Tunnel full waveform inversion method based on spectral element method forward and physical information neural network

Abstract

This invention relates to a tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network, comprising: collecting engineering geological data and obtaining physical property parameters; establishing a two-dimensional finite element model and obtaining spatial information; constructing a wavefield model and performing wavefield forward modeling simulation to obtain synthetic full-waveform data; collecting measured waveform data and preprocessing it, performing time registration and normalization processing on the preprocessed measured waveform data and synthetic full-waveform data, and extracting multiple features; constructing a parameter training set; constructing and training a physical information neural network model, constructing a comprehensive loss function during the training process; and using the physical information neural network model to perform full-waveform inversion. This invention deeply integrates finite element forward modeling, full-waveform inversion, and physical information neural network, taking into account both data fitting and physical consistency, significantly improving the quantitative accuracy and reliability of tunnel advanced geological prediction.

Description

Technical Field

[0001] This invention relates to the field of tunnel physical exploration technology, specifically a tunnel full waveform inversion method based on spectral element forward modeling and physical information neural network. Background Technology

[0002] With the large-scale construction of mountain tunnels, urban rail tunnels, and deep-buried tunnels, the adverse geological conditions ahead of the tunnel face (such as fault fracture zones, weak interlayers, water-rich zones, and karst caves) have increasingly posed a significant threat to construction safety and schedule during tunnel excavation. Failure to accurately assess the geological conditions ahead of the tunnel before excavation can easily lead to disasters such as water inrush, mudslides, collapses, and large deformations, resulting in serious safety risks and economic losses. Therefore, how to conduct safe, reliable, and high-resolution tunnel geological prediction has always been one of the core technical issues in the field of tunnel and underground engineering.

[0003] As one of the important technologies for tunnel physical exploration, seismic exploration plays an important role in the detection of adverse geological conditions in tunnels. Reflection wave seismic exploration uses a seismic source to generate seismic waves and uses a geophone to receive the reflected waves caused by lithological interfaces, faults or fracture zones, thereby generating seismic records. After processing, these seismic records can be used for imaging and locating adverse geological bodies.

[0004] With advancements in geophysical exploration and numerical simulation technologies, full waveform inversion (FWI) has gradually become an important tool for detecting underground structures. FWI, by retrieving complete seismic wave data, can accurately infer the physical properties of the subsurface medium, such as formation velocity and density, thus providing more detailed information about the subsurface structure. However, traditional FWI methods primarily rely on classical numerical calculation methods, such as the finite difference method and the finite element method. These methods require large amounts of high-quality seismic wave data and complex numerical simulations to invert the physical properties of the subsurface medium. Therefore, they suffer from high computational complexity and long processing times. Especially in tunnel construction environments with significant noise interference, seismic wave data is often incomplete, and data distribution is often uneven under complex geological conditions, leading to a significant reduction in inversion accuracy and failing to meet practical needs.

[0005] Some scholars have proposed a method that combines deep neural networks and physical models to establish a non-mapping relationship between seismic data and wave velocity models, demonstrating an inversion effect superior to traditional FWI. However, it is still unable to effectively address the inversion problems of rapidly changing geological conditions and complex geological bodies during tunnel construction. Furthermore, it lacks efficient processing capabilities when faced with noise, incomplete, and low-quality data, which significantly limits the inversion accuracy when predicting adverse geological bodies. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a systematic tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network. This method effectively improves the accuracy and real-time performance of tunnel advanced geological prediction, can handle noise and incomplete or low-quality data in complex geological environments, and enhances the ability to accurately predict adverse geological bodies.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network includes the following steps: S1. Collect engineering geological data and obtain physical property parameters; S2. Based on engineering geological data, establish a two-dimensional finite element model and conduct working condition simulation to obtain spatial information; S3. Construct a wavefield model in the spectral element method software, perform forward wavefield modeling simulation, and obtain synthetic full waveform data; S4. Deploy seismic sources and receivers to conduct geophysical tests, collect measured waveform data, preprocess the measured waveform data, perform time registration and normalization processing on the preprocessed measured waveform data and the synthetic full waveform data, and extract multiple features from the processed waveform. S5. Construct a parameter training set with waveform features and spatial information as input layers and physical property parameters and geological body information as output layers; S6. Using the parameter training set as input, construct and train the physical information neural network model, and construct a comprehensive loss function during the training process; S7. Use a physical information neural network model to perform full waveform inversion of physical property parameters and geological body information of the geological area to be predicted; S8. By comparing the revealed geological conditions with the inversion results, determine whether to retain or retrain the physical information neural network model.

[0008] Furthermore, in S4, a seismic source is set at the center of the tunnel face, and multiple receivers are deployed on the tunnel wall and around the tunnel face. The receivers are deployed along both sides of the tunnel surrounding rock and along the vertical axis.

[0009] Furthermore, in step S6, the comprehensive loss function of the physical information neural network model... for:

[0010] in, For the residual terms of the data; This represents the residual term of the wave equation; This represents the residual term in the mechanical equilibrium equation; For regularization terms; These are the weighting coefficients for each loss term.

[0011] Furthermore, in step S6, the residual term of the data for:

[0012] in, The number of sample points involved in the fitting; The locations of the measurement points are based on actual measurements and survey data; For the predicted waveform characteristics and physical property parameters; These are the measured waveform characteristics and physical properties.

[0013] Furthermore, in step S6, the residual term of the wave equation for: ; in, This represents the number of constraint points for the wave equation. Physical constraint points The fluctuation residuals are expressed as:

[0014] in, Density; Physical constraint points The predicted displacement; Physical constraint points Predicted internal stress; Physical constraint points External forces at the location.

[0015] Furthermore, in step S6, the residual terms of the mechanical equilibrium equations for:

[0016] in, This represents the number of constraint points in the mechanical equilibrium equations. Constraint points The mechanical equilibrium residuals on the surface are expressed as:

[0017] in, Constraint points Predicted internal stress; Constraint points Volume force at the point.

[0018] Furthermore, in step S6, the regularization term for:

[0019] in, These are the weight parameters for each layer of the neural network; This represents the number of layers in the neural network.

[0020] Furthermore, in step S6, the training process includes: iteratively training the physical information neural network model using a stochastic gradient descent optimization algorithm, continuously updating the network's weights and biases, so that the comprehensive loss function gradually decreases and tends to converge.

[0021] Furthermore, in step S7, the inversion process includes: collecting new measured waveform data for the area to be predicted, performing data preprocessing and multi-feature extraction, inputting the extracted waveform features and corresponding spatial information into the trained physical information neural network model, and outputting the target physical property parameters and geological body information of the area.

[0022] Furthermore, in step S8, the judgment process includes: when the inversion result is consistent with the actual geological conditions, retaining the parameters of the existing physical information neural network model; when the inversion result deviates from the actual geological conditions, adding new engineering geological data and corresponding waveform data to the parameter training set, and retraining the physical information neural network model.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention deeply integrates finite element forward modeling, full waveform inversion, and physical information neural networks, balancing data fitting and physical consistency. It addresses the pain points of traditional tunnel full waveform inversion technology from multiple dimensions, including data processing, model training, inversion application, and engineering adaptation. Compared to existing technologies, in addition to significantly improving the quantitative accuracy and reliability of tunnel advanced geological prediction, fully utilizing full waveform information to enhance inversion accuracy, reducing overfitting and enhancing model generalization ability through physical constraints, and reducing computational load to adapt to high-frequency construction prediction, it also possesses the following technical advantages and practical engineering value: 1. Possesses strong anti-interference capabilities and can efficiently process low-quality data from tunnel construction. This invention effectively reduces noise interference in the tunnel construction environment by preprocessing measured waveform data, including DC removal, trend removal, and amplitude normalization, combined with time registration, normalization, and multi-feature extraction of synthetic and measured waveforms. Simultaneously, by embedding physical constraints into the neural network training process, the model no longer relies solely on high-quality, complete seismic wave data. Even when faced with common problems in construction such as incomplete or unevenly distributed waveform data, it can still stably output inversion results, solving the technical problem of significantly reduced inversion accuracy due to data quality issues in traditional full-waveform inversion.

[0024] 2. Achieve dynamic iterative updates of the model to adapt to the dynamic geological changes during tunnel construction. This invention designs a comparison and model update process between the inversion results and the actual geological conditions revealed. When there is a deviation between the inversion results and the actual geology, new engineering geological data and waveform data can be added to the training set to retrain the physical information neural network model. The parameters of the finite element and spectral element models can also be adjusted as needed. This dynamic update mechanism allows the model to continuously adapt to the rapid changes in geological conditions ahead of the tunnel face during tunnel excavation, solving the problems of existing methods being unable to cope with dynamic geological characteristics and lacking real-time performance, thus ensuring the timeliness and accuracy of advanced geological prediction.

[0025] 3. Targeted selection of inversion parameters improves the accuracy of identifying adverse geological bodies. This invention allows for flexible selection of target physical property parameters based on different types of adverse geological bodies during the inversion process. For example, resistivity can be selected for water-bearing bodies, density for fractured zones, and resistivity or density for karst caves as core inversion parameters. This achieves "targeted" inversion for identifying adverse geological bodies. Compared to the traditional method of uniformly extracting general physical property parameters through full-waveform inversion, this design makes the inversion results more closely match the physical characteristics of different adverse geological bodies, significantly improving the identification accuracy and distinguishability of typical adverse geological bodies in tunnels such as fault fractured zones, water-rich zones, and karst caves.

[0026] 4. The combination of refined modeling and wavefield simulation provides a high-quality data foundation for inversion. When establishing a two-dimensional finite element model using ABAQUS, this invention refines the mesh around the tunnel and the surrounding working conditions, and simulates various working conditions such as aquifers, fracture zones, and faults under different strata, obtaining refined spatial information on the surrounding rock stress and strain fields. The refined processing of finite element forward modeling and spectral element wavefield simulation provides high-fidelity training data for the physical information neural network model, solidifying the foundation for inversion accuracy from the data source.

[0027] 5. Multi-constraint integrated loss function design avoids model bias in purely data-driven models. The integrated loss function constructed in this invention integrates data residual terms, wave equation residual terms, mechanical equilibrium equation residual terms, and regularization terms, constraining model training from four dimensions. Specifically, the data residual terms ensure that the model output closely matches the measured data; the wave equation and mechanical equilibrium equation residual terms ensure that the inversion results strictly follow the laws of seismic wave propagation and rock mechanics equilibrium; and the regularization term effectively prevents model overfitting. This design overcomes the problems of pure data-driven neural networks being prone to physical law deviations and the poor integration of physical constraints and data-driven approaches in existing PINN applications. It ensures that the model output closely matches both measured data and the physical laws of engineering practice, significantly improving the rationality and reliability of the inversion results.

[0028] 6. Providing refined engineering results to directly guide tunnel construction practice. This invention utilizes a trained model to invert and obtain the physical property parameter field (P-wave velocity, S-wave velocity, density, water content, etc.) and probability distribution of adverse geological bodies ahead of the tunnel face. This allows for the further generation of advanced geological prediction profiles and risk zoning ahead of the tunnel. Compared to traditional inversion methods that only output single physical property parameters, this type of result better meets the engineering needs of tunnel construction, providing specific and intuitive geological basis for construction support design and the formulation of treatment measures for adverse geological bodies. It achieves seamless integration from geological inversion to engineering application, enhancing the engineering practicality of the technological results.

[0029] 7. Significantly improves computational efficiency, adapting to the high-frequency advance prediction needs of tunnel construction. This invention trains synthetic data and measured processing data from finite element forward modeling using a physical information neural network, transforming the complex iterative numerical simulation calculations of traditional full-waveform inversion into rapid inference calculations by the neural network, significantly reducing computational complexity and workload. The trained model can quickly invert the area to be predicted ahead of the tunnel face, enabling high-frequency advance geological prediction during tunnel excavation, timely tracking of excavation progress and feedback of geological conditions ahead, effectively reducing the risk of construction disasters such as water inrush, mudslides, and collapses.

[0030] 8. The model has strong generalization ability and can be adapted to tunnel engineering scenarios with different complex geological conditions. The physical information neural network model of this invention constructs the input layer based on waveform features and spatial information, and uses physical property parameters and geological body information as the output layer. It is trained by combining a comprehensive loss function with multiple constraints, and supports iterative updates based on supplementary data from actual engineering projects. This model is not limited to the geological conditions of specific tunnels. After a small amount of adaptive training, it can be applied to different types of tunnels, such as mountain tunnels, urban rail tunnels, and deep-buried tunnels, as well as tunnel engineering projects with different complex geological conditions such as faults, fracture zones, karst caves, and water-rich zones. It has good engineering generalization ability and can significantly reduce the development cost of advanced geological prediction models in different tunnel engineering projects. Attached Figure Description

[0031] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present application, the following description will be provided in conjunction with the accompanying drawings. Figure 1 The technical solutions in the embodiments of this application have been clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0033] This embodiment provides a tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network. This method is used to invert the geological structure behind the tunnel and identify anomalies such as water-bearing karst caves, fissures, fracture zones, and fault zones. The method first establishes a tunnel-surrounding rock finite element model and performs full-waveform forward modeling of the wave equation to obtain the synthetic waveform of the area in front of the tunnel. On-site excitation sources and receivers are set up to collect measured waveforms, and the measured and synthetic waveforms are denoised, registered, and multi-attribute features are extracted. Combined with physical parameters obtained from advanced drilling, downhole television, and geological sketching, a parameter training sample set is constructed. Further, a physical information neural network is established with spatial coordinates, finite element forward modeling results, and waveform multi-attributes as inputs, and wave velocity, density, water content, fissure degree, and adverse geological body category as outputs. The loss function simultaneously introduces the full-waveform fitting error, the residuals of the wave equation and the mechanical equilibrium equation, and a regularization term to train the tunnel full-waveform inversion model. The model is used to invert the waveform data of the area to be predicted to obtain the physical parameter field in front of the tunnel face and the probability distribution of adverse geological bodies, thereby realizing advanced geological prediction. Specifically, the method includes the following steps: S1. Collect engineering geological data and physical property parameters: Collect geological exploration data and engineering overview of the tunnel construction area, including route alignment, tunnel depth, tunnel cross-section, design surrounding rock grade, and existing exploration reports, to provide basic support for advanced geological prediction of the tunnel; simultaneously, obtain initial values ​​of physical property parameters for various rocks through drilling and laboratory tests, and determine the spatial location and scale of adverse geological conditions such as fault fracture zones, aquifers, and karst caves. Specifically, physical property parameters include rock type, mineral composition, porosity, natural density, elastic modulus, Poisson's ratio, etc., and estimate the corresponding P-wave and S-wave velocity ranges based on empirical formulas, providing basic input for subsequent model parameter selection. More specifically, the empirical formulas are as follows:

[0034] in, P-wave velocity; For S-wave velocity; It is the elastic modulus; Natural density; It is Poisson's ratio.

[0035] S2. Based on engineering geological data, establish a two-dimensional finite element model and perform working condition simulation to obtain spatial information: Based on engineering geological data and preliminary exploration results, an initial two-dimensional profile finite element model was established in the finite element software (ABAQUS), including information on tunnel excavation, surrounding rock stratification, and adverse geological bodies (such as faults, fracture zones, aquifers, and karst caves). Preferably, the initial two-dimensional profile finite element model has an axial length of 100 meters and a vertical length of 50 meters. The excavated tunnel cross-section is set to a horseshoe shape with a net width of 15 meters, a net height of 20 meters, and a tunnel depth of 25 meters. The surrounding rock is meshed using CPE4R plane strain elements, with mesh refinement near the tunnel and working conditions. The element side length is controlled within 0.5-1m. Based on exploration and laboratory test results, corresponding elastic modulus, Poisson's ratio, and density parameters are assigned to different strata. The top, bottom, left, and right boundaries of the two-dimensional profile finite element model are set as absorbing boundaries, and the tunnel surrounding rock is set as a free boundary.

[0036] The finite element method (ABAQUS) software was used to perform working condition simulations, obtaining spatial information characteristics such as the stress field and strain field of the surrounding rock, providing input files for subsequent wavefield simulations. Specifically, the initial stress field of the surrounding rock before tunnel excavation was obtained by setting the initial geostress field and performing static analysis; then, simulations of different working conditions were performed, including aquifers, fracture zones, faults, and irregular chambers under different strata, to obtain the stress field distribution after excavation. The calculated nodal coordinates, element division, and stress-strain field results were exported in .inp file format as input for subsequent wavefield simulations and prior information for the physical information neural network.

[0037] S3. Construct a wavefield model, perform forward wavefield modeling, and obtain the synthesized waveform: Based on the model geometry and layering information obtained in step S2, a tunnel-surrounding rock wavefield model was constructed in the spectral element method software (SPECFEM2D). The ABAQUS finite element mesh and material parameters were converted into spectral element format on the Linux system using self-written code. At the same time, appropriate mesh refinement was performed around the tunnel and near the working condition to ensure that the minimum element size meets 1 / 8 to 1 / 10 of the highest frequency wavefield.

[0038] Forward wave modeling was performed to obtain time-varying composite full waveform data at each receiver. Furthermore, wave field analysis was conducted on the composite full waveform data to identify and analyze components such as reflected and scattered waves. The influence of key physical properties such as P-wave velocity field, S-wave velocity field, and displacement field on the wave field response (response sensitivity) was qualitatively analyzed, providing a reference for subsequent feature selection and sensitivity analysis.

[0039] S4. Deploy seismic sources and receivers to conduct geophysical tests, collect measured waveform data, preprocess the measured waveform data, perform time registration and normalization on the preprocessed measured waveform data and the synthetic full waveform data, and extract multiple features from the processed waveform: A seismic source was positioned at the center of the tunnel face to deploy either an impact or mechanical seismic source, using a Ricker wavelet with a dominant frequency of approximately 200Hz. Multiple receivers (geophones) were deployed on the tunnel walls and around the tunnel face, arranged along both sides of the surrounding rock and the vertical axis, with a spacing of 5m, covering an area of ​​20-30m in front of the tunnel face. A multi-channel seismograph was used to simultaneously record the seismic waveforms from each receiver. Geophysical exploration was conducted within a certain distance in front of the tunnel face to collect raw measured waveform data of the surrounding rock.

[0040] The original measured waveform data was preprocessed, including DC removal, trend removal, and amplitude normalization, to reduce construction noise and environmental interference. The preprocessed measured waveform data and the synthesized full waveform data obtained through wavefield forward modeling in spectral element method software were then denoised, time-registered, and normalized to ensure the geometric and temporal correspondence between the two as much as possible. Based on this, multiple features were extracted from the processed waveform, including peak amplitude, first wave arrival time, arrival time of major reflection events, and instantaneous frequency. All extracted features were standardized to eliminate differences in dimensions and orders of magnitude, forming a unified feature vector.

[0041] S5. Construct a parameter training set with waveform features and spatial information as input layers and physical property parameters and geological body information as output layers: By combining advanced drilling and geological sketching, physical parameters such as lithology, surrounding rock classification, P-wave velocity, S-wave velocity, density, and water content within a certain range in front of and around the tunnel face are confirmed. The spatial location and scale of adverse geological features such as fault fracture zones, aquifers, and karst caves are also determined. The obtained physical parameters are matched with their corresponding spatial locations, and then matched with the multi-attribute waveform features extracted in step S4. This forms a high-quality parameter training dataset with waveform features and spatial information as the input layer, and physical parameters and geological body information (geological body type, distribution probability) as the output layer. This dataset is used to train the physical information neural network to accurately invert the distribution of physical parameters and adverse geological bodies in front of the tunnel face.

[0042] S6. Using the parameter training set as input, construct and train the physical information neural network model, and construct a comprehensive loss function during the training process: The parameter training set obtained in step S5 is used as input to construct a multi-layer feedforward neural network as the main body of the physical information neural network. Multi-attribute waveform features and spatial information (stress field, strain field, spatial coordinates) are used as input layers, while target physical property parameters (P-wave velocity, S-wave velocity, density, water content, etc.) and geological body information (adverse geological body type and distribution probability) are used as output layers. Physical constraints are introduced during training by incorporating wave equations and mechanical equilibrium equations into the loss function to ensure that the output physical property parameters and adverse geological body distribution conform to actual physical laws.

[0043] To simultaneously satisfy both data constraints and physical constraints, a comprehensive loss function is constructed during the training process of the physical information neural network. The comprehensive loss function includes the following components: ① Residual term of the data (error between measured data and predicted data): At measurement points with available measured and survey data At this point, the waveform features and physical property parameters obtained from the physical information neural network model are: The corresponding measured and survey data are Then the residual terms of the data Represented as:

[0044] in, This represents the number of sample points used in the fitting process.

[0045] The residual terms of the data are used to constrain the physical property parameters of the network output, which can ensure that the inversion results not only conform to the known physical property parameters, but also reasonably interpret the observed waveform data.

[0046] ② Residual terms of the wave equation (wave field physical constraints): The tunnel-surrounding rock wave field can be expressed in the general form of the two-dimensional elastic wave equation as follows:

[0047] in, This indicates how the position of the wave field changes over time; Density; For displacement; Internal stress; It is an external force.

[0048] Then physical constraint points Fluctuation residuals for:

[0049] Then the residual term of the wave equation (the constraint loss of the wave equation) for:

[0050] in, This represents the number of constraint points in the wave equation.

[0051] The residual terms of the wave equation are used to constrain the inversion results to meet the propagation laws of seismic waves.

[0052] ③ The residual terms of the mechanical equilibrium equations: constraint points Mechanical equilibrium residuals for:

[0053] in, Constraint points Volume forces, such as gravity, inertial force, and centrifugal force, are external force terms in the mechanical equilibrium equations.

[0054] The residual terms of the mechanical equilibrium equations for:

[0055] in, This represents the number of constraint points in the mechanical equilibrium equations.

[0056] By incorporating wave equations and mechanical equilibrium equations into the loss function, we ensure that the output physical properties and the distribution of adverse geological bodies conform to actual physical laws.

[0057] ④ Regularization term: To prevent overfitting in neural networks, a weight regularization term is introduced. as follows:

[0058] in, These are the weight parameters for each layer of the neural network; This represents the number of layers in the neural network.

[0059] In summary, the comprehensive loss function of the physical information neural network model It can be represented as:

[0060] in, These are the weighting coefficients for each loss term.

[0061] The physical information neural network model is iteratively trained using a stochastic gradient descent optimization algorithm, continuously updating the network's weights and biases to achieve the desired comprehensive loss function. It gradually decreases and tends to converge.

[0062] After training, a full waveform inversion model of the tunnel with good geological prediction capability under the geological conditions of this tunnel is obtained. It can be used for high-precision inversion and advanced prediction of physical parameters in front of the tunnel face and the distribution of adverse geological bodies.

[0063] S7. Full waveform inversion of geological physical parameters and geological body information ahead of the tunnel: During tunneling ahead of the tunnel face, vibration sources and receivers are deployed in the area to be predicted (40-80m ahead of the tunnel face) to collect new measured waveform data. Multi-feature extraction and data preprocessing are performed according to step S4 to obtain waveform features of the same type as those in the training phase. These features and their corresponding spatial coordinates are input into a trained physical information neural network model (i.e., the tunnel full waveform inversion model). The model outputs the target physical property parameters (such as P-wave velocity, S-wave velocity, density, water content, etc.) and the distribution locations of adverse geological bodies at each grid point in the area. Based on the inverted physical property parameters and the distribution of adverse geological bodies, an advanced geological prediction profile and risk zoning are formed ahead of the tunnel, providing a basis for construction support design and preliminary treatment measures.

[0064] Furthermore, when inverting different geological conditions, different parameters are selected as target parameters. Specifically, when encountering aquatic bodies, resistivity is used as the target parameter; when encountering fracture zones, density is used as the target parameter; and when encountering karst caves, either resistivity or density is used as the target parameter.

[0065] S8. Result Comparison and Model Update: By comparing the revealed geological conditions with the inversion results obtained from S7: when the inversion results are consistent with the actual geological conditions, the parameters of the existing model are retained; when there are deviations, the newly added geological data and corresponding waveform data are added to the training dataset, the physical information neural network model is retrained, and the parameter settings of the finite element and spectral element models are adjusted when necessary.

[0066] Although the present invention has been described using the above preferred embodiments, it is not intended to limit the scope of protection of the present invention. Any changes and modifications made by those skilled in the art to the above embodiments without departing from the spirit and scope of the present invention shall still fall within the scope of protection of the present invention.

Claims

1. A tunneling full-waveform inversion method based on spectral element forward modeling and physical information neural network, characterized in that, Includes the following steps: S1. Collect engineering geological data and obtain physical property parameters; S2. Based on engineering geological data, establish a two-dimensional finite element model and conduct working condition simulation to obtain spatial information; S3. Construct a wavefield model in the spectral element method software, perform forward wavefield modeling simulation, and obtain synthetic full waveform data; S4. Deploy seismic sources and receivers to conduct geophysical tests, collect measured waveform data, preprocess the measured waveform data, perform time registration and normalization processing on the preprocessed measured waveform data and the synthetic full waveform data, and extract multiple features from the processed waveform. S5. Construct a parameter training set with waveform features and spatial information as input layers and physical property parameters and geological body information as output layers; S6. Using the parameter training set as input, construct and train the physical information neural network model, and construct a comprehensive loss function during the training process; S7. Use a physical information neural network model to perform full waveform inversion of physical property parameters and geological body information of the geological area to be predicted; S8. By comparing the revealed geological conditions with the inversion results, determine whether to retain or retrain the physical information neural network model.

2. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 1, characterized in that, In S4, a seismic source is set at the center of the tunnel face, and multiple receivers are set up on the tunnel wall and around the tunnel face. The receivers are arranged along both sides of the tunnel surrounding rock and in the direction of the vertical axis.

3. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 1, characterized in that, In step S6, the comprehensive loss function of the physical information neural network model for: ， in, For the residual terms of the data; This represents the residual term of the wave equation; This represents the residual term in the mechanical equilibrium equation; For regularization terms; These are the weighting coefficients for each loss term.

4. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 3, characterized in that, In step S6, the residual term of the data for: ， in, The number of sample points involved in the fitting; The locations of the measurement points are based on actual measurements and survey data; For the predicted waveform characteristics and physical property parameters; These are the measured waveform characteristics and physical properties.

5. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 3, characterized in that, In step S6, the residual term of the wave equation for: ， in, This represents the number of constraint points for the wave equation. Physical constraint points The fluctuation residuals are expressed as: ， in, Density; Physical constraint points The predicted displacement; Physical constraint points Predicted internal stress; Physical constraint points External forces at the location.

6. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 3, characterized in that, In step S6, the residual term of the mechanical equilibrium equation for: ， in, This represents the number of constraint points in the mechanical equilibrium equations. Constraint points The mechanical equilibrium residuals on the surface are expressed as: ， in, Constraint points Predicted internal stress; Constraint points Volume force at the point.

7. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 3, characterized in that, In step S6, the regularization term for: ， in, These are the weight parameters for each layer of the neural network; This represents the number of layers in the neural network.

8. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 1, characterized in that, In step S6, the training process includes: iteratively training the physical information neural network model using a stochastic gradient descent optimization algorithm, continuously updating the network's weights and biases, so that the comprehensive loss function gradually decreases and tends to converge.

9. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 1, characterized in that, In step S7, the inversion process includes: collecting new measured waveform data for the area to be predicted, performing data preprocessing and multi-feature extraction, inputting the extracted waveform features and corresponding spatial information into the trained physical information neural network model, and outputting the target physical property parameters and geological body information of the area.

10. The tunnel full-waveform inversion method based on spectral element forward modeling and physical information neural network according to claim 1, characterized in that, In step S8, the judgment process includes: when the inversion result is consistent with the actual geological conditions, retaining the parameters of the existing physical information neural network model; when the inversion result deviates from the actual geological conditions, adding new engineering geological data and corresponding waveform data to the parameter training set, and retraining the physical information neural network model.

Images