Multi-modal gradient iterative inversion method and system based on target mask labeling

Abstract

The application belongs to the field of tunnel engineering advanced geological prediction and geophysical joint inversion technology, and provides a multi-modal gradient iterative inversion method and system based on target mask marking. According to the parameter variation characteristics of the inversion result itself, the abnormal region category is adaptively determined, the influence of small amplitude fluctuation, noise or local inversion error on the abnormal body identification result is reduced, and the accuracy and stability of the abnormal body region grading calibration are improved. Based on the abnormal body region identification result, a space mask is constructed, the abnormal region and the background region are marked, and the fused gradient is weighted and corrected by using the space mask. It not only avoids the loss of overall geological structure constraint information caused by directly setting the background region to zero, but also can enhance the attention degree of the abnormal target area in the network training or model updating process, improve the boundary description ability and inversion imaging precision of complex abnormal bodies such as water-bearing body, fault fracture zone and irregular water-conducting channel.

Description

Technical Field

[0001] This invention belongs to the field of advanced geological prediction and geophysical joint inversion technology for tunnel engineering, and particularly relates to a multimodal gradient iterative inversion method and system based on target mask marking. Background Technology

[0002] Conducting advanced geological forecasting ahead of tunnel construction and accurately identifying and imaging complex geological bodies in front of the tunnel face are important technical means to ensure safe tunnel construction. In recent years, multimodal learning and deep learning methods have been gradually introduced into multi-source geophysical joint inversion. By extracting and fusing features from different physical field data or inversion results, the ability to identify complex geological bodies can be improved to a certain extent.

[0003] However, under complex geological conditions, anomalies in the inversion results may exhibit parameter variations of different intensities, ranges, and morphologies. Existing methods typically use fixed thresholds, prior knowledge, or empirical parameters to identify anomalous regions, making it difficult to adaptively determine the number and boundaries of anomaly categories based on the parameter fluctuation characteristics of the inversion results themselves. Furthermore, existing inversion or network training processes fail to adequately distinguish the differences in importance between anomaly regions and background regions during the inversion process. When noise, small fluctuations, or local errors exist in the inversion results, fixed or empirical threshold methods are prone to misclassification, omission, or overclassification of anomalies, affecting the stability of subsequent target region identification. For key anomaly target areas such as water-bearing bodies, fault fracture zones, and karst conduits, the lack of masking and weighting mechanisms based on spatial location and anomaly category makes the network training or model update process susceptible to interference from background regional information, resulting in insufficient clarity in the delineation of anomaly boundaries and morphologies. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a multimodal gradient iterative inversion method and system based on target mask labeling. This invention adaptively determines the anomaly region category based on the parameter variation characteristics of the inversion results themselves, avoiding the problem of relying on manual experience with fixed thresholds. It reduces the impact of small fluctuations, noise, or local inversion errors on the anomaly identification results, improving the accuracy and stability of anomaly region classification and labeling. Based on the anomaly region identification results, a spatial mask is constructed to label the anomaly region and the background region. The spatial mask is then used to weight and correct the fused gradient, making the weight of the anomaly region higher than that of the background region. This avoids directly setting the background region to zero, which would lead to the loss of overall geological structure constraint information, and enhances the attention paid to the anomaly target area during network training or model updates. This improves the boundary characterization ability and inversion imaging accuracy of complex anomalies such as aquifers, fault fracture zones, and irregular water-conducting channels.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a multimodal gradient iterative inversion method based on target mask marking, comprising: Resistivity data, seismic wave data, and ground-penetrating radar data in front of the tunnel face are obtained through forward modeling. Based on resistivity data, seismic wave data, and ground penetrating radar data, the multi-parameter model increment and fused gradient are determined through inversion and fusion. Anomalies are identified in the inverted area based on the multi-parameter model increment and fused gradient, and the anomaly area is classified and calibrated by interval division and threshold factor assignment. A spatial mask is constructed based on the results of anomaly region identification. The anomaly region and the background region are marked. The spatial mask is then used to perform weighted correction on the fused gradient, so that the weight of the anomaly region is higher than that of the background region. The inversion and fusion, anomaly identification and weighted correction are continuously optimized until the preset convergence conditions are met, and the final inversion result is determined.

[0006] Furthermore, the forward modeling includes: assigning resistivity, seismic wave velocity, and radar relative permittivity to the model mesh respectively, forming resistivity parameter volume, seismic wave velocity parameter volume, and radar relative permittivity parameter volume, and then performing forward modeling using the resistivity method, seismic wave method, and ground penetrating radar method respectively to obtain the corresponding resistivity data, seismic wave data, and ground penetrating radar data.

[0007] Furthermore, the determination of the multi-parameter model increment and the fused gradient through inversion and fusion includes: for resistivity data, determining the resistivity model parameter increment based on smooth constraint least squares inversion; for seismic wave data, determining the gradient of the seismic wave velocity model parameter and the seismic wave velocity model parameter increment based on the seismic wave full waveform inversion objective function; for ground penetrating radar data, determining the gradient of the radar relative permittivity model parameter and the radar relative permittivity model parameter increment based on the ground penetrating radar full waveform inversion objective function.

[0008] Furthermore, the resistivity model parameter increments are: ; in, This represents the increment of resistivity model parameters; This represents the sensitivity matrix corresponding to the resistivity method. Represents the smoothness constraint matrix; Indicates the smoothing constraint weight parameters; This represents resistivity observation data; This represents the forward modeling data corresponding to the current resistivity model; The increments of the seismic wave velocity model parameters are: ; ; in, This represents the objective function for the full waveform inversion of seismic waves; Indicates the parameters of the seismic wave velocity model; Represents seismic wave observation data, Represents the forward modeling operator for seismic waves; This represents the gradient of the objective function for seismic wave full waveform inversion with respect to the parameters of the seismic wave velocity model; This indicates the update step size of the seismic wave velocity model; Represents a preconditioning operator; The parameter increments for the radar relative permittivity model are: ; ; in, This represents the target function for inverting the full waveform of a ground-penetrating radar. This represents the model parameters of the radar relative permittivity; This represents ground-penetrating radar observation data; This represents the forward modeling operator for ground-penetrating radar; This represents the gradient of the ground-penetrating radar full-waveform inversion objective function with respect to the radar relative permittivity model parameters; This indicates the update step size of the radar relative permittivity model; This represents the preconditioning operator.

[0009] Furthermore, the model is updated based on the fused gradient and the multi-parameter model increment to obtain the updated model for the current round. The updated model is then used as the initial model for the next round of inversion, and the traditional inversion iteration is performed again to obtain the resistivity gradient, seismic wave velocity gradient, and radar relative permittivity gradient corresponding to the next round of iteration.

[0010] Furthermore, the step of classifying and labeling anomaly regions through interval division and threshold factor assignment includes: obtaining the minimum and maximum parameter values ​​in the current joint inversion results, and determining the parameter variation range. Based on the location of parameter changes in the current joint inversion results, profile data is extracted along the tunnel detection direction, and the parameter differences between adjacent sampling points are determined to form a difference fluctuation sequence; based on the mean and standard deviation of the difference fluctuation sequence; Local peaks in the difference fluctuation sequence are used as candidate category boundaries, and based on the threshold factor... z Candidate peaks are filtered; when a candidate peak is greater than a threshold factor... z If the peak value is within a certain range, it is considered a valid peak; otherwise, it is considered a minor local fluctuation and is discarded. The number of anomaly categories is determined based on the number of valid peaks; Determining the number of anomaly categories n Then, the parameter intervals of the current joint inversion results are divided according to the minimum parameter value and the parameter variation range, and different parameter intervals are assigned hierarchical labels.

[0011] Furthermore, the step of using the spatial mask to perform weighted correction on the fused gradient includes: constructing a spatial mask matrix based on the hierarchical labeling results of the anomalous body region, labeling the anomalous body region as the target region, and labeling the background region as the non-target region; and counting the number of grids in the anomalous body region according to the spatial mask matrix, constructing an anomalous region proportion adjustment factor in combination with the total number of model grids, and generating a target region weight matrix.

[0012] Furthermore, the anomalous body region The weight varies with each grid point in the spatial mask matrix. Increases with the increase of: ; ; in, N is Define the total number of model grids; This represents the number of grid cells in the spatial mask matrix that are marked as anomalous regions. This is an adjustment factor for the proportion of the anomalous region in the overall model space; This represents the enhancement coefficient of the abnormal region.

[0013] Furthermore, the weight matrix Introducing a weighted loss function into the loss function of the multimodal fusion model: ; in, L Represents the weighted loss function; This indicates element-wise multiplication; This represents the fused gradient or the increment of model parameters output by the multimodal fusion model. This represents the target gradient or the increment of the target model parameters.

[0014] Secondly, the present invention also provides a multimodal gradient iterative inversion system based on target mask marking, comprising: The forward modeling module is configured to acquire resistivity data, seismic wave data, and ground-penetrating radar data in front of the tunnel face through forward modeling. The hierarchical calibration module is configured to: determine the multi-parameter model increment and fused gradient based on resistivity data, seismic wave data and ground penetrating radar data through inversion and fusion; identify anomalies in the inverted area based on the multi-parameter model increment and fused gradient; and perform hierarchical calibration of the anomaly area through interval division and threshold factor assignment. The weighted correction module is configured to: construct a spatial mask based on the abnormal body region identification result, mark the abnormal region and the background region, and use the spatial mask to perform weighted correction on the fused gradient so that the weight of the abnormal body region is higher than the weight of the background region. The inversion module is configured to continuously optimize the inversion and fusion, anomaly identification and weighted correction until the preset convergence conditions are met, and to determine the final inversion result.

[0015] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multimodal gradient iterative inversion method based on target mask marking described in the first aspect.

[0016] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the multimodal gradient iterative inversion method based on target mask marking described in the first aspect.

[0017] Fifthly, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the multimodal gradient iterative inversion method based on target mask marking described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention first determines the multi-parameter model increment and fused gradient based on resistivity data, seismic wave data, and ground-penetrating radar data through inversion and fusion. Then, based on the multi-parameter model increment and fused gradient, anomalies are identified in the inverted region, and the anomaly regions are classified and labeled through interval division and threshold factor assignment. Next, a spatial mask is constructed based on the anomaly region identification results to mark the anomaly regions and background regions. The fused gradient is then weighted using the spatial mask, ensuring that the weight of the anomaly regions is higher than that of the background regions. Finally, the inversion and fusion, anomaly identification, and weighted correction are continuously optimized until a preset convergence condition is met, determining the final inversion result. It can adaptively determine the anomaly region category based on the parameter change characteristics of the inversion results themselves, avoiding the problem of relying on human experience for fixed thresholds, reducing the impact of small fluctuations, noise or local inversion errors on the anomaly identification results, and improving the accuracy and stability of anomaly region classification and labeling. Based on the anomaly region identification results, a spatial mask is constructed to mark the anomaly region and the background region, and the spatial mask is used to perform weighted correction on the fused gradient, so that the weight of the anomaly region is higher than that of the background region. This not only avoids the loss of overall geological structure constraint information caused by directly setting the background region to zero, but also enhances the attention to the anomaly target area during network training or model update, improving the boundary characterization ability and inversion imaging accuracy of complex anomalies such as water-bearing bodies, fault fracture zones and irregular water-conducting channels. Attached Figure Description

[0019] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0020] Figure 1 This is a schematic diagram of the overall process of the method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the iterative mechanism in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the adaptive determination of anomaly categories and hierarchical mask calibration process in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the spatial mask construction and gradient weighting correction mechanism based on the target region in Embodiment 1 of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. 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 application pertains.

[0023] Example 1: Under complex geological conditions in tunnel construction, adverse geological bodies such as fault fracture zones, water-rich areas, and karst development zones typically exhibit spatial superposition and non-uniform development characteristics, exhibiting high concealment and suddenness. As a result, tunnel construction is prone to engineering disasters such as water and mud inrushes, collapses, and surrounding rock instability, posing a significant threat to construction safety. Therefore, conducting advanced geological forecasting ahead of tunnel construction and accurately identifying and imaging complex geological bodies ahead of the tunnel face are crucial technical means to ensure safe tunnel construction.

[0024] Currently, various geophysical exploration methods, such as seismic wave methods, resistivity methods, and ground-penetrating radar (GPR), are commonly used in engineering to comprehensively explore geological bodies ahead. Different geophysical methods correspond to different physical field responses and have varying sensitivities to targets such as fault structures, water-rich anomalies, karst bodies, and fracture zones. Seismic wave methods are more sensitive to differences in wave impedance and structural interfaces, resistivity methods are more sensitive to water-rich and conductive anomalies, and GPR has high resolution for near-range local anomalies and shallow structures. Due to differences in detection mechanisms, response scales, spatial resolution, and result representation, a single geophysical method often cannot fully characterize the spatial morphology and physical properties of complex geological bodies. Therefore, using multi-source geophysical data for joint inversion has become an important direction for improving the accuracy and reliability of tunnel advanced geological prediction.

[0025] In recent years, multimodal learning and deep learning methods have been gradually introduced into multi-source geophysical joint inversion. By extracting and fusing features from different physical field data or inversion results, the ability to identify complex geological bodies can be improved to some extent. However, existing multimodal joint inversion methods still have the following shortcomings in engineering applications: Multi-source gradient information is difficult to sustain in inversion iterations. Existing methods often focus on a one-time fusion of multi-source observation data or inversion results, with the fused result typically used directly as the final prediction or interim input, failing to effectively embed it into the subsequent model update process of traditional inversion. This results in multi-source gradient information failing to play a sustained role in subsequent iterations, and a lack of effective connection between the physical constraints of traditional inversion and the feature fusion capabilities of deep learning models, affecting the gradual approximation and stable imaging of complex geological bodies.

[0026] The adaptive capabilities of anomaly category identification and classification are insufficient. Under complex geological conditions, anomalies in the inversion results may exhibit parameter variations of different intensities, ranges, and morphologies. Existing methods typically use fixed thresholds, prior knowledge, or empirical parameters to identify anomalous regions, making it difficult to adaptively determine the number and boundaries of anomaly categories based on the parameter fluctuation characteristics of the inversion results themselves. When noise, small fluctuations, or local errors exist in the inversion results, fixed or empirical threshold methods are prone to misclassification, omission, or overclassification of anomalies, affecting the stability of subsequent target area identification.

[0027] The anomaly target area lacks a differentiated weighted optimization mechanism. Existing inversion or network training processes typically employ uniform loss constraints or global update strategies for the entire inversion region, failing to adequately distinguish the importance differences between the anomaly region and the background region during the inversion process. For key anomaly target areas such as water-bearing bodies, fault fracture zones, and karst channels, the lack of masking and weight adjustment mechanisms based on spatial location and anomaly category makes the network training or model update process susceptible to interference from background regional information, resulting in insufficient clarity in the delineation of anomaly boundaries and morphology.

[0028] In summary, existing multi-source joint inversion methods for tunnels still need to address issues such as the continuous participation of multi-source gradient fusion results in model updates, adaptive identification and hierarchical calibration of anomaly regions, and differentiated weighted optimization of anomaly target areas, in order to improve the accuracy, stability, and engineering applicability of joint inversion results under complex geological conditions.

[0029] To address the aforementioned issues, this embodiment provides a multimodal gradient iterative inversion method based on target mask labeling. It extracts the resistivity gradient, seismic wave velocity gradient, and radar relative permittivity gradient obtained through traditional inversion and inputs them into a multimodal fusion model. This constructs an inversion mechanism combining multi-source gradient fusion with traditional iterative updates, enabling the fused gradient or model parameter increments to continuously participate in model updates. To address the instability in identifying complex anomaly regions, a threshold factor setting method based on fluctuations in adjacent data differences is proposed. The number of anomaly categories and category boundaries are determined by the number of effective peaks, achieving adaptive hierarchical calibration of anomaly regions. Furthermore, a spatial mask and target area weight matrix are constructed based on the anomaly hierarchical calibration results. This allows anomaly regions to receive higher weights during network training or model updates, while background regions retain lower weights, thereby improving the accuracy, stability, and adaptability of multi-source joint inversion results under complex geological conditions. Figure 1 As shown, the method includes: S1. Set the three-dimensional model size and mesh size of the common area for joint inversion in front of the tunnel face, and assign resistivity, seismic wave velocity and radar relative permittivity to the model mesh according to the differences in physical properties of the surrounding rock background, fault fracture zone, water-bearing anomaly and other adverse geological bodies, forming resistivity parameter volume, seismic wave velocity parameter volume and radar relative permittivity parameter volume.

[0030] Optionally, the spatial extent, number of grids, and grid spacing of the joint inversion common area in front of the tunnel face are first defined. The spatial dimensions of the 3D model can be expressed as follows: The number of grids can be expressed as The corresponding grid spacings are as follows: , and In this embodiment, the joint inversion common area can be selected as... The number of grids mentioned above is only an example and can be adjusted according to the detection distance in front of the tunnel face, computing resources, and the resolution of multi-source data.

[0031] The three-dimensional model is discretized using a unified mesh, and the three-parameter physical properties are assigned within the same mesh system, ensuring a one-to-one correspondence between different physical field parameters in space. Specifically, the resistivity parameter is preferably selected from a range of values. The preferred range for seismic wave velocity parameters is... The preferred range for the radar relative permittivity is... Among them, resistivity parameter is used to characterize the difference in conductivity of the medium, seismic wave velocity parameter is used to characterize the difference in elastic wave propagation velocity of the medium, and radar relative permittivity is used to characterize the difference in electromagnetic wave propagation characteristics of the medium.

[0032] Based on the differences in physical properties of the surrounding rock background, fault fracture zones, water-bearing anomalies, and other adverse geological bodies, resistivity, seismic wave velocity, and radar relative permittivity are assigned to each grid cell to form a resistivity parameter body. Seismic wave velocity parameters and radar relative permittivity parameter body .

[0033] Optionally, during the actual construction, the 3D model is first discretized into a mesh, and the geological type of each mesh cell is determined, such as the surrounding rock background, fault fracture zone, water-bearing anomaly, etc. Then, based on the range of physical property parameters corresponding to different geological bodies, resistivity, seismic wave velocity, and radar relative permittivity are assigned to each mesh cell. The resistivity values ​​on all mesh cells together constitute the resistivity parameter volume, the seismic wave velocity values ​​on all mesh cells together constitute the seismic wave velocity parameter volume, and the radar relative permittivity values ​​on all mesh cells together constitute the radar relative permittivity parameter volume. The parameter volume can be understood as the parameter distribution of a certain physical property parameter in the 3D mesh space, also called a 3D parameter field or 3D parameter model; or it can be directly represented by resistivity. Seismic wave velocity Radar relative permittivity express.

[0034] S2. Based on the resistivity parameter volume, seismic wave velocity parameter volume, and radar relative permittivity parameter volume obtained in step S1, forward modeling simulations are performed using the resistivity method, seismic wave method, and ground-penetrating radar method, respectively, to obtain the corresponding resistivity data, seismic wave data, and ground-penetrating radar data; and using a uniform background model as the initial model, traditional inversion is performed on each type of data to obtain gradient information reflecting the direction of model update, and the corresponding model parameter increments are further calculated.

[0035] S2.1 In step S2, based on the resistivity parameter volume, seismic wave velocity parameter volume, and radar relative permittivity parameter volume obtained in step 1, multi-source geophysical data are acquired. As an optional implementation method, the resistivity method uses the finite element method for forward modeling to obtain resistivity data; the seismic wave method uses the wave equation for forward modeling to obtain seismic wave data; and the ground-penetrating radar method uses forward modeling based on the electromagnetic field equation to obtain ground-penetrating radar data.

[0036] S2.2 Using a uniform background model as the initial model, conventional inversion is performed on the resistivity data, seismic wave data, and ground-penetrating radar data to obtain gradient information reflecting the model update direction, and the corresponding model parameter increments are further calculated. Specifically, the resistivity data is processed using a least-squares inversion method based on smoothness constraints, while the seismic wave data and ground-penetrating radar data are processed using a full-waveform inversion method.

[0037] For ease of description, the increments of seismic wave velocity model parameters, resistivity model parameters, and radar relative permittivity model parameters obtained through traditional inversion are respectively denoted as... , and Among them, subscript s Indicates seismic wave velocity parameters, subscript r Represents resistivity parameters, subscript dThis represents the radar relative permittivity parameter.

[0038] For resistivity data, the resistivity model parameter increments are solved using smooth constraint least squares inversion, and these parameter increments can be expressed as: ; in, This represents the increment of resistivity model parameters; This represents the sensitivity matrix corresponding to the resistivity method. Represents the smoothness constraint matrix; Indicates the smoothness constraint weight parameters; This represents resistivity observation data; This represents the forward modeling data corresponding to the current resistivity model.

[0039] For seismic wave data, the full waveform inversion method is used to solve for the parameter increments of the seismic wave velocity model. The objective function for the full waveform inversion of seismic waves can be expressed as: ; in, This represents the objective function for the full waveform inversion of seismic waves; Indicates the parameters of the seismic wave velocity model; Represents seismic wave observation data, This represents the forward modeling operator for seismic waves.

[0040] The gradient of the seismic wave velocity model parameters is obtained based on the objective function. Based on this, the increments of the seismic wave velocity model parameters are determined: ; in, This represents the gradient of the objective function for seismic wave full waveform inversion with respect to the parameters of the seismic wave velocity model; This indicates the update step size of the seismic wave velocity model; This represents a preconditioning operator or an approximate Hessian inverse operator, when no preconditioning is used. Unit operators are available.

[0041] For ground-penetrating radar (GPR) data, the full-waveform inversion method is used to solve for the parameter increments of the radar's relative permittivity model. The objective function for the GPR full-waveform inversion can be expressed as: ; in, This represents the target function for inverting the full waveform of a ground-penetrating radar. This represents the model parameters of the radar relative permittivity; This represents ground-penetrating radar observation data; This represents the forward modeling operator for ground-penetrating radar.

[0042] The gradient of the radar relative permittivity model parameters is obtained based on the objective function. Based on this, the parameter increments of the radar relative permittivity model are determined: ; in, This represents the gradient of the ground-penetrating radar full-waveform inversion objective function with respect to the radar relative permittivity model parameters; This indicates the update step size of the radar relative permittivity model; This indicates a preconditioning operator or an approximate Hessian inverse operator; when preconditioning is not used... Unit operators are available.

[0043] In one alternative implementation, to facilitate subsequent network training and validation, the sample data generated in steps S1 and S2 are divided into training set, validation set, and test set, with a preferred division ratio of [ratio to be inserted here]. The labels in the training samples can be the parameter difference between the real three-parameter geological model and the homogeneous background model, or the target gradient field or parameter increment of the target model corresponding to the real model.

[0044] S3. Increment the gradient information obtained in step S2 or the seismic wave velocity model parameters obtained from the gradient information. Increment of resistivity model parameters Increment of radar relative permittivity model parameters Input a multimodal fusion model, perform feature extraction, scale alignment, and fusion optimization on the model update information corresponding to different physical fields, and obtain the optimized multi-parameter model increment. Or the gradient after fusion.

[0045] In step S3, the gradient information obtained in step S2, or the model parameter increments derived from the gradient information, are input into the multimodal fusion model. In this embodiment, the seismic wave velocity model parameter increments are used. Increment of resistivity model parameters Increment of radar relative permittivity model parameters This serves as the input to the multimodal fusion model. The multimodal fusion model employs a three-branch parallel encoding and fusion decoding structure, including incremental input branches for seismic wave velocity model parameters, resistivity model parameters, and radar relative permittivity model parameters. By extracting features, aligning scales, and fusing features from the model update information of the three branches, a unified fusion feature is obtained. This feature is then decoded and output as an optimized multi-parameter model increment or fused gradient with the same spatial dimensions as the input.

[0046] As an optional implementation, the multimodal fusion model has three input channels, corresponding to the increments of seismic wave velocity model parameters, resistivity model parameters, and radar relative permittivity model parameters, respectively. The output of the multimodal fusion model is either an optimized multi-parameter model increment for model updates or a fused gradient tensor. The optimized multi-parameter model increment or the fused gradient is used to characterize the collaborative information of different detection methods in the model update direction and serves as the basis for subsequent model parameter updates.

[0047] The optimized multi-parameter model increment can be expressed as: ; in, , and These represent the increments of the seismic wave velocity model parameters, resistivity model parameters, and radar relative permittivity model parameters obtained through traditional inversion, respectively. , and These represent the parameter increments of the seismic wave velocity model, resistivity model, and radar relative permittivity model after optimization by the multimodal fusion model, respectively. This represents a multimodal fusion model; This represents the network parameters of the multimodal fusion model.

[0048] The multimodal fusion model can be understood as a multimodal gradient fusion network based on a U-Net structure. Optionally, an improved U-Net network incorporating multi-branch input, multi-scale feature extraction, spatial attention modules, and channel attention mechanisms can be used. This network takes the seismic wave velocity gradient, resistivity gradient, and radar relative permittivity gradient obtained through traditional inversion, or the corresponding model parameter increments, as input. The encoder extracts gradient features from different physical fields. The radar data branch includes a multi-scale feature extraction module that uses convolutional kernels with different receptive fields to extract deep detail information and shallow large-scale information. Subsequently, the network fuses features from different scales and physical fields through skip connections and feature concatenation operations in the U-Net structure, and adjusts the weights of key spatial regions and different physical property parameter channels using spatial attention modules and channel attention mechanisms. Finally, the decoder recovers the spatial scale step by step, outputting the optimized multi-parameter model increment or the fused gradient. Therefore, the multimodal fusion model... This can represent a multimodal gradient fusion network based on U-Net. θ This represents the trainable parameters in the network, including the weight and bias parameters in the encoder, decoder, skip connection fusion layer, spatial attention module, channel attention module, and output layer.

[0049] As a further defined implementation, the mean squared error loss can be used as the basic loss term during the initial training, and a weight decay term can be combined to constrain the network parameters. In this embodiment, to facilitate repeated implementation, the learning rate for the initial training can be set to... Batch size can be selected The number of training rounds can be selected. Rounds; if the validation set loss no longer decreases within a preset number of consecutive rounds, early termination may be triggered. The above parameters are merely examples and do not constitute a limitation on the scope of protection of this invention.

[0050] S4. Based on the fused gradient or the optimized multi-parameter model increment, update the uniform background model or the previous iteration model to obtain the updated model for the current round. Use the updated model as the initial model for the next round of traditional inversion and repeat the traditional inversion iteration to obtain the resistivity gradient, seismic wave velocity gradient and radar relative permittivity gradient corresponding to the next iteration.

[0051] In step 4, after the multimodal fusion model outputs the fused gradient or the optimized model parameter increment, the uniform background model or the previous iteration model is updated based on the fused gradient or the model parameter increment to obtain the updated model for the current iteration. The update process can be represented as follows: ; in, Indicates the first i The updated model obtained in the next iteration; This represents the previous iteration model or a uniform background model; This represents the increment of model parameters after processing by the multimodal fusion model.

[0052] As an alternative implementation, gradient descent can also be used to update the model parameters: ; in, Indicates the update step size; This represents the fused gradient used for model updates or the fused gradient after target region weighting.

[0053] Specifically, in the first iteration, the... For a uniform background model; in subsequent iterations, the... The updated model is obtained from the previous round of joint inversion. This updated model is used as the initial model for the next round of traditional inversion. Traditional inversion is then performed again on the resistivity data, seismic wave data, and ground-penetrating radar data to obtain the resistivity gradient, seismic wave velocity gradient, and radar relative permittivity gradient corresponding to the next iteration. This forms a cyclic iterative inversion mechanism that combines multi-mode gradient fusion with traditional inversion, such as... Figure 2As shown.

[0054] S5. In the joint inversion results of the first iteration or subsequent iterations, identify anomalous regions based on the distribution of model parameters or gradient response characteristics; determine the number of anomalous categories and category boundaries based on the parameter variation range, fluctuation characteristics of adjacent data differences and threshold factors in the current joint inversion results, and sub-mask the identified anomalous regions.

[0055] In step S5, in the joint inversion results of the initial iteration or subsequent iterations, anomalies are identified in the inversion region based on the model parameter distribution or gradient response characteristics, and the identified anomaly regions are then classified and labeled. Specifically, the minimum parameter value in the current joint inversion result is obtained. and maximum parameter value And calculate the range of parameter changes: ; in, This indicates the range of parameter variations in the current joint inversion results.

[0056] Based on the locations of significant parameter changes in the current joint inversion results, profile data is extracted along the tunnel detection direction, and the parameter differences between adjacent sampling points are calculated to form a difference fluctuation sequence. The mean of this difference fluctuation sequence is then used as the basis for further analysis. and standard deviation Construct the threshold factor: ; in, z Indicates the threshold factor; a and b This is an adjustable coefficient. Preferably, it is an empirical constant. a and b The values ​​are 0.2 and 0.3 respectively, and can also be adjusted according to different types of detection data, model scale, noise level and actual engineering needs.

[0057] The local peaks in the difference fluctuation sequence are used as candidate category boundaries, and the threshold factor is used as the basis for classification. z Candidate peaks are filtered; when a candidate peak is greater than the threshold factor... z When the peak value is less than or equal to the threshold factor, it is determined as a valid peak; when the peak value is less than or equal to the threshold factor, it is determined as a valid peak. z When this occurs, it is considered as a local small fluctuation, noise disturbance, or local inversion error and is discarded.

[0058] The number of anomaly categories is determined based on the number of valid peaks. n Specifically, let the number of effective peaks be... Since a valid peak corresponds to a significant parameter abrupt change, it can serve as a boundary between adjacent categories, thus reducing the number of anomaly categories. n It can be determined according to the following formula: ; in, Indicates the number of valid peaks; n This indicates the number of anomaly categories or the number of parameter partitions in the inversion result. When When the current inversion result shows no significant parameter abrupt change, the inversion area ahead can be treated as a homogeneous surrounding rock background or an area without obvious anomalies; when When a key parameter abruptly occurs, the inversion results are divided into two categories. When, it indicates the presence of two major parameter abrupt changes, the inversion results are divided into 3 categories; when The inversion results were then divided into four categories.

[0059] In a preferred embodiment, the number of anomaly categories n The maximum number is 4. This is because, within the limited detection range ahead of the tunnel face, the types of media exhibiting significant changes are usually not numerous; too many identified parameter abrupt changes may contain noise, minor fluctuations, or local inversion errors. Limiting the number of anomaly categories to no more than 4 allows for the expression of major anomaly differences while avoiding over-classification that could lead to instability in the mask calibration results.

[0060] Determining the number of anomaly categories n Afterwards, according to and The parameter range of the current joint inversion results is divided, and different parameter ranges are assigned hierarchical mask labels.

[0061] when n= At time 2, the hierarchical marking method is as follows: ; when n= At time 3, the hierarchical marking method is as follows: ; when n= At time 4, the hierarchical marking method is as follows: ; in, This indicates the hierarchical labeling results for the corresponding grid points; The reference mask marker representing the anomalous body region, , and This represents the hierarchical mask markers corresponding to different anomaly levels, with 0 indicating the background area or unmarked area.

[0062] As an alternative implementation, when an anomaly exhibits low-value anomalies, different hierarchical labels can be assigned sequentially from the low parameter value range to the high parameter value range; when an anomaly exhibits high-value anomalies, hierarchical labeling can be performed in the opposite direction; when an anomaly exhibits a deviation from the background anomaly, hierarchical labeling can also be performed based on the magnitude of the parameter deviation from the background value. Through the above methods, the number of anomaly categories can be adaptively determined based on the parameter change characteristics of the current joint inversion result, and hierarchical mask labeling can be performed on anomaly regions within different parameter ranges, such as... Figure 3 As shown.

[0063] S6. Based on the hierarchical mask marking results of the anomalous body region, construct a spatial mask matrix, mark the anomalous body region as the target region, and mark the background region as the non-target region; and count the number of grids in the anomalous body region according to the spatial mask matrix, construct an anomalous region proportion adjustment factor in combination with the total number of model grids, and further generate the target region weight matrix.

[0064] In step S6, as Figure 4 As shown, a spatial mask matrix is ​​constructed based on the results of anomaly region identification. M Specifically, the grid points identified as anomalous regions are masked, and the grid points corresponding to the background regions are marked as non-target regions, thus forming a spatial mask matrix with the same spatial size as the inversion model.

[0065] Let the total number of model meshes be N The number of grids marked as anomalous regions in the spatial mask matrix is Then, construct an adjustment factor for the proportion of abnormal regions: ; in, This represents the adjustment factor for the proportion of the anomalous region in the overall model space.

[0066] In one alternative implementation, each grid point in the spatial mask matrix is ​​denoted as... ,in i This indicates the grid point number. For the background area, For anomalous regions, This is a mask marker value greater than 0. If a binary mask is used, the anomaly region corresponds to... If a hierarchical mask is used, the anomaly region corresponds to .

[0067] Based on the aforementioned regulatory factor Spatial mask value and the enhancement coefficient of the abnormal region , construct the firsti Weight values ​​corresponding to each grid point: ; in, Indicates the first i The weight values ​​of each grid point; This represents the enhancement coefficient of the anomalous region, and This can be adjusted according to specific circumstances; preferably... A value of 2 or 4 can be used to make the network pay more attention to outlier regions during training, fine-tuning, or model updates.

[0068] When the i When each grid point is located in the background area Its weight is: ; The background region is given a lower weight to reduce its impact on network training or gradient updates, but the background region is not completely removed.

[0069] When the i Each grid point is located within the anomalous body region and is masked using a binary mask. When, its weight is: ; At this point, the anomalous region receives a higher weight than the background region, making it contribute more during network training or gradient update.

[0070] When using a hierarchical mask, the weight of the anomaly region varies. Increases with the increase of: ; Therefore, different anomaly levels correspond to different weights, achieving differentiated weighting for strong and weak anomaly regions.

[0071] S7. Introduce the target region weight matrix into the loss function of the multimodal fusion model, or use the target region weight matrix to weight and correct the gradient after fusion or the incremental parameters of the optimized model, so that the anomalous body region gets a higher weight and the background region retains a lower weight, thereby improving the contribution of the anomalous body region in the subsequent network training or model update process.

[0072] In step S7, the target region weight matrix obtained in step S6 is used to perform weighted correction on the loss function or the gradient after fusion of the multimodal fusion model.

[0073] In one alternative implementation, the weight matrix Introducing a weighted loss function into the loss function of the multimodal fusion model: ; in, L Represents the weighted loss function; This indicates element-wise multiplication; This represents the fused gradient or the increment of model parameters output by the multimodal fusion model. This represents the target gradient or the increment of the target model parameters.

[0074] In another alternative implementation, the weight matrix can also be used to weight and correct the fused gradient: ; in, This represents the fused gradient after weight correction. Through this method, the gradient corresponding to the anomalous region is enhanced, while the gradient of the background region is retained with lower weights, thereby achieving focused optimization based on the target region.

[0075] S8. Repeat the traditional inversion, gradient fusion, model update, anomaly identification, hierarchical mask calibration, spatial mask construction and target region weighted optimization process, so that the spatial mask matrix and target region weight matrix are dynamically updated with the inversion results until the preset maximum number of iterations is reached, or the change in the inversion results of two adjacent rounds is less than the preset threshold, or the loss function tends to stabilize.

[0076] In step S8, the inversion process is performed using a cyclic iterative method. Specifically, in the first... i In the first iteration, the... i- The updated model serves as the initial model for traditional inversion, obtaining resistivity gradients, seismic wave velocity gradients, and radar relative permittivity gradients, or calculating the corresponding model parameter increments from these gradients. These gradient information or model parameter increments are input into the multimodal fusion model to obtain the fused gradients or model parameter increments, which are then used to update the current model. Subsequently, anomaly identification is performed based on the joint inversion results obtained in the current iteration, determining the anomaly regions and their category ranges. A spatial mask and target area weight matrix are constructed based on the anomaly identification results. The target area weight matrix is ​​used for weighting the loss function of the next round of multimodal fusion model or for gradient correction after fusion, ensuring that the anomaly regions receive greater attention in subsequent iterations.

[0077] Repeat the traditional inversion, gradient fusion, model update, anomaly identification, threshold factor determination, hierarchical mask calibration, spatial mask construction, and target region weighted optimization process until the preset maximum number of iterations is reached, or the change in the inversion results of two adjacent rounds is less than the preset threshold, or the loss function tends to stabilize, and output the final inversion result.

[0078] As an alternative implementation method, when the rate of change of inversion error between two adjacent iterations is less than a preset threshold... Stop iteration when the number of iterations reaches its maximum value; alternatively, stop iteration when the number of iterations reaches its maximum value. The iteration terminates at the specified time. Preferably, Desirable Magnitude Desirable The Mask matrix, threshold factor, and weight distribution can be updated synchronously in each iteration to achieve progressively focused optimization of the abnormal regions.

[0079] S9. Output the final inversion result, and visualize and quantitatively evaluate the final inversion result.

[0080] In step S9, the final inversion results are output, visualized, and quantitatively evaluated. For simulated samples with real labels, evaluation can be performed using indicators such as root mean square error, mean absolute error, structural similarity, Dice coefficient of anomaly regions, cross-union ratio, boundary positioning deviation, and data mismatch reduction rate. Simultaneously, the consistency of the three-parameter inversion results, residual convergence curves, spatial mask calibration results, and target area weight distribution can be combined to comprehensively analyze the accuracy, stability, and anomaly identification performance of the joint inversion results. Through the above steps, this invention enables continuous iterative updating of multi-source gradient fusion results, adaptive determination of anomaly categories, hierarchical mask calibration, and target area-based focused optimization under complex geological conditions, thereby improving the accuracy and stability of multi-source joint inversion results.

[0081] This embodiment proposes a joint iterative inversion strategy for tunnel multi-source data based on a gradient fusion network. The resistivity gradient, seismic wave velocity gradient, and radar relative permittivity gradient obtained from traditional inversion are input into a multi-modal fusion model for feature extraction, alignment, and fusion. The fused gradient or corresponding incremental model parameters are then used to update the current model, which is then used as the initial model for the next round of traditional inversion. This allows multi-source gradient information to continuously participate in the model update process, avoiding the multi-source fusion result being merely a one-time output. This improves the stability of the joint inversion results under complex geological conditions of tunnels and enhances its adaptability to complex and unseen models.

[0082] This embodiment proposes a threshold factor setting method based on the fluctuation of adjacent data differences. It extracts profile data from the current joint inversion results, calculates the parameter differences between adjacent sampling points, and constructs a threshold factor by combining the mean, standard deviation, and adjustable coefficient of the difference sequence. Then, it determines the number of anomaly categories and category boundaries based on the number of effective peaks. Therefore, it can adaptively determine the anomaly region category based on the parameter variation characteristics of the inversion results themselves, avoiding the problem of relying on manual experience with fixed thresholds, reducing the impact of small fluctuations, noise, or local inversion errors on the anomaly identification results, and improving the accuracy and stability of anomaly region classification and labeling.

[0083] This embodiment proposes a focused optimization method based on target region mask marking and weight matrix. A spatial mask matrix is ​​constructed according to the hierarchical calibration results of the anomaly region, and a target region weight matrix is ​​generated by combining anomaly region proportion adjustment factors and anomaly region enhancement coefficients. This results in a higher weight for the anomaly region and a lower weight for the background region. Therefore, this avoids the loss of overall geological structure constraint information caused by directly setting the background region to zero, and enhances the attention paid to the anomaly target area during network training or model updates, improving the boundary characterization ability and inversion imaging accuracy of complex anomalies such as aquifers, fault fracture zones, and irregular water-conducting channels.

[0084] This embodiment continuously executes traditional inversion, gradient fusion, model updating, anomaly identification, threshold factor determination, hierarchical mask calibration, spatial mask construction, and target area weighted optimization during the inversion process, forming a dynamic closed-loop iterative inversion process. Through this closed-loop mechanism, the physical constraints of traditional inversion, the nonlinear feature expression capability of the multimodal fusion model, and the target area weighted optimization strategy can work synergistically to improve the overall stability, refinement, and engineering applicability of joint inversion of multi-source tunnel data.

[0085] Example 2: This embodiment provides a multimodal gradient iterative inversion system based on target mask marking, including: The forward modeling module is configured to acquire resistivity data, seismic wave data, and ground-penetrating radar data in front of the tunnel face through forward modeling. The hierarchical calibration module is configured to: determine the multi-parameter model increment and fused gradient based on resistivity data, seismic wave data and ground penetrating radar data through inversion and fusion; identify anomalies in the inverted area based on the multi-parameter model increment and fused gradient; and perform hierarchical calibration of the anomaly area through interval division and threshold factor assignment. The weighted correction module is configured to: construct a spatial mask based on the abnormal body region identification result, mark the abnormal region and the background region, and use the spatial mask to perform weighted correction on the fused gradient so that the weight of the abnormal body region is higher than the weight of the background region. The inversion module is configured to continuously optimize the inversion and fusion, anomaly identification and weighted correction until the preset convergence conditions are met, and to determine the final inversion result.

[0086] The working method of the system is the same as that of the multimodal gradient iterative inversion method based on target mask marking in Example 1, and will not be repeated here.

[0087] In another embodiment, the system includes: The 3D model size and property assignment module is configured to set the spatial range, number of grids, and grid spacing of the common area for joint inversion in front of the tunnel face, and to assign three-parameter properties under a unified 3D grid system. Based on the differences in physical properties of the surrounding rock background, fault fracture zone, water-bearing anomaly, and other adverse geological bodies, resistivity, seismic wave velocity, and radar relative permittivity are assigned to the model grid respectively, forming resistivity parameter volume, seismic wave velocity parameter volume, and radar relative permittivity parameter volume, providing a unified three-parameter physical property model for subsequent multi-source data forward modeling, model training, and inversion testing. The multi-source data acquisition and traditional inversion module is configured to perform forward modeling using the resistivity method, seismic wave method, and ground-penetrating radar method based on the resistivity parameter volume, seismic wave velocity parameter volume, and radar relative permittivity parameter volume, respectively, to obtain the corresponding resistivity data, seismic wave data, and ground-penetrating radar data; and using a uniform background model as the initial model, perform traditional inversion on the resistivity data, seismic wave data, and ground-penetrating radar data respectively to obtain gradient information reflecting the model update direction, and further obtain the seismic wave velocity model parameter increment, resistivity model parameter increment, and radar relative permittivity model parameter increment; The multimodal gradient fusion module is configured to input the gradient information obtained by the traditional inversion or the increments of seismic wave velocity model parameters, resistivity model parameters and radar relative permittivity model parameters obtained from the gradient information into the multimodal fusion model, and perform feature extraction, scale alignment and fusion decoding on the model update information corresponding to different physical fields to obtain the fused gradient or the optimized multi-parameter model increment. The model update and iterative inversion module is configured to update the uniform background model or the previous iteration model based on the fused gradient or the optimized multi-parameter model increment, to obtain the updated model of the current round, and use the updated model as the initial model for the next round of traditional inversion, and perform traditional inversion iteration again to obtain the gradient information and model parameter increment corresponding to the next round of iteration, thereby forming a cyclic iterative inversion mechanism that combines multimodal gradient fusion and traditional inversion. The anomaly identification and classification module is configured to identify anomaly regions in the joint inversion results of the first or subsequent iterations based on the distribution of model parameters or gradient response characteristics; and to screen effective peaks, determine the number of anomaly categories and category boundaries, and classify and mask the identified anomaly regions according to the parameter variation range, fluctuation characteristics of adjacent data differences and threshold factors in the current joint inversion results. The spatial mask construction module is configured to construct a spatial mask matrix based on the hierarchical mask labeling results of the anomalous body region, label the anomalous body region as the target region, and label the background region as the non-target region, so as to characterize the position, range and anomalous level of the anomalous body region in the inversion model space. The target region weight matrix construction module is configured to count the number of grids in the anomalous body region based on the spatial mask matrix, and construct an anomalous region proportion adjustment factor in combination with the total number of model grids; further, it constructs the target region weight matrix based on the anomalous region proportion adjustment factor, the spatial mask matrix, and the anomalous region enhancement coefficient, so that the anomalous body region receives a higher weight and the background region retains a lower weight. The target region weighted optimization module is configured to introduce the target region weight matrix into the loss function of the multimodal fusion model, or to use the target region weight matrix to perform weighted correction on the gradient after fusion or the incremental model parameters after optimization, so that the model update information corresponding to the anomalous body region is enhanced, and the model update information corresponding to the background region is retained with a lower weight, thereby improving the contribution of the anomalous body region in the subsequent network training or model update process, and realizing focused optimization based on the target region. The dynamic closed-loop iterative output module is configured to continuously execute traditional inversion, gradient fusion, model update, anomaly identification, threshold factor determination, hierarchical mask calibration, spatial mask construction, target area weight matrix construction, and target area weighted optimization process during the inversion process until the preset maximum number of iterations is reached, or the change in the inversion results of two adjacent rounds is less than the preset threshold, or the loss function tends to stabilize, and then outputs the final inversion result. The results output and evaluation module is configured to visualize, export, and quantitatively evaluate the final inversion results. The output content includes at least the inversion result volume data, spatial mask map, target area weight matrix map, threshold factor map, iteration log, and evaluation indicators.

[0088] In this embodiment, the modules communicate through a unified data interface. This unified data interface includes at least a model parameter interface, a forward data interface, a gradient information interface, a fused gradient interface, a model parameter increment interface, a threshold factor interface, a spatial mask interface, a target region weight matrix interface, an evaluation index interface, and a result output interface. As an optional implementation, each interface is associated using a unified task identifier (model_id) and an iteration identifier (iter_id) to ensure data traceability and backtracking between different modules.

[0089] In one alternative implementation, the system further includes a comprehensive database for uniformly storing model data, forward modeling data, traditional inversion gradient information, fused gradient data, model parameter increments, threshold factors, spatial mask matrices, target region weight matrices, training states, and evaluation results. As a further defined implementation, the comprehensive database employs a combined storage method of "relational metadata table + volume data file," wherein the metadata table records sample numbers, model parameter ranges, iteration cycles, training states, evaluation metrics, and result indexes, and the volume data file stores the three-dimensional parameter volume, forward modeling data volume, gradient tensor, model parameter increments, spatial mask matrix, target region weight matrix, and final inversion results. Preferably, the volume data file uses HDF5, NPZ, or an equivalent three-dimensional array storage format.

[0090] In one alternative implementation, the system can be deployed on a standalone workstation or a server cluster. For ease of implementation, recommended hardware configurations include: a central processing unit with at least 16 cores, at least 128GB of RAM, at least one graphics processing unit with at least 24GB of video memory, and at least 2TB of solid-state storage. Recommended software environments include the Ubuntu operating system, Python runtime environment, deep learning frameworks, numerical computing libraries, and a driver environment supporting GPU acceleration. The above hardware and software configurations are merely examples and do not limit the specific implementation of the invention.

[0091] In this embodiment, the system operation flow is as follows: the 3D model size and property assignment module generates the model mesh and the three-parameter property parameter volume and writes them into the comprehensive database; the multi-source data acquisition and traditional inversion module reads the three-parameter property parameter volume, generates multi-source forward modeling data, and outputs traditional inversion gradient information and corresponding model parameter increments using a uniform background model as the initial model; the multi-modal gradient fusion module reads the gradient information or model parameter increments and outputs the fused gradient or optimized multi-parameter model increments; the model update and iterative inversion module updates the model parameters based on the fused gradient or optimized multi-parameter model increments, and uses the updated model as the initial model for the next round of traditional inversion.

[0092] Subsequently, the anomaly identification and classification module identifies anomalies based on the current joint inversion results. It determines the number of anomaly categories and their boundaries through fluctuations in adjacent data differences, threshold factors, and the number of effective peaks, and generates a classification mask labeling result. The spatial mask construction module generates a spatial mask matrix based on the classification mask labeling result. The target region weight matrix construction module generates a target region weight matrix based on the spatial mask matrix, anomaly region proportion adjustment factor, and anomaly region enhancement coefficient. The target region weighted optimization module uses the target region weight matrix for weighting the loss function of the multimodal fusion model, or for weighted correction of the gradient and model parameter increments after fusion. The dynamic closed-loop iterative output module determines whether the convergence condition is met. If not, it continues to execute the next round of traditional inversion, gradient fusion, model update, and target region weighted optimization. If the condition is met, the final inversion result is sent to the result output and evaluation module.

[0093] As a further defined implementation, the results output and evaluation module can statistically analyze the inversion error, residual convergence curve, anomaly area localization accuracy, structural similarity, anomaly area Dice coefficient, crossover ratio, boundary localization deviation, and data mismatch reduction rate for each iteration, and generate a visual report for use in model training monitoring, inversion effect analysis, and result comparison and verification.

[0094] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the multimodal gradient iterative inversion method based on target mask marking described in Embodiment 1.

[0095] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the multimodal gradient iterative inversion method based on target mask marking described in Embodiment 1.

[0096] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the multimodal gradient iterative inversion method based on target mask marking described in Embodiment 1.

[0097] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A multimodal gradient iterative inversion method based on target mask marking, characterized in that, include: Resistivity data, seismic wave data, and ground-penetrating radar data in front of the tunnel face are obtained through forward modeling. Based on resistivity data, seismic wave data, and ground-penetrating radar data, the multi-parameter model increment and fused gradient are determined through inversion and fusion. Anomalies are identified in the inversion region based on the multi-parameter model increment and the fused gradient, and the anomaly region is classified and labeled by interval division and threshold factor assignment. A spatial mask is constructed based on the results of anomaly region identification. The anomaly region and the background region are marked. The spatial mask is then used to perform weighted correction on the fused gradient, so that the weight of the anomaly region is higher than that of the background region. The inversion and fusion, anomaly identification and weighted correction are continuously optimized until the preset convergence conditions are met, and the final inversion result is determined.

2. The multimodal gradient iterative inversion method based on target mask marking as described in claim 1, characterized in that, The forward modeling simulation includes: assigning resistivity, seismic wave velocity, and radar relative permittivity to the model mesh respectively, forming resistivity parameter volume, seismic wave velocity parameter volume, and radar relative permittivity parameter volume, and then performing forward modeling simulations using the resistivity method, seismic wave method, and ground penetrating radar method respectively to obtain the corresponding resistivity data, seismic wave data, and ground penetrating radar data.

3. The multimodal gradient iterative inversion method based on target mask marking as described in claim 2, characterized in that, The determination of multi-parameter model increments and fused gradients through inversion and fusion includes: for resistivity data, determining resistivity model parameter increments based on smooth constraint least squares inversion; for seismic wave data, determining the gradient and increment of seismic wave velocity model parameters based on the seismic wave full waveform inversion objective function; and for ground penetrating radar data, determining the gradient and increment of radar relative permittivity model parameters based on the ground penetrating radar full waveform inversion objective function.

4. The multimodal gradient iterative inversion method based on target mask marking as described in claim 3, characterized in that, The parameter increments for the resistivity model are: ； in, This represents the increment of resistivity model parameters; This represents the sensitivity matrix corresponding to the resistivity method. Represents the smoothness constraint matrix; Indicates the smoothing constraint weight parameters; This represents resistivity observation data; This represents the forward modeling data corresponding to the current resistivity model; The increments of the seismic wave velocity model parameters are: ； ； in, This represents the objective function for the full waveform inversion of seismic waves; Indicates the parameters of the seismic wave velocity model; Represents seismic wave observation data, Represents the forward modeling operator for seismic waves; This represents the gradient of the objective function for seismic wave full waveform inversion with respect to the parameters of the seismic wave velocity model; This indicates the update step size of the seismic wave velocity model; Represents a preconditioning operator; The parameter increments for the radar relative permittivity model are: ； ； in, This represents the target function for inverting the full waveform of a ground-penetrating radar. This represents the model parameters of the radar relative permittivity; This represents ground-penetrating radar observation data; This represents the forward modeling operator for ground-penetrating radar; This represents the gradient of the ground-penetrating radar full-waveform inversion objective function with respect to the radar relative permittivity model parameters; This indicates the update step size of the radar relative permittivity model; This represents the preconditioning operator.

5. The multimodal gradient iterative inversion method based on target mask marking as described in claim 3, characterized in that, The model is updated based on the fused gradient and the multi-parameter model increment to obtain the updated model for the current round. The updated model is then used as the initial model for the next round of inversion. The traditional inversion iteration is repeated to obtain the resistivity gradient, seismic wave velocity gradient, and radar relative permittivity gradient corresponding to the next round of iteration.

6. The multimodal gradient iterative inversion method based on target mask marking as described in claim 1, characterized in that, The step of classifying and labeling anomaly regions through interval division and threshold factor assignment includes: obtaining the minimum and maximum parameter values ​​in the current joint inversion results, and determining the parameter variation range. Based on the location of parameter changes in the current joint inversion results, profile data is extracted along the tunnel detection direction, and the parameter differences between adjacent sampling points are determined to form a difference fluctuation sequence; based on the mean and standard deviation of the difference fluctuation sequence; Local peaks in the difference fluctuation sequence are used as candidate category boundaries, and based on the threshold factor... z Candidate peaks are filtered; when a candidate peak is greater than a threshold factor... z If the peak value is within a certain range, it is considered a valid peak; otherwise, it is considered a minor local fluctuation and is discarded. The number of anomaly categories is determined based on the number of valid peaks; Determining the number of anomaly categories n Then, the parameter intervals of the current joint inversion results are divided according to the minimum parameter value and the parameter variation range, and different parameter intervals are assigned hierarchical labels.

7. The multimodal gradient iterative inversion method based on target mask marking as described in claim 6, characterized in that, The step of using the spatial mask to perform weighted correction on the fused gradient includes: constructing a spatial mask matrix based on the hierarchical labeling results of the anomalous body region, labeling the anomalous body region as the target region, and labeling the background region as the non-target region; and counting the number of grids in the anomalous body region according to the spatial mask matrix, constructing an anomalous region proportion adjustment factor in combination with the total number of model grids, and generating a target region weight matrix.

8. The multimodal gradient iterative inversion method based on target mask marking as described in claim 1, characterized in that, Abnormal body region The weight varies with each grid point in the spatial mask matrix. Increases with the increase of: ； ； in, N is Define the total number of model grids; This represents the number of grid cells in the spatial mask matrix that are marked as anomalous regions. This is an adjustment factor for the proportion of the anomalous region in the overall model space; This represents the enhancement coefficient of the abnormal region.

9. The multimodal gradient iterative inversion method based on target mask marking as described in claim 8, characterized in that, weight matrix Introducing a weighted loss function into the loss function of the multimodal fusion model: ； in, L Represents the weighted loss function; This indicates element-wise multiplication; This represents the fused gradient or the increment of model parameters output by the multimodal fusion model. This represents the target gradient or the increment of the target model parameters.

10. A multimodal gradient iterative inversion system based on target mask marking, characterized in that, include: The forward modeling module is configured to acquire resistivity data, seismic wave data, and ground-penetrating radar data in front of the tunnel face through forward modeling. The hierarchical calibration module is configured to determine the multi-parameter model increment and fused gradient by inversion and fusion based on resistivity data, seismic wave data, and ground penetrating radar data. Anomalies are identified in the inversion region based on the multi-parameter model increment and the fused gradient, and the anomaly region is classified and labeled by interval division and threshold factor assignment. The weighted correction module is configured to: construct a spatial mask based on the abnormal body region identification result, mark the abnormal region and the background region, and use the spatial mask to perform weighted correction on the fused gradient so that the weight of the abnormal body region is higher than the weight of the background region. The inversion module is configured to continuously optimize the inversion and fusion, anomaly identification and weighted correction until the preset convergence conditions are met, and to determine the final inversion result.

Images