, strong noise resistance and time-series adaptive control capability is formally assembled.

[0126] The preset error correction term refers to a set of mathematical compensation factors (equivalent to adding a safety regulator to the prediction results) that are pre-set at the output of the fusion model based on historical engineering experience or residual distribution patterns to smooth out sudden changes and offset inherent biases of the system.

[0127] The multi-model dynamic weighted fusion model refers to the final completed integrated prediction model composed of a heterogeneous group of SVR, LSTM and LightGBM algorithms with optimal hyperparameter configuration, whose speaking ratio dynamically switches with the construction period and includes error correction.

[0128] This invention first divides the training and testing sets and introduces SVR, LSTM, and LightGBM, which have different mathematical preferences in terms of small sample size, temporal evolution, and tabular features, to construct a heterogeneous model group. Then, guided by minimizing the MAPE of the fusion model, the optimal hyperparameter combination of each model is automatically locked by combining the improved Grey Wolf optimization algorithm. The fusion weights are refined by combining the performance of the static test set and the dynamic temporal prediction errors of different construction periods. With the help of preset error correction terms, the construction of a multi-model dynamic weighted fusion model is completed. This effectively solves the problem that traditional single machine learning algorithms or fixed-weight fusion models are prone to getting stuck in local optima, losing sight of the bigger picture, and making predictions inaccurate when facing the complex and variable dam grouting life cycle. It realizes the deep complementary advantages of multiple heterogeneous models at the algorithm mechanism level, and provides a solid and adaptive intelligent model foundation for subsequent stable and high-precision post-grouting permeability prediction throughout the entire process.

[0129] In an optional embodiment of the present invention, the step of controlling the output prediction value of the fusion model includes: The construction unit data to be predicted in the adaptive normalized dataset is input into the multi-model dynamic weighted fusion model to obtain the final post-irrigation permeability prediction value.

[0130] In this embodiment of the invention, the construction unit data to be predicted in the adaptive normalized dataset can be input into the multi-model dynamic weighted fusion model to obtain the final predicted value of post-grouting permeability. The real-time on-site test data obtained after previous cleaning, normalization, and standardization is used as an independent variable to drive the fusion model that has been trained and is configured with optimal hyperparameters and dynamic weights. Through the black-box network and matrix inference inside the algorithm, the real-time online inversion of the final seepage prevention quality index of underground hidden rock mass after grouting construction is realized.

[0131] An adaptive normalized dataset refers to a pure numerical normalized dataset that, in the preceding steps, performs improved Z-score normalization, maximum and minimum value normalization, and global standardization normalization on the construction parameters, geological parameters, and environmental parameters in the fully filled dataset, respectively. The output is a pure numerical normalized dataset with completely unified global feature scale and free from interference from large dimensional numbers.

[0132] The data for the construction unit to be predicted refers to the real-time feature dataset of a specific borehole section that is currently undergoing grouting construction or has just been completed at the dam site, and whose final seepage prevention performance (post-grouting permeability) has not yet been measured by physical pressure water test, and which urgently needs to be intelligently analyzed.

[0133] The multi-model dynamic weighted fusion model refers to the final completed integrated prediction model composed of a heterogeneous group of algorithms, namely Support Vector Regression (SVR), Improved Long Short-Term Memory Network (LSTM), and Lightweight Gradient Boosting Tree (LightGBM), with its speaking ratio dynamically switching according to the construction period and with residual error correction.

[0134] The final post-grouting permeability prediction value refers to the final permeability value (such as Lu value) that the multi-model dynamic weighted fusion model outputs after receiving input features, undergoing multi-path algorithm collaborative deduction, dynamic weighted summation, and residual correction. This value characterizes the theoretically achievable permeability value of the construction section after the grout solidifies.

[0135] This invention directly inputs the real-time on-site construction unit data to be predicted, after adaptive normalization processing, into a multi-model dynamic weighted fusion model that has completed hyperparameter locking. This effectively eliminates the dimensional barriers and microscopic data noise between multi-source input parameters, enabling high-precision, digital, and intelligent real-time online inversion and deduction of the final seepage prevention quality indicators of underground rock mass concealed grouting projects without waiting for lengthy physical pressure water tests. This provides the most crucial quantitative data basis for the immediate evaluation of subsequent dam construction site quality control.

[0136] In an optional embodiment of the present invention, the step of generating quality analysis results for cement grouting construction based on the predicted values ​​includes: The performance of the predicted post-irrigation permeability is evaluated using mean square error, mean absolute error, mean absolute percentage error, and prediction stability coefficient to obtain model performance indicators. When it is determined that the model performance index has not reached the preset accuracy or the engineering geological environment has changed, the multi-model dynamic weighted fusion model is triggered to enter an adaptive update process including incremental training and transfer learning, so as to freeze the core parameters of the multi-model dynamic weighted fusion model and adjust the target network layer weights of the multi-model dynamic weighted fusion model to obtain an adaptively optimized fusion model. The adaptively optimized fusion model, the optimized feature matrix, and the post-grouting permeability prediction value are integrated and input into an intelligent visualization system developed based on a graphical user interface component and a dynamic chart engine to generate a visualized cement grouting construction quality control and decision-making suggestion report.

[0137] In this embodiment of the invention, the mean square error, mean absolute error, mean absolute percentage error, and prediction stability coefficient can be used to evaluate the performance of the predicted post-irrigation permeability value, thereby obtaining model performance indicators. This allows for the construction of a comprehensive, multi-dimensional monitoring radar for model accuracy and stability. By performing objective mathematical assessments on the predicted output values, the reliability of the current fusion model under the current construction conditions can be quantitatively evaluated in real time, providing a scientific trigger criterion for whether the model needs to evolve in the future.

[0138] The post-irrigation permeability prediction value refers to the intelligent inference value output by the multi-model dynamic weighted fusion model at the output end, which is used to characterize the final seepage prevention quality status of the dam's hidden works.

[0139] Mean squared error (MSE) is the expected value of the square of the difference between the predicted value and the true value. Because it is extremely sensitive to outlier extreme errors, it is mainly used here to amplify and capture the severe prediction distortion that occasionally occurs in the fusion model.

[0140] Mean absolute error (MAE) refers to the average of the absolute values ​​of the difference between the predicted value and the actual value. It can objectively and truthfully reflect the actual absolute deviation of the predicted value in terms of physical dimensions.

[0141] Mean Absolute Percentage Error (MAPE) refers to the average absolute value of the relative error. It visually represents the relative degree of prediction deviation in percentage form and is a global accuracy indicator that is not affected by the absolute size of physical quantities.

[0142] The Predictive Stability Coefficient (PS) is a characteristic coefficient used to quantitatively evaluate the magnitude of predictive fluctuations and robustness against noise in a model as it progresses over a continuous time series. It is used to prevent the model from exhibiting drastic oscillations where it is extremely accurate at one moment and extremely poor at the next.

[0143] Model performance metrics refer to the comprehensive quantitative evaluation results of the current AI prediction architecture's "accuracy + robustness," which are composed of the above four mathematical evaluation metrics.

[0144] In this embodiment of the invention, when the model performance index fails to reach the preset accuracy or the engineering geological environment changes, the multi-model dynamic weighted fusion model is triggered to enter an adaptive update process including incremental training and transfer learning. This process freezes the core parameters of the multi-model dynamic weighted fusion model and adjusts the target network layer weights to obtain an adaptively optimized fusion model. This completely solves the problem of "incompatibility" and "increasing inaccuracy" of intelligent prediction architectures when dams cross work areas, cross strata, or face sudden complex fault fracture zones. Through a dynamic evolution mechanism of "freezing general knowledge and fine-tuning the antenna network," the model can be upgraded and adapted to the environment with very little new field data.

[0145] Preset accuracy refers to a set of threshold boundaries for indicators set in advance based on the minimum tolerance of the dam design institute and the engineering supervision party for seepage prevention quality control (for example, the preset accuracy threshold can be MAPE≤8% for sequence I / II, ≤12% for sequence III, PS≥0.85, etc. If the model error exceeds the standard by the preset accuracy threshold, it is judged as not meeting the standard).

[0146] Variation in the engineering geological environment refers to the objective phenomenon that, as the grouting section advances deeper underground or into different curtain depths, it encounters unexpected abrupt changes in rock strata, giant karst caves, underground rivers, or fracture zones, resulting in a complete alteration of the physical and mechanical background.

[0147] Incremental training and transfer learning: an advanced AI secondary evolution training mechanism. "Incremental training" refers to adding new data streams generated in the field without overhauling the original model; "transfer learning" refers to applying the general physical laws (knowledge) learned by the model in the existing geological environment to a completely new heterogeneous environment.

[0148] The adaptive update process refers to an online model upgrade and retraining mechanism that automatically sounds an alarm and executes the process when the triggering conditions are met.

[0149] Freezing core parameters means keeping the weight matrix of the underlying deep kernel network responsible for extracting general physical and hydraulic diffusion laws in the fusion model fixed during the secondary training process, thus locking its general knowledge base.

[0150] Adjusting the weights of the target network layer means, with the core parameters locked, only opening the parameters of the top / specific network layer that directly connects to the new geological conditions and is responsible for the output mapping, and using a small number of new samples to perform targeted fine-tuning and weight reconstruction.

[0151] The adaptively optimized fusion model refers to an upgraded closed-loop fusion prediction model that has completed incremental migration and evolution, retained general mechanism knowledge, and deeply adapted to the new geological environment after the current mutation.

[0152] In this embodiment of the invention, the adaptively optimized fusion model, the optimized feature matrix, and the predicted post-grouting permeability can be integrated and input into an intelligent visualization system developed based on a graphical user interface component and a dynamic chart engine. This generates a visualized report on cement grouting construction quality control and decision-making suggestions, bridging the final mile from technical solutions to engineering implementation. The complex and obscure calculation results of the large digital model at the back end are translated into concrete graphics and control strategies that can be directly understood by the chief engineer, quality inspector, and supervisor on site through a modern front-end rendering architecture, formally forming the ultimate productivity closed loop of "data + features + prediction + closed-loop update + decision guidance".

[0153] The optimized feature matrix refers to a standardized, high-density structured dataset that perfectly integrates basic input parameters, industry engineering mechanism interactions, and quality output labels, outputted from the preceding steps.

[0154] Graphical User Interface (GUI) components refer to standardized front-end control units (such as windows, buttons, input boxes, layout containers, etc.) used to build human-computer interaction interfaces, and are responsible for carrying the overall operating framework of intelligent systems.

[0155] A dynamic chart engine refers to the underlying core of data visualization (such as a data-driven chart library and rendering engine) specifically designed for high-frequency rendering and smooth display of highly interactive charts such as dynamic time-series curves, 3D geological cloud maps, and real-time pressure-flow scatter plots on web pages or software terminals.

[0156] The intelligent visualization system refers to a digital software platform that integrates the aforementioned interface components and chart engine, allowing on-site construction management personnel to directly interact with, monitor, and evaluate the dam grouting quality.

[0157] The visualized cement grouting construction quality control and decision-making recommendation report refers to a concrete engineering terminal report that is automatically generated by an intelligent visualization system. It includes intuitive traffic light quality warnings, charts showing the future trend of permeability of each borehole section, and specific action recommendations for process control (such as adjusting grouting pressure and thickening grout ratio) for abnormal borehole sections.

[0158] This invention implements comprehensive accuracy monitoring of predicted values ​​by introducing MSE, MAE, MAPE, and prediction stability coefficients. When predictions exceed limits or geological changes occur, it adaptively triggers incremental and transfer learning mechanisms that "freeze the core and fine-tune the target layer." Finally, it seamlessly integrates and outputs the model, features, and prediction results by combining graphical user interface components and a dynamic chart engine. This effectively solves the industry-wide catastrophic pain point of intelligent prediction models experiencing a sharp drop in generalization performance when facing cross-work areas or sudden complex geological environments, and being unable to be continuously applied across scenarios. It achieves on-orbit closed-loop self-evolution of the prediction architecture and transforms digital results into visualized cement grouting decision reports with extremely high visual readability that can directly guide on-site quality control.

[0159] To enable those skilled in the art to better understand the embodiments of the present invention, an example is used below to illustrate the embodiments of the present invention.

[0160] Step 1: Multi-dimensional data collection and correlation fusion First, basic construction data was collected, strictly following the requirements of the "Technical Specification for Cement Grouting Construction of Hydraulic Structures": Grouting process data was collected every 5 minutes, with average values ​​recorded, including operation type, borehole number, borehole sequence, section number, time, grouting volume (Tol_q, unit L), flow rate (q, unit L / min), pressure (p, unit MPa), grouting inlet volume (q_in, unit L / min), grout return volume (q_out, unit L / min), and cement injection volume (cet, unit Kg). Grouting statistics were compiled using a section statistics table as the core, recording... Hole number, hole sequence, segment number, top of grouting hole segment (unit: m), bottom of grouting hole segment (unit: m), segment length (unit: m), permeability before grouting (unit: Lu), initial injection rate (unit: L / min), final injection rate (unit: L / min), grouting volume (unit: L), ash injection volume (unit: Kg), unit ash injection volume (unit: Kg / m), grouting pressure (unit: MPa), pure grouting time (unit: min); grouting end data record: segment length, grouting volume, ash injection volume, construction start and end time, grouting time, and permeability before grouting (unit: Lu).

[0161] Dynamic geological parameters are collected in real time using logging-while-drilling equipment, including rock mass integrity coefficient (Kv, unitless, range 0-1, representing the degree of rock mass integrity, 1 for completely intact, 0 for completely broken), fracture attitude (strike α, unit °, range 0-360°; dip β, unit °, range 0-90°; dip γ, unit °, range 0-360°), and rock mass weathering degree (W, graded 1-5, grade 1 for unweathered, grade 2 for slightly weathered, grade 3 for moderately weathered, grade 4 for strongly weathered, and grade 5 for completely weathered). At the same time, environmental parameters during construction are collected using environmental monitoring equipment, including ambient temperature (T, unit °C, accuracy ±0.1 °C), ambient humidity (RH, unit %, accuracy ±1%), and wind speed in the construction area (v_w, unit m / s, accuracy ±0.1 m / s).

[0162] Using "hole number-hole sequence-segment number-collection timestamp" as a unique four-dimensional identifier, basic construction data, dynamic geological data, and environmental data are linked and fused using Python's Pandas library to ensure that all parameters in the same construction period correspond one-to-one, forming a multi-source dataset, where the timestamp is accurate to the second and the data deviation is ≤1 minute.

[0163] Step 2: Data Filtering and Sorting The multi-source datasets were cross-validated using Python to filter complete data for construction units 7-18. The filtering criteria were: missing data rate of single borehole segment ≤3% (missing rate = number of missing fields / total number of fields × 100%), no invalid data caused by equipment failure (such as parameter sudden change to 0 due to sensor power failure) or signal interference (such as irregular pressure fluctuations that are not related to flow rate), and the timestamp deviation between dynamic geological data and construction data ≤1min.

[0164] The filtered data is divided into sequences according to the hole sequence (Sequence I, Sequence II, Sequence III). The smallest data unit is the "hole segment". Each data entry generates a unique identifier in the format of "Construction Unit Number - Hole Number - Hole Sequence - Segment Number - Timestamp" (e.g., "7-XW1-Z-1-20240520103000" represents data collected at 10:30:00 on May 20, 2024, for Unit 7, Hole XW1-Z-1, Sequence I, Segment 1).

[0165] Only data for the "grouting" operation type is retained, and sorted using Python in the order of "Construction Unit Number → Borehole Number → Segment Number → Grouting Volume". The grouting process data is then uniformly organized into 20 rows × 16 columns (including the original 9 construction fields + 3 dynamic geological fields + 3 environmental fields). Rows shorter than 20 are filled with the weighted average of adjacent data within the same borehole segment. The formula for calculating the weighted average is...

[0166] , where xi is the adjacent valid data, wi is the weight (distributed inversely according to time distance, i.e., wi=1 / di, di is the time difference between the current missing time and the adjacent data time, in min), and more than 20 rows of valid data rows with a mean difference of less than 3% are retained (mean difference = |current row data - mean of data in this well segment| / mean of data in this well segment × 100%).

[0167] Step 3: Enhanced Data Preprocessing (1) Outlier handling A dual mechanism of "physical constraints + statistical detection" is adopted: First, reasonable ranges for parameters are set based on the physical principles of grouting engineering, specifically: grouting pressure 0.1-5MPa, flow rate 0.1-20L / min, rock mass integrity coefficient 0≤Kv≤1, ambient temperature -10-45℃, ambient humidity 10%-95%, and wind speed 0-10m / s. Data exceeding these ranges are directly marked as extreme outliers. For data within the physical constraints, "box plot + improved 3σ criterion" is used with Python's SciPy library for detection. The box plot calculates the lower quartile (Q1), upper quartile (Q3), and interquartile range (IQR = Q3-Q1). Data < Q1-1.5×IQR or > Q3+1.5×IQR are marked as statistical outlier candidates. The improved 3σ criterion is then used for verification. The calculation formula is as follows: Where x is the data to be detected, μ is the sample mean of the parameter, and σ is the sample standard deviation, a value satisfying this formula is determined to be a statistical outlier. Extreme outliers are directly deleted from the corresponding borehole segment data, while statistical outliers are replaced by the median under the same geological conditions (consistent rock weathering degree and fracture density) for that borehole sequence.

[0168] (2) Handling missing values When construction parameters are missing, if a single borehole segment is missing ≤ 2 fields, the median of the corresponding parameters for the same sequence of boreholes is used to fill the gap; if more than 2 fields are missing, the borehole segment is discarded. When dynamic geological data is missing, "interpolation between adjacent boreholes + geological stratification correction" is used, and the formula is as follows:

[0169] Where xfill is the geological parameter value to be filled, xleft is the corresponding geological parameter of the adjacent borehole on the left, xright is the corresponding geological parameter of the adjacent borehole on the right, dleft is the distance between the current borehole and the adjacent borehole on the left (in meters, measured by the construction drawings), dright is the distance between the current borehole and the adjacent borehole on the right (in meters), and Kw is the weathering degree correction coefficient (Kw=1.0 when W=1, Kw=0.95 when W=2, Kw=0.9 when W=3, Kw=0.85 when W=4, and Kw=0.8 when W=5); when environmental data is missing, the average environmental parameter value of the same construction unit and the same time period (interval ≤1h) is used to fill the missing data.

[0170] (3) Data normalization The "parameter type adaptive normalization" method is adopted: construction parameters (flow rate, pressure, grouting volume, etc.) are normalized using a modified Z-Score, with the formula as follows: Where x′ is the normalized data, x is the original data, μ is the sample mean of the parameter, σ is the sample standard deviation, and ε=10. 8 (To avoid failure when σ=0); Dynamic geological parameters (rock mass integrity coefficient, fracture density, etc.) are normalized using Min-Max, and the formula is as follows:

[0171] Where xmin and xmax are the sample minimum and maximum values ​​of the parameter; environmental parameters (temperature, humidity, etc.) are standardized and normalized, and the formula is: σall is the global standard deviation of this environmental parameter (calculated based on data from all construction units). All normalization operations are implemented using Python's Scikit-learn library, and the same μ, σ, xmin, xmax, and σall are used to calculate the predicted data.

[0172] Step 4: Feature Engineering and Input / Output Determination Pearson correlation analysis was performed using SPSS software to calculate the correlation coefficients between each feature and post-irrigation permeability, retaining features with an absolute correlation value ≥ 0.2. Then, the mutual information entropy between each feature and post-irrigation permeability was calculated using the Python Scikit-learn library, selecting features with an entropy value ≥ 0.1. The formula for mutual information entropy is:

[0173] Where X is the characteristic variable, Y is the post-grouting permeability, P(x,y) is the joint probability distribution, and P(x) and P(y) are the marginal probability distributions; three engineers with more than 10 years of grouting engineering experience were invited to evaluate the engineering significance of the screened characteristics, and characteristics with an evaluation pass rate of ≥80% were retained.

[0174] Structural engineering characteristics: amount of mortar injected per unit time (Unit: kg / min, cet is the amount of ash injected, tpure is the pure filling time); Pressure-flow coordination coefficient (unit MPa) (min / L, p is the grouting pressure, q is the flow rate); Crack development - ash injection amount matching coefficient (Unit Article) Kg / m², ρ is the fracture density, and L is the segment length.

[0175] The input parameters are defined as follows: 18 items: section length (m), pre-grouting permeability (Lu), initial injection rate (L / min), final injection rate (L / min), grouting pressure (MPa), grouting volume (L), grout volume (Kg), pure grouting time (min), rock mass compressive strength (MPa), fracture density (fractures / m), groundwater level (m), groundwater velocity (m / s), rock mass integrity coefficient (unitless), fracture orientation (°), fracture dip angle (°), ambient temperature (°C), grout volume per unit time (Kg / min), and pressure-flow synergy coefficient (MPa). The output parameter is the post-irrigation permeability (unit: Lu). The "pre-irrigation permeability of the next sequence replaces the post-irrigation permeability of the previous sequence" (only within the same construction unit). The calculation steps are as follows: the pre-irrigation permeability of the borehole section is taken from the last value of the pressure test data of that borehole section; the pre-irrigation permeability of a single borehole is the average pre-irrigation permeability of all borehole sections in that borehole, rounded to two decimal places; the pre-irrigation permeability of a process is the average pre-irrigation permeability of all single boreholes in that process, rounded to two decimal places.

[0176] Step 5: Basic Model Training (1) The kernel function for training the IGWO-SVR core model uses the radial basis function (RBF), and the formula is as follows:

[0177] Where g is the kernel function parameter (g > 0), x and z are the processed sample data vectors, and ||x| z|| is the Euclidean distance; the SVR regression function is...

[0178] Where ai* and ai are Lagrange multipliers (0≤ai≤C, 0≤ai*≤C, C is the penalty coefficient), b is the bias vector, and m is the number of samples.

[0179] The training / test set split ratios were set to I-order (318 / 43), II-order (284 / 33), and III-order (624 / 60), with a batch size of 32. Ten-fold cross-validation was used for training via Python's Scikit-learn library, which involved dividing the training data into 10 subsets, using one subset as the validation set in turn, and the remaining 9 subsets as the training set. The average loss from the 10 training iterations was taken as the final loss.

[0180] (2) Improve LSTM model training The model structure is as follows: input layer (18 neurons, corresponding to 18 input parameters) → LSTM layer (I / II order is [64-32], i.e., 64 neurons in the first layer and 32 neurons in the second layer; III order is [128-64], i.e., 128 neurons in the first layer and 64 neurons in the second layer) → fully connected layer (32 neurons) → output layer (1 neuron, corresponding to post-irrigation permeability). The activation function is ReLU (ReLU(x)=max(0,x)), and the training parameters are set as follows: learning rate I order = 0.0002, II order = 0.0004, III order = 0.0003, number of iterations Num_epochs = 800, loss function is MAE, and training is implemented using the PyTorch framework.

[0181] (3) LightGBM model training The parameters are set as follows: max_depth=8 (maximum depth of decision tree), learning_rate=0.08 (learning rate), n_estimators=150 (number of decision trees), subsample=0.85 (sample sampling rate), colsample_bytree=0.85 (feature sampling rate), num_leaves=31 (number of leaf nodes); training is performed using five-fold cross-validation with Python's LightGBM library, and an early stopping mechanism is added, i.e., training stops if the MAE of the validation set does not decrease for 10 consecutive rounds.

[0182] Step 6: Improve and optimize the IGWO algorithm (1) Algorithm improvement The formula for correcting the nonlinear convergence factor is as follows:

[0183] Where p = 2.5 (decay order), t is the current iteration number, Maxiter = 500 (maximum iteration number), and e is the natural constant (≈2.71828).

[0184] The adaptive weight update formula is:

[0185]

[0186]

[0187] Where MAPEα, MAPEβ, and MAPEδ are the mean absolute percentage errors of the models corresponding to α, β, and δ gray wolves, respectively.

[0188] yk is the actual value. The predicted value is n, where n is the number of samples. In the later stages of iteration (t > 0.7 * Maxiter), a chaotic perturbation is added to the optimal position of the gray wolf population, as shown in the formula:

[0189] Where Xopt is the current optimal position, Logistic(r) = 4r(1) r) (r is a random number in the range [0,1]).

[0190] (2) Parameter optimization The optimization objective is to minimize the MAPE of the fusion model. The search range is set as follows: the penalty coefficient C of SVR is in the range of [0.001, 5000], the kernel function parameter g is in the range of [0.001, 5000], the learning rate of LSTM is in the range of [0.0001, 0.001], the number of neurons in LSTM layer is in the range of [32, 256], and the max_depth of LightGBM is in the range of [3, 10], and the learning_rate is in the range of [0.01, 0.2].

[0191] The population size was set to I / II order = 40, III order = 30, and the number of iterations Maxiter = 500. The IGWO algorithm was implemented using Python programming to output the optimal parameter combination for each model.

[0192] Step 7: Construction of Dynamic Weighted Fusion Model Static weights are calculated based on the MAPE test set for each model, using the following formula:

[0193] Where k=1 corresponds to IGWO-SVR, k=2 corresponds to improved LSTM, and k=3 corresponds to LightGBM; The time-series weights are dynamically adjusted based on the prediction errors of the model at different construction stages, as shown in the formula:

[0194] Where MAPEk,t is the prediction error of the k-th model in the t-th construction period (each period is 2 hours); the fusion weight is wk=0.6×wks+0.4×wkt (static weight accounts for 60%, and time series weight accounts for 40%).

[0195] The fusion prediction formula is

[0196] in This is the final predicted value. Let be the predicted value of the k-th model, and Δy be the error correction term.

[0197] (yi is the true value of the i-th sample, (where n is the number of samples and the predicted value of the k-th model for the i-th sample) Model performance is evaluated using mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and prediction stability coefficient (PS), with the following formulas:

[0198]

[0199]

[0200] (std is the standard deviation, mean is the mean) The requirements for the fused model are: MAPE ≤ 8% (Sequence I / II), ≤ 12% (Sequence III), and PS ≥ 0.85.

[0201] Step 8: Dynamic Adaptive Optimization When the predicted data for the new project meets the criteria of "MAPE > threshold" (8% for sequence I / II, 12% for sequence III) or "PS < 0.85", an adaptive update is triggered. The effective data of the new project (after preprocessing in steps 1-3) is added to the original training set according to borehole sequence and geological type (fractured area, moderately permeable area, weakly permeable area). The core parameters of the model (C and g optimized by IGWO, and the LSTM layer weights of LSTM) are frozen, and only the parameters of the output layer and fully connected layer are fine-tuned. The learning rate is reduced to 1 / 5 of the original learning rate.

[0202] When the new engineering geological type differs significantly from the original training set (mean difference in rock mass integrity coefficient > 0.3), transfer learning is employed. The original model is used as a pre-trained model, 70% of the network layer weights are frozen, and the remaining 30% of the layers are trained using the new engineering data to accelerate model adaptation. The fusion weights are recalculated to generate a fusion model adapted to the new engineering project. The model version is saved (.pth format), and an adaptation log (including data source, parameter adjustments, and performance metrics) is recorded.

[0203] Step 9: Implementation and Application of Intelligent Visualization System This system utilizes Python's PyQt5 library and ECharts visualization library to develop an intelligent visualization system. The system's functional modules operate as follows: Clicking the "Data Import" button allows you to select a multi-source data file in Excel / CSV format. The system automatically verifies data integrity, marks missing fields and outliers, and generates a data quality report (including missing rate, number of outliers, and data distribution statistics). Clicking "Feature Analysis" generates a feature importance heatmap, parameter correlation network diagram, and time-series feature trend chart, and labels core influencing factors (the top 5 features by importance score). Inputting the dynamic geological parameters of the current construction unit allows the system to automatically match the optimal construction parameters (grouting) under similar historical geological conditions. The system displays the pressure and injection rate range, and outputs a matching score (0-100 points). Click "Model Import" to import the fusion model file (.pth format) of the corresponding borehole sequence, select the target borehole sequence, and click "Start Prediction." The system outputs the predicted values, error values, and prediction stability coefficients of the fusion model and each basic model in real time. Based on the prediction results and geological analysis, the system automatically generates construction adjustment suggestions (e.g., if the predicted permeability is too high, it is recommended to increase the grouting pressure by 0.2-0.5MPa or extend the pure grouting time by 5-10 minutes). Click "Export Report" to export a PDF prediction report, which includes data preprocessing results, feature analysis charts, prediction results, error analysis, construction suggestions, etc.

[0204] Step 10: Engineering Application Verification Three new construction units with different geological conditions (fractured zone, moderately permeable zone, and weakly permeable zone) were selected. Each unit contained 20 grouting holes with 156, 142, and 138 hole segments, respectively. Data acquisition and preprocessing were completed according to steps 1-3. The preprocessed data was input into the fusion model, and the predicted permeability values ​​after grouting for each sequence were output. The relative error between the predicted and actual values ​​and the prediction stability coefficient (PS) were calculated.

[0205]

[0206] The verification results are required to be: the relative error of sequence I holes ≤ 6%, sequence II holes ≤ 4%, sequence III holes ≤ 10%, and the prediction stability coefficient PS ≥ 0.88. If a sequence does not meet the requirements, the adaptive optimization in step 8 is triggered, and the data of that unit is supplemented for incremental training and transfer learning until the error and stability meet the requirements.

[0207] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0208] Reference Figure 2 The diagram shows a structural block diagram of a grouting construction quality analysis device provided in an embodiment of the present invention, which may specifically include the following modules: Data acquisition module 201 is used to acquire basic construction data, dynamic geological parameters, and construction environment parameters; The core feature acquisition module 202 is used to acquire core features that are strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters. The engineering interaction feature calculation module 203 is used to determine engineering mechanism information and calculate engineering interaction features based on the engineering mechanism information; The optimized feature matrix construction module 204 is used to construct an optimized feature matrix through the associated core features and the engineering interaction features; The initial base model group training module 205 is used to train multiple initial base model groups composed of heterogeneous base models using the optimized feature matrix; The fusion model construction module 206 is used to train the initial basic model group through hyperparameter optimization and temporal dynamic weighting operations to construct the fusion model. The prediction output module 207 is used to control the fusion model to output prediction values ​​and generate quality analysis results of cement grouting construction based on the prediction values.

[0209] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0210] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 3As shown, it includes a processor 301, a communication interface 302, a memory 303, and a communication bus 304, wherein the processor 301, the communication interface 302, and the memory 303 communicate with each other through the communication bus 304. Memory 303 is used to store computer programs; When the processor 301 executes the program stored in the memory 303, it implements any of the grouting construction quality analysis methods described in the above embodiments: The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0211] The communication interface is used for communication between the aforementioned terminal and other devices.

[0212] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0213] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0214] like Figure 4 As shown, in another embodiment of the present invention, a computer-readable storage medium 401 is also provided, which stores instructions that, when run on a computer, cause the computer to execute the grouting construction quality analysis method described in the above embodiment.

[0215] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described grouting construction quality analysis method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0216] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0217] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0218] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0219] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for analyzing the quality of grouting construction, characterized in that, include: Acquire basic construction data, dynamic geological parameters, and construction environment parameters; Based on the basic construction data, the dynamic geological parameters, and the construction environment parameters, core features that are strongly correlated with post-irrigation permeability are obtained. Determine the engineering mechanism information and calculate the engineering interaction characteristics based on the engineering mechanism information; An optimized feature matrix is ​​constructed by combining the associated core features and the engineering interaction features; The optimized feature matrix is ​​used to train multiple initial base model groups composed of heterogeneous base models; The initial base model group is trained by hyperparameter optimization and temporal dynamic weighting operations to construct a fusion model; The fusion model is controlled to output predicted values, and quality analysis results of cement grouting construction are generated based on the predicted values.

2. The method according to claim 1, characterized in that, Before the step of obtaining the core features strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters, the method further includes: Spatiotemporal correlation mapping is performed on the basic construction data, the dynamic geological parameters, and the construction environment parameters to obtain a multi-source fusion dataset; The multi-source fusion dataset is filtered and sorted according to the data integrity and validity standards. After removing invalid data, a sorted dataset with hole segment identifiers is generated according to the hole order. Based on the borehole segment identifier, the grouting process data in the time series dataset is normalized in two dimensions to obtain the normalized time series dataset.

3. The method according to claim 1 or 2, characterized in that, The step of obtaining core features strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters includes: By utilizing physical constraints based on engineering principles and statistical detection based on box plots and an improved three-standard-deviation criterion, outliers in the normalized time-series dataset are identified and processed to obtain an outlier-corrected dataset. Based on the parameter type, the missing values ​​of the outlier correction dataset are filled using differentiated strategies such as filling with the median under the same conditions, interpolating adjacent wells and correcting for geological stratification, and filling with the mean of the environment during the same period, to obtain a complete filled dataset. Improved Z-score normalization is performed on the target construction parameters of the complete filled dataset, maximum and minimum value normalization is applied to the target geological parameters of the complete filled dataset, and global standardization normalization is applied to the target construction environment parameters of the complete filled dataset to eliminate the dimensional differences and data noise of the target construction parameters, target geological parameters, and target construction environment parameters, thus obtaining an adaptive normalized dataset. Through a triple screening mechanism of Pearson correlation analysis, mutual information entropy calculation, and expert engineering significance verification, core features strongly correlated with post-irrigation permeability are extracted from the adaptive normalized dataset.

4. The method according to claim 3, characterized in that, The engineering mechanism information includes the amount of ash injected per unit time, the pressure-flow synergy coefficient, and the crack development-ash injection volume matching coefficient. The steps of determining the engineering mechanism information and calculating the engineering interaction characteristics based on the engineering mechanism information include: The core features are used to calculate engineering interaction features including the amount of ash injected per unit time, the synergy coefficient of pressure and flow rate, and the matching coefficient between crack development and ash injection amount. The core features include input parameters and output parameters, and the step of constructing an optimized feature matrix through the associated core features and the engineering interaction features includes: An optimized feature matrix is ​​constructed by concatenating the input parameters, the output parameters, and the engineering interaction features.

5. The method according to claim 4, characterized in that, The step of training an initial base model group consisting of multiple heterogeneous base models using the optimized feature matrix includes: The optimized feature matrix is ​​divided into a training set and a test set. The training set is used to train and generate an initial base model set including a support vector regression model, an improved long short-term memory network model, and a lightweight gradient boosting tree model. With the objective function of minimizing the average absolute percentage error of the fusion model composed of the initial basic model group, the improved Grey Wolf optimization algorithm is used to optimize the core hyperparameters of each model in the initial basic model group, and the optimal parameter combination of each model in the initial basic model group is output. The optimal parameter combination is configured into each model in the initial basic model group, and the fusion weight is calculated based on the average absolute percentage error of each model in the initial basic model group on the test set and the time-series prediction error of different construction periods. Based on the fusion weights and the preset error correction terms, a multi-model dynamic weighted fusion model is constructed.

6. The method according to claim 5, characterized in that, The step of controlling the output prediction value of the fusion model includes: The construction unit data to be predicted in the adaptive normalized dataset is input into the multi-model dynamic weighted fusion model to obtain the final post-irrigation permeability prediction value.

7. The method according to claim 6, characterized in that, The steps for generating quality analysis results for cement grouting construction based on the predicted values ​​include: The performance of the predicted post-irrigation permeability is evaluated using mean square error, mean absolute error, mean absolute percentage error, and prediction stability coefficient to obtain model performance indicators. When it is determined that the model performance index has not reached the preset accuracy or the engineering geological environment has changed, the multi-model dynamic weighted fusion model is triggered to enter an adaptive update process including incremental training and transfer learning, so as to freeze the core parameters of the multi-model dynamic weighted fusion model and adjust the target network layer weights of the multi-model dynamic weighted fusion model to obtain an adaptively optimized fusion model. The adaptively optimized fusion model, the optimized feature matrix, and the post-grouting permeability prediction value are integrated and input into an intelligent visualization system developed based on a graphical user interface component and a dynamic chart engine to generate a visualized report on cement grouting construction quality control and decision-making suggestions.

8. A grouting construction quality analysis device, characterized in that, include: The data acquisition module is used to acquire basic construction data, dynamic geological parameters, and construction environment parameters. The core feature acquisition module is used to acquire core features that are strongly correlated with post-irrigation permeability based on the basic construction data, the dynamic geological parameters, and the construction environment parameters. The engineering interaction feature calculation module is used to determine engineering mechanism information and calculate engineering interaction features based on the engineering mechanism information. An optimized feature matrix construction module is used to construct an optimized feature matrix through the associated core features and the engineering interaction features; The initial base model group training module is used to train multiple initial base model groups composed of heterogeneous base models using the optimized feature matrix; The fusion model construction module is used to train the initial basic model group through hyperparameter optimization and temporal dynamic weighting operations to construct the fusion model; The prediction output module is used to control the fusion model to output prediction values ​​and generate quality analysis results of cement grouting construction based on the prediction values.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the method as described in claims 1-7.

Images