Earthquake wave velocity intelligent inversion method and system for tunnel advanced detection

Abstract

The present application belongs to the field of earth exploration, and provides a seismic wave velocity intelligent inversion method and system for tunnel advanced detection, establishes a coupling relationship and a constraint mapping relationship between P-wave velocity, S-wave velocity and density in line with geological laws, and accordingly constructs an elastic parameter correlation prior data set and a simulation sample for multi-parameter inversion, establishes a tunnel seismic multi-parameter inversion network based on model domain-data domain-imaging domain three-domain fusion, realizes coupling representation and collaborative constraint of three-domain information through an attention weighted cross fusion mechanism, and uses an interface-parameter two-dimensional progressive training strategy to gradually reveal the law of stratum information in front according to the tunnel advanced detection seismic response from near to far, so that the applicability, stability and engineering reliability of the multi-parameter inversion, rock mass parameter prediction and surrounding rock classification results under different work points and different geological conditions are improved.

Description

Technical Field

[0001] This invention belongs to the field of earth exploration, specifically relating to an intelligent inversion method and system for seismic wave velocity for tunnel advance detection. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The tunnel seismic wave method for advanced detection mainly includes data acquisition, data processing, wave velocity acquisition, structural imaging, and result interpretation. Among these, accurate wave velocity acquisition is a crucial intermediate step, determining the accurate location and morphology of adverse geological structures ahead of the tunnel. Therefore, obtaining a relatively accurate wave velocity distribution ahead of the tunnel is an important research objective.

[0004] Existing methods for obtaining wave velocity ahead of tunnels can be roughly divided into three categories: The first category includes methods such as migration velocity analysis and tomography. These methods mainly utilize the arrival time information of reflected signals in tunnel detection data to obtain the approximate distribution of wave velocity ahead of the tunnel, but their resolution and accuracy are relatively low. Therefore, researchers subsequently developed a full-waveform inversion method based on the wave equation. This method utilizes the full wavefield information contained in seismic data to obtain a more accurate wave velocity distribution ahead of the tunnel, and has achieved initial success. However, it also faces problems such as strong inversion uncertainty and heavy reliance on the initial model, which restricts its further development. The third category is wave velocity inversion methods based on deep learning. As a data-driven approach, it trains a wave velocity inversion network using massive data-model pairs to achieve end-to-end rapid inversion of subsequent observation data.

[0005] For deep learning-based wave velocity inversion methods, leveraging the strong nonlinear representation capabilities and end-to-end fast inference advantages of deep neural networks, they have achieved superior inversion accuracy and efficiency compared to traditional full-waveform inversion methods on simulated data. However, their stability and reliability in actual tunnel engineering scenarios still need improvement. The following challenges remain in realizing the practical application of deep learning-based tunnel seismic wave velocity inversion: Current deep learning-based tunnel wave velocity inversion methods mostly employ idealized acoustic wave assumptions or empirical parameter assignments, making it difficult to fully integrate exploration data, stratigraphic structure, lithological characteristics, and relevant geological parameters to establish elastic parameter models that conform to actual geological laws. In particular, they lack effective characterization of the intrinsic relationships between parameters such as P-wave velocity, S-wave velocity, and density. This not only limits the ability to characterize propagation mechanisms under complex surrounding rock conditions but also restricts the accuracy and applicability of multi-parameter elastic wave inversion and subsequent predictions of rock mass parameters such as Poisson's ratio and elastic modulus.

[0006] Secondly, in the inversion of multiple parameters such as P-wave velocity, S-wave velocity, and density, the strong coupling and significant crosstalk between parameters make network learning difficult and computationally intensive, easily affecting inversion accuracy and generalization ability, leading to false anomalies. Meanwhile, existing methods often focus on single-type information-driven approaches, failing to effectively integrate multi-source information such as low-frequency model constraints, original seismic wavefield responses, and high-frequency structural imaging, resulting in insufficient input constraints. Furthermore, the lack of effective decoupling mechanisms for different parameters at the output end makes it difficult to stably learn the complex input-output mapping relationships between multiple parameters. This problem becomes even more pronounced under conditions of limited tunnel observation, noise interference, and insufficient data coverage.

[0007] On the other hand, existing wave velocity inversion networks are mostly based on idealized or typical models to construct simulation training sets. The geological structure complexity, parameter combination relationships, and observation conditions in the simulation data often differ significantly from the real tunnel environment, easily causing a mismatch between the training distribution and the measured distribution. This leads to biases or even false anomalies in the network's multi-parameter inversion based on actual measurements. In addition, existing methods generally lack the ability to continuously correct and adapt by incorporating construction process information. It is difficult to maintain the consistency and stability of the correlation between P-wave velocity, S-wave velocity, density, and their derived rock mass parameters under different sections, geological conditions, and observation systems, thus affecting the reliability of the inversion results and their engineering application value. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes an intelligent inversion method and system for seismic wave velocity in advance tunnel detection. This invention improves the applicability, stability, and engineering reliability of multi-parameter inversion, rock mass parameter prediction, and surrounding rock classification results under different work sites and geological conditions.

[0009] According to some embodiments, the present invention adopts the following technical solution: A smart inversion method for seismic wave velocity for tunnel advance detection includes the following steps: Based on the current digital information database of tunnel engineering, geological morphology information, lithological characteristics, surrounding rock grade, geological structural attributes and elastic parameters are extracted. The coupling relationship, joint distribution relationship and constraint mapping relationship between lithology, surrounding rock grade or geological structural attributes and P-wave velocity model, S-wave velocity model and density model are established in accordance with geological laws. Based on the geological morphology information and the coupling relationship, joint distribution relationship and constraint mapping relationship, a tunnel geological model morphology is established, an elastic parameter association prior model is formed, and an elastic parameter association prior dataset for tunnel advanced detection is constructed. By combining construction process parameters, advanced drilling, face excavation and exposure, and surrounding rock classification information, the elastic parameters are associated with prior datasets for constraint correction and optimization. Based on the observation system and construction conditions of tunnel advance detection, the elastic parameter-related prior model is subjected to forward modeling simulation to form a model domain-data domain-imaging domain dataset; A three-domain collaborative multi-parameter inversion network based on attention-weighted cross-fusion of model domain, data domain, and imaging domain is constructed. Low-frequency inversion results are used as model domain information, raw seismic data as data domain information, and high-frequency imaging results as imaging domain information. The network is then input into the three-domain collaborative multi-parameter inversion network. Through the coupled representation and collaborative constraints of the three-domain information, independent decoupled prediction branches are constructed for P-wave velocity, S-wave velocity, and density to obtain multi-parameter prediction results corresponding to the seismic observation data. Based on the gradual revelation of the geological response ahead of the tunnel from near to far in the interface dimension, progressive focusing training is carried out on different interface layers; based on the inversion difficulty and coupling degree of P-wave velocity, S-wave velocity and density in the parameter dimension, the three-domain collaborative multi-parameter inversion network is optimized in stages according to the method of progressive decoupling learning from coupled representation. The model loss function is calculated by comparing the P-wave velocity, S-wave velocity, and density output by the three-domain collaborative multi-parameter inversion network with the actual values ​​of the model. The gradient is calculated and backpropagated by combining the data consistency loss between the seismic observation data and the forward model data of the prediction model, as well as the parameter decoupling constraint loss used to suppress multi-parameter crosstalk, and the parameters of the three-domain collaborative multi-parameter inversion network are optimized and updated. The seismic observation data were processed using a three-domain collaborative multi-parameter inversion network with updated parameters to obtain the inversion results of P-wave velocity, S-wave velocity, and density. Based on the inversion results, the surrounding rock classification category and its probability are obtained by using a pre-trained rock mass parameter prediction and surrounding rock classification network.

[0010] As one implementation method, the method also includes a target section migration and adaptation step: a small sample actual dataset is constructed by combining the advanced drilling, face excavation and exposure, surrounding rock classification and construction parameter information of the target section; the small sample actual dataset includes the original acquisition data of the corresponding survey line segment, low-frequency inversion results, high-frequency imaging results, and one or more of the local parameter constraint information and category labels obtained from drilling, exposure or surrounding rock classification; using the three-domain collaborative multi-parameter inversion network with updated parameters as the initial network, the encoder, three-domain fusion layer, multi-parameter decoupling output layer or parameter mapping layer are frozen, partially frozen or constrained and fine-tuned, and the network parameters are updated in combination with the small sample actual dataset to achieve target section migration and adaptation.

[0011] As one implementation method, the digital information database of the current tunnel project is obtained through pre-construction, and the construction process includes: acquiring the exploration data, preliminary geological data and related geological parameter information of the current tunnel project during the tunnel route selection, extracting the effective geological information along the tunnel route, and forming a digital information database of on-site geological data.

[0012] As one implementation method, the elastic parameter-related information includes the distribution range of longitudinal wave velocity Vp, transverse wave velocity Vs and density ρ, the joint distribution characteristics of elastic parameters corresponding to different surrounding rock grades, and the constraint mapping relationship between lithology and elastic parameters.

[0013] As one implementation method, the structural morphology of the tunnel geological model includes the interface of the tunnel geological model, the morphology of anomalies, the number of anomalies, and the spatial combination relationship between different structural units.

[0014] As one implementation method, the process of establishing the geological model structure of the tunnel also includes: for key detection target areas, arranging and combining uncertain parameters to construct multiple sets of adverse geological morphology models and corresponding elastic parameter correlation prior models.

[0015] As one implementation method, the operation of constraint correction and optimization of the elasticity parameter associated prior dataset includes: Using rock mass grade, local wave velocity values, or density constraint information obtained from tunnel pre-drilling, parameter relationship correction is performed on similar areas in the digital information database of field geological data. By utilizing parameters from the tunnel construction process, we can analyze the quality of the rock mass ahead and the changing trend of the surrounding rock, and update the parameter constraints for similar areas in the digital information database of on-site geological data. Based on the parameters of the rock mass collected on-site, the numerical relationships between P-wave velocity, S-wave velocity, density, and derived rock mass parameters were analyzed, and the parameter mapping relationships in the digital information database of on-site geological data were adjusted.

[0016] As an alternative implementation, the model domain-data domain-imaging domain dataset includes: Original seismic acquisition data corresponding to prior models associated with different elastic parameters; Low-frequency inversion results, primarily based on low-frequency information; High-frequency imaging results, primarily based on high-frequency information.

[0017] As an alternative implementation, the three-domain collaborative multi-parameter inversion network includes a model domain encoder, a data domain encoder, an imaging domain encoder, and a multi-parameter decoupled decoder based on attention-weighted cross-fusion. The model domain encoder extracts background constraint features using low-frequency inversion results as input, the data domain encoder extracts wavefield response features using seismic observation data as input, and the imaging domain encoder extracts structural boundary and anomalous body morphology features using high-frequency imaging results as input. The multi-parameter decoupled decoder, based on a shared fusion backbone, sets up P-wave velocity prediction branches, S-wave velocity prediction branches, and density prediction branches, enabling the network to achieve multi-parameter decoupled output while maintaining multi-domain coupled representation capabilities.

[0018] As a further defined implementation, the density prediction branch introduces an amplitude-sensitive feature fusion mechanism and a parameter constraint mechanism to enhance the network's ability to identify density changes, taking into account the strong correlation between density parameters and waveform amplitude and impedance changes.

[0019] As an alternative implementation method, the process of progressively focusing training on different interface layers based on the gradually revealed pattern of the geological response ahead of the tunnel from near to far includes: Samples from the model domain-data domain-imaging domain dataset are retrieved and, before being input into the three-domain collaborative multi-parameter inversion network, are masked or weighted according to the interface layer progressive focusing strategy for the data and model regions corresponding to different layers. The processed samples are input into the three-domain collaborative multi-parameter inversion network. The three-domain collaborative multi-parameter inversion network is trained by progressively focusing from near to far in the interface dimension and by staged optimization from joint representation to sub-parameter identification in the parameter dimension. Once the update focus area and parameter learning stage are determined, the corresponding training strategy is continued until the set stopping condition is met, resulting in a trained three-domain collaborative multi-parameter inversion network.

[0020] As an optional implementation, the model loss function includes a model loss function, a data consistency loss function, and a parameter decoupling constraint loss function; wherein, the model loss function is used to constrain the error between the predicted P-wave velocity, S-wave velocity, and density and the actual model, the data consistency loss function is used to constrain the error between the forward modeling data of the predicted model and the actual observation data, and the parameter decoupling constraint loss function is used to suppress crosstalk between P-wave velocity, S-wave velocity, and density, and improve the stability and physical consistency of the multi-parameter inversion results.

[0021] As an alternative implementation, the rock mass parameter prediction and surrounding rock classification network adopts a multi-task structure that combines a rock mass parameter prediction branch and a surrounding rock classification branch. The rock mass parameter prediction branch is oriented towards the continuous value regression of Poisson's ratio and elastic modulus parameters, while the surrounding rock classification branch is oriented towards surrounding rock category discrimination. At the same time, physical constraints are used to jointly constrain the intrinsic correlation between P-wave velocity, S-wave velocity, density and derived parameters of Poisson's ratio and elastic modulus, thereby realizing the extension from elastic parameter inversion to rock mass parameter prediction-surround rock classification.

[0022] As an alternative implementation, the small sample actual dataset includes: the original acquisition data of the corresponding survey line segment, low-frequency inversion results, high-frequency imaging results, and several of the local parameter constraint information and category labels obtained from drilling, exposure or surrounding rock classification.

[0023] As an alternative implementation method, the process of continuously optimizing the three-domain collaborative multi-parameter inversion network by using the updated parameter three-domain collaborative multi-parameter inversion network as the initial network and adapting it to the target segment includes freezing, partially freezing, or constraining fine-tuning some encoders, three-domain fusion layers, multi-parameter decoupling output layers, and parameter mapping layers that are significantly different from the target segment of the three-domain collaborative multi-parameter inversion network, and updating the network parameters in combination with local constraint information in a small sample actual dataset.

[0024] A seismic wave velocity intelligent inversion system for tunnel advanced detection includes: The elastic parameter association prior construction module is used to extract geological morphology information, lithological characteristics, surrounding rock grade, geological structure attributes, and elastic parameter-related information based on the current digital information database of tunnel engineering. It establishes coupling relationships, joint distribution relationships, and constraint mapping relationships between lithology, surrounding rock grade, or geological structure attributes and P-wave velocity models, S-wave velocity models, and density models that conform to geological laws. Based on the geological morphology information and the coupling relationships, joint distribution relationships, and constraint mapping relationships, it establishes the structural morphology of the tunnel geological model, forms the elastic parameter association prior model, and constructs an elastic parameter association prior dataset for tunnel advanced detection. The three-domain dataset construction module is used to combine construction process parameters, advanced drilling, tunnel face excavation and exposure, and surrounding rock classification information to constrain, correct and optimize the elastic parameter associated prior dataset; based on the tunnel advanced detection observation system and construction conditions, the elastic parameter associated prior model is forward-modeled to form a model domain-data domain-imaging domain dataset. The three-domain collaborative multi-parameter inversion module is used to construct a three-domain collaborative multi-parameter inversion network based on attention-weighted cross-fusion of the model domain, data domain, and imaging domain. Low-frequency inversion results are used as model domain information, raw seismic data are used as data domain information, and high-frequency imaging results are used as imaging domain information. Through the coupled representation and collaborative constraints of the three-domain information, independent decoupled prediction branches are constructed for P-wave velocity, S-wave velocity, and density to obtain multi-parameter prediction results corresponding to the seismic observation data. The interface-parameter dual-dimensional progressive training module is used to progressively focus on different interface layers based on the gradual revelation of the geological response ahead of the tunnel from near to far in the interface dimension; and to optimize the three-domain collaborative multi-parameter inversion network in stages according to the inversion difficulty and coupling degree of P-wave velocity, S-wave velocity and density in the parameter dimension, following the approach of moving from coupled representation to gradual decoupling learning. The rock mass parameter prediction and surrounding rock classification module is used to calculate the model loss function using the P-wave velocity, S-wave velocity, and density output by the three-domain collaborative multi-parameter inversion network and the model's true values. It also incorporates the data consistency loss between seismic observation data and the forward modeling data, as well as the parameter decoupling constraint loss used to suppress multi-parameter crosstalk, to calculate the gradient and perform gradient backpropagation, optimizing and updating the parameters of the three-domain collaborative multi-parameter inversion network. The updated three-domain collaborative multi-parameter inversion network is then used to process the seismic observation data to obtain the inversion results for P-wave velocity, S-wave velocity, and density. Based on these inversion results, a pre-trained rock mass parameter prediction and surrounding rock classification network is used to obtain the surrounding rock classification category and its probability.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention establishes a geologically consistent coupling and constraint mapping relationship between P-wave velocity, S-wave velocity, and density by digitally representing and constraining information such as exploration data, stratigraphic structure, lithological characteristics, surrounding rock grade, and geological structural attributes. Based on this, it constructs an elastic parameter association prior dataset and simulation samples for multi-parameter inversion. By incorporating geological laws into the construction process of multi-parameter combination relationships, the training samples are made closer to the actual engineering geological environment in terms of geological structure, parameter coupling relationships, and observation conditions. This reduces the distribution difference between simulated and measured data from the source, improves the applicability, stability, and reliability of deep learning methods for multi-parameter inversion of P-wave velocity, S-wave velocity, and density in real-world scenarios, and provides a more reliable geological prior basis for subsequent rock mass parameter prediction and surrounding rock identification.

[0026] This invention constructs a multi-parameter inversion network based on the fusion of the model domain, data domain, and imaging domain. Low-frequency inversion results, raw seismic data, and high-frequency imaging results are input into the network as background constraint information, wavefield response information, and structural characterization information, respectively. The three-domain information is coupled and synergistically constrained through an attention-weighted cross-fusion mechanism. At the output end, independent decoupled prediction branches are constructed for P-wave velocity, S-wave velocity, and density. Combined with a parameter crosstalk suppression constraint mechanism, the mutual interference and false anomaly risk in the multi-parameter inversion process are reduced, thereby improving the accuracy, stability, and reliability of the multi-parameter inversion results.

[0027] This invention proposes a two-dimensional progressive training strategy based on interfaces and parameters. Addressing the issues of interface response aliasing, strong parameter coupling, and training instability in multi-parameter seismic inversion of tunnels, this strategy focuses on learning different interface layers layer by layer, following the gradual revelation of the strata response ahead of the tunnel from near to far. Furthermore, considering the inversion difficulty and coupling degree of parameters such as P-wave velocity, S-wave velocity, and density, the network is optimized in stages, progressing from coupled representation to gradual decoupling learning. Through the synergistic advancement of progressive training at the interface dimension and joint representation to individual parameter identification at the parameter dimension, the network can gradually learn the response relationships between different layers and different attribute parameters. This reduces the nonlinearity and crosstalk effects of multi-parameter inversion, improving network training convergence and the accuracy, stability, and reliability of multi-parameter inversion results.

[0028] Based on the obtained P-wave velocity, S-wave velocity, and density inversion results, this invention further constructs a rock mass parameter prediction and surrounding rock classification output mechanism to achieve coordinated prediction of rock mass parameters such as Poisson's ratio and elastic modulus, as well as surrounding rock categories. This method does not independently estimate subsequent parameters; instead, it fully inherits the implicit correlation between P-wave velocity, S-wave velocity, and density, and combines the physical constraints and geological laws between rock mass mechanical parameters to establish a unified mapping relationship from basic elastic parameters to higher-order rock mass parameters and surrounding rock classification results. This enhances the physical consistency, engineering interpretability, and reliability of the prediction results.

[0029] To address the issue of insufficient generalization ability of deep learning multi-parameter inversion models under different contract sections, geological conditions, and observation systems, this invention proposes a target contract section migration and adaptation strategy oriented towards preserving multi-parameter correlations. Based on a small number of measured samples and local constraint information from the target contract section, this strategy fine-tunes the constraints of the three-domain fusion feature layer, the multi-parameter decoupling output layer, and the key parameter mapping layer. This allows the network to adapt to differences in the geological environment and observation conditions of the target contract section while maintaining and correcting the implicit correlations between P-wave velocity, S-wave velocity, density, and their derived rock mass parameters. This reduces the risk of parameter mismatch and correlation distortion when applied across contract sections, and improves the stability, reliability, and engineering applicability of multi-parameter inversion and subsequent rock mass parameter prediction results in real-world scenarios.

[0030] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0032] Figure 1 This is a flowchart of the overall process for the intelligent seismic wave velocity inversion method in this embodiment; Figure 2 This is a flowchart of the elastic parameter association prior construction and simulation sample library generation in this embodiment; Figure 3 This is a schematic diagram of the three-domain collaborative multi-parameter inversion network and two-dimensional progressive training in this embodiment; Figure 4 This is a flowchart of the rock mass parameter prediction and surrounding rock classification network and target segment migration in this embodiment. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0036] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0037] Example 1 This embodiment provides a smart inversion method for seismic wave velocity for tunnel advance detection, such as... Figure 1 As shown, it specifically includes the following steps: Step 1: Construct a prior dataset of elastic parameters based on constraints from geological exploration information; Specifically, the project acquires survey data, preliminary geological data, and related geological parameters from the current tunnel engineering route selection phase, extracts effective geological information along the tunnel route and digitally expresses it, and constructs a digital database of on-site geological data.

[0038] Furthermore, the lithological category, surrounding rock grade, stratigraphic structure, tectonic attributes, and adverse geological body annotation information are expressed in a matrix to form a multi-channel geological feature tensor: ; in, to These respectively represent lithology, surrounding rock grade, structural attributes, anomaly annotation, and other geological constraint information. Indicates the location of the model mesh.

[0039] Based on the aforementioned geological feature tensor, a relationship between geological properties and P-wave velocity is established. transverse wave velocity and density The mapping relationship between them: ; in, , and These represent the elastic parameter mapping functions determined by geological constraints. To ensure that different parameter combinations conform to geological laws, parameter association prior constraint terms can be further constructed: ; In this embodiment, the model region can be set to, for example, 80m × 150m, discretized using a 1m × 1m grid; the front region is used to characterize the tunnel excavation space, and the rear region is used to characterize the area to be explored ahead. For different geological conditions, the interface undulations, anomaly morphology, number of anomalies, and spatial combination relationships are parameterized to form multiple sets of elastic parameter-correlated prior models. For example... Figure 2 As shown, by combining construction process parameters, advanced drilling, face excavation and exposure, and surrounding rock classification information, the above-mentioned associated prior model is continuously constrained, corrected and optimized to obtain an elastic parameter associated prior dataset that is closer to the real engineering scenario.

[0040] Step 2: Construct a three-domain dataset consisting of the model domain, data domain, and imaging domain; Based on the elastic parameter association prior model obtained in step 1, forward modeling is performed using the tunnel advance detection observation system and construction conditions to generate original seismic observation data corresponding to different prior models. For a two-dimensional isotropic elastic medium, its displacement field satisfies: ; in, It is a displacement vector. and Let Lamé constant be . For density, This is the source term. The corresponding P-wave velocity and S-wave velocity satisfy: ; In this embodiment, a Ricker wavelet is used as the seismic source, and the elastic wave propagation process is simulated using the finite-difference time-domain method. Raw seismic data is recorded at the detector location, forming the data domain input. Subsequently, a small number of iterations of full-waveform inversion are performed on the raw seismic data using low-frequency signals to obtain the low-frequency inversion result, which serves as the model domain input. Then, migration imaging is performed based on high-frequency signals to obtain the high-frequency imaging result, which serves as the imaging domain input. This results in a three-domain dataset: model domain, data domain, and imaging domain. ; in, Represents the raw earthquake data. This indicates the low-frequency inversion result. This indicates the results of high-frequency imaging. This represents the label elasticity parameter model.

[0041] Step 3: As Figure 3 As shown, a three-domain collaborative multi-parameter inversion network of model domain, data domain, and imaging domain is constructed, and multi-parameter decoupled prediction is performed; A three-domain collaborative multi-parameter inversion network based on attention-weighted cross-fusion—model domain, data domain, and imaging domain—is constructed. This network includes a model domain encoder, a data domain encoder, an imaging domain encoder, and a multi-parameter decoupled decoder. The model domain encoder extracts background constraint features from low-frequency inversion results; the data domain encoder extracts wavefield response features from raw seismic data; and the imaging domain encoder extracts structural boundaries and anomaly morphology features from high-frequency imaging results. The multi-parameter decoupled decoder, based on a shared fusion backbone, includes separate branches for predicting P-wave velocity (Vp), S-wave velocity (Vs), and density (ρ) to achieve collaborative inversion and decoupled output of multiple parameters.

[0042] In this embodiment, the low-frequency inversion results, the original seismic data, and the high-frequency imaging results are used as model domain information, data domain information, and imaging domain information, respectively, and are input into the network. The network includes a model domain encoder, a data domain encoder, an imaging domain encoder, and a multi-parameter decoupling decoder.

[0043] Among them, the model domain encoder retrieves low-frequency inversion results. As input, extract background constraint features: ; The data domain encoder uses raw seismic data Using this as input, extract wavefield response features: ; Imaging domain encoder with high-frequency imaging results As input, extract structural boundary and anomalous body morphological features: ; Furthermore, an attention-weighted cross-fusion mechanism is used to achieve coupled representation and collaborative constraints of information from the three domains, enabling the network to form a multi-source physical information fusion representation for multi-parameter inversion at the input end. Simultaneously, independent prediction branches are constructed at the output end for Vp, Vs, and ρ. ; in, , and These represent the P-wave velocity branch, the S-wave velocity branch, and the density branch, respectively. , These represent the predicted P-wave velocity, S-wave velocity, and density, respectively. This approach achieves coupled representation of three-domain information at the input end and decoupled prediction of multiple parameters at the output end, thereby mitigating crosstalk issues in multi-parameter inversion. Preferably, the density ρ prediction branch further incorporates a feature fusion and constraint mechanism sensitive to waveform amplitude and impedance changes to enhance the network's ability to identify density variations.

[0044] Step 4: Train the three-domain collaborative multi-parameter inversion network using an interface-parameter dual-dimensional progressive training strategy; Samples from the three-domain dataset are retrieved and, before being input into the three-domain collaborative multi-parameter inversion network, undergo two-dimensional progressive training according to the interface and parameter dimensions. In the interface dimension, based on the pattern of gradually revealing the stratigraphic information ahead from near to far in the seismic response of tunnel advance detection, progressive focusing constraints are applied to the data and model regions corresponding to different layers, enabling the network to learn the response characteristics of different interface layers from shallow to deep and from local to global. In the parameter dimension, based on the inversion difficulty and coupling degree of P-wave velocity Vp, S-wave velocity Vs, and density ρ, the network is optimized in stages according to a method from multi-parameter coupled representation to gradual decoupling and identification.

[0045] In this embodiment, the initial training focuses on learning the near-field interface and dominant response parameters. The training then gradually expands to deeper layers and more parameter response relationships. Finally, it transitions to joint optimization of the entire interface and all parameters. At the interface level, based on the characteristic that tunnel seismic response gradually reveals information about the preceding medium from near to far, a progressive masking matrix is ​​set for different depth regions. ； in, This indicates the current training phase. In the early stages of training... The initial focus is on the near-field interface region; as training progresses, the focused region gradually expands to deeper layers and the entire model region. In terms of parameters, according to... , and To assess the inversion difficulty and coupling degree, staged weights are used to control the training intensity of different parameter branches: , and Indicates the first The weighting coefficients for P-wave velocity, S-wave velocity, and density branching are defined for each stage. The total loss function in this embodiment is: ; in, The model loss function, Let the data consistency loss function be... For parameter decoupling constraint loss, For physical constraint loss, Associate prior constraint terms with the elastic parameters defined in step S1. to These are the weighting coefficients.

[0046] The model loss function is: ; in, Represents element-wise product. , , This represents the structural similarity weight.

[0047] The data consistency loss function is: ; in, Indicates the forward operand operator. This indicates observed earthquake data. This represents the time window or effective response weight matrix corresponding to the current training phase.

[0048] To suppress multi-parameter crosstalk, the parameter decoupling constraint loss is: ; in, , Indicates different parameter branches at the th The stage's feature representation reduces crosstalk by decreasing the correlation between branches with different parameters.

[0049] Physical constraint losses can further constrain parameters to satisfy basic elasticity laws: ; ; This ensures that the prediction results meet the requirements. And Poisson's ratio is within a reasonable range.

[0050] Step 5: As Figure 4 As shown, rock mass parameters are predicted, surrounding rock is classified, and target segment migration and adaptation are performed based on the inversion results.

[0051] The trained three-domain collaborative multi-parameter inversion network is used to predict the P-wave velocity from the test set or measured data. transverse wave velocity and density The inversion results were obtained. Based on this, a rock mass parameter prediction and surrounding rock classification network was further constructed to... , and As input, rock mass parameters such as Poisson's ratio and elastic modulus are calculated. Poisson's ratio and elastic modulus are calculated using the following formula: ; ; The surrounding rock classification branch uses Softmax to output the probability of different surrounding rock categories: ; in, High-level features shared by the rock mass parameter prediction branch and the classification branch. and These are the classification layer weights and biases, respectively.

[0052] Furthermore, considering the differences in geological environment and observation systems across different contract sections, this embodiment adopts a target contract section migration adaptation strategy oriented towards multi-parameter correlation preservation. Specifically, based on a small number of measured samples from the target contract section, advanced drilling data, tunnel face excavation exposure information, and surrounding rock classification information, the three-domain fusion feature layer, multi-parameter decoupling output layer, and key parameter mapping layer are constrained and fine-tuned. To maintain and correct the implicit relationships between multiple parameters, a parameter correlation preservation loss is constructed: ; in, This represents the parameter correlation matrix obtained from the target segment prediction. This represents the parameter correlation matrix in the source domain or during the pre-training phase. This represents the Frobenius norm.

[0053] The corresponding total migration adaptation loss is: ; in, This indicates the loss of data consistency in the target segment. This represents the local parameter constraint loss of the target section. , and These are the weighting coefficients.

[0054] Through the above steps, a smart seismic wave velocity inversion processing workflow can be realized, from the construction of elastic parameter association priors, generation of three-domain datasets, three-domain collaborative multi-parameter inversion, interface-parameter two-dimensional progressive training, to rock mass parameter prediction, surrounding rock classification, and target segment migration adaptation. Final test results show that the method proposed in this embodiment can improve the accuracy and stability of P-wave velocity, S-wave velocity, and density inversion while maintaining the consistency of geological constraints and multi-parameter associations, and enhance the engineering applicability of subsequent rock mass parameter prediction and surrounding rock classification results.

[0055] The main network parameters and hardware requirements in this embodiment are as follows: computation is performed on a graphics card with four 40GB VRAM chips. The network is built on the PyTorch platform, the Adam optimizer batch size is 16, and the initial learning rate is... It is trained using an exponential decay strategy.

[0056] Example 2 This embodiment provides a seismic wave velocity intelligent inversion system for tunnel advance detection, which specifically includes the following modules: The elastic parameter association prior construction module is used to extract effective geological information along the tunnel and express it digitally, and to establish a geological model that conforms to the multi-parameter coupling relationship; The three-domain dataset construction module is used to perform forward modeling based on the elastic parameter associated prior dataset to generate raw seismic acquisition data; and to obtain low-frequency inversion results and high-frequency imaging results based on the raw seismic acquisition data, thereby forming a three-domain dataset of model domain, data domain, and imaging domain. The three-domain collaborative multi-parameter inversion module is used to input low-frequency inversion results, raw seismic data and high-frequency imaging results into the model domain-data domain-imaging domain three-domain collaborative multi-parameter inversion network, and decouple and predict P-wave velocity Vp, S-wave velocity Vs and density ρ at the output end to obtain multi-parameter inversion results. The interface-parameter two-dimensional progressive training module is used to retrieve samples from the model domain, data domain, and imaging domain datasets, and to perform progressive training in stages according to the interface dimension and parameter dimension until the set stopping condition is reached, so as to obtain the trained three-domain collaborative multi-parameter inversion network. The rock mass parameter prediction and surrounding rock classification module takes the longitudinal wave velocity Vp, transverse wave velocity Vs and density ρ output by the three-domain collaborative multi-parameter inversion network as input, and further predicts rock mass parameters such as Poisson's ratio μ and elastic modulus E, and outputs the surrounding rock classification category and its probability.

[0057] The target section migration and adaptation module is based on the trained three-domain collaborative multi-parameter inversion network. It performs constraint fine-tuning on key parameters to adapt the network to the differences in geological environment and observation conditions of the target section, and corrects the implicit correlation between P-wave velocity, S-wave velocity, density and their derived rock mass parameters, thereby improving the inversion accuracy, stability and engineering applicability of the network on the measured data of the target section.

[0058] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart inversion method for seismic wave velocity for tunnel advance detection, characterized in that, Includes the following steps: Based on the current digital information database of tunnel engineering, geological morphology information, lithological characteristics, surrounding rock grade, geological structural attributes and elastic parameters are extracted. The coupling relationship, joint distribution relationship and constraint mapping relationship between lithology, surrounding rock grade or geological structural attributes and P-wave velocity model, S-wave velocity model and density model are established in accordance with geological laws. Based on the geological morphology information and the coupling relationship, joint distribution relationship and constraint mapping relationship, a tunnel geological model morphology is established, an elastic parameter association prior model is formed, and an elastic parameter association prior dataset for tunnel advanced detection is constructed. By combining construction process parameters, advanced drilling, face excavation and exposure, and surrounding rock classification information, the elastic parameters are associated with prior datasets for constraint correction and optimization. Based on the observation system and construction conditions of tunnel advance detection, the elastic parameter-related prior model is subjected to forward modeling simulation to form a model domain-data domain-imaging domain dataset; A three-domain collaborative multi-parameter inversion network based on attention-weighted cross-fusion of model domain, data domain, and imaging domain is constructed. Low-frequency inversion results are used as model domain information, raw seismic data as data domain information, and high-frequency imaging results as imaging domain information. The network is then input into the three-domain collaborative multi-parameter inversion network. Through the coupled representation and collaborative constraints of the three-domain information, independent decoupled prediction branches are constructed for P-wave velocity, S-wave velocity, and density to obtain multi-parameter prediction results corresponding to the seismic observation data. Based on the pattern gradually revealed by the geological response in front of the tunnel from near to far at the interface dimension, progressive focusing training is carried out on different interface layers. Based on the inversion difficulty and coupling degree of P-wave velocity, S-wave velocity and density in the parameter dimension, the three-domain collaborative multi-parameter inversion network is optimized in stages according to the method of gradual decoupling learning from coupled representation; The model loss function is calculated by comparing the P-wave velocity, S-wave velocity, and density output by the three-domain collaborative multi-parameter inversion network with the actual values ​​of the model. The gradient is calculated and backpropagated by combining the data consistency loss between the seismic observation data and the forward model data of the prediction model, as well as the parameter decoupling constraint loss used to suppress multi-parameter crosstalk, and the parameters of the three-domain collaborative multi-parameter inversion network are optimized and updated. The seismic observation data were processed using a three-domain collaborative multi-parameter inversion network with updated parameters to obtain the inversion results of P-wave velocity, S-wave velocity, and density. Based on the inversion results, the surrounding rock classification category and its probability are obtained by using a pre-trained rock mass parameter prediction and surrounding rock classification network.

2. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, It also includes a target section migration and adaptation step: combining the advanced drilling, face excavation and exposure, surrounding rock classification and construction parameter information of the target section, a small sample actual dataset is constructed; the small sample actual dataset includes the original acquisition data of the corresponding survey line segment, low-frequency inversion results, high-frequency imaging results, and one or more of the local parameter constraint information and category labels obtained from drilling, exposure or surrounding rock classification; using the three-domain collaborative multi-parameter inversion network with updated parameters as the initial network, some encoders, three-domain fusion layers, multi-parameter decoupling output layers or parameter mapping layers are frozen, partially frozen or constrained and fine-tuned, and the network parameters are updated in combination with the small sample actual dataset to achieve target section migration and adaptation.

3. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, The current digital information database for tunnel engineering is pre-built, and the construction process includes: acquiring the survey data, preliminary geological data and related geological parameter information of the current tunnel engineering during the tunnel route selection period, extracting the effective geological information along the tunnel route, and forming a digital information database of on-site geological data.

4. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, The information related to the elastic parameters includes the distribution range of P-wave velocity Vp, S-wave velocity Vs and density ρ, the joint distribution characteristics of elastic parameters corresponding to different surrounding rock grades, and the constraint mapping relationship between lithology and elastic parameters. The geological model structure includes the interface, anomaly morphology, number of anomalies, and spatial combination relationship between different structural units. Alternatively, the process of establishing a geological model of a tunnel may also include: for key detection target areas, arranging and combining uncertain parameters to construct multiple sets of adverse geological morphology models and corresponding elastic parameter-related prior models.

5. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, The operations for constraint correction and optimization of the elasticity parameter associated prior dataset include: Using rock mass grade, local wave velocity values, or density constraint information obtained from tunnel pre-drilling, parameter relationship correction is performed on similar areas in the digital information database of field geological data. By utilizing parameters from the tunnel construction process, we can analyze the quality of the rock mass ahead and the changing trend of the surrounding rock, and update the parameter constraints for similar areas in the digital information database of on-site geological data. Based on the parameters of the rock mass collected on-site, the numerical relationships between P-wave velocity, S-wave velocity, density, and derived rock mass parameters were analyzed, and the parameter mapping relationships in the digital information database of on-site geological data were adjusted.

6. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, The model domain—data domain—imaging domain dataset includes: Original seismic acquisition data corresponding to prior models associated with different elastic parameters; Low-frequency inversion results, primarily based on low-frequency information; High-frequency imaging results, primarily based on high-frequency information.

7. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, The three-domain collaborative multi-parameter inversion network includes a model domain encoder, a data domain encoder, an imaging domain encoder, and a multi-parameter decoupled decoder based on attention-weighted cross-fusion. The model domain encoder extracts background constraint features using low-frequency inversion results as input, the data domain encoder extracts wavefield response features using seismic observation data as input, and the imaging domain encoder extracts structural boundary and anomalous body morphology features using high-frequency imaging results as input. The multi-parameter decoupled decoder, based on a shared fusion backbone, sets up P-wave velocity prediction branches, S-wave velocity prediction branches, and density prediction branches, enabling the network to achieve multi-parameter decoupled output while maintaining multi-domain coupled representation capabilities.

8. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 7, characterized in that, The density prediction branch addresses the strong correlation between density parameters and waveform amplitude and impedance changes by introducing an amplitude-sensitive feature fusion mechanism and a parameter constraint mechanism to enhance the network's ability to identify density changes.

9. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, Based on the gradual revelation of the geological response ahead of the tunnel from near to far at the interface dimension, the process of progressively focusing training on different interface layers includes: Samples from the model domain-data domain-imaging domain dataset are retrieved and, before being input into the three-domain collaborative multi-parameter inversion network, are masked or weighted according to the interface layer progressive focusing strategy for the data and model regions corresponding to different layers. The processed samples are input into the three-domain collaborative multi-parameter inversion network. The three-domain collaborative multi-parameter inversion network is trained by progressively focusing from near to far in the interface dimension and by staged optimization from joint representation to sub-parameter identification in the parameter dimension. Determine the update focus area and parameter learning stage, and continue to execute the corresponding training strategy until the set stopping condition is met, thus obtaining the trained three-domain collaborative multi-parameter inversion network.

10. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, The model loss function includes a model loss function, a data consistency loss function, and a parameter decoupling constraint loss function. The model loss function is used to constrain the error between the predicted P-wave velocity, S-wave velocity, and density and the actual model. The data consistency loss function is used to constrain the error between the forward modeling data of the predicted model and the actual observation data. The parameter decoupling constraint loss function is used to suppress crosstalk between P-wave velocity, S-wave velocity, and density, and improve the stability and physical consistency of the multi-parameter inversion results.

11. The intelligent inversion method for seismic wave velocity for tunnel advance detection as described in claim 1, characterized in that, The rock mass parameter prediction and surrounding rock classification network adopts a multi-task structure that combines a rock mass parameter prediction branch and a surrounding rock classification branch. The rock mass parameter prediction branch is oriented towards the continuous value regression of Poisson's ratio and elastic modulus parameters, while the surrounding rock classification branch is oriented towards surrounding rock category discrimination. At the same time, physical constraints are used to jointly constrain the intrinsic correlation between P-wave velocity, S-wave velocity, density and derived parameters of Poisson's ratio and elastic modulus, thereby realizing the extension from elastic parameter inversion to rock mass parameter prediction-surround rock classification.

12. A seismic wave velocity intelligent inversion system for tunnel advanced detection, characterized in that, include: The elastic parameter association prior construction module is used to extract geological morphology information, lithological characteristics, surrounding rock grade, geological structure attributes, and elastic parameter-related information based on the current digital information database of tunnel engineering. It establishes coupling relationships, joint distribution relationships, and constraint mapping relationships between lithology, surrounding rock grade, or geological structure attributes and P-wave velocity models, S-wave velocity models, and density models that conform to geological laws. Based on the geological morphology information and the coupling relationships, joint distribution relationships, and constraint mapping relationships, it establishes the structural morphology of the tunnel geological model, forms the elastic parameter association prior model, and constructs an elastic parameter association prior dataset for tunnel advanced detection. The three-domain dataset construction module is used to combine construction process parameters, advanced drilling, tunnel face excavation and exposure, and surrounding rock classification information to constrain, correct and optimize the elastic parameter associated prior dataset; based on the tunnel advanced detection observation system and construction conditions, the elastic parameter associated prior model is forward-modeled to form a model domain-data domain-imaging domain dataset. The three-domain collaborative multi-parameter inversion module is used to construct a three-domain collaborative multi-parameter inversion network based on attention-weighted cross-fusion of the model domain, data domain, and imaging domain. Low-frequency inversion results are used as model domain information, raw seismic data are used as data domain information, and high-frequency imaging results are used as imaging domain information. Through the coupled representation and collaborative constraints of the three-domain information, independent decoupled prediction branches are constructed for P-wave velocity, S-wave velocity, and density to obtain multi-parameter prediction results corresponding to the seismic observation data. The interface-parameter dual-dimensional progressive training module is used to progressively focus on different interface layers based on the pattern gradually revealed by the geological response ahead of the tunnel from near to far in the interface dimension. Based on the inversion difficulty and coupling degree of P-wave velocity, S-wave velocity and density in the parameter dimension, the three-domain collaborative multi-parameter inversion network is optimized in stages according to the method of gradual decoupling learning from coupled representation; The rock mass parameter prediction and surrounding rock classification module is used to calculate the model loss function using the P-wave velocity, S-wave velocity, and density output by the three-domain collaborative multi-parameter inversion network and the model's true values. It also incorporates the data consistency loss between seismic observation data and the forward modeling data, as well as the parameter decoupling constraint loss used to suppress multi-parameter crosstalk, to calculate the gradient and perform gradient backpropagation, optimizing and updating the parameters of the three-domain collaborative multi-parameter inversion network. The updated three-domain collaborative multi-parameter inversion network is then used to process the seismic observation data to obtain the inversion results for P-wave velocity, S-wave velocity, and density. Based on these inversion results, a pre-trained rock mass parameter prediction and surrounding rock classification network is used to obtain the surrounding rock classification category and its probability.

Images