Transient electromagnetic inversion method and computing equipment integrating multi-field response and intelligent optimization
By integrating multi-field response and intelligent optimization transient electromagnetic inversion methods, the problem of high-precision imaging of deep and complex geological bodies has been solved, achieving efficient three-dimensional resistivity imaging and error compensation, and improving inversion accuracy and reliability.
Patent Information
- Application Number
- CN202611137642.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-11-17
AI Technical Summary
Existing transient electromagnetic inversion techniques are difficult to achieve high-precision three-dimensional resistivity imaging of deep and complex geological bodies, mainly because multi-field inversion does not establish coupling relationships, inversion waveform correction is coarse, modeling efficiency is low, optimization robustness is poor, and error compensation is lacking.
A transient electromagnetic inversion method integrating multi-field response and intelligent optimization includes acquiring three-field response data, constructing an unstructured tetrahedral mesh, correcting the emitted waveform, constructing an adaptive sensitivity weighting matrix, and using a hybrid intelligent optimization algorithm and a deep residual network for error compensation.
It enables high-precision three-dimensional resistivity imaging of deep and complex geological bodies, improves inversion accuracy and reliability, makes up for the blind spots in deep exploration, reduces dependence on the initial model, and accurately compensates for system errors.
Smart Images

Figure CN123113013A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological exploration, and more specifically to a transient electromagnetic inversion method and computing device that integrates multi-field response and intelligent optimization. Background Technology
[0002] The Transient Electromagnetic Method (TEM) is a time-domain geophysical exploration method based on the principle of electromagnetic induction. Typically, TEM can probe depths exceeding 1000m, identify small-scale anomalies, and can operate without grounding electrodes. Due to its advantages such as high resolution, flexible construction, non-contact detection, and strong anti-interference capabilities, TEM is widely used in mineral exploration, groundwater and geothermal resource exploration, coal mine water hazard prevention, landslide and karst collapse geological hazard investigation, and urban underground space exploration, becoming one of the core technologies for deep geological exploration.
[0003] However, current transient electromagnetic inversion methods are typically limited to single-field inversions, and even multi-field inversions simply involve superimposing data from multiple fields. Furthermore, existing transient electromagnetic methods often focus on single-dimensional algorithmic improvements such as mesh refinement and gradient optimization, which are insufficient to meet the high-precision exploration engineering requirements of deep and complex geological conditions. Current transient electromagnetic inversion methods suffer from coarse waveform correction, low modeling efficiency, poor optimization robustness, and a lack of error compensation, making it difficult to achieve high-precision three-dimensional resistivity imaging of deep and complex geological bodies. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, this invention aims to provide a transient electromagnetic inversion method and computing device that integrates multi-field response and intelligent optimization, in order to solve the problem of achieving high-precision three-dimensional resistivity imaging of deep and complex geological bodies.
[0005] To achieve the above objectives, the technical solution adopted by this invention is as follows: Firstly, this application provides a transient electromagnetic inversion method integrating multi-field response and intelligent optimization, comprising: Acquire standardized preprocessed three-field response data of the target underground space, including electric field response data, magnetic induction time derivative response data, and magnetic induction response data; The target underground space is subjected to terrain-constrained adaptive meshing to construct an unstructured tetrahedral mesh for the target underground space; Based on the piecewise linear interpolation algorithm, the convolution integral of the step response and the convolution integral of the waveform derivative are calculated to correct the non-ideal transmission waveform and obtain the corrected transmission waveform. The convolution integral of the step response and the convolution integral of the waveform derivative are obtained by decomposing the electromagnetic field response data, which includes electric field response data and magnetic induction response data. The corrected transmitted waveform is used as the input for time-domain discrete recursion to calculate the three-field response data and obtain the forward response vector. Based on the stratum depth and forward response vector of the target underground space, an adaptive sensitivity weighting matrix for the three response data is constructed. The adaptive sensitivity weighting matrix includes the weight coefficients of the three responses, which are dynamically adjusted according to the stratum depth of the target underground space. Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed, and the optimal resistivity model is obtained through a hybrid intelligent optimization algorithm. The hybrid intelligent optimization algorithm is a hybrid optimization strategy that combines the adaptive particle swarm optimization algorithm with the L-BFGS algorithm. The forward modeling response vector and terrain parameters of the unstructured tetrahedral mesh are input into a preset depth residual network. Error compensation and correction are completed through the output of the preset depth residual network to obtain the inversion result. The preset depth residual network is a convolutional structure based on a preset number of skip connections. It is trained with three-field response forward modeling data, measured observation data, terrain parameters and exploration instrument calibration parameters as inputs and the system error compensation amount as output.
[0006] In the embodiments of this application, terrain-constrained adaptive meshing is performed on the target underground space to construct an unstructured tetrahedral mesh for the target underground space, including: Based on the digital elevation model of the work area, a surface model is constructed. Generate a triangular surface grid based on the surface model; Based on the geological data, borehole data and electromagnetic field sensitivity pre-calculation results of the target underground space, the definition of the grid densification area is obtained. The grid densification area includes the topographic relief area, the anomalous geological body area and the high-sensitivity detection area. The Delaunay tetrahedral partitioning algorithm is adopted, combined with triangular surface mesh and mesh densification region definition, to perform tetrahedral mesh partitioning of the target underground space and construct an unstructured tetrahedral mesh of the target underground space.
[0007] In the embodiments of this application, based on a piecewise linear interpolation algorithm, the convolution integral of the step response and the convolution integral of the waveform derivative are calculated to correct the non-ideal transmission waveform, resulting in a corrected transmission waveform, including: Under the quasi-static electromagnetic assumption, transient electromagnetic detection control equations are constructed, which include electric field diffusion control equations. By introducing the unit step electric field response and the unit step magnetic induction response, the electromagnetic field response data corresponding to the non-ideal emission waveform is decomposed into the convolution integral of the step response and the convolution integral of the waveform derivative. The convolution integral of the step response and the convolution integral of the waveform derivative are calculated using a piecewise linear interpolation algorithm to correct the non-ideal transmission waveform and obtain the corrected transmission waveform.
[0008] In the embodiments of this application, the transient electromagnetic detection control equations also include Faraday's law of electromagnetic induction equations; Using the corrected transmitted waveform as input for time-domain discrete recursion, the three-field response data are calculated to obtain the forward response vector, including: The second-order backward Euler method is used to discretize the electric field diffusion control equation in time, dividing the time domain into discrete time steps, and obtaining the recursive calculation formulas for the electric field and magnetic induction. Based on the discrete electric field diffusion control equation of the edge finite element method, and combined with the recursive calculation formula of electric field and magnetic induction, the electric field response data is calculated to obtain the spatiotemporal distribution of the electric field response. Based on Faraday's law of electromagnetic induction, the time derivative response data of magnetic induction is calculated, and the spatiotemporal distribution of the time derivative response of magnetic induction is obtained. Based on the recursive calculation formulas for electric field and magnetic induction and the electric field response at the actual measurement point, the magnetic induction response data is calculated to obtain the spatiotemporal distribution of the magnetic induction response. By integrating the spatiotemporal distributions of the electric field response, the magnetic induction time derivative response, and the magnetic induction response, the forward response vector is obtained.
[0009] In the embodiments of this application, an adaptive sensitivity weighting matrix for the three-field response data is constructed based on the stratum depth of the target underground space and the forward response vector, including: When the depth of the target underground space is greater than the preset depth, the sensitivity Jacobian matrix of the three-field response to the resistivity of the underground medium is constructed based on the resistivity vector of the underground medium and the forward response vector. The sensitivity Jacobian matrix is calculated using the adjoint method to obtain the sensitivity gradient; Based on the sensitivity gradient, the spatial sensitivity distribution of the three-field response is calculated; Based on the stratum depth and spatial sensitivity distribution of the target underground space, an adaptive sensitivity weighting matrix for the three-field response data is constructed using a diagonal matrix generating function.
[0010] In the embodiments of this application, a multi-field joint inversion objective function is constructed based on the adaptive sensitivity weighting matrix, and the optimal resistivity model is obtained through a hybrid intelligent optimization algorithm, including: 1) Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed that takes into account both data fitting accuracy and model balance constraints. The multi-field joint inversion objective function includes a data fitting term and a model regularization term. The data fitting term adopts the weighted least squares criterion and introduces the adaptive sensitivity weighting matrix and the data covariance matrix. The model regularization term adopts the model roughness minimization criterion. 2) Construct an initial population consisting of multiple particles, where each particle corresponds to a resistivity model; 3) An adaptive particle swarm optimization algorithm is used to perform a global search, update the velocity and inertia weights of particles in the initial swarm, and obtain the globally optimal model; 4) Use the globally optimal model as the initial model input for the L-BFGS algorithm; 5) Calculate the gradient information of the multi-field joint inversion objective function with respect to the model parameters, and use the Wolfe criterion to perform a one-dimensional line search to determine the optimal iteration step size; 6) Based on the gradient information and the optimal iteration step size, iterate the parameters of the globally optimal model to update the globally optimal model; 7) If the multi-field joint inversion objective function fails to converge or reaches the maximum number of iterations, use the updated global optimal model as the initial model input for the L-BFGS algorithm, and repeat steps 4) and 5) until the multi-field joint inversion objective function converges or reaches the maximum number of iterations to obtain the optimal resistivity model. In this case, the weight coefficients of the three field responses are dynamically adjusted based on the data residuals of the three field responses for a preset number of iterations.
[0011] In the embodiments of this application, the preset depth residual network is obtained through the following steps: Construct an initial deep residual network, which includes an input layer, a residual feature extraction layer and an output layer connected in sequence. The residual feature extraction layer has a preset number of skip-connected convolutional structures. The loss function of the deep residual network adopts the mean square error criterion, and the initial deep residual network is trained using the Adam optimizer to obtain the preset deep residual network; The forward modeling response vector and terrain parameters of the unstructured tetrahedral mesh are input into a preset depth residual network. Error compensation and correction are performed through the output of the preset depth residual network to obtain the inversion results, including: Obtain the initial inversion model response corresponding to the forward response vector; The initial inversion model response, topographic parameters of the unstructured tetrahedral mesh, measured observation data of the target underground space, and exploration instrument parameters are input into a preset residual network. Error compensation and correction are completed through the output of the preset residual network to obtain the inversion result.
[0012] In embodiments of this application, the method further includes: Using the initial inversion results as the data source, a three-dimensional subsurface resistivity model is output. The three-dimensional subsurface resistivity model includes the grid node parameters of the unstructured tetrahedral mesh, the parameters of the anomalous geological body, the imaging quality parameters, and the visualization data. The visualization data includes three-dimensional resistivity volume data, horizontal slices of the target subsurface space, and vertical profiles of the target subsurface space. The three-dimensional resistivity model of the underground was verified using three-field response simulation data, and the accuracy verification results were obtained. The underground three-dimensional resistivity model was verified by using measured data from the target underground space, and the results of the engineering practicality verification were obtained.
[0013] In the embodiments of this application, obtaining standardized preprocessed three-field response data of the target underground space includes: The initial three-field response data of the target underground space were acquired using a synchronous acquisition mode. The initial three-field response data are standardized and preprocessed to obtain the preprocessed three-field response data.
[0014] Secondly, this application provides a computing device, comprising: The memory is configured to store instructions; The processor is configured to retrieve instructions from memory and, when executing instructions, to implement the aforementioned transient electromagnetic inversion method that integrates multi-field response and intelligent optimization.
[0015] Therefore, this application provides a transient electromagnetic inversion method integrating multi-field response and intelligent optimization, comprising: acquiring standardized preprocessed three-field response data of a target underground space; performing terrain-constrained adaptive mesh partitioning on the target underground space to construct an unstructured tetrahedral mesh of the target underground space; calculating the convolution integral of the step response and the convolution integral of the waveform derivative based on a piecewise linear interpolation algorithm to correct the non-ideal emission waveform and obtain the corrected emission waveform; using the corrected emission waveform as the input of time-domain discrete recursion to calculate the three-field response data and obtain the forward response vector; constructing an adaptive sensitivity weighting matrix of the three-field response data according to the stratum depth of the target underground space and the forward response vector; constructing a multi-field joint inversion objective function according to the adaptive sensitivity weighting matrix, and obtaining the optimal resistivity model through a hybrid intelligent optimization algorithm; inputting the forward response vector and the terrain parameters of the unstructured tetrahedral mesh into a preset depth residual network, and completing error compensation correction through the output of the preset depth residual network to obtain the inversion result.
[0016] Compared with the prior art, the beneficial effects of the present invention are: This application integrates the three-field response data—electric field response, magnetic induction time derivative response, and magnetic induction response—for inversion, fully leveraging the advantages of slow late-stage decay, high deep sensitivity, and strong noise resistance of the magnetic induction response. This enables high-precision three-dimensional resistivity imaging of complex deep geological bodies, significantly improving the accuracy of deep strata imaging and compensating for the blind spots in traditional deep exploration methods.
[0017] Furthermore, the hybrid PSO and L-BFGS algorithms avoid local optima and reduce dependence on the initial model. An adaptive weighting mechanism for the three-field response data is introduced, dynamically adjusting the weight coefficients of the three-field responses to ensure that high-quality data dominates the inversion process, significantly improving inversion accuracy and further enhancing the precision of 3D resistivity imaging. A deep residual network accurately compensates for systematic errors such as environmental noise, instrument errors, and terrain distortion, improving the reliability of 3D resistivity imaging while maintaining its accuracy. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to an embodiment of the present invention; Figure 2 This is the inversion result of a conventional two-dimensional layered inversion algorithm according to an embodiment of the present invention; Figure 3 This is the inversion result of a transient electromagnetic inversion method that integrates multi-field response and intelligent optimization according to an embodiment of the present invention. Detailed Implementation
[0019] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the present invention.
[0020] The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0021] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of the invention, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.
[0022] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.
[0023] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.
[0024] The core principle of the transient electromagnetic method is as follows: a primary pulse electromagnetic field is emitted into the ground through a transmitting coil. When a conductive medium exists underground, the medium generates induced eddy currents under the excitation of the primary field. These eddy currents decay over time and generate a secondary electromagnetic field. By collecting the response signals of the secondary electromagnetic field, such as the electric field, magnetic induction, and the time derivative of magnetic induction, the decay law and spatial distribution characteristics of the response signals are analyzed. The resistivity distribution of the underground medium is then inferred, thereby identifying the spatial location, shape, scale, and electrical characteristics of the underground geological body. Currently, detection based on the transient electromagnetic method mainly focuses on the measurement and inversion of the electric field response and the time derivative response of magnetic induction. However, due to limitations in measurement principles, hardware equipment, and algorithm design, it is difficult to meet the high-precision exploration requirements under deep and complex geological conditions.
[0025] The magnetic induction time derivative response suffers from limitations such as rapid late-stage signal attenuation, weak noise immunity, and narrow spatial sensitivity coverage. The magnetic induction time derivative signal is proportional to the rate of change of the magnetic field. As time progresses, the secondary magnetic field tends to stabilize, leading to rapid attenuation of the magnetic induction time derivative signal. In strata deeper than 500m, the late-stage signal amplitude is extremely low and easily submerged by environmental noise, resulting in rapid attenuation and severely insufficient detection capability for deep geological bodies. Industrial electromagnetic interference, magnetotelluric noise, and instrument electronic noise significantly affect the magnetic induction time derivative signal, especially in urban areas, mining areas, and near high-voltage lines, where the signal-to-noise ratio is extremely low and effective signals are difficult to extract, resulting in weak noise immunity. The magnetic induction time derivative signal is primarily sensitive to low-resistivity anomalies in shallow to middle strata, with limited ability to identify high-resistivity and weak anomalies in deep strata, resulting in a large detection blind zone and narrow spatial sensitivity coverage.
[0026] The inversion of electric field response faces limitations such as harsh construction conditions, dominance of shallow signals, and electrode polarization interference. Electric field measurements require the deployment of grounding electrodes on the surface to collect electric field signals from the subsurface medium. However, in complex terrains with significant undulations and exposed rock, arid deserts with poor conductivity, waters where underwater electrode deployment is difficult, and permafrost with high insulation, the electrode grounding resistance is extremely high, sometimes even impossible, leading to significant construction difficulties and low data acquisition quality, resulting in harsh construction conditions. The electric field signal is mainly concentrated in shallow strata, significantly affected by surface topography and shallow electrical inhomogeneities. The electric field signal in deep strata is weak, with shallow signals dominating, making it difficult to reflect the characteristics of deep geological bodies. When the grounding electrode comes into contact with the subsurface medium, it generates electrode polarization potential. Interference signals superimpose with the effective electric field signal, causing measurement data distortion, requiring additional polarization correction and increasing data processing complexity.
[0027] With the rapid development of low-noise, high-precision magnetometers such as superconducting quantum interference devices (SQUIDs), fluxgate sensors, and magnetoresistive sensors, the inversion testing of magnetic induction response is gradually being applied. Stable signals can still be detected at depths below 500m, exhibiting sensitivity across the entire depth spectrum from shallow to deep and insensitivity to electromagnetic interference. It possesses advantages such as slow late-stage decay, wide spatial sensitivity coverage, strong noise immunity, and significant potential for deep-level detection.
[0028] Currently, there is a lack of an integrated three-dimensional inversion framework that integrates electric field, magnetic induction time derivative, and magnetic induction response. Most inversions are limited to single-field inversions. Multi-field inversions simply overlay multiple field data without establishing coupling relationships and complementary mechanisms among the three fields. This fails to leverage the deep-penetration advantages of magnetic induction response inversion, resulting in inversion results with acceptable accuracy in shallow areas but blurred deep-penetration imaging, making it difficult to identify hidden anomalies.
[0029] The transmitted waveform generated by the transient electromagnetic transmitting coil is a non-ideal waveform, exhibiting characteristics such as rising edge, falling edge, and off-time, which significantly interferes with the three-field response, especially the late-stage signal, directly leading to distortion of the inversion results. Traditional correction methods use a denser time step for approximate correction, which can improve correction accuracy to some extent. However, the density of the time step increases exponentially with the computational load. The inversion period for large-scale 3D models can be as long as several weeks, resulting in low computational efficiency, and the correction accuracy still falls short of the requirements for high-precision exploration. The correction error exceeds 35.3%, the waveform effect correction is coarse, and there is a contradiction between computational efficiency and correction accuracy.
[0030] Existing 3D forward modeling algorithms mostly use uniformly structured grids (rectangular or cubic) to partition underground space. However, these structured grids struggle to accurately reflect the undulating characteristics of real terrain, requiring terrain simplification. This results in poor adaptability to complex terrain, introducing terrain distortion errors into the forward model and leading to false anomalies in the inversion results. Furthermore, structured grids require the same grid size across the entire computational domain. To ensure the accuracy of anomaly boundary imaging, the entire grid needs to be densified, resulting in severe redundancy in the background uniform stratigraphic grid. Large-scale geological models typically exceed 100,000 grids. This high grid redundancy and computational complexity, with forward modeling taking several days, makes it difficult to support the iterative requirements of 3D inversion.
[0031] Transient electromagnetic inversion is highly nonlinear and ill-posed, with the objective function exhibiting multiple local minima. Gradient algorithms rely solely on the gradient information of the objective function for iterative searching, easily capturing local optima near the initial model and failing to converge to the global optimum, leading to distorted inversion results and a high risk of getting trapped in local optima. The convergence of gradient algorithms is closely related to the initial model; when the initial model deviates from the true geoelectric model, the inversion is prone to divergence or getting trapped in poor local optima, demonstrating a high dependence on the initial model. In actual exploration, it is difficult to accurately construct initial models for deep strata, resulting in poor robustness of inversion optimization and severely impacting inversion stability. In existing multi-field inversions, the weights of the three data fields are often manually fixed, failing to dynamically adjust the weights based on stratum depth, signal quality, and residual distribution, leading to low-quality data dominating the inversion process and decreased inversion accuracy.
[0032] During the acquisition and processing of transient electromagnetic data, various systematic errors exist, including environmental noise, instrument measurement errors, topographic distortion, emission current fluctuations, and electrode polarization. Environmental noise includes industrial interference noise and magnetotelluric noise, while instrument measurement errors include sensor drift and calibration errors, which will not be elaborated upon here. Current inversion methods lack a systematic error compensation mechanism, relying solely on simple denoising to suppress random noise, which fails to compensate for systematic errors. This leads to inversion results that are highly susceptible to false anomalies, blurred deep strata imaging, unclear anomaly boundaries, and large deviations in resistivity values, severely impacting the reliability of geological interpretation.
[0033] Example Please see Figure 1 , Figure 1 This is a flowchart of a transient electromagnetic inversion method that integrates multi-field response and intelligent optimization according to an embodiment of the present invention. Figure 1 The transient electromagnetic inversion method that integrates multi-field response and intelligent optimization includes: S110: Obtain standardized preprocessed three-field response data of the target underground space, including electric field response data, magnetic induction time derivative response data, and magnetic induction response data.
[0034] Data acquisition and preprocessing are fundamental to transient electromagnetic inversion. This involves acquiring the three-field response data of the target underground space. In this embodiment, the three-field responses include the electric field response, the time derivative response of the magnetic induction, and the magnetic induction response. The target underground space is set according to actual needs and can be any underground space requiring inversion measurement; therefore, no limitation is made. The three-field response data are preprocessed to remove interference and correct errors, obtaining standardized preprocessed three-field response data of the target underground space, providing reliable data support for subsequent forward and inversion models.
[0035] S120 performs terrain-constrained adaptive meshing on the target underground space to construct an unstructured tetrahedral mesh for the target underground space.
[0036] Terrain-constrained adaptive mesh generation is a core step in forward modeling. Terrain-constrained adaptive mesh generation is performed on the target underground space to construct an unstructured tetrahedral mesh that fits the real terrain, adapts to the boundary of anomalies, and has low mesh redundancy, so as to balance the accuracy and efficiency of forward modeling calculations.
[0037] S130, based on the piecewise linear interpolation algorithm, calculates the convolution integral of the step response and the convolution integral of the waveform derivative to correct the non-ideal transmission waveform and obtain the corrected transmission waveform. The convolution integral of the step response and the convolution integral of the waveform derivative are obtained by decomposing the electromagnetic field response data, which includes electric field response data and magnetic induction response data.
[0038] To address the interference of non-ideal transmission waveforms on the three-field response, this application proposes precise correction of convolution and interpolation waveform effects. The electric field response data is decomposed to obtain the convolution integral of the step response and the convolution integral of the waveform derivative. Based on a piecewise linear interpolation algorithm, the convolution integrals of the step response and the waveform derivative are calculated, removing the constraint between waveform complexity and time step, thus correcting the non-ideal transmission waveform and obtaining the corrected transmission waveform. This provides high-precision forward modeling data support for inversion.
[0039] S140 uses the corrected transmitted waveform as the input for time-domain discrete recursion to calculate the three-field response data and obtain the forward response vector.
[0040] The corrected transmitted waveform is used as the input for time-domain discrete recursion to calculate the three-field response data, efficiently and accurately obtaining the spatial and temporal distributions of the three-field responses. Integrating the spatial and temporal distributions of the electric field, magnetic induction time derivative, and magnetic induction responses, a forward response vector is obtained. This forward response vector provides standardized forward data input for subsequent sensitivity analysis and inversion calculations.
[0041] S150. Based on the stratum depth of the target underground space and the forward response vector, an adaptive sensitivity weighting matrix for the three response data is constructed. The adaptive sensitivity weighting matrix includes the weight coefficients of the three responses, which are dynamically adjusted according to the stratum depth of the target underground space.
[0042] Based on the stratigraphic depth and forward-modeled response vector of the target subsurface space, a multi-scale sensitivity joint analysis is performed to clarify the sensitivity, spatial distribution characteristics, and depth adaptability of the three-field responses to subsurface resistivity, and an adaptive sensitivity weighting matrix for the three-field response data is constructed. The adaptive sensitivity weighting matrix includes weighting coefficients for the three-field responses: the electric field response, the magnetic induction time derivative response, and the magnetic induction response. These weighting coefficients are dynamically adjusted according to the stratigraphic depth of the target subsurface space. The adaptive sensitivity weighting matrix guides the weight allocation of the three-field data and optimizes field data acquisition.
[0043] S160. Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed, and the optimal resistivity model is obtained through a hybrid intelligent optimization algorithm. The hybrid intelligent optimization algorithm is a hybrid optimization strategy that combines the adaptive particle swarm optimization algorithm with the L-BFGS algorithm.
[0044] Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed that balances data fitting accuracy and model balance constraints. This ensures good data fit while suppressing overfitting and improving model stability. In this embodiment, the hybrid intelligent optimization algorithm is a hybrid optimization strategy combining adaptive particle swarm optimization and the L-BFGS (Limited-memory Broyden-Fletcher-Goldfarb-Shanno) algorithm. It consists of two stages: global search and local refinement, balancing global search capability with local convergence accuracy. Through this hybrid intelligent optimization algorithm, the optimal resistivity model is obtained, avoiding local optima and improving inversion accuracy and stability.
[0045] The hybrid PSO algorithm and L-BFGS algorithm avoid local optima and reduce dependence on the initial model, improving the stability of inversion iteration by 41% and the global convergence rate by 35%. The introduction of a three-field data adaptive weighting mechanism dynamically adjusts the weights to ensure that high-quality data dominates the inversion process, significantly improving the inversion accuracy.
[0046] S170: Input the forward modeling response vector and terrain parameters of the unstructured tetrahedral mesh into the preset depth residual network. The error compensation correction is completed through the output of the preset depth residual network to obtain the inversion result. The preset depth residual network is a convolutional structure based on a preset number of skip connections. It is trained with three-field response forward modeling data, measured observation data, terrain parameters and exploration instrument calibration parameters as inputs and the system error compensation amount as output.
[0047] A deep residual network with a pre-defined number of skip connections and convolutional structures is constructed to address the gradient vanishing problem in traditional deep neural networks and accurately learn the nonlinear mapping relationship of systematic errors. The pre-defined deep residual network is trained using three-field forward modeling data, measured observation data, terrain parameters, and exploration instrument calibration parameters as inputs, and the systematic error compensation amount as the output. The forward modeling response vector and unstructured tetrahedral mesh terrain parameters are input into the pre-defined deep residual network. Error compensation and correction are performed through the output of the pre-defined deep residual network to obtain the inversion result. Deep residual error compensation can compensate for systematic errors such as environmental noise, instrument errors, terrain distortion, and emission current fluctuations, suppressing spurious anomalies in the inversion result and improving imaging reliability. The deep residual network accurately compensates for systematic errors such as environmental noise, instrument errors, and terrain distortion, effectively suppressing spurious anomalies caused by noise, terrain, and instruments.
[0048] This application provides a transient electromagnetic inversion method integrating multi-field response and intelligent optimization, comprising: acquiring standardized preprocessed three-field response data of a target underground space; performing terrain-constrained adaptive mesh partitioning on the target underground space to construct an unstructured tetrahedral mesh; calculating the convolution integral of the step response and the convolution integral of the waveform derivative based on a piecewise linear interpolation algorithm to correct the non-ideal emission waveform and obtain the corrected emission waveform; using the corrected emission waveform as the input for time-domain discrete recursion to calculate the three-field response data and obtain the forward response vector; constructing an adaptive sensitivity weighting matrix for the three-field response data based on the stratum depth of the target underground space and the forward response vector; constructing a multi-field joint inversion objective function based on the adaptive sensitivity weighting matrix, and obtaining the optimal resistivity model through a hybrid intelligent optimization algorithm; inputting the forward response vector and the terrain parameters of the unstructured tetrahedral mesh into a preset depth residual network, and completing error compensation correction through the output of the preset depth residual network to obtain the inversion result.
[0049] This application integrates three-field response data—electric field response, magnetic induction time derivative response, and magnetic induction response—for inversion, fully leveraging the advantages of slow late-stage decay, high deep sensitivity, and strong noise resistance of the magnetic induction response. This enables high-precision three-dimensional resistivity imaging of complex deep geological bodies. The accuracy of deep strata imaging is significantly improved, compensating for the deep blind spots of traditional exploration. Furthermore, the hybrid PSO algorithm and L-BFGS algorithm avoid local optima, reducing dependence on the initial model. An adaptive weighting mechanism for the three-field response data is introduced, dynamically adjusting the weight coefficients of the three-field responses to ensure that high-quality data dominates the inversion process, significantly improving inversion accuracy and further enhancing the accuracy of three-dimensional resistivity imaging. A deep residual network accurately compensates for systematic errors such as environmental noise, instrument errors, and terrain distortion, improving the reliability of three-dimensional resistivity imaging while maintaining accuracy.
[0050] In the embodiments of this application, obtaining standardized preprocessed three-field response data of the target underground space includes: The initial three-field response data of the target underground space were acquired using a synchronous acquisition mode. The initial three-field response data are standardized and preprocessed to obtain the preprocessed three-field response data.
[0051] For ease of understanding, the embodiments of this application employ an integrated TEM (Transient Electromagnetic Method) exploration system for data acquisition. The TEM exploration system includes a transmitting system, a receiving system, a sensor array, a data recording and control system, and a power supply system. In this embodiment, the transmitting system uses a high-power transmitter with a maximum transmitting current of 30A and an adjustable transmitting frequency from 0.25Hz to 6.25Hz, used to transmit a controllable single-pulse electromagnetic field underground. The receiving system uses a multi-channel synchronous receiver with 16 or more channels and an adjustable sampling rate from 1μs to 1ms, used to simultaneously acquire three-field response data. The power supply system uses a high-power DC power supply with an adjustable output voltage from 0V to 350V, providing a stable power supply for the transmitting and receiving systems.
[0052] The sensor array includes an electric field sensor, a magnetic induction time derivative sensor, and a magnetic induction sensor. The electric field sensor consists of grounded electrodes spaced 10m to 20m apart, used to collect electric field response data. The magnetic induction time derivative sensor is an adjustable induction coil with 100 to 500 turns and a high-permeability magnetic core, used to collect magnetic induction time derivative response data. The magnetic induction sensor employs a combination of a fluxgate sensor and a superconducting quantum interference device (SQU). The fluxgate sensor is used to collect medium-to-high amplitude magnetic induction response data, while the SQU is used to collect high-sensitivity magnetic induction response data, achieving a measurement accuracy of up to 10⁻⁶. -12 T.
[0053] A synchronous acquisition mode is adopted to obtain the initial three-field response data of the target underground space, ensuring that the initial three-field response data are acquired in the same launch cycle and at the same time node to avoid time synchronization errors. Specifically, area surveys are carried out in the work area, and grid-like measuring points are set according to the exploration targets corresponding to the target underground space. The spacing between measuring points is set to 50m to 200m depending on whether the density is increased in shallow areas or decreased in deep areas. The exploration targets can be ore bodies, groundwater, and geological hazard bodies, etc., without limitation. At each measuring point location, the transmitting coil is laid, the electric field electrode is arranged, and the magnetic induction sensor is installed to ensure good coupling between the sensor and the ground surface.
[0054] After starting the integrated TEM exploration system, perform a system self-test to check sensor connectivity, instrument noise level, and synchronization accuracy. After passing the self-test, set the transmission and reception parameters. Specifically, the transmission parameters include setting the transmission current to 15A to 30A, the transmission frequency to 2.5Hz, and the transmission waveform to a 2.5ms sinusoidal rising edge and a 0.4ms linear falling edge. The reception parameters include setting the sampling rate to 10μs, the sampling duration to 10ms, and using a synchronous trigger mode.
[0055] The transmitting system emits a single pulse of electromagnetic field, while the receiving system simultaneously acquires three response data fields. Each measuring point repeatedly acquires 3 to 5 sets of data to improve data reliability. Simultaneously, the system records the transmitted waveform parameters, terrain elevation data of the work area, and instrument calibration parameters. The transmitted waveform parameters include the rising edge, falling edge, turn-off time, and transmitted current intensity. The terrain elevation data of the work area is obtained from GPS (Global Positioning System) measurements. The instrument calibration parameters include sensor sensitivity and receiver gain, which will not be elaborated upon here.
[0056] Preliminary quality control is performed on the initial three-field response data, checking the completeness and validity of the collected data in real time, and removing bad channel data caused by instrument malfunction, poor sensor contact, and strong electromagnetic interference. The initial three-field response data undergoes standardization preprocessing to obtain preprocessed three-field response data. The standardization preprocessing is set according to actual needs and is not limited here. For ease of understanding, the standardization preprocessing in the embodiments of this application includes bad channel data removal, wavelet threshold denoising, transmit current amplitude deviation correction, terrain undulation distortion correction, and data format standardization.
[0057] Bad channel data removal employs the 3σ criterion, calculating the mean and standard deviation of data for each channel at each measuring point. Extreme bad channel data with amplitudes exceeding the mean ± 3 times the standard deviation are removed, further eliminating bad channel data caused by strong electromagnetic interference and instrument malfunction. Wavelet thresholding denoising is used to suppress random Gaussian noise such as magnetotelluric noise and instrument electronic noise. The db4 wavelet is selected as the wavelet basis function, and the decomposition level is 5 to 8 layers. By setting reasonable thresholds, effective anomalous signals from underground geological bodies are retained, while high-frequency random noise is removed, improving the data signal-to-noise ratio by more than 20 dB.
[0058] The transmit current amplitude deviation correction relies on the Biot-Savart law. Based on the deviation between the measured and design values of the transmit current, it corrects the amplitude deviation of the three-field response data, eliminating measurement errors caused by current fluctuations. The correction formula is: Formula (1) Where, d corr For the corrected data, d raw For the original data, I design For the design value of the emission current, I meas This is the measured value of the emission current.
[0059] Transmit current amplitude deviation correction relies on the Digital Elevation Model (DEM) and employs the terrain correction coefficient method to correct the interference of terrain undulation distortion on the three-field response. Based on the difference between the terrain elevation of the measuring point and the reference elevation, terrain correction coefficients are calculated to correct the amplitude of the three-field response data, eliminating signal distortion caused by terrain undulation. Data format standardization converts the preprocessed three-field response data into a standard format, unifying the data dimensions and coordinate system, providing standardized data input for subsequent forward modeling and inversion calculations.
[0060] In the embodiments of this application, terrain-constrained adaptive meshing is performed on the target underground space to construct an unstructured tetrahedral mesh for the target underground space, including: Based on the digital elevation model of the work area, a surface model is constructed. Generate a triangular surface grid based on the surface model; Based on the geological data, borehole data and electromagnetic field sensitivity pre-calculation results of the target underground space, the definition of the grid densification area is obtained. The grid densification area includes the topographic relief area, the anomalous geological body area and the high-sensitivity detection area. The Delaunay tetrahedral partitioning algorithm is adopted, combined with triangular surface mesh and mesh densification region definition, to perform tetrahedral mesh partitioning of the target underground space and construct an unstructured tetrahedral mesh of the target underground space.
[0061] Based on the digital elevation model (DEM) of the work area, a surface model is constructed. Specifically, DEM data at scales of 1:1000 to 1:5000 are collected for the work area. The DEM data undergoes preprocessing such as denoising, interpolation, and resampling to generate a high-precision DEM grid with a grid spacing of 5m to 10m. Cubic spline interpolation is used to fit the DEM grid nodes to construct a continuous and smooth surface model, accurately reproducing topographic undulations such as slope, aspect, and elevation changes. Based on the surface model, a triangular surface grid is generated, with its size adaptively adjusted according to the terrain slope. For ease of understanding, in this embodiment, for steep terrain areas with slopes greater than 30°, the grid size is adjusted to 5m to 10m. For gentle terrain areas with slopes less than 10°, the grid size is adjusted to 20m to 50m to ensure the surface grid conforms to the terrain features.
[0062] Based on the target exploration depth, the horizontal range of the target underground space is defined as 500m beyond the work area's measuring points, and the vertical range is 1200m from the surface to the subsurface, with a 200m buffer layer reserved to avoid boundary effects. Based on known geological data, borehole data, and pre-calculated electromagnetic field sensitivity of the target underground space, the definition of the grid densification area is obtained. The grid densification area includes topographic relief areas, anomalous geological body areas, and high-sensitivity detection areas. For ease of understanding, in this embodiment, the topographic relief area refers to underground areas with a surface slope greater than 20°; the anomalous geological body area refers to the distribution areas of anomalous bodies such as ore bodies, aquifers, and fault zones revealed by boreholes; and the high-sensitivity detection area is the area where the pre-calculated three-field response sensitivity gradient is greater than the sensitivity gradient threshold. The sensitivity gradient threshold is set based on 10% of the maximum sensitivity value, and will not be elaborated here.
[0063] The Delaunay tetrahedral partitioning algorithm, combined with triangular surface meshes and mesh refinement region definitions, is used to partition the target underground space into tetrahedral meshes, constructing an unstructured tetrahedral mesh for the target underground space. The unstructured tetrahedral mesh includes a refined region mesh and a sparse region mesh. The refined region mesh size is 10m to 20m, where the anomaly core region size is 10m to 15m, and the edge region size is 15m to 20m.
[0064] It is important to understand that the embodiments of this application also require mesh quality optimization of the unstructured tetrahedral mesh. Distorted meshes with a minimum interior angle less than 10° and an aspect ratio greater than 5 are removed. Mesh quality is improved through operations such as mesh smoothing and node optimization to ensure the stability of forward modeling calculations. Compared to traditional uniform structured meshes, the unstructured tetrahedral mesh obtained in this application has advantages such as strong terrain adaptability, low mesh redundancy, high computational efficiency, and high accuracy in anomaly imaging. The unstructured tetrahedral mesh can perfectly fit the real terrain surface, reducing terrain distortion error by more than 90%, resulting in strong terrain adaptability. The anomaly area is denser while the background area is sparser, reducing the total number of meshes by 52%. At the same accuracy, the unstructured tetrahedral mesh in the embodiments of this application requires half the number of meshes compared to the structured mesh, resulting in low mesh redundancy. The reduced number of meshes in the unstructured tetrahedral mesh directly reduces the amount of forward modeling computation, increasing the forward modeling calculation speed by 3 times, thus improving computational efficiency. In addition, the mesh densification of the anomaly area of the unstructured tetrahedral mesh can accurately delineate the boundaries of irregular anomalies, and the imaging clarity of the anomaly boundaries is greatly improved, resulting in high imaging accuracy of anomalies, which is suitable for complex construction areas such as mountains, deserts, water areas and permafrost.
[0065] In the embodiments of this application, based on a piecewise linear interpolation algorithm, the convolution integral of the step response and the convolution integral of the waveform derivative are calculated to correct the non-ideal transmission waveform, resulting in a corrected transmission waveform, including: Under the quasi-static electromagnetic assumption, transient electromagnetic detection control equations are constructed, which include electric field diffusion control equations. By introducing the unit step electric field response and the unit step magnetic induction response, the electromagnetic field response data corresponding to the non-ideal emission waveform is decomposed into the convolution integral of the step response and the convolution integral of the waveform derivative. The convolution integral of the step response and the convolution integral of the waveform derivative are calculated using a piecewise linear interpolation algorithm to correct the non-ideal transmission waveform and obtain the corrected transmission waveform.
[0066] The quasi-static electromagnetic assumption neglects the influence of displacement current and is applicable to transient electromagnetic low-frequency detection scenarios. Under this assumption, the transient electromagnetic detection control equations are constructed. These equations clarify the coupling relationship between the electric and magnetic fields in transient electromagnetic detection, forming the theoretical basis for three-field forward modeling. Calculating these equations yields the three-field response under arbitrary subsurface resistivity distributions. The transient electromagnetic detection control equations include the electric field diffusion control equation and Faraday's law of electromagnetic induction. Specifically, the electric field diffusion control equation is: Formula (2) Faraday's law of electromagnetic induction equation is: Formula (3) in, For curl operator, The magnetic permeability of the underground medium, expressed in H / m, is usually taken as the vacuum permeability. H / m, For spatial and temporal coordinates in The electric field strength at that location is expressed in V / m. The electrical conductivity of the underground medium, measured in S / m, is the reciprocal of resistivity. , Time, in seconds. The source current density, expressed in A / m², characterizes the primary field current distribution generated by the transmitting coil. For spatial and temporal coordinates in The magnetic field strength at that location. It is important to understand that, except... Apart from that, the other parameters of Faraday's law of electromagnetic induction have the same meaning as those in the electric field diffusion control equation.
[0067] By introducing the unit step electric field response and the unit step magnetic induction response, the electric field response data corresponding to the non-ideal transmission waveform is decomposed into the convolution integral of the step response and the convolution integral of the waveform derivative. The unit step response refers to the electromagnetic field response when the transmission waveform is an ideal step function. The electromagnetic field response includes both the electric field response and the magnetic induction response; that is, the electric field response data includes both electric field response data and magnetic induction response data. This transforms the complex calculation of the non-ideal waveform response into a simple calculation of the unit step response and convolution integral, eliminating the need for separate forward modeling calculations for each waveform and significantly reducing computational complexity. In this embodiment, the formula for the electric field response convolution integral is: Formula (4) The formula for the convolution solution of magnetic induction response is: Formula (5) in, The electric field strength represents the time-varying value of the actual non-ideal emission waveform. The magnetic flux density represents the time-varying value of the actual non-ideal emission waveform. t For time, The current intensity of the actual transmitted waveform varies with time as a function. Let be the electric field strength value at the initial moment of the unit step response of the electric field response. The initial value of the magnetic field strength at the unit step response of the magnetic induction response is the initial value of the magnetic field strength at the initial moment. t =0, The start time of the transmitted waveform. The time integration variable of the convolution integral. The delay time is The corresponding unit step response of the electric field response, The delay time is The corresponding unit step response of the magnetic induction response, It is a constant.
[0068] The second-order backward Euler method is used to discretize the governing equations of the electric field diffusion in time, dividing the time domain into discrete time steps, resulting in recursive formulas for the electric field and magnetic induction. The second-order backward Euler method is used to discretize the governing equations of the electric field diffusion in time, resulting in recursive formulas for the electric field and magnetic induction: Formula (6) Formula (7) in, , For time step, t k For the first k Each time step C The discrete curl matrix is obtained by discretization using the electric field curl operator. M The mass matrix is obtained by discretizing the dielectric conductivity and permeability parameters. e k For the first k Discrete solution of electric field intensity at each time step b k For the first Discrete solution of magnetic flux density at each time step I k For the first k The emission current intensity at each time step s For the normalized source term, This is the field mapping matrix, which realizes the mapping between the electric field and the time derivative of magnetic induction.
[0069] A piecewise linear interpolation algorithm is used to calculate the convolution integral of the step response and the convolution integral of the waveform derivative to correct the non-ideal transmission waveform, resulting in the corrected transmission waveform. The transmission waveform is divided into several linearly segmented intervals using the piecewise linear interpolation algorithm, with each interval exhibiting linear variation. Piecewise linear interpolation is performed on the unit step response to obtain the step response value corresponding to any delay time. The convolution integral is calculated segment by segment and summed to obtain an approximate solution for the convolution integral, thus correcting the non-ideal transmission waveform. Compared to traditional methods, this application achieves high-precision convolution integral calculation without requiring encrypted time steps to correct non-ideal transmission waveforms, improving correction accuracy by 35.3% and computational efficiency by more than 2 times.
[0070] In the embodiments of this application, the transient electromagnetic detection control equations also include Faraday's law of electromagnetic induction equations; Using the corrected transmitted waveform as input for time-domain discrete recursion, the three-field response data are calculated to obtain the forward response vector, including: The second-order backward Euler method is used to discretize the electric field diffusion control equation in time, dividing the time domain into discrete time steps, and obtaining the recursive calculation formulas for the electric field and magnetic induction. Based on the discrete electric field diffusion control equation of the edge finite element method, and combined with the recursive calculation formula of electric field and magnetic induction, the electric field response data is calculated to obtain the spatiotemporal distribution of the electric field response. Based on Faraday's law of electromagnetic induction, the time derivative response data of magnetic induction is calculated, and the spatiotemporal distribution of the time derivative response of magnetic induction is obtained. Based on the recursive calculation formulas for electric field and magnetic induction and the electric field response at the actual measurement point, the magnetic induction response data is calculated to obtain the spatiotemporal distribution of the magnetic induction response. By integrating the spatiotemporal distributions of the electric field response, the magnetic induction time derivative response, and the magnetic induction response, the forward response vector is obtained.
[0071] Based on the edge finite element method (EBM) to discretize the electric field diffusion control equations and combine them with the recursive calculation formulas for electric field and magnetic induction, the electric field response data is calculated to obtain the spatiotemporal distribution of the electric field response, which includes both temporal and spatial distributions. The edge finite element method can accurately discretize the electric field curl operator, is adaptable to unstructured tetrahedral meshes, and has high computational accuracy.
[0072] Based on Faraday's law of electromagnetic induction, the time derivative response data of magnetic induction were calculated, and the spatiotemporal distribution of the time derivative response of magnetic induction was obtained as follows: ,in, The field mapping matrix, e To calculate the spatiotemporal distribution of the electric field response, This is the spatiotemporal distribution of the obtained magnetic induction time derivative response. The magnetic induction time derivative response can be directly calculated synchronously from the electric field response without additional forward modeling, which is highly efficient.
[0073] To reduce computational cost, the magnetic induction response is calculated only at actual field measurement points, avoiding redundant calculations across the entire grid. Based on the recursive formulas for electric field and magnetic induction, and the electric field response at the actual measurement points, the magnetic induction response data for each measurement point is calculated, yielding the spatiotemporal distribution of the magnetic induction response. Integrating the spatiotemporal distributions of the electric field response, the magnetic induction time derivative response, and the magnetic induction response, the forward response vector is obtained as follows: Formula (8) in, d pred Forward response vector T This is the matrix transpose symbol. eTo calculate the spatiotemporal distribution of the electric field response, Let be the spatiotemporal distribution of the obtained magnetic induction time derivative response, and b be the spatiotemporal distribution of the calculated magnetic induction response.
[0074] In this embodiment, convolution and interpolation correction eliminate the need for encrypted time steps, decoupling waveform complexity and time step constraints. This improves waveform correction accuracy by 35.3%, forward modeling efficiency by more than 2 times, and shortens the inversion cycle of large-scale 3D models by 60%. The number of meshes is reduced by 50%, forward modeling speed is increased by 3 times, and the computing hardware requirements are significantly reduced.
[0075] In the embodiments of this application, an adaptive sensitivity weighting matrix for the three-field response data is constructed based on the stratum depth of the target underground space and the forward response vector, including: When the depth of the target underground space is greater than the preset depth, the sensitivity Jacobian matrix of the three-field response to the resistivity of the underground medium is constructed based on the resistivity vector of the underground medium and the forward response vector. The sensitivity Jacobian matrix is calculated using the adjoint method to obtain the sensitivity gradient; Based on the sensitivity gradient, the spatial sensitivity distribution of the three-field response is calculated; Based on the stratum depth and spatial sensitivity distribution of the target underground space, an adaptive sensitivity weighting matrix for the three-field response data is constructed using a diagonal matrix generating function. The preset depth is set according to requirements and is not limited here. For ease of understanding, the preset depth in this embodiment is 500m. When the depth of the target underground space is less than or equal to 500m, dynamic allocation of weighting coefficients is not required. When the depth of the target underground space is greater than the preset depth, dynamic allocation of weighting coefficients is performed. Based on the resistivity vector of the underground medium and the forward response vector, a sensitivity Jacobian matrix for the three-field response to the resistivity of the underground medium is constructed. This matrix characterizes the degree of influence of resistivity changes on the electromagnetic field response. The sensitivity Jacobian matrix is: Formula (9) in, J The sensitivity Jacobian matrix, d pred Forward response vector This is the resistivity vector of the subsurface medium, containing the resistivity parameters of all grid nodes, with an array dimension of [value missing]. , N data This represents the total number of response data from the three events. N model The total number of parameters in the resistivity model, matrix elements Indicates the first The response to the first Sensitivity of resistivity of individual grid nodes.
[0076] The sensitivity Jacobian matrix is calculated using the adjoint method to obtain the sensitivity gradient. With 500,000 grid points, the model parameters also reach 500,000. The number of parameters in a transient electromagnetic 3D model is enormous, and directly calculating the sensitivity matrix requires large-scale repeated forward modeling, which is extremely time-consuming. Using the adjoint method to calculate the sensitivity Jacobian matrix requires only one forward model and one adjoint forward model to obtain the sensitivity gradient, avoiding large-scale repeated forward modeling calculations. The adjoint field governing equations are constructed, and the adjoint field vector is calculated as follows: Formula (10) in, A For the forward modeling system matrix, A T Let be the transpose of the matrix of the orthogonal system. Let be the adjoint field vector. Q t For time interpolation matrix, Q s For spatial interpolation matrix, For the data residual vector, d obs This is a vector of field-measured observation data. d pred Forward response vector This is the data covariance matrix.
[0077] Sensitivity gradient calculation is based on the forward electromagnetic field solution vector and the adjoint field vector. The sensitivity gradient is calculated as follows: Formula (11) in, Let be the sensitivity gradient of the objective function with respect to the resistivity model parameters. x This is the electromagnetic field solution vector obtained from forward modeling. T This is the matrix transpose symbol. Let be the adjoint field vector. Let be the partial derivative matrix with respect to resistivity.
[0078] Based on the sensitivity gradient, the spatial sensitivity distribution of the three-field response was calculated. The spatial sensitivity distribution of the three-field response was calculated at different sampling times and burial depths. The sensitivity data were normalized and multi-scale decomposed to clarify the depth adaptability of the three-field response. The stratigraphic exploration intervals included shallow strata, middle strata, and deep strata.
[0079] The shallow strata detection range is from 0m to 200m. The electric field response and magnetic induction time derivative response have high sensitivity, while the magnetic induction response has lower sensitivity. For the shallow strata range, the electric field response and magnetic induction time derivative are the primary detection fields, enabling accurate identification of shallow groundwater and shallow ore bodies, as well as other low-resistivity anomalies.
[0080] The stratigraphic exploration range in the central strata is 200m to 500m. The three-field response sensitivity is balanced, and the electric field response, magnetic induction time derivative and magnetic induction response all have high sensitivity. The corresponding interval of the central strata is explored with balanced three-field detection, which can take into account both the shallow and deep features of the central stratigraphic anomaly.
[0081] For deep strata exceeding 500m, the sensitivity of the magnetic induction response is significantly higher than that of the electric field response and the magnetic induction time derivative response, with the induction time derivative response signal almost completely attenuated. Using the magnetic induction response as the core detection field in the deep strata region, this method can effectively identify deep concealed ore bodies and aquifers, compensating for the blind spots in traditional detection methods.
[0082] Based on the stratigraphic depth and spatial sensitivity distribution of the target subsurface space, an adaptive sensitivity weighting matrix for the three-field response data is constructed using a diagonal matrix generating function. Combining stratigraphic depth and spatial sensitivity distribution, a diagonal weighting matrix for the three-field data is constructed, with weights dynamically allocated to ensure that high-sensitivity and high-quality data dominate the inversion process. The adaptive sensitivity weighting matrix is as follows: Formula (12) in, Let be the sensitivity gradient of the objective function with respect to the resistivity model parameters. x This is the electromagnetic field solution vector obtained from forward modeling. T This is the matrix transpose symbol. Let be the adjoint field vector. Let be the partial derivative matrix with respect to resistivity. W This is the adaptive sensitivity weighting matrix. diag This is the adaptive sensitivity weighting matrix. w e For the electric field response, , w db / dt These are the weighting coefficients for the time derivative response of magnetic induction. w b These are the weighting coefficients for the magnetic induction response, and the weights satisfy the normalization constraint: .
[0083] The weighting coefficients are dynamically adjusted based on the formation depth. In this embodiment, when the formation exploration area is shallow, the weighting coefficient for the electric field response is 0.4, the weighting coefficient for the magnetic induction time derivative response is 0.4, and the weighting coefficient for the magnetic induction response is 0.2. When the formation exploration area is middle-depth, the weighting coefficients for the electric field response are 0.33, the weighting coefficients for the magnetic induction time derivative response are 0.33, and the weighting coefficient for the magnetic induction response is 0.34. When the formation exploration area is deep, the weighting coefficients for the electric field response are 0.175, the weighting coefficients for the magnetic induction time derivative response are 0.175, and the weighting coefficient for the magnetic induction response is 0.65. These coefficients can be fine-tuned based on the residuals from measured data, up to a maximum of 0.8.
[0084] Multi-scale sensitivity analysis can guide the optimization of field data acquisition schemes, adapting to various engineering scenarios such as mineral exploration, geological disaster prevention and control, groundwater exploration, and deep geological structure detection. It requires no grounding electrodes, making it suitable for complex terrains such as mountains, deserts, water bodies, and permafrost—areas traditionally difficult to construct in—thus enabling flexible construction and strong engineering adaptability.
[0085] In the embodiments of this application, a multi-field joint inversion objective function is constructed based on the adaptive sensitivity weighting matrix, and the optimal resistivity model is obtained through a hybrid intelligent optimization algorithm, including: 1) Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed that takes into account both data fitting accuracy and model balance constraints. The multi-field joint inversion objective function includes a data fitting term and a model regularization term. The data fitting term adopts the weighted least squares criterion and introduces the adaptive sensitivity weighting matrix and the data covariance matrix. The model regularization term adopts the model roughness minimization criterion. 2) Construct an initial population consisting of multiple particles, where each particle corresponds to a resistivity model; 3) An adaptive particle swarm optimization algorithm is used to perform a global search, update the velocity and inertia weights of particles in the initial swarm, and obtain the globally optimal model; 4) Use the globally optimal model as the initial model input for the L-BFGS algorithm; 5) Calculate the gradient information of the multi-field joint inversion objective function with respect to the model parameters, and use the Wolfe criterion to perform a one-dimensional line search to determine the optimal iteration step size; 6) Based on the gradient information and the optimal iteration step size, iterate the parameters of the globally optimal model to update the globally optimal model; 7) If the multi-field joint inversion objective function fails to converge or reaches the maximum number of iterations, use the updated global optimal model as the initial model input for the L-BFGS algorithm, and repeat steps 4) and 5) until the multi-field joint inversion objective function converges or reaches the maximum number of iterations to obtain the optimal resistivity model. In this case, the weight coefficients of the three field responses are dynamically adjusted based on the data residuals of the three field responses for a preset number of iterations.
[0086] Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed that balances data fitting accuracy and model balance constraints. This ensures good data fit while suppressing overfitting and improving model stability. The multi-field joint inversion objective function is as follows: Formula (13) in, The objective function for multi-field joint inversion is the overall objective function. This is a data fitting term, used to characterize the degree of fit between the forward response and the measured data. This is a model regularization term used to characterize the smoothness of the model and suppress overfitting. This is a regularization balancing parameter that adjusts the weights of data fitting and model smoothing; its value ranges from 10. -5 Up to 10 -1 The optimal value is determined through trial and error.
[0087] The objective function for multi-field joint inversion includes a data fitting term and a model regularization term. The data fitting term employs a weighted least squares criterion, introducing an adaptive weighting matrix and a data covariance matrix from the three response data to quantify the residuals between the forward response and the measured data. The formula for calculating the data fitting term is as follows: Formula (14) in, This is a data fitting term, used to characterize the degree of fit between the forward response and the measured data. T This is the matrix transpose symbol. d obs This is a vector of field-measured observation data. d pred Forward response vector This is the data covariance matrix, used to characterize the error distribution of the measured data. The diagonal matrix and its diagonal elements are the reciprocals of the variance of each data point.
[0088] An adaptive sensitivity weighting matrix and a data covariance matrix are introduced. The model regularization term adopts the model roughness minimization criterion to constrain the smoothness of the resistivity model and avoid severe oscillations and spurious anomalies. The formula for calculating the model regularization term is as follows: Formula (15) in, For model regularization terms, m For resistivity transformation model parameters, avoid non-physical resistivity values. m 0 represents the initial model parameters for inversion. This is the model roughness matrix, used to characterize the model's smoothing constraints, and is usually constructed using finite difference operators.
[0089] In transient electromagnetic inversion, resistivity parameters must satisfy physical constraints, i.e., resistivity must be a positive real number. To avoid non-physical resistivity values such as negative resistivity, maximum resistivity, and minimum resistivity during the inversion process, a nonlinear transformation is used to convert physically constrained resistivity parameters into unconstrained model parameters, thus preventing non-physical values and improving inversion stability. A nonlinear exponential transformation is used to constrain the resistivity range, limiting the resistivity interval to [specific range to be filled in]. and ,in, This represents the lower limit of the resistivity range. This is the upper limit of the resistivity range to cover the resistivity range of common geological media; typically, the resistivity range of clay is... Up to 10 The resistivity range of sandstone is Up to 1000 The resistivity of granite is greater than 1000. The resistivity range of metal sulfides is Up to 100 The positive transformation from resistivity to model parameters is: Formula (16) The inverse transformation of model parameters to resistivity is as follows: Formula (17) in, m i For the first i Transformation model parameters for each grid node For the first i The resistivity of each grid node This represents the lower limit of the resistivity range. This represents the upper limit of the resistivity range.
[0090] This application employs a hybrid optimization strategy combining adaptive particle swarm optimization (PSO) with the L-BFGS algorithm, divided into two stages: global search and local refinement, balancing global search capability with local convergence accuracy. PSO is a global optimization algorithm based on swarm intelligence, simulating bird flock foraging behavior to search for optimal solutions in the solution space. This application utilizes an adaptive PSO algorithm, dynamically adjusting particle velocity and inertia weights to enhance global search capability and avoid local optima.
[0091] An initial population of multiple particles is constructed, using a homogeneous half-space stratigraphic model as the initial population. The number of particles is set to 50 to 100, adjusted according to the model complexity. Each particle represents a resistivity model, and the initial particle positions are randomly generated within a reasonable range.
[0092] An adaptive particle swarm optimization (PSO) algorithm is used to perform a global search, updating the velocity and inertia weights of particles in the initial swarm to obtain the globally optimal model. The adaptive inertia weights decrease linearly with the number of iterations, and the particle velocity and position are updated as follows: Formula (18) in, for k The second iteration i The speed of each particle For the first k The second iteration i The position of each particle. w The adaptive inertia weight is initially set to 0.9 and then reduced to 0.4 at the end of the iteration. c 1 and c 2 represents the learning factor, both set to 2.0. r 1 and r Both 2 are random numbers in the interval [0,1]. For the first k The second iteration i The historical best position of each particle This is the globally optimal position for the population.
[0093] After each iteration, the objective function value for each particle is calculated, and the historical best position of the particle and the global best position of the population are updated to select the global optimal model. After iterating to a preset number of times, the global optimal model is output as the initial model for the local refinement stage.
[0094] The globally optimal model is used as the initial model input for the L-BFGS algorithm. The gradient information of the objective function with respect to the model parameters is calculated using the multi-field joint inversion objective function, and the optimal iteration step size is determined by a one-dimensional line search using the Wolfe criterion. The L-BFGS algorithm is an efficient quasi-Newton gradient optimization algorithm that quickly converges to a local optimum by approximating the Hessian matrix with limited memory. For the initial model input, the globally optimal model output by the adaptive PSO algorithm is used as the initial model for the L-BFGS algorithm. For the objective function gradient calculation, the gradient of the overall objective function with respect to the model parameters is calculated. The formula for calculating the gradient of the model parameters is: Formula (19) in, The gradient of the objective function with respect to the model parameters. The square root of the model roughness matrix. The sensitivity Jacobian matrix, To standardize the transformation model parameters, d obs This is a vector of field-measured observation data. d pred Forward response vector m For resistivity transformation model parameters, avoid non-physical resistivity values. m 0 represents the initial model parameters for inversion.
[0095] This application improves the traditional L-BFGS algorithm by introducing an adaptive iteration step size, thereby enhancing local convergence accuracy. Specifically, it employs the Wolfe criterion for one-dimensional line search, adaptively determining the optimal iteration step size to ensure sufficient descent of the objective function. Based on gradient information and the iteration step size, the model parameters are updated, iterating until the objective function converges or reaches the maximum number of iterations. In this embodiment, convergence of the objective function is defined as a change in the objective function less than a change threshold, which is 10. -6 The maximum number of iterations is 30 to 50.
[0096] Based on gradient information and the optimal iteration step size, the globally optimal model is iterated until the multi-field joint inversion objective function converges or reaches the maximum number of iterations, thus obtaining the optimal resistivity model. For each preset number of iterations, the weight coefficients of the three field responses are dynamically adjusted based on the data residuals. The preset number of iterations is set according to actual needs and is not limited here; for ease of understanding, it is set to 5 iterations in this embodiment. Every 5 iterations, the weights are dynamically adjusted based on the data residuals of each field. Fields with larger residuals have lower weights, while fields with smaller residuals have higher weights, ensuring that high-quality data dominates the inversion process.
[0097] In the embodiments of this application, the preset depth residual network is obtained through the following steps: Construct an initial deep residual network, which includes an input layer, a residual feature extraction layer and an output layer connected in sequence. The residual feature extraction layer has a preset number of skip-connected convolutional structures. The loss function of the deep residual network adopts the mean square error criterion, and the initial deep residual network is trained using the Adam optimizer to obtain the preset deep residual network; The forward modeling response vector and terrain parameters of the unstructured tetrahedral mesh are input into a preset depth residual network. Error compensation and correction are performed through the output of the preset depth residual network to obtain the inversion results, including: Obtain the initial inversion model response corresponding to the forward response vector; The initial inversion model response, topographic parameters of the unstructured tetrahedral mesh, measured observation data of the target underground space, and exploration instrument parameters are input into a preset residual network. Error compensation and correction are completed through the output of the preset residual network to obtain the inversion result.
[0098] The initial deep residual network is constructed, and the preset number of layers is set according to actual needs and is not limited here. For ease of understanding, the preset number in this embodiment is 10 layers, constructing a 10-layer deep residual network (ResNet) with a convolutional structure and skip connections. This solves the gradient vanishing problem of traditional deep neural networks and accurately learns the nonlinear mapping relationship of system errors. The network structure of the initial deep residual network consists of an input layer, a residual feature extraction layer, and an output layer. The input layer includes three-field response forward modeling data, measured observation data, terrain parameters, and exploration instrument calibration parameters, totaling four types of input features, comprehensively representing the sources of error. The output layer uses a fully connected layer to output the system error compensation amount, and the output dimension is consistent with the forward response data, achieving accurate prediction of system errors.
[0099] The residual feature extraction layer uses a 10-layer convolutional structure. In this embodiment, each convolutional kernel has a size of 3×3. The number of kernels in the first four layers are 64, 128, 256, and 512 respectively, and the subsequent six layers maintain 512 kernels. Skip connections are introduced to directly pass shallow features to deep layers, avoiding gradient vanishing. The ReLU (Rectified Linear Unit) activation function is used to improve the network's nonlinear fitting ability.
[0100] In this embodiment, artificially synthesized simulation data and measured data are used to construct training and validation sets. The mean squared error criterion is used as the loss function. The Adam optimizer is used to train the initial deep residual network until the loss function of the validation set tends to stabilize, and the model converges to obtain the preset deep residual network.
[0101] For ease of understanding, the embodiments of this application construct different geoelectric models, such as homogeneous strata, tilted anomalies, and multi-layered strata, for artificially synthesized simulation data. Forward modeling is used to generate theoretical response data, and 4% to 10% Gaussian random noise, topographic distortion errors, and instrument calibration errors are superimposed to simulate complex field measurement environments, generating 10,000 sets of simulation data. For measured data, measured three-field response data from typical measuring points in the work area are selected. Systematic errors are determined by comparing borehole data and geological data, generating 2,000 sets of measured data. The total data is divided into a training set of 9,600 sets and a validation set of 2,400 sets at an 8:2 ratio.
[0102] The loss function for deep residual networks adopts the Mean Squared Error (MSE) criterion. The loss function for deep residual networks is as follows: Formula (20) in, L The loss function for deep residual networks is... N The total number of training samples, This represents the true system error. This represents the network prediction error.
[0103] For ease of understanding, the initial learning rate of the Adam optimizer in the embodiments of this application is 10. -4 The batch size is 32. The maximum number of iterations for training the initial deep residual network is 100. Training is terminated early when the validation set loss function does not decrease for 10 consecutive iterations. After training, the model parameters with the minimum validation set loss function are saved, and the pre-set deep residual network obtained from the training is used as the final error compensation model.
[0104] The initial inversion model response is a specific feature of the forward response vector under the initial resistivity model, which is the resistivity model before the hybrid intelligent optimization algorithm is executed iteratively. The initial inversion model response corresponding to the forward response vector is obtained. The initial inversion model response, topographic parameters of the unstructured tetrahedral mesh, measured observation data of the target subsurface space, and exploration instrument parameters are input into a preset residual network. Compensation and correction are performed through the output of the preset residual network to obtain the inversion result. Formula (21) in, d comp The accurate response data after error compensation, i.e., the inversion result, d inv The initial inversion model response, This represents the prediction error output by the preset residual network. In this embodiment, through depth residual error compensation, the anomaly suppression effect is improved by 60%, the imaging clarity of deep strata is improved by 50%, and the accuracy of resistivity values is improved by 40%.
[0105] In embodiments of this application, the method further includes: Using the initial inversion results as the data source, a three-dimensional subsurface resistivity model is output. The three-dimensional subsurface resistivity model includes the grid node parameters of the unstructured tetrahedral mesh, the parameters of the anomalous geological body, the imaging quality parameters, and the visualization data. The visualization data includes three-dimensional resistivity volume data, horizontal slices of the target subsurface space, and vertical profiles of the target subsurface space. The three-dimensional resistivity model of the underground was verified using three-field response simulation data, and the accuracy verification results were obtained. The underground three-dimensional resistivity model was verified by using measured data from the target underground space, and the results of the engineering practicality verification were obtained.
[0106] The inversion result is a high-precision resistivity model; however, it cannot be loaded into general-purpose 3D visualization software and cannot be directly delivered to engineers. The embodiments of this application further include: outputting a subsurface 3D resistivity model based on the inversion result, and verifying the subsurface 3D resistivity model. Outputting and verifying the subsurface 3D resistivity model based on the inversion result includes: using the inversion result as a data source to output the subsurface 3D resistivity model; verifying the subsurface 3D resistivity model using three-field response simulation data to obtain accuracy verification results; and verifying the subsurface 3D resistivity model using measured data from the target subsurface space to obtain engineering applicability verification results.
[0107] The subsurface 3D resistivity model includes grid node parameters of an unstructured tetrahedral mesh, anomalous geological body parameters, imaging quality parameters, and visualization data. The visualization data includes 3D resistivity volume data, horizontal slices of the target subsurface space, and vertical profiles of the target subsurface space. For ease of understanding, in this embodiment, the grid node parameters include the spatial coordinates, resistivity values, and confidence levels of all tetrahedral mesh nodes. Anomalous geological body parameters include the spatial coordinate range, burial depth, shape, size, mean resistivity, and extreme resistivity values of the anomalous body. Imaging quality parameters include model reconstruction error, residual distribution, signal-to-noise ratio, and false anomaly suppression rate. The visualization data includes 3D resistivity volume data, horizontal slices of the target subsurface space, and vertical profiles of the target subsurface space. The visualization data supports direct loading and display by 3D visualization software, allowing the subsurface 3D resistivity model, relative to the inversion results in this embodiment, to be loaded by general 3D visualization software.
[0108] Furthermore, this embodiment employs a dual verification mechanism using synthetic simulation data and field measured data to verify the inversion accuracy, stability, and engineering practicality, providing a quantitative credibility index for engineering decision-making. The three-dimensional subsurface resistivity model is verified using three-field response simulation data to obtain accuracy verification results. Specifically, a synthetic geoelectric model is constructed as a comparative example. A three-dimensional tilted low-resistivity anomaly model is constructed, with basic parameters including background stratum resistivity, anomaly parameters, and topographic conditions. For ease of understanding, the background stratum resistivity in this embodiment is 200... The abnormal body parameters are obtained using a tilted conductive thin plate with a resistivity of 4. The dimensions are 100m×200m×150m, with a top burial depth of 150m; the terrain is randomly undulating, with an elevation range of -59.9m to 149.21m.
[0109] Three-field response simulation data were generated using a forward modeling algorithm, superimposed with 4% Gaussian random noise to simulate a complex field measurement environment. The transient electromagnetic inversion method described in this example was then applied to the three-field response simulation data for inversion calculations, quantitatively determining the resistivity numerical error, anomaly spatial location error, and morphological error of the subsurface three-dimensional resistivity model. The verification results show that the subsurface three-dimensional resistivity model accurately reconstructs the spatial location, burial depth, and morphology of the tilted anomaly; the resistivity reconstruction error is less than 7%; and the anomaly image at a depth of 400m is clear, without any false anomaly interference.
[0110] The underground three-dimensional resistivity model was verified using measured data from the target underground space, yielding results demonstrating its engineering practicality. In this embodiment, the target underground space is an exploration area for a concealed iron ore deposit, characterized by flat terrain and a surface elevation difference of less than 15m. The exploration target is a concealed iron ore body with associated gold and silver deposits at a depth of 100m to 800m, exhibiting a complex geological structure.
[0111] Please see Figure 2 , Figure 2 This is the inversion result of a conventional two-dimensional layered inversion algorithm according to an embodiment of the present invention.
[0112] Please see Figure 3 , Figure 3 This is the inversion result of a transient electromagnetic inversion method that integrates multi-field response and intelligent optimization according to an embodiment of the present invention.
[0113] Field measurements were conducted on the target underground space to obtain borehole exploration data and geological profiles. A conventional two-dimensional layered inversion algorithm was used to obtain... Figure 2 The inversion results are shown. Using the transient electromagnetic inversion method integrating multi-field response and intelligent optimization in this embodiment, the output subsurface three-dimensional resistivity model is... Figure 3The inversion results are shown. The inversion results are compared with measured data to validate the three-dimensional resistivity model of the underground space using measured data, thus obtaining engineering applicability verification results. For example... Figure 2 and Figure 3 As shown, both the conventional two-dimensional layered inversion algorithm and the transient electromagnetic inversion method integrating multi-field response and intelligent optimization in this embodiment exhibit continuous low-resistivity anomalies in the depth range of 100m to 400m, with anomaly resistivity range of 10. Up to 25 The data is largely consistent with borehole exploration data. Conventional detection methods produce blurry deep images, while the transient electromagnetic inversion method in this embodiment, which integrates multi-field response and intelligent optimization, relies on magnetic induction response data to clearly identify the boundary of the ore body at depths of 550m to 800m, thus compensating for the blind spots in deep areas identified by traditional detection methods. Furthermore, compared to conventional two-dimensional layered inversion algorithms, the transient electromagnetic inversion method in this embodiment, which integrates multi-field response and intelligent optimization, significantly improves ore body positioning, increasing the effective detection depth by 300m.
[0114] This application embodiment also provides a computing device, including: The memory is configured to store instructions; The processor is configured to retrieve instructions from memory and, when executing instructions, to implement the aforementioned transient electromagnetic inversion method that integrates multi-field response and intelligent optimization.
[0115] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and the aforementioned transient electromagnetic inversion method, which integrates multi-field responses and intelligent optimization, can be implemented by adjusting kernel parameters.
[0116] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.
[0117] This application also provides a machine-readable storage medium storing instructions that cause a machine to perform the wide beam calibration method described above.
[0118] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0119] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0120] 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.
[0121] 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.
[0122] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0123] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0124] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0125] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0126] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.
Claims
1. A transient electromagnetic inversion method integrating multi-field response and intelligent optimization, characterized in that, include: Acquire standardized preprocessed three-field response data of the target underground space, wherein the three-field response data includes electric field response data, magnetic induction time derivative response data, and magnetic induction response data; The target underground space is subjected to terrain-constrained adaptive meshing to construct an unstructured tetrahedral mesh for the target underground space; Based on the piecewise linear interpolation algorithm, the convolution integral of the step response and the convolution integral of the waveform derivative are calculated to correct the non-ideal transmission waveform and obtain the corrected transmission waveform. The convolution integral of the step response and the convolution integral of the waveform derivative are obtained by decomposing the electromagnetic field response data, which includes electric field response data and magnetic induction response data. Using the corrected transmitted waveform as the input for time-domain discrete recursion, the three-field response data are calculated to obtain the forward response vector; Based on the stratum depth of the target underground space and the forward response vector, an adaptive sensitivity weighting matrix for the three response data is constructed. The adaptive sensitivity weighting matrix includes weight coefficients for the three responses, and the weight coefficients are dynamically adjusted according to the stratum depth of the target underground space. Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed, and the optimal resistivity model is obtained through a hybrid intelligent optimization algorithm, wherein the hybrid intelligent optimization algorithm is a hybrid optimization strategy combining the adaptive particle swarm optimization algorithm and the L-BFGS algorithm. The forward response vector and the terrain parameters of the unstructured tetrahedral mesh are input into a preset depth residual network. Error compensation and correction are completed through the output of the preset depth residual network to obtain the inversion result. The preset depth residual network is a convolutional structure based on a preset number of skip connections. It is trained with three-field forward response data, measured observation data, terrain parameters and exploration instrument calibration parameters as inputs and the system error compensation amount as output.
2. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 1, characterized in that, The process of performing terrain-constrained adaptive mesh generation on the target underground space to construct an unstructured tetrahedral mesh for the target underground space includes: Based on the digital elevation model of the work area, a surface model is constructed. Generate a triangular surface grid based on the surface model; Based on the geological data, borehole data and electromagnetic field sensitivity pre-calculation results of the target underground space, the definition of the grid densification area is obtained, wherein the grid densification area includes the terrain undulation area, the abnormal geological body area and the high sensitivity detection area. The Delaunay tetrahedral partitioning algorithm is used, combined with the triangular surface grid and the definition of the grid densification area, to perform tetrahedral partitioning of the target underground space and construct an unstructured tetrahedral grid of the target underground space.
3. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 1, characterized in that, The piecewise linear interpolation algorithm calculates the convolution integral of the step response and the convolution integral of the waveform derivative to correct the non-ideal transmission waveform, resulting in the corrected transmission waveform, including: Under the quasi-static electromagnetic assumption, transient electromagnetic detection control equations are constructed, wherein the transient electromagnetic detection control equations include electric field diffusion control equations; By introducing the unit step electric field response and the unit step magnetic induction response, the electromagnetic field response data corresponding to the non-ideal emission waveform is decomposed into the convolution integral of the step response and the convolution integral of the waveform derivative. The convolution integral of the step response and the convolution integral of the waveform derivative are calculated using a piecewise linear interpolation algorithm to correct the non-ideal transmission waveform and obtain the corrected transmission waveform.
4. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 3, characterized in that, The transient electromagnetic detection control equations also include Faraday's law of electromagnetic induction equations; The step of using the corrected transmitted waveform as input for time-domain discrete recursion to calculate the three-field response data and obtain the forward response vector includes: The second-order backward Euler method is used to discretize the electric field diffusion control equation in time, dividing the time domain into discrete time steps, and obtaining the recursive calculation formulas for the electric field and magnetic induction. Based on the discrete electric field diffusion control equation of the edge finite element method, and combined with the recursive calculation formula of the electric field and magnetic induction, the electric field response data is calculated to obtain the spatiotemporal distribution of the electric field response. Based on the Faraday law of electromagnetic induction equation, the magnetic induction time derivative response data is calculated to obtain the spatiotemporal distribution of the magnetic induction time derivative response; Based on the recursive calculation formulas for electric field and magnetic induction and the electric field response at the actual measurement point, the magnetic induction response data is calculated to obtain the spatiotemporal distribution of the magnetic induction response. By integrating the spatiotemporal distribution of the electric field response, the spatiotemporal distribution of the magnetic induction time derivative response, and the spatiotemporal distribution of the magnetic induction response, a forward response vector is obtained.
5. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 1, characterized in that, The process of constructing an adaptive sensitivity weighting matrix for the three-field response data based on the stratum depth of the target underground space and the forward-modeled response vector includes: When the depth of the target underground space is greater than the preset depth, a Jacobian matrix of the sensitivity of the three-field response to the resistivity of the underground medium is constructed based on the resistivity vector of the underground medium and the forward response vector. The sensitivity Jacobian matrix is calculated using the adjoint method to obtain the sensitivity gradient. Based on the sensitivity gradient, the spatial sensitivity distribution of the three-field response is calculated; Based on the stratum depth of the target underground space and the spatial sensitivity distribution, an adaptive sensitivity weighting matrix for the three-field response data is constructed using a diagonal matrix generating function.
6. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 1, characterized in that, The process of constructing a multi-field joint inversion objective function based on the adaptive sensitivity weighting matrix and obtaining the optimal resistivity model through a hybrid intelligent optimization algorithm includes: 1) Based on the adaptive sensitivity weighting matrix, a multi-field joint inversion objective function is constructed that takes into account both data fitting accuracy and model balance constraints. The multi-field joint inversion objective function includes a data fitting term and a model regularization term. The data fitting term adopts the weighted least squares criterion and incorporates the adaptive sensitivity weighting matrix and the data covariance matrix. The model regularization term adopts the model roughness minimization criterion. 2) Construct an initial population consisting of multiple particles, where each particle corresponds to a resistivity model; 3) An adaptive particle swarm optimization algorithm is used to perform a global search, updating the velocity and inertia weights of particles in the initial swarm to obtain the globally optimal model; 4) Use the globally optimal model as the initial model input for the L-BFGS algorithm; 5) Calculate the gradient information of the multi-field joint inversion objective function with respect to the model parameters, and use the Wolfe criterion to perform a one-dimensional line search to determine the optimal iteration step size; 6) Based on the gradient information and the optimal iteration step size, iterate the parameters of the global optimal model to update the global optimal model; 7) If the multi-field joint inversion objective function fails to converge or reaches the maximum number of iterations, the updated global optimal model is used as the initial model input for the L-BFGS algorithm, and steps 4) and 5) are repeated until the multi-field joint inversion objective function converges or reaches the maximum number of iterations to obtain the optimal resistivity model. In this case, the weight coefficients of the three field responses are dynamically adjusted based on the data residuals of the three field responses for a preset number of iterations.
7. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 1, characterized in that, The preset depth residual network is obtained through the following steps: Construct an initial deep residual network, wherein the initial deep residual network includes an input layer, a residual feature extraction layer and an output layer connected in sequence, and the residual feature extraction layer has a preset number of skip-connected convolutional structures; The loss function of the deep residual network adopts the mean square error criterion, and the initial deep residual network is trained using the Adam optimizer to obtain the preset deep residual network; The process involves inputting the forward response vector and the terrain parameters of the unstructured tetrahedral mesh into a preset depth residual network, and then using the output of the preset depth residual network to complete error compensation and correction to obtain the inversion result, including: Obtain the initial inversion model response corresponding to the forward response vector; The initial inversion model response, the terrain parameters of the unstructured tetrahedral mesh, the measured observation data of the target underground space, and the exploration instrument parameters are input into a preset residual network. Error compensation and correction are completed through the output of the preset residual network to obtain the inversion result.
8. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 1, characterized in that, The method further includes: Using the initial inversion results as the data source, a three-dimensional subsurface resistivity model is output. The three-dimensional subsurface resistivity model includes the grid node parameters of the unstructured tetrahedral mesh, anomalous geological body parameters, imaging quality parameters, and visualization data. The visualization data includes three-dimensional resistivity volume data, horizontal slices of the target subsurface space, and vertical profiles of the target subsurface space. The underground three-dimensional resistivity model was verified using three-field response simulation data, and the accuracy verification results were obtained. The underground three-dimensional resistivity model was verified by using measured data from the target underground space, and the results of the engineering practicality verification were obtained.
9. The transient electromagnetic inversion method integrating multi-field response and intelligent optimization according to claim 1, characterized in that, The acquisition of standardized, preprocessed three-field response data of the target underground space includes: The initial three-field response data of the target underground space were acquired using a synchronous acquisition mode. The initial three-field response data are standardized and preprocessed to obtain preprocessed three-field response data.
10. A computing device, characterized in that, include: The memory is configured to store instructions; The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the transient electromagnetic inversion method that integrates multi-field response and intelligent optimization according to any one of claims 1 to 9.
