Tunnel earthquake-electricity joint inversion method and system based on mixed attention mechanism

Abstract

The invention belongs to the technical field of geophysical exploration, and provides a tunnel earthquake-electricity joint inversion method and system based on a mixed attention mechanism, and the method comprises the steps: generating a three-dimensional model corresponding to the wave velocity and resistivity attributes according to the historical unfavorable geological type, and carrying out the imaging processing, respectively obtaining a corresponding seismic reflection imaging result and a resistivity inversion imaging result; three-dimensional structure position information is embedded after splicing; and constructing and training a multi-parameter inversion network, and processing the target construction exploration data by using the trained multi-parameter inversion network to obtain a wave velocity and resistivity inversion result of the target area. According to the method, serialization segmentation processing is carried out on the earthquake-electricity data, a joint inversion task is reexpressed as a sequence-to-sequence prediction problem, and position information and a weight relation of the earthquake-electricity data are captured through position coding and construction of a network structure based on a mixed attention mechanism; therefore, more accurate wave velocity-resistivity joint inversion is realized.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration technology, specifically relating to a tunnel seismic-electric joint inversion method and system based on a hybrid attention mechanism. 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] Tunnel face prediction is an effective means of identifying adverse geological conditions ahead of a tunnel. Accurate geological prediction during tunnel construction can effectively reduce construction risks and avoid accidents such as landslides, rockfalls, and sudden water or mud inrushes. Currently, common tunnel face prediction methods include the tunnel seismic wave method and the tunnel resistivity method. The tunnel seismic wave method detects anomalies by utilizing the propagation characteristics of seismic waves in the underground medium, and is suitable for detecting large-scale geological structures and deep anomalies. The resistivity method detects anomalies by measuring the resistivity changes of the surrounding rock, and can effectively identify low-resistivity bodies such as water bodies and karst caves. Both methods have their own detection sensitivities and advantages; the seismic wave method can provide deeper information, while the resistivity method performs better in shallow exploration. In complex geological environments, the tunnel seismic wave method and the resistivity method are often used simultaneously, and their results are integrated for comprehensive analysis to improve the accuracy of the detection.

[0004] According to the inventors, traditional tunnel comprehensive advance prediction methods typically rely on manual experience to judge seismic wave and resistivity detection results, or on constrained inversion based on rock property relationships or structural similarities. The former depends on expert intuition and experience, is highly subjective, and easily affected by operator judgment differences; the latter mainly relies on the relationship of physical parameters or the structural similarity of different detection methods for inversion, but this type of method has high regional requirements and lacks sufficient flexibility in complex geological environments. Overall, traditional methods often fail to fully utilize the advantages of multi-geophysical data when facing various complex geological bodies, limiting inversion accuracy and adaptability. In recent years, deep learning methods, due to their excellent feature extraction capabilities, have been widely applied in multi-source data fusion analysis, providing a new solution for multi-parameter tunnel inversion.

[0005] However, the following challenges remain in introducing deep learning methods for seismic-electrical joint inversion in practical tunnel engineering: Existing deep learning methods for solving multi-parameter joint inversion problems mostly employ convolutional neural networks (CNNs). By extracting local features from the data and gradually fusing them, CNNs can capture local spatial information relatively well. However, CNNs have a small receptive field, making it difficult to effectively model long-distance dependencies. This poses a challenge for data that needs to be processed on a large scale or over long distances, such as the joint inversion of tunnel seismic wave methods and resistivity methods.

[0006] In actual tunnel advance prediction projects, the prediction starting mileage of the long-distance seismic wave method and the short-distance resistivity method may be inconsistent, that is, the location of the common inversion area of ​​multiple geophysical fields is not fixed, and the existing deep learning network architecture has difficulty learning the relative spatial relationship between different data sources.

[0007] In joint inversion, different detection methods require different weights for physical parameters when facing different adverse geological conditions. The core challenge of using deep learning methods to solve joint inversion problems is to design a network structure that can effectively integrate various physical parameters and reasonably allocate their importance in the inversion. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a tunnel seismic-electric joint inversion method and system based on a hybrid attention mechanism. This invention re-formulates the joint inversion task as a sequence-to-sequence prediction problem by serializing and segmenting the seismic-electric data. Furthermore, by using position encoding and constructing a network structure based on a hybrid attention mechanism, it captures the position information and weight relationships of the seismic-electric data, thereby achieving a more accurate wave velocity-resistivity joint inversion.

[0009] According to some embodiments, the present invention adopts the following technical solution: A tunnel seismic-electric joint inversion method based on a hybrid attention mechanism includes the following steps: Based on historical adverse geological types, generate three-dimensional models corresponding to wave velocity and resistivity properties, and construct different types of sub-unit libraries respectively; Models from each sub-unit are randomly selected and spliced ​​together to form multiple three-dimensional integrated adverse geological models, which serve as sample data. Imaging processing was performed on the three-dimensional model in the three-dimensional integrated adverse geological model to obtain the corresponding seismic reflection imaging results and resistivity inversion imaging results, respectively. The seismic reflection imaging results and resistivity inversion imaging results are stitched together and then embedded with three-dimensional structural location information; A multi-parameter inversion network was constructed and trained. The multi-parameter inversion network includes a 3D block embedding module, an encoder, and a decoder. The 3D block embedding module divides the input data into fixed-size 3D blocks and generates a sequence of embedding vectors with positional encoding through linear mapping. The encoder consists of multiple cascaded hybrid attention modules and downsampling modules, used to extract deep cross-scale joint features of seismic-electric data step by step. The decoder consists of multiple hybrid attention modules and upsampling modules, which use skip connections to stitch together features at different scales and restore spatial resolution step by step. The trained multi-parameter inversion network is used to process the target construction exploration data to obtain the wave velocity and resistivity inversion results of the target area.

[0010] As an alternative implementation method, adverse geological types include faults, karst caves and aquifers. Fault sub-unit libraries, karst cave sub-unit libraries and empty model sub-unit libraries are constructed respectively. Each sub-unit library contains multiple three-dimensional models with different structures, locations and scales. As an alternative implementation method, the process of randomly selecting models from each sub-unit for splicing includes: randomly selecting four geological sub-units from the fault sub-unit library, the karst sub-unit library, and the empty model sub-unit library with a probability of 1:1:2 for splicing to form a predetermined number of undesirable geological bodies of the same size.

[0011] As an alternative implementation, the three-dimensional model includes a three-dimensional wave velocity model and a three-dimensional resistivity model.

[0012] As an optional implementation method, the process of imaging the three-dimensional model in the three-dimensional integrated adverse geological model includes: performing forward modeling and diffraction superposition imaging on the three-dimensional wave velocity model in the three-dimensional integrated adverse geological model to obtain the corresponding seismic reflection imaging results; and performing forward modeling and least squares inversion on the three-dimensional resistivity model in the three-dimensional integrated adverse geological model to obtain the resistivity inversion imaging results.

[0013] As an alternative implementation, the process of embedding three-dimensional structural location information after stitching together seismic reflection imaging results and resistivity inversion imaging results includes: after stitching together seismic reflection imaging results and resistivity inversion imaging results in the channel dimension, dividing them into three-dimensional embedding blocks of fixed size along the length, width, and height directions; flattening and linearly mapping the three-dimensional embedding blocks to obtain a high-dimensional embedding vector sequence; and embedding the original three-dimensional structural location information into the sequence representation through sinusoidal position coding to maintain the spatial consistency of seismic-electric data. As an alternative implementation, the multi-parameter inversion network is constructed based on a spatial-channel hybrid attention mechanism. The multi-parameter inversion network takes the serialized seismic-electric data as input, captures the relative spatial positional relationship between different detection methods through the spatial attention mechanism, and adaptively adjusts the contribution weight of different physical property parameters in the inversion process through the channel attention mechanism, thereby achieving the coordinated inversion of wave velocity and resistivity.

[0014] As an alternative implementation, during training, the gradient is calculated based on the output wave velocity and resistivity inversion results, and gradient backpropagation is performed to optimize and update the parameters of the multi-parameter inversion network.

[0015] As an alternative implementation, the three-dimensional block embedding module is used to transform the multi-channel input volume data of seismic wave method and resistivity method into an embedding vector sequence suitable for Transformer network processing, realizing the mapping from the three-dimensional spatial domain to the sequence domain. Specifically, the seismic imaging data and resistivity imaging data are stitched together in the channel dimension to form multi-channel inversion input volume data in a unified format. Subsequently, the input volume data is non-overlappingly segmented along the three spatial dimensions with a preset block size to obtain several three-dimensional embedding blocks. Each embedding block contains a joint expression of local seismic and resistivity information. Each embedding block is flattened and dimension-increasingly transformed using a linear projection layer to map it to a fixed-length high-dimensional embedding vector space to generate the initial token representation required by Transformer.

[0016] As an alternative implementation, to preserve the spatial position information of the embedded vector in the original 3D volume space, a sinusoidal position code is added to the output of the 3D block embedding module, calculated as follows: ,

[0017] in, This represents the index position of the patch in the sequence. Dimension index for location encoding The dimension of the embedded vector. and These represent the even-dimensional and odd-dimensional positional coding components, respectively. The positional coding generates a corresponding coding vector based on the three-dimensional coordinate position of the embedded block, and adds it to the embedded vector in the channel dimension to achieve spatial topological information embedding. In this way, the subsequent coding end can perceive the spatial relationship between different embedded vectors, ensuring that spatial structural information is not lost during the seismoelectric joint inversion process, and providing a priori reference for extracting geometric correlations.

[0018] As an alternative implementation, the hybrid attention module includes two parallel structures: a spatial attention branch and a channel attention branch. These two branches share a query matrix and a key matrix. They calculate the attention responses for the spatial and channel dimensions using different numerical matrices, and perform feature fusion at the output. This allows for the simultaneous capture of the spatial structure dependency and channel property correlation of seismic-electric data. The calculation process includes three parts, with the spatial attention branch calculated as follows:

[0019] The channel attention mechanism branch calculation is as follows:

[0020] Spatial-channel output feature fusion calculation is as follows:

[0021] in and These represent the query matrix and key matrix obtained from the input features through linear transformation, respectively. This is the projection form of the key matrix after downsampling in spatial dimension. and Let represent the numerical matrices in the spatial attention branch and the channel attention branch, respectively, where d is the scaling factor. Here is the weight normalization function. and These represent the output features of spatial attention and channel attention, respectively. Represents 3D convolution. Channel compression convolution, the final result The comprehensive output features of the hybrid attention module include feature representations enhanced by cross-space and cross-channel correlation.

[0022] As an alternative implementation, the training of the multi-parameter inversion network is constrained by an objective function: the neural network is optimized using the mean squared error as the model mismatch loss, and the mean squared error loss function is defined as follows:

[0023] in This represents the predicted wave velocity and resistivity output by the multi-parameter inversion network. Wave velocity, resistivity value, Calculate the L2 norm.

[0024] A tunnel seismic-electric joint inversion system based on a hybrid attention mechanism includes: The database construction module is configured to generate three-dimensional models corresponding to wave velocity and resistivity properties based on historical adverse geological types, and to construct different types of sub-unit libraries respectively; The sample model building module is configured to randomly extract models from each sub-unit and stitch them together to form multiple three-dimensional integrated adverse geological models as sample data. The imaging module is configured to perform imaging processing on the three-dimensional model in the three-dimensional integrated adverse geological model, and obtain the corresponding seismic reflection imaging results and resistivity inversion imaging results respectively. The data stitching module is configured to stitch together the seismic reflection imaging results and resistivity inversion imaging results and then embed the three-dimensional structural location information. The model building and training module is configured to build and train a multi-parameter inversion network. The multi-parameter inversion network includes a 3D block embedding module, an encoder, and a decoder. The 3D block embedding module divides the input data into fixed-size 3D blocks and generates a sequence of embedding vectors with positional encoding through linear mapping. The encoder consists of multiple cascaded hybrid attention modules and downsampling modules, used to extract deep cross-scale joint features of seismic-electric data step by step. The decoder consists of multiple hybrid attention modules and upsampling modules, which use skip connections to stitch together features at different scales and restore spatial resolution step by step. The joint inversion module is configured to process the target construction exploration data using a trained multi-parameter inversion network to obtain the wave velocity and resistivity inversion results of the target area.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the problem that traditional deep learning multi-parameter inversion methods, which commonly employ convolutional neural networks and struggle to model long-range dependencies of large-scale structures in seismic and electrical imaging results, proposes a Transformer-based MP-InvNet multi-parameter inversion network. By employing 3D block embedding and sinusoidal position encoding, seismic and electrical data are mapped to a unified sequence space. Furthermore, a multi-level hybrid attention mechanism is used to achieve global feature association, enabling the model to comprehensively understand complex geological structures such as faults, fracture zones, and aquifers at a larger scale. This effectively overcomes the problem of insufficient receptive field in convolutional networks, thereby significantly improving the structural identification capability and inversion accuracy of complex geological bodies ahead of tunnels.

[0026] This invention addresses the challenges in tunnel engineering where seismic wave methods and resistivity methods have different detection distances, inconsistent starting mileages, and non-fixed common inversion areas, making it difficult to establish spatial correspondences. By explicitly injecting the relative positional information of seismic-electric data in three-dimensional space through position encoding, MP-InvNet can automatically learn the spatial alignment relationships between different geophysical methods within the Transformer's self-attention mechanism. Simultaneously, a hybrid attention mechanism progressively fuses features at different scales in a multi-level encoding-decoding structure, ensuring consistent representation of the deep penetration information of seismic wave methods and the shallow, high-resolution information of resistivity methods. This improves the accuracy of seismic-electric structure correspondence and the spatial consistency of the inversion results.

[0027] This invention addresses the problem of varying sensitivity of different geophysical methods to different adverse geological bodies and the difficulty of effectively allocating weights for various physical property parameters in traditional joint inversion methods. It introduces a spatial-channel hybrid attention mechanism. Spatial attention captures the spatial dependencies of geological structures across regions, while channel attention adaptively adjusts the importance of wave velocity and resistivity in different geological scenarios, enabling complementary enhancement of the characteristics of seismic wave methods and resistivity methods. This mechanism can automatically focus on more discriminative physical property information based on the type of adverse geological body, achieving efficient collaborative inversion of wave velocity and resistivity, and improving the generalization ability of the inversion model under different geological environments. This provides more reliable multi-parameter integrated information for advanced tunnel geological prediction.

[0028] 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

[0029] 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.

[0030] Figure 1 A flowchart illustrating one embodiment; Figure 2 This is a schematic diagram of a multi-parameter inversion network processing procedure in one embodiment. Figure 3 This is a schematic diagram of the joint inversion results of one embodiment. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] 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.

[0033] 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.

[0034] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0035] Example 1 A tunnel seismic-electric joint inversion method based on a space-channel hybrid attention mechanism, such as Figure 1 As shown, it includes the following steps: Step S1: Construct a 3D geological model database. Based on common adverse geological types in tunnel engineering, such as faults, karst caves, and water-bearing bodies, generate a 3D geological sub-unit library corresponding to wave velocity and resistivity attributes. Each geological sub-unit is 30m×30m×30m in size. Construct fault sub-unit libraries, karst cave sub-unit libraries, and empty model sub-unit libraries respectively. Each geological sub-unit contains fault models with different locations, orientations, and dip angles, or karst cave models with different locations and scales, as well as empty models with different wave velocity and resistivity background attributes. Each geological sub-unit contains two attributes: wave velocity and resistivity. Four geological sub-units were randomly selected in a ratio of 1:1:2 and spliced ​​together to form a three-dimensional comprehensive adverse geological model of 30m×120m×30m. A total of 8,000 sets of samples were constructed for training and validation of the seismic-electric joint inversion method.

[0036] Step S2: Perform seismic and electrical numerical simulations. Forward modeling and diffraction stacking imaging are performed on the 3D wave velocity model to obtain the corresponding seismic reflection imaging results. Forward modeling and least-squares inversion are performed on the 3D resistivity model to obtain the 3D resistivity inversion imaging results. The seismic and resistivity images obtained in this step serve as preliminary feature inputs for the seismic-electrical joint inversion.

[0037] Step S3: Construct a multi-parameter inversion network MP-InvNet based on a spatial-channel hybrid attention mechanism module. This network includes a 3D patch embedding module, an encoder module, and a decoder module. The 3D patch embedding module takes multi-channel data (after stitching seismic imaging data and resistivity inversion results in the channel dimension) as input, divides the input volume data into fixed-size 3D blocks, and generates a sequence of embedding vectors with positional encoding through linear mapping. The multi-level hybrid attention encoder module takes the above embedding vector sequence as input and extracts cross-scale hybrid representations of seismic and electrical features through multiple levels of attention encoding blocks. The decoder module takes the feature sequences output from each encoding layer as input, realizes the step-by-step decoding of seismic-electrical features through hierarchical skip links and attention mechanisms, and restores spatial resolution through upsampling operations, finally obtaining the joint inversion results of the wave velocity model and resistivity model.

[0038] Specifically, the serialized seismic-electric input 3D block embedding process involves stitching seismic imaging results and resistivity inversion results together along the channel dimension and then inputting them into the 3D block embedding module. This module divides the input volume data into fixed-size 3D embedding blocks along the length, width, and height directions, flattens and linearly maps the embedding blocks to obtain a high-dimensional embedding vector sequence, and embeds the original 3D structural position information into the sequence representation through sinusoidal position encoding to maintain the spatial consistency of the seismic-electric data.

[0039] The 3D block embedding module transforms multi-channel input volume data from seismic wave and resistivity methods into embedding vector sequences suitable for Transformer network processing, achieving a mapping from the 3D spatial domain to the sequence domain. Specifically, seismic imaging data and resistivity imaging data are concatenated along the channel dimension to form multi-channel inversion input volume data in a unified format. Subsequently, the input volume data is non-overlappingly segmented along the three spatial dimensions with a preset block size, resulting in several 3D embedding blocks. Each embedding block contains a joint representation of local seismic and resistivity information. Each embedding block is flattened and dimension-increasingly transformed using a linear projection layer, mapping it to a fixed-length high-dimensional embedding vector space to generate the initial token representation required by the Transformer.

[0040] To preserve the spatial position information of the embedded vector in the original 3D volume space, sinusoidal positional encoding is added to the output of the 3D block embedding module. The calculation formula is as follows: ,

[0041] in, This represents the index position of the patch in the sequence. Dimension index for location encoding The dimension of the embedded vector. and These represent the even-dimensional and odd-dimensional positional encoding components, respectively. Positional encoding generates corresponding encoding vectors based on the 3D coordinates of the embedded blocks, and these vectors are added to the embedding vectors in the channel dimension to embed spatial topological information. This method enables subsequent multi-level hybrid attention encoders to perceive the spatial relationships between different embedding vectors, ensuring that spatial structural information is not lost during the seismoelectric joint inversion process, and providing prior references for MP-InvNet to extract geometric correlations.

[0042] Step S4: Construct and train the MP-InvNet multi-parameter inversion network, such as... Figure 2As shown, the MP-InvNet network in this embodiment includes a symmetrical multi-level hybrid attention encoding and decoding structure: the encoding end consists of four layers of cascaded hybrid attention modules and downsampling modules, used to extract deep cross-scale joint features of seismic-electric data step by step; the decoding end consists of four layers of hybrid attention modules and upsampling modules, which realize the concatenation of features at different scales through skip connections and restore spatial resolution step by step.

[0043] The hybrid attention module includes two parallel structures: a spatial attention branch and a channel attention branch. The two branches share the query matrix and the key matrix. They calculate the attention responses in the spatial dimension and the channel dimension respectively through different numerical matrices, and perform feature fusion at the output end, thereby simultaneously capturing the spatial structure dependence and channel physical property correlation of seismic-electric data.

[0044] The hybrid attention module consists of spatial attention and channel attention branches, used to simultaneously extract importance weights across location dependencies and across material property parameters. Spatial attention output. The calculation is as follows:

[0045] The channel attention mechanism branch calculation is as follows:

[0046] Spatial-channel output feature fusion calculation is as follows:

[0047] in and These represent the query matrix and key matrix obtained from the input features through a linear transformation, respectively. This is the projection form of the key matrix after downsampling in spatial dimension. and These represent the numerical matrices in the spatial attention branch and the channel attention branch, respectively, where d is the scaling factor. Here is the weight normalization function. and These represent the output features of spatial attention and channel attention, respectively. Represents 3D convolution. Channel compression convolution, the final result The comprehensive output features of the hybrid attention module include feature representations enhanced by both cross-spatial and cross-channel correlation. Mean squared error loss (MSE) is used as the training objective function to perform backpropagation and iterative optimization of the MP-InvNet network parameters. The objective function is:

[0048] Step S5: Apply the trained multi-parameter inversion network MP-InvNet, based on a spatial-channel hybrid attention mechanism, to the test set data to verify the seismic-electric joint inversion performance. Partial inversion results on the test set are shown below. Figure 3 As shown, by introducing spatial attention and channel attention into the network to achieve deep fusion of cross-scale seismic-electric features, MP-InvNet achieves significantly better joint inversion results than networks driven by a single geophysical input or purely data features. Especially in complex and challenging geological areas such as faults, karst caves, and aquifers, this invention can more accurately recover the spatial location and physical property distribution of anomalies, verifying the effectiveness and superiority of the proposed hybrid attention mechanism in seismic-electric joint inversion.

[0049] The main network parameters and hardware conditions in this embodiment are as follows: computation is implemented using two NVIDIA A100 Tensor Core GPUs. The network is built on the PyTorch platform, using the Adam optimizer, with a batch size of 50, a learning rate of 0.0005, and exponential decay during the training process.

[0050] Example 2 A tunnel seismic-electric joint inversion system based on a hybrid attention mechanism includes: The database construction module is configured to generate three-dimensional models corresponding to wave velocity and resistivity properties based on historical adverse geological types, and to construct different types of sub-unit libraries respectively; The sample model building module is configured to randomly extract models from each sub-unit and stitch them together to form multiple three-dimensional integrated adverse geological models as sample data. The imaging module is configured to perform imaging processing on the three-dimensional model in the three-dimensional integrated adverse geological model, and obtain the corresponding seismic reflection imaging results and resistivity inversion imaging results respectively. The data stitching module is configured to stitch together the seismic reflection imaging results and resistivity inversion imaging results and then embed the three-dimensional structural location information. The model building and training module is configured to build and train a multi-parameter inversion network. The multi-parameter inversion network includes a 3D block embedding module, an encoder, and a decoder. The 3D block embedding module divides the input data into fixed-size 3D blocks and generates a sequence of embedding vectors with positional encoding through linear mapping. The encoder consists of multiple cascaded hybrid attention modules and downsampling modules, used to extract deep cross-scale joint features of seismic-electric data step by step. The decoder consists of multiple hybrid attention modules and upsampling modules, which use skip connections to stitch together features at different scales and restore spatial resolution step by step. The joint inversion module is configured to process the target construction exploration data using a trained multi-parameter inversion network to obtain the wave velocity and resistivity inversion results of the target area.

[0051] 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.).

[0052] 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.

[0053] 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.

[0054] 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.

[0055] 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 tunnel seismic-electric joint inversion method based on a hybrid attention mechanism, characterized in that, Includes the following steps: Based on historical adverse geological types, generate three-dimensional models corresponding to wave velocity and resistivity properties, and construct different types of sub-unit libraries respectively; Models from each sub-unit are randomly selected and spliced ​​together to form multiple three-dimensional integrated adverse geological models, which serve as sample data. Imaging processing was performed on the three-dimensional model in the three-dimensional integrated adverse geological model to obtain the corresponding seismic reflection imaging results and resistivity inversion imaging results, respectively. The seismic reflection imaging results and resistivity inversion imaging results are stitched together and then embedded with three-dimensional structural location information; A multi-parameter inversion network was constructed and trained. The multi-parameter inversion network includes a 3D block embedding module, an encoder, and a decoder. The 3D block embedding module divides the input data into fixed-size 3D blocks and generates a sequence of embedding vectors with positional encoding through linear mapping. The encoder consists of multiple cascaded hybrid attention modules and downsampling modules, used to extract deep cross-scale joint features of seismic-electric data step by step. The decoder consists of multiple hybrid attention modules and upsampling modules, which use skip connections to stitch together features at different scales and restore spatial resolution step by step. The trained multi-parameter inversion network is used to process the target construction exploration data to obtain the wave velocity and resistivity inversion results of the target area.

2. The tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, Unfavorable geological types include faults, karst caves, and aquifers. Fault sub-unit libraries, karst cave sub-unit libraries, and empty model sub-unit libraries are constructed respectively. Each sub-unit library contains multiple three-dimensional models with different structures, locations, and scales. The process of randomly selecting models from each sub-unit and splicing them together includes: randomly selecting four geological sub-units from the fault sub-unit library, the karst sub-unit library, and the empty model sub-unit library with a probability of 1:1:2 and splicing them together to form a predetermined number of undesirable geological bodies of the same size.

3. The tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, The imaging process of the three-dimensional model in the three-dimensional integrated adverse geological model includes: performing forward modeling and diffraction superposition imaging on the three-dimensional wave velocity model in the three-dimensional integrated adverse geological model to obtain the corresponding seismic reflection imaging results; and performing forward modeling and least squares inversion on the three-dimensional resistivity model in the three-dimensional integrated adverse geological model to obtain the resistivity inversion imaging results.

4. The tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, The process of embedding three-dimensional structural location information after stitching together seismic reflection imaging results and resistivity inversion imaging results includes: stitching the seismic reflection imaging results and resistivity inversion imaging results together in the channel dimension and dividing them into three-dimensional embedding blocks of fixed size along the length, width and height directions; flattening and linearly mapping the three-dimensional embedding blocks to obtain a high-dimensional embedding vector sequence; and embedding the original three-dimensional structural location information into the sequence representation through sinusoidal position encoding to maintain the spatial consistency of seismic-electric data.

5. The tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, The multi-parameter inversion network is constructed based on a space-channel hybrid attention mechanism. The multi-parameter inversion network takes the serialized seismic-electric data as input, captures the relative spatial positional relationship between different detection methods through the space attention mechanism, and adaptively adjusts the contribution weight of different physical property parameters in the inversion process through the channel attention mechanism, thereby achieving the coordinated inversion of wave velocity and resistivity.

6. The tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, The 3D block embedding module is used to transform the multi-channel input volume data of seismic wave method and resistivity method into an embedding vector sequence suitable for Transformer network processing, realizing the mapping from the 3D spatial domain to the sequence domain. Specifically, the seismic imaging data and resistivity imaging data are spliced ​​in the channel dimension to form a unified format of multi-channel inversion input volume data. Then, the input volume data is non-overlappingly segmented along the three spatial dimensions with a preset block size to obtain several 3D embedding blocks. Each embedding block contains a joint expression of local seismic and resistivity information. Each embedding block is flattened and dimension-increasingly transformed using a linear projection layer to map it to a fixed-length high-dimensional embedding vector space to generate the initial token representation required by Transformer.

7. The tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, To preserve the spatial position information of the embedded vector in the original 3D volume space, a sinusoidal position code is added to the output of the 3D block embedding module. The calculation formula is as follows: , in, This is the index position of the patch in the sequence. Dimension index for location encoding The dimension of the embedded vector. and These represent the even-dimensional and odd-dimensional positional coding components, respectively. The positional coding generates a corresponding coding vector based on the three-dimensional coordinate position of the embedded block, and adds it to the embedded vector in the channel dimension to achieve spatial topological information embedding. In this way, the subsequent coding end can perceive the spatial relationship between different embedded vectors, ensuring that spatial structural information is not lost during the seismoelectric joint inversion process, and providing a priori reference for extracting geometric correlations.

8. The tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, The hybrid attention module comprises two parallel structures: a spatial attention branch and a channel attention branch. These two branches share the query matrix and the key matrix. They calculate the attention responses for the spatial and channel dimensions using different numerical matrices, and perform feature fusion at the output. This allows for the simultaneous capture of the spatial structural dependence and channel property correlation of seismic-electric data. The calculation process includes three parts, with the spatial attention branch calculated as follows: The channel attention mechanism branch calculation is as follows: Spatial-channel output feature fusion calculation is as follows: in and These represent the query matrix and key matrix obtained from the input features through linear transformation, respectively. This is the projection form of the key matrix after downsampling in spatial dimension. and Let represent the numerical matrices in the spatial attention branch and the channel attention branch, respectively, where d is the scaling factor. Here is the weight normalization function. and These represent the output features of spatial attention and channel attention, respectively. Represents 3D convolution. Channel compression convolution, the final result The integrated output features of the hybrid attention module include feature representations enhanced by cross-space and cross-channel correlation.

9. A tunnel seismic-electric joint inversion method based on a hybrid attention mechanism as described in claim 1, characterized in that, During training, based on the output wave velocity and resistivity inversion results, the gradient is calculated based on the loss function and gradient backpropagation is performed to optimize and update the parameters of the multi-parameter inversion network. The objective function is designed to constrain the training of the multi-parameter inversion network: the mean squared error is used as the model mismatch loss to optimize the neural network, and the mean squared error loss function is defined as follows: in This represents the predicted wave velocity and resistivity output by the multi-parameter inversion network. Wave velocity, resistivity value, Calculate the L2 norm.

10. A tunnel seismic-electric joint inversion system based on a hybrid attention mechanism, characterized in that, include: The database construction module is configured to generate three-dimensional models corresponding to wave velocity and resistivity properties based on historical adverse geological types, and to construct different types of sub-unit libraries respectively; The sample model building module is configured to randomly extract models from each sub-unit and stitch them together to form multiple three-dimensional integrated adverse geological models as sample data. The imaging module is configured to perform imaging processing on the three-dimensional model in the three-dimensional integrated adverse geological model, and obtain the corresponding seismic reflection imaging results and resistivity inversion imaging results respectively. The data stitching module is configured to stitch together the seismic reflection imaging results and resistivity inversion imaging results and then embed the three-dimensional structural location information. The model building and training module is configured to build and train a multi-parameter inversion network. The multi-parameter inversion network includes a 3D block embedding module, an encoder, and a decoder. The 3D block embedding module divides the input data into fixed-size 3D blocks and generates a sequence of embedding vectors with positional encoding through linear mapping. The encoder consists of multiple cascaded hybrid attention modules and downsampling modules, used to extract deep cross-scale joint features of seismic-electric data step by step. The decoder consists of multiple hybrid attention modules and upsampling modules, which use skip connections to stitch together features at different scales and restore spatial resolution step by step. The joint inversion module is configured to process the target construction exploration data using a trained multi-parameter inversion network to obtain the wave velocity and resistivity inversion results of the target area.

Images