A grouting defect intelligent identification method and system for earth-rock dam engineering

Abstract

This invention relates to the field of intelligent monitoring technology, specifically to an intelligent identification method and system for grouting defects in earth-rock dam engineering. The specific implementation steps include: First, collecting multimodal data of the earth-rock dam, then synchronizing and aligning the multimodal data to form a structured multimodal dataset. Next, using a grouting defect feature quantization model, the dataset is processed hierarchically and in parallel. At the feature fusion layer, grouting anomaly feature vectors from the multimodal data are fused and spatially verified. Quantization is performed within a specific numerical range to obtain a quantized vector representing the anomaly. Finally, based on a defect physical field inversion model, the quantized vector of the grouting defect is used as input. Using a physical information neural network and physical law constraints, the physical properties of the defect area are iteratively solved, and the deviation value between the defect area and the physical properties of the normal grout body is calculated. Based on the deviation value, a multiphysics coupling diagnostic method is used to identify the type of defect.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, specifically to an intelligent identification method and system for grouting defects in earth-rock dam engineering. Background Technology

[0002] In earth-rock dam engineering, grouting technology is a key step in ensuring the integrity of the dam body and foundation, as well as its seepage prevention performance. However, due to complex geological conditions and uncertainties in construction techniques, internal defects such as voids, cracks, and looseness may still form after grouting. These defects are potential safety hazards and must be detected and repaired in a timely manner.

[0003] Traditional defect identification methods primarily rely on post-grouting geophysical exploration techniques and human experience. Engineers typically use ground-penetrating radar or acoustic flaw detectors for passive detection after grouting is completed, and then infer the presence of defects through subjective interpretation. However, this method has several drawbacks: low signal-to-noise ratio, making it susceptible to environmental interference leading to false alarms; heavy reliance on personal experience, lacking objectivity and prone to missed detections; and inability to obtain precise physical properties of defects, making quantitative analysis difficult.

[0004] Faced with insufficient automation capabilities, current attempts have focused on using machine learning to analyze geophysical data in hopes of achieving automated identification. To address the issues of low signal-to-noise ratio and susceptibility to interference, existing efforts have involved integrating multi-source sensors into dams for long-term monitoring. However, these approaches fail to fundamentally solve the problems. Pure machine learning models lack sufficient understanding of physical principles and cannot effectively handle complex and variable geological data. Furthermore, multi-source sensors are typically isolated and lack the ability for spatiotemporal synchronization and data fusion, thus hindering the formation of a complete and collaborative diagnostic system.

[0005] This solution was proposed against this backdrop, aiming to achieve automated, high-precision, and quantifiable intelligent identification of grouting defects through the synergistic work of three technologies: multimodal data acquisition, intelligent feature analysis, and physical information neural networks, thereby fundamentally overcoming the limitations of existing technologies.

[0006] To address this, a method and system for intelligent identification of grouting defects in earth-rock dam engineering is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent identification method and system for grouting defects in earth-rock dam engineering. To address the problems existing in the prior art, this invention first uses probes and sensors to collect electromagnetic and acoustic wave reflection data inside the dam body, and combines this with construction parameters and material property parameters to obtain multimodal data. Subsequently, this multimodal data is synchronized and aligned to form a structured multimodal dataset. Next, a grouting defect feature quantification model is constructed based on spatial cross-verification. This model performs hierarchical and parallel processing on the dataset, extracting and fusing grouting anomaly feature vectors from the multimodal data. The model performs spatial cross-verification on the fused features at the feature fusion layer and quantizes them through specific values, ultimately obtaining a quantified vector representing the anomaly. Finally, this method constructs a defect physical field inversion model based on physical field propagation theory. This model utilizes a physical information neural network and physical law constraints to iteratively optimize and solve for the physical properties of the defective region. Simultaneously, the model obtains the physical properties of the normal grout body by collecting data from the defect-free region, and calculates the deviation value between the physical properties of the defective region and the normal grout body. Finally, based on the deviation value, a multiphysics coupling diagnostic method is used to identify the type of defect.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method and system for intelligent identification of grouting defects in earth-rock dam engineering, the method comprising:

[0009] Multimodal data of earth-rock dams are collected using probes and sensors, and the multimodal data are synchronized and aligned to obtain a structured multimodal dataset;

[0010] A grouting defect feature quantification model is constructed based on spatial mutual verification. The structured multimodal dataset is used as input and processed in a hierarchical and parallel manner to extract grouting abnormal feature vectors. The grouting abnormal feature vectors are fused through the grouting defect feature quantification model, and spatial mutual verification is performed on the fused grouting abnormal feature vectors in space. At the same time, the grouting defect quantification vector is obtained by numerical quantification of defects.

[0011] A physical field inversion model for grouting defects is constructed. The quantized vector of grouting defects is used as input, and physical laws are used as constraints. The physical properties of the defect area are solved iteratively using a physical information neural network. At the same time, the physical properties of the normal grout body are obtained through the physical field inversion model. The deviation between the physical properties of the defect area and the physical properties of the normal grout body is calculated. Based on the deviation, the type of defect is identified using a multi-physics field coupling diagnostic method.

[0012] Preferably, the process of collecting multimodal data of earth-rock dams using probes and sensors, and synchronizing and aligning the multimodal data to obtain a structured multimodal dataset includes: using probes to emit electromagnetic waves and sound waves at the same location, and collecting the reflection data of electromagnetic waves and sound waves inside the dam body; simultaneously using a high-precision positioning system to record the three-dimensional spatial coordinates of the reflection data; collecting construction parameters and grouting material characteristic parameters through embedded sensors, including grouting pressure, grouting flow rate, and drilling depth, and grouting material characteristic parameters including grout viscosity and conductivity; and achieving synchronization and alignment based on the three-dimensional spatial information of the data to obtain a structured multimodal dataset.

[0013] Preferably, the specific implementation process of constructing a grouting defect feature quantification model based on spatial mutual verification includes: extracting grouting anomaly feature vectors of different types of data using different sub-modules based on recurrent neural networks and fully connected networks respectively; constructing a feature fusion module of the grouting defect feature quantification model based on an attention mechanism feature fusion layer, fusing the grouting anomaly feature vectors, and adjusting the weights of the grouting anomaly feature vectors through spatial mutual verification.

[0014] Preferably, a structured multimodal dataset is used as input and processed in a hierarchical and parallel manner to extract grouting anomaly feature vectors. The grouting anomaly feature vectors are then fused using a grouting defect feature quantification model, and spatial verification is performed on the fused grouting anomaly feature vectors. Specifically, the process includes: inputting the structured multimodal dataset into different sub-modules of the grouting defect feature quantification model based on the different structures of electromagnetic wave and sound wave reflection data, construction parameters, and material property parameters, and extracting grouting anomaly feature vectors respectively; inputting the grouting anomaly feature vectors into the feature fusion layer based on an attention mechanism in the model, learning and initially adjusting the weights of the feature vectors, and fusing the grouting anomaly feature vectors; and further adjusting the weights of the feature vectors based on the correlation of grouting anomaly feature vectors from different modalities on the same spatial coordinates to obtain a fused feature vector with intrinsic correlation.

[0015] Preferably, the specific implementation process of obtaining the grouting defect quantification vector through numerical quantification of defects includes: the grouting defect feature quantification model is trained on different types of defect data, including leakage, holes and cracks, non-compact and uneven filling, and grouting blind spots, and learns the mapping relationship of the weights of the fusion feature vector with inherent correlation, and quantifies the fusion feature vector into a grouting defect quantification vector with specific values.

[0016] Preferably, the specific implementation process of constructing a defect physical field inversion model based on grouting defects includes: mapping the input grouting defect quantization vector into initial values ​​of physical parameters that can be processed within the defect physical field inversion model for inversion through a fully connected layer; and iteratively solving the physical properties of the region based on the propagation mode of physical waves and the initial values ​​of physical parameters using a physical information neural network.

[0017] Preferably, the specific implementation process of using the grouting defect quantization vector as input, physical laws as constraints, and physical information neural network to iteratively solve the physical properties of the defect area includes: inputting the grouting defect quantization vector into the defect physical field inversion model, mapping the grouting defect quantization vector to the initial value of physical parameters, using the acoustic wave equation and electromagnetic wave equation as physical losses as constraints in the physical information neural network, and using the initial value of physical parameters to iteratively solve the physical properties of the defect area.

[0018] Preferably, the specific implementation process of obtaining the physical properties of the normal grout body through the defect physical field inversion model and calculating the deviation value between the physical properties of the defect area and the physical properties of the normal grout body includes: collecting multimodal data of the defect-free grout area; obtaining the physical properties of the defect-free grout area through the defect physical field inversion model; obtaining the benchmark values ​​of different physical properties of the normal grout body through statistical analysis of the physical properties; subtracting the physical property value of the defect area from the benchmark value and taking the absolute value to obtain the deviation value between the different physical properties of the defect area and the benchmark value.

[0019] Preferably, the specific implementation process of identifying the type of defect based on the deviation value using the multiphysics coupling diagnostic method includes: calculating the percentage of the deviation value between the different physical properties of the defect area and the benchmark value relative to the benchmark value, thus obtaining the deviation value percentage; by learning from real data of different types of defects, defining a multiphysics coupling coefficient for each defect type, wherein the multiphysics coupling coefficient is calculated according to a formula obtained by weighted summation of the deviation values ​​of different physical properties; substituting the deviation values ​​of different physical properties into the formula to calculate the coupling coefficient value for each defect type, and taking the defect type with the largest coupling coefficient value as the diagnostic result.

[0020] According to the described standardized scheme, an intelligent identification system for grouting defects in earth-rock dam engineering is integrated, including a multimodal data acquisition module, an intelligent feature analysis module, and an intelligent diagnosis module, wherein:

[0021] Multimodal data acquisition module: used to acquire multimodal data of earth-rock dams. It uses a high-precision positioning system to synchronously record the three-dimensional spatial coordinates of each data point. Through the central data processing unit, the data is formatted and noise filtered, and synchronized and aligned based on the spatial information of the data to obtain a structured multimodal dataset.

[0022] Intelligent Feature Analysis Module: Used to extract and fuse abnormal features from structured multimodal datasets, and perform cross-verification in space, transforming the cross-verified grouting abnormal feature vector into a grouting defect quantification vector;

[0023] Intelligent diagnostic module: Based on physical information neural network, it analyzes the physical field anomalies reflected by the quantification vector of grouting defects and, combined with the constraints of physical laws, infers the physical properties that cause the anomalies. By comparing with the different physical property benchmarks of normal grouting bodies, it uses multi-physics field coupling diagnostic method to identify the type of defects.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. By hierarchically and parallelly processing multimodal data, the model can extract abnormal features from multiple dimensions and spatially verify these features, ensuring that a defect is only confirmed when abnormal information from different sources corroborates each other in the same location. This multi-verification mechanism fundamentally solves the problem of false positives and false negatives caused by the low signal-to-noise ratio of a single data source in traditional methods. It not only identifies the existence of defects but also transforms abstract abnormal patterns into calculable and analyzable grouting defect quantification vectors through specific numerical quantification, providing an objective standard for measuring the severity of defects. Traditional methods require engineers to manually interpret maps from different sources, which is time-consuming, labor-intensive, and prone to errors. The model of this invention can automatically complete the hierarchical processing, feature extraction, and fusion of data, automating the originally complex manual analysis process and greatly improving diagnostic efficiency.

[0026] 2. This invention utilizes a physical information neural network to inversely deduce the specific physical properties of the defect area from abstract feature vectors. This transforms the diagnostic result from an abstract classification into one based on understandable physical attributes. Engineers can clearly see the physical nature of the defect, thus increasing their confidence in the diagnostic results. This scheme, using the same model, can not only solve for the physical properties of the defect area but also establish a benchmark for the physical properties of normal grouting bodies. By calculating the deviation between the two, the system can accurately quantify the severity and physical properties of the defect. Physical laws are used as constraints on the model. This ensures that the model's inversion results not only match the observation data but are also physically rigorous and valid. This fundamentally improves the scientific rigor and reliability of the diagnostic results, avoiding erroneous conclusions that are physically invalid.

[0027] 3. This invention utilizes a multiphysics coupling diagnostic method, transforming physical principles into quantifiable mathematical models. This transforms the diagnostic results from abstract classifications into explicit physical logic and computational processes, providing not only diagnostic results but also reliable and traceable diagnostic evidence, significantly increasing engineers' trust in the results. It eliminates reliance on single deviation values ​​by calculating the coupling coefficients of all relevant and irrelevant physical attributes, creating a unique physical fingerprint for each defect. A diagnosis is only confirmed when this fingerprint highly matches the theoretical model, fundamentally solving the problem of misjudgment caused by single attributes and greatly improving the accuracy and reliability of diagnostic results. Attached Figure Description

[0028] Figure 1 A flowchart illustrating a method and system for intelligent identification of grouting defects in earth-rock dam engineering.

[0029] Figure 2 This is a schematic diagram of a grouting defect intelligent identification system for earth-rock dam engineering.

[0030] Figure 3 This is a schematic diagram of the grouting defect feature quantification model proposed in this invention application. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-3 This invention proposes an intelligent identification method and system for grouting defects in earth-rock dam engineering, the technical solution of which is as follows:

[0033] A method and system for intelligent identification of grouting defects in earth-rock dam engineering, with reference to Figure 1 The specific implementation steps of the method proposed in this invention include:

[0034] Multimodal data of earth-rock dams are collected using probes and sensors, and the multimodal data are synchronized and aligned to obtain a structured multimodal dataset;

[0035] A quantification model for grouting defect features is constructed based on spatial mutual verification to extract grouting anomaly feature vectors. The quantification model for grouting defect features fuses the grouting anomaly feature vectors and performs spatial mutual verification on the fused grouting anomaly feature vectors.

[0036] The grouting anomaly feature vector after spatial mutual verification is quantized into a grouting defect quantization vector with specific numerical values.

[0037] Based on grouting defects, a physical field inversion model for defects is constructed. The quantization vector of grouting defects is used as input, physical laws are used as constraints, and physical information neural networks are used to iteratively solve the physical properties of the defect area.

[0038] The physical properties of the normal grout are obtained by using the defect physical field inversion model. The deviation between the physical properties of the defect area and the physical properties of the normal grout is calculated. Based on the deviation, the type of defect is identified by the multiphysics coupling diagnostic method.

[0039] Example 1

[0040] This embodiment provides a specific application of an intelligent identification method and system for grouting defects in earth-rock dam engineering. Its typical application scenario is the quality inspection of consolidation grouting engineering of an underground power station on the right bank of a dam, realizing the automated identification of grouting defects.

[0041] Furthermore, multimodal data of the earth-rock dam are collected using probes and sensors, and the multimodal data is synchronized and aligned to obtain a structured multimodal dataset. The specific process is as follows:

[0042] Ground-penetrating radar (GPR) and a high-frequency acoustic transducer were used to probe the vicinity of the grouting hole, and a high-precision positioning system was used to record the three-dimensional spatial coordinates of the reflected data. Simultaneously, embedded sensors recorded the grouting pressure, flow rate, grout viscosity, and conductivity in real time. The GPR antenna frequency was 500MHz, the acoustic transducer frequency was 50kHz, the RTK-GPS positioning accuracy was ±5mm, and the grout viscosity sensor output range was 1-100mPa·s. Synchronization and alignment were achieved based on the spatial information of this data to obtain a structured multimodal dataset.

[0043] Specifically, the multimodal data is synchronized and aligned to obtain a structured multimodal dataset. The processing steps are as follows:

[0044] Construction parameters, including grouting pressure, grouting flow rate, and drilling depth, as well as grouting material characteristic parameters, including grout viscosity and conductivity, are collected by embedded sensors. The grouting pressure, grouting flow rate, grout viscosity, and conductivity collected at each moment are correlated with the depth and three-dimensional position information of the drill rod at the same moment via RTK-GPS. Data with spatial coordinates from different devices are transmitted to the central data processing unit, where the data is formatted and noise filtered, and synchronization and alignment are achieved based on the spatial information of the data to obtain a structured multimodal dataset.

[0045] By linking construction parameters and grouting material properties with the depth and three-dimensional position of the drill rod at the same time, spatial consistency of the data is ensured from the source, laying a solid foundation for subsequent intelligent analysis and avoiding errors caused by data mismatch.

[0046] By using a unified high-precision positioning system (RTK-GPS), the precise synchronization and alignment of all data in three-dimensional space is ensured. This allows the model to simultaneously utilize evidence from two different physical fields—electromagnetic waves and sound waves—for spatial cross-verification during diagnostics, thereby significantly improving the accuracy and reliability of diagnostic results and effectively avoiding misjudgments and missed reports caused by a single data source.

[0047] Furthermore, a quantification model for grouting defect features is constructed based on spatial mutual verification to extract grouting anomaly feature vectors. The quantification model fuses the grouting anomaly feature vectors and performs spatial mutual verification on the fused grouting anomaly feature vectors. The specific process is as follows:

[0048] The collected and aligned structured multimodal datasets are input into a well-designed grouting defect feature quantification model. Upon receiving the structured multimodal datasets, the model's internal architecture distributes data according to data type, achieving hierarchical and parallel processing. Specifically, electromagnetic wave and acoustic echo data are input into corresponding sub-modules to extract geophysical features of signal attenuation and echo intensity reduction, respectively. Construction parameters and grouting material characteristic parameters are also input into corresponding sub-modules to extract parameter features of grouting pressure drop and grouting material properties.

[0049] The grouting anomaly feature vectors extracted from each submodule are input into a feature fusion layer based on an attention mechanism. The vectors are concatenated into a longer vector. The network learns to assign weights to each input feature to reflect its relative importance in identifying defects. The grouting anomaly feature vectors from geophysical exploration and parameters are deeply fused to form a unified high-dimensional feature vector.

[0050] The model performs spatial verification based on the spatial coordinates carried by these feature vectors. For example, when radar and acoustic features both show anomalies in the same spatial coordinates (X: 345.21, Y: 567.89, Z: 257.89), and the grouting pressure associated with that point is also lower than normal, the model will give these features higher weights to confirm the authenticity of the anomaly.

[0051] By processing data in parallel using different types of sub-modules and performing spatial cross-validation in the feature fusion layer based on an attention mechanism, the model can automatically learn and assign weights, effectively eliminating interference and noise from massive amounts of data, and significantly improving the accuracy and efficiency of feature extraction.

[0052] By inputting raw data with different structures into the corresponding sub-modules and readjusting the weights based on their correlation on the same spatial coordinates, the final fusion features are ensured to have a high degree of intrinsic correlation and credibility, providing a solid basis for the final diagnosis.

[0053] Furthermore, the spatially verified grouting anomaly feature vector is quantified into a grouting defect quantification vector with specific numerical values. The specific process includes:

[0054] The grouting defect feature quantification model is trained using defect data such as leakage, pores and cracks, incomplete and uneven filling, and grouting blind spots. It learns the mapping relationship of fusion feature weights with inherent correlation. The model quantifies the mutually verified abnormal patterns into specific values. For example, radar signal attenuation is mapped to the value 60, acoustic echo anomaly is mapped to the value 55, and pressure is mapped to the value 68. Finally, a grouting defect quantification vector is output, such as [radar attenuation: 60, acoustic echo: 55, pressure: 68].

[0055] Specifically, the grouting defect quantification vector is obtained through specific numerical quantification of defects, and the processing procedure is as follows:

[0056] The grouting defect feature quantification model receives the feature vectors of training samples as input. The training samples contain the grouting defect quantification vectors after spatiotemporal verification and the correct values ​​corresponding to the feature vectors. The grouting defect quantification vectors are calculated through the hierarchical structure inside the model to obtain a predicted value. Based on the predicted value, the backpropagation algorithm is used to calculate the contribution of each parameter to the loss value. Based on the results of the backpropagation algorithm, the parameters are fine-tuned using gradient descent to minimize the loss value.

[0057] The model can minimize the gap between the predicted and the true values, ensuring that the final learned mapping relationship is highly accurate and reliable, providing a solid foundation for subsequent physical property inversion, and fundamentally improving the accuracy and reliability of the overall diagnosis.

[0058] By training on different types of defect data, the grouting defect feature quantification model has learned a quantification method that maps abstract features to specific numerical values, thus providing an objective standard for measuring the severity of defects and providing calculable input for subsequent physical inversion.

[0059] Furthermore, a physical field inversion model for grouting defects is constructed based on the defects. The quantized vector of the grouting defects is used as input, and physical laws are used as constraints. A physical information neural network is used to iteratively solve for the physical properties of the defect region. The specific process includes:

[0060] The quantized vector of grouting defects is input into the defect physics inversion model. A fully connected layer maps the input quantized vector into initial physical parameter values ​​that the model can process for inversion. The model uses the built-in physical laws of the electromagnetic wave equation and the acoustic wave equation as physical losses for constraint. Through iterative inversion using a neural network of physical information within the model, the physical properties at coordinates (X: 345.21, Y: 567.89, Z: 257.89) are solved. The inversion results show that the dielectric constant of this region is close to that of dry air, the density is 0.1 g / cm³, and the acoustic impedance is 0.01 Mrayl.

[0061] By mapping the quantized vector of grouting defects to initial values ​​of physical parameters, and using physical laws as constraints through a defect physics field inversion model, the model can iteratively solve for the physical properties of the defect region. This solves the problem of the inability to perform physical inversion in traditional methods, making the diagnostic results interpretable, scientific, and reliable.

[0062] Furthermore, the physical properties of the normal grout body are obtained through the defect physical field inversion model. The deviation between the physical properties of the defect area and the physical properties of the normal grout body is calculated. Based on the deviation, the type of defect is identified using a multiphysics coupling diagnostic method. The specific process includes:

[0063] A known defect-free area was selected for data collection, and the data was input into a defect physical field inversion model. The model inverted the physical properties of the area, and through statistical analysis, the baseline range of physical properties for a normal grouting body was obtained. An area suspected of having a defect was probed, and the deviation of its physical properties from the baseline was calculated. The model inverted a defect area at coordinates (X: 345.21, Y: 567.89, Z: 257.89), and the inversion result was compared with the baseline value to calculate the percentage deviation.

[0064] The preset defect types include cavities and water-rich cracks, and their respective coupling coefficient formulas are defined. Cavities are characterized by high deviations in density and acoustic impedance, but low deviations in conductivity; therefore, density and acoustic impedance have high weights in the formula, while conductivity has low weights. Water-rich cracks are characterized by extremely high conductivity deviations, while density and acoustic impedance deviations are relatively low. Therefore, conductivity has high weights in the formula, while density and acoustic impedance have low weights. The calculated percentage deviations of different physical properties are substituted into the formulas to calculate the coupling coefficient values ​​for cavities and water-rich cracks. The defect type with the highest coupling coefficient value is taken as the diagnostic result.

[0065] Specifically, the determination of the weights in the coupling coefficient formula is processed as follows:

[0066] The coupling coefficient formula is embedded as a physical loss into the machine learning framework. By training on historical engineering data, which includes the physical attribute deviation values ​​of defective areas and the correct defect type labels, a loss function is defined to measure the gap between the model's diagnostic results and the true labels, and at the same time to measure the degree of agreement between the model's prediction results and the coupling coefficient formula. Using backpropagation and gradient descent optimization algorithms, the model automatically adjusts the weights, minimizes the loss function, and learns the weight allocation of the coupling coefficient formula.

[0067] Machine learning can capture subtle correlations and complex patterns in data that are difficult for humans to perceive, thus finding combinations with more precise weights than those set manually. The process of determining weights is automated, eliminating the need for tedious manual theoretical analysis and calibration.

[0068] By calculating the percentage deviation between the inversion results and the baseline value, a quantitative comparison can be made directly with the physical property baseline value of the normal grout body. This allows for an intuitive and data-driven presentation of the severity of the anomaly, making the diagnostic conclusions more credible and convincing. It goes beyond simply judging "whether there is a defect"; it can accurately calculate how much deviation there is between the physical properties of the defective area and the normal baseline value.

[0069] By using a weighted summation of multidimensional bias percentages, the model can identify defect types in an interpretable, non-black-box manner, which has a strong generalization ability when faced with scarce data and novel defect types.

[0070] Example 2

[0071] In this application embodiment, a method and system for intelligent identification of grouting defects in earth-rock dam engineering is applied to the quality inspection of the anti-seepage curtain grouting project of a large hydropower station dam foundation; see reference Figure 2 This is a structural diagram of an intelligent identification method and system for grouting defects in earth-rock dam engineering. The intelligent identification method and system for grouting defects in earth-rock dam engineering includes: a multimodal data acquisition module, an intelligent feature analysis module, and an intelligent diagnosis module.

[0072] Furthermore, the multimodal data acquisition module includes probes for emitting and receiving electromagnetic and acoustic waves, as well as embedded sensors. The probes emit electromagnetic and acoustic waves at the same location, collecting reflected electromagnetic and acoustic wave data from within the dam body. Simultaneously, a high-precision positioning system records the three-dimensional spatial coordinates of the reflected data. The embedded sensors collect construction parameters and grouting material characteristic parameters, correlating the pressure, flow rate, viscosity, and conductivity values ​​collected at each moment with the depth and three-dimensional position information of the drill rod at the same moment. Then, the central data processing unit formats and filters the data for noise, and synchronizes and aligns it based on the spatial information of the data to obtain a structured multimodal dataset.

[0073] Furthermore, the intelligent feature analysis module includes different sub-modules based on recurrent neural networks and fully connected networks, and a feature fusion module based on an attention mechanism. This module transforms raw, complex multimodal data into high-value information that can be understood and analyzed by subsequent models. The sub-modules extract grouting anomaly feature vectors from different types of data. The feature fusion module then performs deep fusion of the preliminary features extracted from different sub-modules and cross-verifies them spatially. This ensures that only when anomaly information from different sources mutually corroborates each other at the same location are they ultimately confirmed as valid defect features. The cross-verified anomaly features are then quantified within a specific numerical range to obtain a quantized vector of the grouting defect.

[0074] Furthermore, the intelligent diagnostic module includes a fully connected layer and a physical information neural network. The fully connected layer maps the input grouting defect quantization vector into initial values ​​of physical parameters that can be processed within the model for inversion. The physical information neural network iteratively solves for the physical properties of the region based on the propagation mode of physical waves and the initial values ​​of physical parameters, using physical laws as constraints, and calculates the deviation value between the physical properties and the physical properties of normal grouting body. After identifying the physical properties of the defect, a multiphysics coupling diagnostic method is used to determine the type of defect based on the deviation pattern of the physical properties.

[0075] By fundamentally improving diagnostic quality and reliability through the intelligent diagnostic module, diagnostic results are no longer inexplicable black boxes or vague qualitative judgments, but rather based on understandable and quantifiable physical properties, greatly enhancing the interpretability and accuracy of the diagnosis. Furthermore, by utilizing a multiphysics coupling diagnostic method, the defect type is determined based on the deviation patterns of physical properties, providing a transparent and reliable basis for diagnosis. This not only ensures the physical rigor of the diagnostic results but also provides a solid scientific foundation for subsequent repairs.

[0076] Example 3

[0077] This invention provides a typical case of low signal-to-noise ratio in the quality inspection of the consolidation grouting project of the seepage-proof core wall of a high-core earth-rock dam, where the physical properties of the uncompacted area are similar to those of the surrounding normal soil. The invention substitutes the percentage deviation values ​​of different physical properties into the coupling coefficient formula to calculate the coupling coefficient value for each defect type. The defect type with the highest coupling coefficient value is taken as the diagnostic result. The specific implementation method can be as follows:

[0078] After acquiring multimodal data from the detection area, the model converts the data into physical properties using a defect physics field inversion model. The inversion results are displayed at coordinates (X: 550.00, Y: 870.50, Z: 5.00): density: 2.1 g / cm³, close to the normal baseline of 2.2 g / cm³; acoustic impedance: 4.0 Mrayl, slightly lower than the normal baseline of 5.0 Mrayl; electrical conductivity: 0.1 S / m, on par with the normal baseline. This defect is characterized by small deviations in density and acoustic impedance, but normal electrical conductivity.

[0079] In the coupling coefficient formula, the deviations in density and acoustic impedance (weighted at 0.4) are decisive factors because they directly reflect the compactness of the material. The deviation in electrical conductivity has a weight of -0.2, meaning that if the conductivity is abnormal, it will actually reduce the likelihood that the region is not compacted, thus enhancing the diagnostic rigor of the model. The model substitutes the inversion results into the formula to calculate a higher coupling coefficient value, and the model ultimately diagnoses the defect as an uncompacted region.

[0080] The multiphysics coupling diagnostic method no longer relies on a single geophysical data point, but instead comprehensively analyzes the deviation patterns of multiple physical properties. This method can extract a "physical fingerprint" sufficient to distinguish defects from seemingly minor physical differences, and can use a certain feature to exclude other defect types, thus achieving high-precision diagnosis in low signal-to-noise ratio environments.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of grouting defects in earth-rock dam engineering, characterized in that, include: Multimodal data of earth-rock dams are collected using probes and sensors, and the multimodal data are synchronized and aligned to obtain a structured multimodal dataset; A quantitative model for grouting defect features is constructed based on spatial mutual verification. The structured multimodal dataset is used as input and processed in a hierarchical and parallel manner to extract grouting anomaly feature vectors. The grouting defect feature vectors are fused using a grouting defect feature quantification model, and the fused grouting defect feature vectors are spatially cross-validated. Simultaneously, grouting defect quantification vectors are obtained through numerical quantification of defects. The process includes: The grouting defect feature quantification model is trained on different types of defect data, including leakage, holes and cracks, incomplete and uneven filling, and grouting blind spots. It learns the mapping relationship of the weights of the fusion feature vector with inherent correlation, and quantifies the fusion feature vector into a grouting defect quantification vector with specific values. A physical field inversion model for grouting defects is constructed. The quantized vector of the grouting defect is used as input, and physical laws are used as constraints. A physical information neural network is used to iteratively solve for the physical properties of the defective region. The physical properties of the normal grout body are obtained through the defect physical field inversion model. The deviation between the physical properties of the defective region and the normal grout body is calculated. Based on the deviation, a multiphysics coupling diagnostic method is used to identify the type of defect. The process includes: The percentage of deviations between different physical properties of the defect area and the benchmark value is calculated to obtain the deviation percentage. By training and learning on real data of different types of defects, the weights of the deviation percentages of each physical property in the weighted summation formula are determined, and the multi-field coupling coefficient calculation formula for each defect type is obtained. The deviation percentage is substituted into the corresponding multi-field coupling coefficient calculation formula to calculate the coupling coefficient value for each defect type, and the defect type with the largest coupling coefficient value is taken as the diagnostic result.

2. The intelligent identification method for grouting defects in earth-rock dam engineering according to claim 1, characterized in that, The specific implementation process of collecting multimodal data of earth-rock dams using probes and sensors, and synchronizing and aligning the multimodal data to obtain a structured multimodal dataset includes: Electromagnetic and acoustic waves are emitted from the same location using probes, and the reflection data of electromagnetic and acoustic waves inside the dam are collected. At the same time, a high-precision positioning system is used to record the three-dimensional spatial coordinates of the reflection data. Construction parameters and grouting material characteristic parameters are collected through embedded sensors. The construction parameters include grouting pressure, grouting flow rate, and drilling depth. The grouting material characteristic parameters include grout viscosity and conductivity. Synchronization and alignment are achieved based on the three-dimensional spatial information of the data to obtain a structured multimodal dataset.

3. The intelligent identification method for grouting defects in earth-rock dam engineering according to claim 1, characterized in that, The specific implementation process of constructing a quantitative model of grouting defect characteristics based on spatial mutual verification includes: Different sub-modules based on recurrent neural networks and fully connected networks are used to extract grouting anomaly feature vectors from different types of data. A feature fusion module based on the attention mechanism is constructed to build a grouting defect feature quantification model, which fuses the grouting anomaly feature vectors and adjusts the weights of the grouting anomaly feature vectors through spatial mutual verification.

4. The intelligent identification method for grouting defects in earth-rock dam engineering according to claim 1, characterized in that, The structured multimodal dataset is used as input and processed hierarchically and in parallel to extract grouting anomaly feature vectors. These feature vectors are then fused using a grouting defect feature quantification model, and spatial cross-validation is performed on the fused vectors. The specific implementation process includes: Based on the different structures of electromagnetic wave and sound wave reflection data, construction parameters, and material property parameters, a structured multimodal dataset is input into different sub-modules of the grouting defect feature quantification model to extract grouting anomaly feature vectors. These feature vectors are then input into the model's attention-based feature fusion layer to learn and initially adjust the weights of the feature vectors, thereby fusing them. Finally, the weights of the feature vectors are adjusted again based on the correlation between the grouting anomaly feature vectors of different modes on the same spatial coordinates to obtain a fused feature vector with intrinsic correlation.

5. The intelligent identification method for grouting defects in earth-rock dam engineering according to claim 1, characterized in that, The specific implementation process of constructing a physical field inversion model for grouting defects includes: Through a fully connected layer, the input grouting defect quantization vector is mapped into initial values ​​of physical parameters that can be processed within the defect physical field inversion model for inversion; the physical information neural network iteratively solves for the physical properties of the region based on the propagation mode of the physical wave and the initial values ​​of the physical parameters.

6. The intelligent identification method for grouting defects in earth-rock dam engineering according to claim 1, characterized in that, The specific implementation process of using a quantized vector of grouting defects as input, physical laws as constraints, and a physical information neural network to iteratively solve the physical properties of the defect region includes: The quantized vector of grouting defects is input into the defect physical field inversion model, and the quantized vector of grouting defects is mapped to the initial value of physical parameters. The physical information neural network uses the acoustic wave equation and electromagnetic wave equation as physical loss constraints, and uses the initial value of physical parameters to iteratively solve the physical properties of the defect region.

7. The intelligent identification method for grouting defects in earth-rock dam engineering according to claim 1, characterized in that, The physical properties of normal grout are obtained through a defect physical field inversion model. The specific implementation process for calculating the deviation between the physical properties of the defective region and the physical properties of the normal grout includes: Multimodal data of the defect-free grouting area are collected, and the physical properties of the defect-free grouting area are obtained through the defect physical field inversion model. By statistically analyzing the physical properties, the baseline values ​​of different physical properties of the normal grouting body are obtained. The physical property values ​​of the defective area are subtracted from the baseline values ​​and the absolute value is taken to obtain the deviation values ​​of different physical properties of the defective area from the baseline values.

8. An intelligent identification system for grouting defects in earth-rock dam engineering, characterized in that, include: Multimodal data acquisition module: used to collect multimodal data of earth-rock dams. It uses a high-precision positioning system to synchronously record the three-dimensional spatial coordinates of each data point. Through the central data processing unit, the data is formatted and noise filtered, and synchronized and aligned based on the spatial information of the data to obtain a structured multimodal dataset. The intelligent feature analysis module is used to extract and fuse abnormal features from structured multimodal datasets, perform spatial cross-verification, and transform the cross-verified grouting anomaly feature vectors into grouting defect quantification vectors. The process includes: The grouting defect feature quantification model is trained on different types of defect data, including leakage, holes and cracks, incomplete and uneven filling, and grouting blind spots. It learns the mapping relationship of the weights of the fusion feature vector with inherent correlation, and quantifies the fusion feature vector into a grouting defect quantification vector with specific values. Intelligent Diagnostic Module: Based on a physical information neural network, this module analyzes the physical field anomalies reflected in the quantification vector of grouting defects and, combined with the constraints of physical laws, infers the physical properties causing the anomalies. By comparing these anomalies with benchmarks of different physical properties of normal grouting bodies, it uses a multi-physics coupling diagnostic method to identify the type of defect. The process includes: The percentage of deviations between different physical properties of the defect area and the benchmark value is calculated to obtain the deviation percentage. By training and learning on real data of different types of defects, the weights of the deviation percentages of each physical property in the weighted summation formula are determined, and the multi-field coupling coefficient calculation formula for each defect type is obtained. The deviation percentage is substituted into the corresponding multi-field coupling coefficient calculation formula to calculate the coupling coefficient value for each defect type, and the defect type with the largest coupling coefficient value is taken as the diagnostic result.

Images