Transient electromagnetic induced polarization response optimization multi-parameter inversion method of combined decoupling characterization learning network

Abstract

The invention discloses a transient electromagnetic induced polarization response optimization multi-parameter inversion method of a joint decoupling characterization learning network. The method comprises the following steps: constructing transient electromagnetic induced polarization response data; building a decoupling deep learning network framework on the basis of a Mama architecture model; the decoupling deep learning network framework comprises an encoder, an induced polarization decoder and a resistivity decoder; the encoder decouples the transient electromagnetic induced polarization response data and extracts a zero-frequency resistivity factor and an induced polarization factor; the induced polarization factor is transmitted to an induced polarization decoder, and the zero-frequency resistivity factor is transmitted to a resistivity decoder; decoding the induced polarization factor by using an induced polarization decoder, decoding the zero-frequency resistivity factor by using a resistivity decoder, and generating a zero-frequency resistivity model and a charging rate parameter model; and taking the zero-frequency resistivity model and the charging rate parameter model as initial inversion models, performing iterative fitting on transient electromagnetic induced polarization response data by adopting a gradient inversion algorithm, and obtaining a final multi-parameter inversion model.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, and in particular to a multi-parameter inversion method for optimizing transient electromagnetic induced response using a joint decoupled characterization learning network. Background Technology

[0002] With the deepening of the national strategic action for breakthroughs in mineral exploration, transient electromagnetic methods (TEMs), due to their high volume sensitivity to well-conducting geological bodies, are now widely used in underground metal mineral exploration. However, in actual exploration, metal minerals are often associated with sulfide minerals, forming metal sulfides. These minerals often possess high excitation polarization characteristics, and the resulting induced polarization effect affects the transient electromagnetic response. In actual measurements, a transient electromagnetic induced polarization response that decays rapidly in the middle stage and exhibits negative anomalies in the late stage is observed. Currently, conventional single-parameter resistivity methods cannot reasonably explain and fit the induced polarization response, greatly limiting the reliability of the obtained inversion results.

[0003] With the increasing duration of effective observation data, induced polarization (IP) responses are being observed more frequently in actual transient electromagnetic (TEM) exploration, leading to a surge in research on the inversion and extraction of IPI responses from transient electromagnetic systems. Existing technologies, such as those presented by Shen Di et al. ("Research on One-Dimensional Forward and Inverse Methods of Ground-to-Air Transient Electromagnetic Systems Considering Excitation Polarization Effects"), Man Kaifeng ("Research on Forward and Inverse Retrieval of Excitation Polarization Parameters in Time-Domain Airborne Electromagnetic Systems"), and Wang Shaojie ("One-Dimensional Inversion and Application of Electrical Source Transient Electromagnetic Systems"), all simulate transient IPI responses by introducing quantitative IPI models and propose their own multi-parameter regularized inversion methods to simultaneously extract zero-frequency resistivity and charge rate. However, the coupling relationships between inversion parameters, significant sensitivity differences, and the inherent strong ill-posedness of multi-parameter inversion methods severely limit the application effectiveness of traditional gradient inversion algorithms.

[0004] For example, the patent publication number CN118534551A, entitled "A Three-Dimensional Airborne Transient Electromagnetic Excitation Parameter Differentiation Constraint Inversion Method," includes: constructing a data difference model based on the correlation coefficient of the induced polarization parameters, and subsequently constructing model constraint terms in the inversion process based on this model to improve the reliability of the multi-parameter inversion results. Another example is the patent publication number CN120068627A, entitled "A Self-Supervised Inversion Method Based on Deep Learning for Transient Electromagnetic Methods," which includes: generating different resistivity models and corresponding electromagnetic response data as a training set through numerical simulation methods, updating the parameters of the inversion network, and enabling the network to recognize the input data and output the corresponding resistivity value.

[0005] The aforementioned techniques suffer from the following problems: Traditional gradient inversion methods all aim to improve the accuracy of inversion results by imposing various constraints on the inversion model, focusing on improving regularization constraints. However, a key manifestation of the strong ill-posedness of inversion is that the reliability of the inversion results heavily depends on the rationality of the initial model construction, which is even more pronounced in multi-parameter inversions. An inappropriate initial model selection, even with regularization constraints, can easily lead to local extrema in the inversion process, producing multi-parameter results that do not reflect reality. Furthermore, when using deep learning methods to directly extract data from transient electromagnetic data, the correctness of the inversion results largely depends on the coverage of the dataset. For complex electromagnetic inversion problems, generalization ability remains a highly controversial issue.

[0006] Therefore, there is an urgent need to propose a logically clear, simple, accurate and reliable multi-parameter inversion method for optimizing transient electromagnetic induced response using a joint decoupled representation learning network. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a multi-parameter inversion method for optimizing transient electromagnetic induced response using a jointly decoupled representation learning network. The technical solution adopted by the present invention is as follows: A multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network includes the following steps: Transient electromagnetic induced response data were constructed based on the geoelectric model; A decoupled deep learning network framework is built based on the Mamba architecture model. The decoupled deep learning network framework includes an encoder, an induced polarization decoder, and a resistivity decoder. The encoder decouples the transient electromagnetic induced polarization response data and extracts the zero-frequency resistivity factor and the excitation polarization factor. The excitation polarization factor is passed to the induced polarization decoder, and the zero-frequency resistivity factor is passed to the resistivity decoder. The excitation polarization factor is decoded using an induced polarization decoder, and the zero-frequency resistivity factor is decoded using a resistivity decoder to generate a zero-frequency resistivity model and a charge rate parameter model. The zero-frequency resistivity model and the charge rate parameter model are used as the initial inversion models. The gradient inversion algorithm is used to iteratively fit the transient electromagnetic induced response data to obtain the final multi-parameter inversion model.

[0008] Compared with the prior art, the present invention has the following beneficial effects: This invention employs a decoupled deep learning network framework based on the Mamba architecture model, which enables efficient processing of transient electromagnetic time series response data.

[0009] This invention is based on the characteristic that transient electromagnetic induced response is formed by the coupling of induction response and polarization response. It cleverly utilizes the feature that building a network architecture is conducive to decoupling learning, and realizes the information encoding and decoupling of transient electromagnetic induced response to generate zero-frequency resistivity factor and excitation polarization factor. These are then read by the subsequent decoding module and used to generate zero-frequency resistivity and charge rate parameter models in layers.

[0010] This invention utilizes an induced polarization decoder to decode the excitation polarization factor and a resistivity decoder to decode the zero-frequency resistivity factor, generating a zero-frequency resistivity model and a charge rate parameter model. Using the zero-frequency resistivity model and charge rate parameter model as the initial inversion model, a gradient inversion algorithm is employed to iteratively fit the transient electromagnetic induced polarization response data. This invention leverages deep learning methods to quickly generate a rationalized initial model, and then uses the gradient inversion method for iterative data fitting, significantly improving the ill-posedness of multi-parameter inversion and enhancing the reliability of the multi-parameter inversion results.

[0011] In summary, this invention has the advantages of clear and simple logic, accuracy and reliability, and has high practical and promotional value in the field of geophysical exploration technology. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of protection. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the decoupled deep learning network framework of the present invention.

[0014] Figure 2 This is a schematic diagram of the network architecture of the encoder of the present invention.

[0015] Figure 3 This is a schematic diagram illustrating the influence of induced polarization response data and the selection of the initial inversion model in this invention.

[0016] Figure 4 This is the inversion result of the multi-model data in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0018] In this embodiment, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0019] The terms "first" and "second," etc., used in the specification and claims of this embodiment are used to distinguish different objects, not to describe a specific order of objects. For example, "first target object" and "second target object," etc., are used to distinguish different target objects, not to describe a specific order of target objects.

[0020] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0021] In the description of the embodiments in this application, unless otherwise stated, "multiple" means two or more. For example, multiple processing units means two or more processing units; multiple systems means two or more systems.

[0022] like Figures 1 to 4 As shown, this embodiment provides a multi-parameter inversion method for optimizing transient electromagnetic induced response using a jointly decoupled representation learning network, which includes the following steps: The first step leverages the advantages of the Mamba network architecture in efficiently processing ultra-long-term time-series data and facilitating decoupled learning. Using Mamba as the basic network unit, a decoupled deep learning network framework with decoupled encoding and corresponding decoding is constructed in layers. The overall structure involves inputting the induced polarization response, which is then decoupled and separated by an encoder, and finally, corresponding decoders are used to obtain multi-parameter models.

[0023] Here, the decoupled deep learning network framework includes an encoder, an induced polarization decoder, and a resistivity decoder. The encoder decouples the transient electromagnetic induced polarization response data and extracts the zero-frequency resistivity factor and the excitation polarization factor. The excitation polarization factor is then passed to the induced polarization decoder, and the zero-frequency resistivity factor is passed to the resistivity decoder. Decoding operations are performed within both the induced polarization decoder and the resistivity decoder.

[0024] In this embodiment, the encoder, induced polarization decoder, and resistivity decoder have identical structures. The encoder includes an input layer, an embedding layer, a Mamba layer, a first linear layer, and an output layer connected in sequence. The encoder's input layer acquires transient electromagnetic induced polarization response data and performs a one-dimensional convolution operation using the encoder's embedding layer to obtain first embedded data. The encoder's Mamba layer performs state-space processing on the first embedded data to obtain first data. The encoder's first linear layer performs fully connected processing on the first data to obtain dual-channel data, which is then split and the zero-frequency resistivity factor and excitation polarization factor are output using the encoder's output layer. The Mamba layer includes a normalization layer, a second linear layer, a third linear layer, a backward state layer, a forward state layer, and a fourth linear layer. In the encoder, the embedding data is normalized using the normalization layer, and linearized using the second and third linear layers to obtain hidden states Z and X. The backward and forward state layers then perform state-space selection on hidden state X to obtain forward-processed data. and back-processing data The hidden state Z is activated using the ReLU activation function to obtain the weight data. Use weighted data Forward processed data respectively and back-processing data Perform matrix addition to obtain the data after state space selection. After selecting data in the state space using the fourth linear layer Perform fully connected processing and add it to the embedded data after normalization layer processing to obtain the first data.

[0025] In addition, in the induced polarization decoder, the input layer of the induced polarization decoder obtains the excitation polarization factor, and performs a one-dimensional convolution operation using the embedding layer of the induced polarization decoder to obtain the second embedded data. The Mamba layer of the induced polarization decoder processes the state space of the second embedded data to obtain the data after the first state space selection. The first linear layer of the induced polarization decoder performs fully connected processing on the data after the first state space selection to obtain single-channel data, and outputs the charging rate parameter model using the output layer of the induced polarization decoder.

[0026] In the resistivity decoder, the input layer of the resistivity decoder obtains the zero-frequency resistivity factor, and the embedding layer of the resistivity decoder performs a one-dimensional convolution operation to obtain the third embedded data. The Mamba layer of the induced polarization decoder performs state space processing on the third embedded data to obtain the second state space selected data. The first linear layer of the induced polarization decoder performs fully connected processing on the second state space selected data to obtain single-channel data, and the output layer of the induced polarization decoder outputs the zero-frequency resistivity model.

[0027] In this embodiment, the forward state layer selects the state space of the hidden state X from the previous order, and the backward state layer selects the state space of the hidden state X from the reverse order, the expression of which is: in, Represent the state space at time t; Represent the state space at time t-1; This represents the output of the state space at time t; A represents the input to the state space at time t; A represents the state space at time t-1. The corresponding learnable weight parameters; B represents the input to the state space at time t. The corresponding learnable weight parameters; C represents the state space at time t. The corresponding learnable weight parameters; D represents the input to the state space at time t. The corresponding second learnable weight parameter.

[0028] The second step, when considering the transient electromagnetic induced response, requires the introduction of a multi-parameter Debye complex resistivity model to describe the induced polarization properties of the subsurface medium. Its expression is as follows: in, Represents complex resistivity; Represents zero-frequency resistivity; This indicates the charging rate, typically ranging from 0 to 0.98. Represents the time constant, typically ranging from 1 to 2. to The second parameter is set to a constant value here because of its low sensitivity. Second.

[0029] Multiple geoelectric models were designed to generate a large amount of transient electromagnetic induced response data as a training set for network training. The response of the transient electromagnetic magnetic field with respect to the time derivative was studied under a one-dimensional layered model. Calculation formula: Where I represents the transmitting current value; Indicates the radius of the transmitting coil; Represents the observed time series; Indicates the imaginary part; Represents the reflection coefficient related to the distribution of underground electrical properties; Represents the integral variable; Represents the first-order Bessel function; Indicates the time-frequency conversion sampling interval; represents the time-frequency transform filter coefficients; n represents the number of time-frequency transform filter coefficients.

[0030] The third step involves extracting the data from the early half of the time channel of the generated large amount of transient electromagnetic induced response training data and using it as input to the encoder for decoupling, thereby simultaneously generating the zero-frequency resistivity factor and the excitation polarization factor.

[0031] The fourth step involves using an induced polarization decoder to decode the excitation polarization factor and a resistivity decoder to decode the zero-frequency resistivity factor, thereby generating a zero-frequency resistivity model and a charge rate parameter model.

[0032] The fifth step involves using the zero-frequency resistivity model and the charge rate parameter model as the initial inversion models, and employing a gradient inversion algorithm to iteratively fit the transient electromagnetic induced response data to obtain the final multi-parameter inversion model. At the k+1th iteration, we have: ; Here, due to the model parameters in the (k+1)th iteration... The formula is quite long and not convenient to show, so here we use... and To split it. Among them, and It has no practical meaning. In the above formula, where, This represents the model parameters for the (k+1)th iteration; Indicates the parameters of the k-th model; Represents a data weighting matrix; Let T denote the Jacobian matrix of the k-th iteration; T denotes the matrix transpose operation. R represents the model smoothness weight matrix; R represents the model vertical smoothness factor. This represents the damping factor in the k-th iteration; Represents the weighting factors that control the data fitting terms and model constraint terms; Represents the identity matrix; Represents observation data; This indicates the forward modeling calculation.

[0033] Case 1 like Figure 3 As shown, a simple three-layer model is designed, with the middle layer being a polarization layer, the surface resistivity being 100Ωm and the thickness being 20m, the zero-frequency resistivity of the middle polarization layer being 10Ωm, the charging rate being 0.5, the time constant being 0.001 seconds, and the bottom layer resistivity being 200Ωm. Figure 3(a) shows the transient electromagnetic induced response generated under this geoelectric model. A significant sign reversal is observed in the later data, exhibiting a large number of negative responses. Subsequently, 3% random noise was added to this data to generate measured data. Inversion calculations were then performed using a uniform layered half-space model with different parameter settings. The initial models included conventional resistivity inversion with a resistivity value of 100 Ωm. Seven initial model settings considering the induced electrostatic effect were used, with combinations of zero-frequency resistivity and charge rate of (100 Ωm, 0.05), (100 Ωm, 0.1), (100 Ωm, 0.2), (500 Ωm, 0.05), (500 Ωm, 0.1), (500 Ωm, 0.2), and (1000 Ωm, 0.05). The fitting results and inversion results of the inversion data from different initial models are shown below. Figure 3 As shown in (b)-(d), it is evident that using a single resistivity model for inversion calculations cannot fit the induced polarization (IP) response data, and the inversion results deviate significantly from the actual model. When using an IPA model to consider IPA inversion, the inversion data can effectively fit the IPA response data. However, it can be seen that under different initial IPA models, the fitting of the inversion data and the multi-parameter inversion models show significant differences. In some cases, it cannot effectively reflect the distribution of underground zero-frequency resistivity and charge rate. Therefore, the rationality of the initial model selection has a significant impact on the reliability of the inversion results. Here, it should be noted that... Figure 3 (a) is a graph of the transient electromagnetic induced response generated under the geoelectric model; Figure 3 (b) shows the data fitting; Figure 3 (c) shows the zero-frequency resistivity inversion result; Figure 3 In the middle (d), the charging rate inversion result is shown.

[0034] Case 2 in, Figure 4 (a) shows a comparison of the zero-frequency resistivity inversion results of the low-resistivity, high-polarization model; Figure 4 Figure 4(b) shows a comparison of the charging rate inversion results for the low-resistivity, high-polarization model; Figure 4(c) shows a comparison of the zero-frequency resistivity inversion results for the high-resistivity, high-polarization model. Figure 4 (d) shows a comparison of the charging rate inversion results of the high-resistivity, high-polarization model. Here, two typical three-layer geoelectric models are designed to verify the inversion strategy. The first is the low-resistivity, high-polarization model, suitable for scenarios such as massive metal sulfide deposits. The surface layer is a non-polarized layer with a resistivity of 100 Ωm, the middle layer is a low-resistivity polarized layer with a zero-frequency resistivity of 10 Ωm and a charging rate of 0.5, and the background is a bottom layer with a resistivity of 200 Ωm. The second model is the high-resistivity, high-polarization model, suitable for scenarios such as disseminated metal sulfides. The surface layer is a non-polarized layer with a resistivity of 100 Ωm, the middle layer is a high-resistivity polarized layer with a zero-frequency resistivity of 500 Ωm and a charging rate of 0.5, and the background is a bottom layer with a resistivity of 200 Ωm. The inversion results of this strategy are compared with those of the traditional inversion model. Figure 4 As shown, for the multi-parameter results of the first model, such as Figure 4 In (a) and (b), we first observe that the inversion results based on the uniform half-space initial model show that the zero-frequency resistivity inversion results indicate the existence of an intermediate low-resistivity layer underground. However, the charge rate results deviate significantly from the actual situation, revealing a shallow, falsely highly polarized layer. Based on early response data, the initial model generated by the decoupled deep learning network can preliminarily reflect the geoelectric distribution underground. Subsequently, this model is used as the initial model to iteratively fit the overall data using the gradient inversion method. It can be seen that the inversion results match the actual model well, especially in accurately characterizing the location of the low-resistivity, highly polarized layer. Regarding the multi-parameter results of the second model, as follows... Figure 4 In (c) and (d), under high resistivity and high polarization conditions, the sensitivity of the response to the polarimetric body is greatly reduced. The inversion results using the uniform half-space initial model, while reflecting the distribution of low and high resistivity underground, deviate significantly from the actual model, especially the charge rate results, which show a polarized half-space pattern, showing a large discrepancy with the actual model. The initial model generated by the decoupled deep learning network initially shows the distribution of an intermediate high-resistivity layer underground, with a polarization distribution appearing in the intermediate layer. Using this as the initial model for gradient inversion, the inversion results effectively reflect the underground resistivity distribution, and the values ​​match the actual model well. Notably, the charge rate results effectively reflect the distribution of a highly polarized layer, significantly improving the accuracy of multi-parameter inversion.

[0035] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any changes made based on the design principles of the present invention, or any non-creative modifications made thereon, shall fall within the scope of protection of the present invention.

Claims

1. A multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network, characterized in that, Includes the following steps: Transient electromagnetic induced response data were constructed based on the geoelectric model; A decoupled deep learning network framework is built based on the Mamba architecture model. The decoupled deep learning network framework includes an encoder, an induced polarization decoder, and a resistivity decoder. The encoder decouples the transient electromagnetic induced polarization response data and extracts the zero-frequency resistivity factor and the excitation polarization factor. The excitation polarization factor is passed to the induced polarization decoder, and the zero-frequency resistivity factor is passed to the resistivity decoder. The excitation polarization factor is decoded using an induced polarization decoder, and the zero-frequency resistivity factor is decoded using a resistivity decoder to generate a zero-frequency resistivity model and a charge rate parameter model. The zero-frequency resistivity model and the charge rate parameter model are used as the initial inversion models. The gradient inversion algorithm is used to iteratively fit the transient electromagnetic induced response data to obtain the final multi-parameter inversion model.

2. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 1, characterized in that, The encoder, induced polarization decoder, and resistivity decoder have the same structure, and the encoder includes an input layer, an embedding layer, a Mamba layer, a first linear layer, and an output layer connected in sequence. The input layer of the encoder acquires transient electromagnetic induced polarization response data and performs a one-dimensional convolution operation using the embedding layer of the encoder to obtain first embedded data. The Mamba layer of the encoder performs state-space processing on the first embedded data to obtain first data. The first linear layer of the encoder performs fully connected processing on the first data to obtain dual-channel data, splits it, and outputs the zero-frequency resistivity factor and excitation polarization factor using the output layer of the encoder.

3. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 2, characterized in that, The input layer of the induced polarization decoder obtains the excitation polarization factor and performs a one-dimensional convolution operation using the embedding layer of the induced polarization decoder to obtain the second embedded data. The Mamba layer of the induced polarization decoder performs state space processing on the second embedded data to obtain the data after the first state space selection. The first linear layer of the induced polarization decoder performs fully connected processing on the data after the first state space selection to obtain single-channel data, and outputs the charge rate parameter model using the output layer of the induced polarization decoder.

4. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 2, characterized in that, The input layer of the resistivity decoder obtains the zero-frequency resistivity factor and performs a one-dimensional convolution operation using the embedding layer of the resistivity decoder to obtain the third embedded data. The Mamba layer of the induced polarization decoder performs state space processing on the third embedded data to obtain the data after the second state space selection. The first linear layer of the induced polarization decoder performs fully connected processing on the data after the second state space selection to obtain single-channel data, and outputs the zero-frequency resistivity model using the output layer of the induced polarization decoder.

5. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 2, 3, or 4, characterized in that, The Mamba layer includes a normalization layer, a second linear layer, a third linear layer, a backward state layer, a forward state layer, and a fourth linear layer. In the encoder, the normalization layer normalizes the embedded data, and the second and third linear layers perform linear processing to obtain hidden state Z and hidden state X. The backward and forward state layers then perform spatial state selection on hidden state X to obtain forward-processed data. and back-processing data The hidden state Z is activated using the ReLU activation function to obtain the weight data. Use weighted data Forward processed data respectively and back-processing data Perform matrix addition to obtain the data after state space selection. After selecting data in the state space using the fourth linear layer The first data is obtained by performing fully connected processing and adding it to the embedded data after normalization layer processing.

6. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 5, characterized in that, The forward state layer selects the state space of the hidden state X from the previous order, and the backward state layer selects the state space of the hidden state X from the reverse order, the expression of which is: ; ； in, Represent the state space at time t; Represent the state space at time t-1; This represents the output of the state space at time t; A represents the input to the state space at time t; A represents the state space at time t-1. The corresponding learnable weight parameters; B represents the input to the state space at time t. The corresponding first learnable weight parameter; C represents the state space at time t. The corresponding learnable weight parameters; D represents the input to the state space at time t. The corresponding second learnable weight parameter.

7. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 1, characterized in that, In constructing the transient electromagnetic induced response data, a multi-parameter Debye complex resistivity model is introduced, the expression of which is: in, Represents complex resistivity; Represents zero-frequency resistivity; Indicates the charging rate; This represents the time constant.

8. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 7, characterized in that, Also includes: The response of transient electromagnetic magnetic field with respect to time derivative under a one-dimensional layered geoelectric model. The expression is: Where I represents the transmitting current value; Indicates the radius of the transmitting coil; Represents the observed time series; Indicates the imaginary part; Represents the reflection coefficient related to the distribution of underground electrical properties; Represents the integral variable; Represents the first-order Bessel function; Indicates the time-frequency conversion sampling interval; represents the time-frequency transform filter coefficients; n represents the number of time-frequency transform filter coefficients.

9. The multi-parameter inversion method for transient electromagnetic induced response optimization using a joint decoupled representation learning network according to claim 7, characterized in that, The zero-frequency resistivity model and charge rate parameter model are used as the initial inversion models. A gradient inversion algorithm is then used to iteratively fit the transient electromagnetic induced response data to obtain the final multi-parameter inversion model. At the k+1th iteration, the following holds: ; ; ; ;in, This represents the model parameters for the (k+1)th iteration; Indicates the parameters of the k-th model; Represents a data weighting matrix; Let T denote the Jacobian matrix of the k-th iteration; T denotes the matrix transpose operation. R represents the model smoothness weight matrix; R represents the model vertical smoothness factor. This represents the damping factor in the k-th iteration; Represents the weighting factors that control the data fitting terms and model constraint terms; Represents the identity matrix; Represents observation data; This indicates the forward modeling calculation.

Images