Joint inversion method based on multi-geophysical data feature fusion and dual joint strategy driving

Abstract

The invention discloses a multi-geophysical data feature fusion and dual joint strategy driven joint inversion method. The method comprises the following steps: acquiring different types of geophysical response data; constructing a first joint inversion network based on an encoder-decoder structure, and extracting and fusing features from the geophysical response data to obtain a preliminary physical property model; inputting the preliminary physical property model into a second joint inversion network, and performing feature interaction fusion on the preliminary physical property model to obtain an optimized physical property model; simulation response data are generated based on the optimized physical property model, and parameters of the first joint inversion network are finely adjusted through transfer learning; and performing inversion on the actually measured geophysical response data by using the first joint inversion network after fine adjustment to obtain a final inversion result. According to the method, the precision, the stability and the boundary recognition capability of an inversion result are remarkably improved, and meanwhile, the robustness of the model under different noise interferences is enhanced through a dual joint strategy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of cross between geophysics and information science, and particularly relates to a joint inversion method driven by multi-geophysical data feature fusion and double joint strategy. BACKGROUND

[0002] The joint inversion technology can break through the limitation of single inversion method, effectively reduce the multi-solution of inversion result, and improve the resolution and reliability of the underground three-dimensional geological model by constructing a collaborative interpretation framework of multi-geophysical data, which has important application value in energy exploration, geological disaster assessment and other fields.

[0003] The traditional joint inversion method mainly realizes data fusion by introducing petrophysical constraints or structural consistency constraints in the inversion objective function. Among them, the petrophysical constraints are based on known physical relationships (such as velocity-porosity, density-resistivity, etc.), and the structural consistency constraints suppress noise interference by assuming spatial continuity of the model. However, this kind of method is severely dependent on prior assumptions, and when the assumptions are not accurate or the data are incomplete, it is easy to lead to distortion or excessive smoothing of the inversion result, which is difficult to adapt to the solution demand of complex nonlinear problems.

[0004] With the development of deep learning technology, the joint inversion method based on data driving shows significant advantages. The existing joint inversion method based on deep learning mainly includes two types: one is simultaneous inversion based on deep learning constraints, which realizes end-to-end inversion mapping through multi-channel network architecture (such as U-Net), has efficient data fusion ability, but its interpretability is poor, and the completeness of the training data is required, which is easy to appear physical meaning distortion when the geological conditions are complex; the second is deep learning constraint inversion under the traditional iterative framework, which embeds the learned physical relationship as a constraint into the iterative process, although it has good robustness, but the computational efficiency is low, and the compatibility of network constraint and iterative framework is difficult to optimize, which is easy to affect the convergence performance.

[0005] In addition, the existing fusion strategy (such as early fusion, mid-fusion and late fusion) mainly depends on a single mode, and cannot fully tap the complex feature correlation and complementarity between different geophysical data, which limits the further improvement of inversion accuracy and robustness. Therefore, there is an urgent need for a joint inversion method that can deeply fuse multi-source features and effectively balance accuracy and efficiency. SUMMARY

[0006] To solve the above technical problems, the present application provides a joint inversion method driven by multi-geophysical data feature fusion and double joint strategy, which comprises: acquiring different types of geophysical response data; constructing a first joint inversion network based on an encoder-decoder structure to extract and fuse features from the geophysical response data to obtain a preliminary physical property model; inputting the preliminary physical property model into a second joint inversion network configured to learn the correlation constraints between different physical property models and to obtain an optimized physical property model by feature interaction fusion on the preliminary physical property model; generating simulated response data based on the optimized physical property model and fine-tuning parameters of the first joint inversion network through transfer learning; using the fine-tuned first joint inversion network to perform inversion on the measured geophysical response data to obtain a final inversion result.

[0007] Optionally, the step of constructing a first joint inversion network based on an encoder-decoder structure to extract and fuse features from the geophysical response data to obtain a preliminary physical property model includes the following steps: inputting different types of geophysical response data into the first joint inversion network; in the encoder stage of the first joint inversion network, extracting high-level features of each data through downsampling and fusing the high-level features at different levels to obtain multi-modal fusion features; in the decoder stage of the first joint inversion network, introducing the multi-modal fusion features through upsampling and combining a skip connection to reconstruct and output the preliminary physical property model.

[0008] Optionally, the step of extracting high-level features of each data through downsampling and fusing the high-level features at different levels to obtain multi-modal fusion features in the encoder stage of the first joint inversion network includes the following steps: after at least one downsampling layer of the encoder, splicing feature maps corresponding to different geophysical data modalities; performing convolution operation on the spliced feature maps to extract the multi-modal fusion features.

[0009] Optionally, the step of inputting the preliminary physical property model into a second joint inversion network to obtain an optimized physical property model by feature interaction fusion on the preliminary physical property model includes the following steps: constructing a second joint inversion network with both input and output being physical property models; inputting the preliminary physical property model into the second joint inversion network; learning the correlation constraints between different physical property models through the second joint inversion network to output the optimized physical property model.

[0010] Optionally, the step of generating simulation response data based on the optimized physical property model and fine-tuning the parameters of the first joint inversion network through transfer learning has the following specific process: Performing forward calculation on the optimized physical property model to generate corresponding simulation geophysical response data; Constructing a loss function according to the simulation geophysical response data and the measured geophysical response data; Fine-tuning the parameters of the first joint inversion network through back propagation based on the loss function.

[0011] Optionally, the method adopts a phased training strategy; In the first phase, the first joint inversion network is trained with geophysical response data as input and real physical property model as output; In the second phase, the second joint inversion network is trained with real physical property model as input and output.

[0012] The present application also proposes a joint inversion system driven by multi-geophysical data feature fusion and double joint strategy, which is used to implement the method, comprising: A data acquisition module for acquiring different types of geophysical response data; A first joint inversion module comprising a first joint inversion network based on an encoder-decoder structure, for extracting and fusing features from the geophysical response data to obtain a preliminary physical property model; A second joint inversion module comprising a second joint inversion network configured to learn the correlation constraints between different physical property models, for obtaining an optimized physical property model by feature interaction fusion on the preliminary physical property model; A network optimization module for generating simulation response data based on the optimized physical property model and fine-tuning the parameters of the first joint inversion network through transfer learning; A final inversion module for inverting the measured geophysical response data using the fine-tuned first joint inversion network to obtain a final inversion result.

[0013] The present application also proposes a computer device comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0014] The present application also proposes a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the method.

[0015] The present application also proposes a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the method.

[0016] Compared with the prior art, the present application has the following advantages and technical effects: The present application realizes multi-level and multi-stage geophysical data collaborative inversion by constructing a first joint inversion network based on an encoder-decoder structure to realize feature extraction and fusion of multi-source geophysical response data, further realizing feature interaction and fusion between physical property models through a second joint inversion network, and finally realizing parameter fine-tuning of the first network based on the optimized physical property model through transfer learning, effectively overcoming the defects of insufficient mining of complex feature relationships of the existing single fusion strategy, significantly improving the precision, stability and boundary recognition ability of the inversion result, and simultaneously enhancing the robustness of the model under different noise interference through the double joint strategy. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application illustrated in the drawings, and their description, are presented to add generic scope to this application. In the drawings: Figure 1 Schematic diagram of the multi-modal feature fusion method of the embodiment of the present application; Figure 2 Schematic diagram of the system structure of the embodiment of the present application; Figure 3 Schematic diagram of a single training data sample and its observation system of the embodiment of the present application, wherein Figs. (a) and (b) respectively show the observation system schematic diagrams of transient electromagnetic and magnetic methods; Figure 4 Schematic diagram of the multi-geophysical data feature fusion and double joint strategy driven joint inversion network structure of the embodiment of the present application; Figure 5 Schematic diagram of the loss decay in the training process of different deep learning inversion methods of the embodiment of the present application; Figure 6 Inversion comparison diagram of different joint inversions in the test Model1 of the embodiment of the present application, wherein (a) and (e) respectively show the slice diagram of Model1 at X=0m, (b) and (f) respectively show the separate deep learning inversion result diagram, (c) and (g) respectively show the joint inversion result diagram based on U-Net, and (d) and (h) respectively show the joint inversion result diagram; Figure 7 Second joint inversion result diagram of the embodiment of the present application, wherein Figs. (a) and (d) respectively show the profile diagram of the real model at X=0m, (b) and (e) respectively show the first joint inversion result diagram, and (c) and (f) respectively show the second joint inversion result diagram; Figure 8 FIGS. (a) and (e) respectively represent a true physical property model diagram; FIGS. (b) and (f) respectively show a result diagram of secondary joint inversion under the condition of 3% Gaussian random noise; (c) and (g) respectively show a result diagram of secondary joint inversion under the condition of 5% Gaussian random noise; (d) and (h) respectively show a result diagram of secondary joint inversion under the condition of 10% Gaussian random noise. DETAILED DESCRIPTION

[0018] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0019] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0020] As a frontier technology in the field of geophysics, joint inversion can break through the inherent bottleneck of traditional single inversion method by constructing a multi-geophysical data collaborative interpretation framework. By establishing a mathematical model with multiple parameter constraints, the advantages of different geophysical data can be fully utilized, and the advantages of different geophysical responses can be complementary. This data fusion strategy not only effectively reduces the multi-solution of the inversion result, but also significantly improves the resolution and reliability of the underground three-dimensional geological model through cross-validation mechanism, providing more scientifically valuable decision-making basis for energy exploration, geological disaster assessment and other fields.

[0021] Traditional joint inversion usually realizes multi-data fusion by introducing petrophysical constraints or structural consistency constraints in the inversion objective function. Among them, petrophysical constraints rely on known physical relationships such as velocity-porosity, density-resistivity, etc., to ensure that the inversion results conform to the actual geological laws; structural consistency constraints effectively suppress unreasonable fluctuations in the inversion results caused by noise by assuming model spatial continuity or smoothness. These constraints not only enhance the consistency between different data sources, but also effectively reduce the uncertainty and non-uniqueness of inversion, providing a more reliable solution for the interpretation of complex geological problems. However, this method relies too much on prior assumptions, and for complex nonlinear problems such as geophysical inversion, if the prior assumptions are not accurate or complete, it is easy to lead to inaccurate or over-smoothed inversion results, especially in the case of insufficient data or large noise.

[0022] With the continuous improvement of computing device performance, deep learning technology can effectively capture the complex correlation between multi-geophysical data due to its strong feature learning ability, opening up a new path for joint inversion of multi-geophysical data. Deep learning can automatically extract complex features from geophysical data due to its strong non-linear mapping ability, realizing efficient fusion of multi-source geophysical data. Currently, joint inversion based on deep learning can be divided into two categories according to the process: simultaneous inversion based on deep learning constraints and deep learning constraint inversion under the traditional iterative framework.

[0023] Simultaneous inversion based on deep learning constraints ingeniously uses the multi-channel input characteristics of the U-Net architecture to simultaneously import different types of geophysical data (such as seismic, electromagnetic, gravity, etc.) into the network. Through the internal feature interaction mechanism of the network, each type of data forms a natural mutual constraint in the feature extraction and mapping process, and finally directly outputs the inversion results of fused multi-source information, realizing the collaborative inversion goal of multi-data from the model level. This multi-channel architecture not only enhances the adaptability of the model to multi-source data, but also provides a flexible framework for mining the potential correlation between different data. Although this method has shown significant results in the joint inversion of seismic, electrical, gravity, and other multi-geophysical data, there are still two main problems: first, it lacks sufficient explainability, making it difficult to determine the specific contribution of each geophysical data to the final result; second, it is insufficient to mine the complex feature relationships between geophysical data, limiting the precision and robustness of the inversion results.

[0024] Deep learning constraint inversion under the traditional iterative framework uses deep learning as a "bridge" to first learn the internal constraint relationships between different geophysical data (such as the correlation of physical parameters and the consistency of spatial distribution) through a deep network, and then embed these learned constraints into the iterative framework of traditional inversion to assist traditional methods in optimizing the inversion direction at each iteration, ultimately generating an enhanced inversion result. This method retains the robustness of traditional iterative inversion, so its inversion results are more reliable in scenarios with strong noise interference or incomplete data, and the precision has also been improved to some extent; but since it still relies on the iterative process of traditional inversion, its computational efficiency is poorer than the first method which directly outputs results through the network end-to-end.

[0025] Although simultaneous inversion based on deep learning constraints achieves direct constraint and efficient fusion of data through the U-Net multi-channel structure, this method relies too much on the modeling ability of the network for complex geological scenarios. When there is a deviation between the training data and the actual geological conditions (such as not covering special structures or physical property anomalies), the inversion result is prone to physical meaning distortion. At the same time, this method lacks sufficient explainability, making it difficult to determine the specific contribution of each geophysical data to the final result.

[0026] Deep learning constrained inversion under the traditional iterative framework, although retaining the robustness of traditional iteration and improving the accuracy, is significantly less computationally efficient than the former due to the nested iterative process of separate inversion. Especially when dealing with large-scale three-dimensional data, the accumulation of iteration times will significantly increase the time cost. At the same time, the compatibility of the constraint relationship learned by the deep learning network with the traditional iterative framework needs to be finely adjusted. If the constraint strength or timing is not properly set, it may cause the problem of slow convergence speed or even local optimal solution in inversion.

[0027] In addition, one of the key challenges of deep learning-based multi-geophysical data joint inversion is how to effectively fuse the data features from different geophysical methods. These data usually have different resolutions, physical properties, and measurement scales, and direct fusion may lead to information loss or inaccurate results. Currently, multi-modal fusion methods in deep learning mainly include early fusion, mid-fusion, late fusion, and hybrid fusion strategies. Early fusion directly concatenates the original data; mid-fusion gradually fuses information during feature extraction; late fusion integrates results at the model output stage; while hybrid fusion combines multiple strategies to balance performance and complexity. However, most existing methods rely on a single fusion strategy and fail to fully capture and exploit the complex relationships and potential complementarity between different geophysical data.

[0028] Embodiment one To overcome the problems in the current multi-geophysical data feature fusion, the embodiment provides a multi-geophysical data feature fusion and dual joint strategy driven joint inversion method, As a specific implementation, the method comprises the following steps: acquiring different types of geophysical response data; building a first joint inversion network based on an encoder-decoder structure to extract and fuse features from the geophysical response data to obtain a preliminary physical property model; inputting the preliminary physical property model into a second joint inversion network, the second joint inversion network being configured to learn the correlation constraints between different physical property models and obtain an optimized physical property model by feature interaction fusion on the preliminary physical property model; generating simulated response data based on the optimized physical property model and fine-tuning the parameters of the first joint inversion network through transfer learning; using the fine-tuned first joint inversion network to invert the measured geophysical response data to obtain the final inversion result.

[0029] The process of acquiring different types of geophysical response data can be implemented as follows: The proposed joint inversion method is used for three-dimensional transient electromagnetic and three-dimensional magnetic joint inversion in this embodiment. In the construction of the training set, the background and the anomaly body are combined in the form of a combination, the anomaly body is set as a cube, and the number is randomly selected as 1 to 3. In order to ensure the consistency of the model structure, the anomaly body of the transient electromagnetic and magnetic method is strictly consistent in shape and spatial position. In terms of physical property setting: the background resistivity of the transient electromagnetic method is randomly selected from (150, 300, 450) Ω·m, the resistivity of the anomaly body is randomly selected from 10~100 Ω·m and 700~800 Ω·m, and the resistivity of the air layer is set to 1×10 8 Ω·m. The background magnetic permeability of the magnetic method is 0SI, the magnetic permeability of the anomaly body is randomly selected from 0.1~1SI, and the magnetic permeability of the air layer is set to 0SI. The observation system magnetic dip angle is 45°, the magnetic declination angle is 0°, and the background field strength is 50000nT.

[0030] The resistivity and magnetic susceptibility model is forward calculated by using the SimPEG open source code. The grid is set as a tensor grid with a size of 24×24×24. The grid size in the measurement area (600m×600m) is 50m×50m. Outside the area, the filling coefficient is expanded by 1.3 to reduce the boundary effect. The transient electromagnetic method uses a large fixed source loop device. The transmitting coil is 300m×300m, the current is 1A, the measurement point distance and the measurement line distance are both 20m, a total of 11×11 measurement points are arranged, and the measurement time is 1×10 -5 s to 1×10 -2 s, a total of 41 time channels. The magnetic observation system is basically the same as the transient electromagnetic method. The point distance is set to 10m, and a total of 21×21 measurement points are arranged. Figure 3 The distribution of a single training data sample and the observation system (projected to the bottom surface). Through the above training set generation method, 24400 resistivity and magnetic susceptibility models are finally generated for joint inversion testing, and the ratio of the training set to the validation set is 9:1.

[0031] The first joint inversion network based on the encoder-decoder structure can be implemented. The step of extracting and fusing features from the geophysical response data to obtain a preliminary physical property model includes the following specific processes: Different types of geophysical response data are input into the first joint inversion network. In the encoder stage of the first joint inversion network, high-level features of each data are extracted through downsampling, and the high-level features are fused at different levels to obtain multi-modal fusion features. In the decoder stage of the first joint inversion network, the multi-modal fusion features are introduced through upsampling and combined with the jump connection, and the preliminary physical property model is reconstructed and output.

[0032] Further, in the encoder stage of the first joint inversion network, the high-level features of each data are extracted by downsampling, and the high-level features are fused at different levels to obtain multi-modal fusion features, and the specific process includes: After at least one down-sampling layer of the encoder, the feature maps corresponding to different geophysical data modalities are spliced; and a convolution operation is performed on the spliced feature maps to extract the multi-modal fusion features.

[0033] Implementable, the preliminary physical property model is input into the second joint inversion network, and the preliminary physical property model is fused by feature interaction to obtain an optimized physical property model, and the specific process includes: A second joint inversion network is constructed, and the input and output of the second joint inversion network are both physical property models; the preliminary physical property model is input into the second joint inversion network; the second joint inversion network learns the correlation constraints between different physical property models, and outputs the optimized physical property model.

[0034] Implementable, based on the optimized physical property model, simulation response data is generated, and the parameters of the first joint inversion network are fine-tuned through transfer learning, and the specific process includes: Forward calculation is performed on the optimized physical property model to generate corresponding simulated geophysical response data; a loss function is constructed according to the simulated geophysical response data and the measured geophysical response data; based on the loss function, the parameters of the first joint inversion network are fine-tuned through back propagation.

[0035] Implementable, the method adopts a phased training strategy; In the first phase, the first joint inversion network is trained with geophysical response data as input and real physical property model as output; in the second phase, the second joint inversion network is trained with real physical property model as input and output.

[0036] As a specific embodiment, the method combines the feature fusion strategies of the up-sampling and down-sampling stages. As shown in Figure 1 In the down-sampling stage, this embodiment gradually extracts low-level to high-level fusion feature representations (plus circles) by layer-by-layer fusion of feature maps (solid circles) of different geophysical data modalities, so that multi-modal data can share information and collaborative learning at different levels. In the up-sampling stage, this embodiment recombines the fusion features with the up-sampled high-level features using skip connection, thereby retaining more spatial detail information. By explicitly introducing feature fusion at multiple stages, not only does the structure of multi-modal fusion become more transparent, but also the model's interpretability is enhanced. At the same time, the multi-level fusion of the up-sampling and down-sampling stages enables the model to capture the complex relationships between geophysical data more deeply, thereby significantly improving the accuracy and reliability of the inversion results.

[0037] Meanwhile, in order to more fully tap the potential feature relationship between multi-geophysical data, the embodiment also proposes a double joint inversion strategy. Feature extraction and fusion are realized in two aspects: on the one hand, feature fusion is performed between geophysical response data; on the other hand, feature interaction fusion is introduced between physical property models. As shown in Figure 2 the figure, first, in the first joint inversion, features are extracted and fused from multi-source geophysical response data by joint inversion network 1 (first joint inversion network), the preliminary joint inversion from response to model is completed, and the preliminary physical property model (multi-geophysical model 1) is generated. On this basis, the second joint inversion is carried out, the preliminary inversion result is input into joint inversion network 2 (i.e. second joint inversion network), and further joint constraints are imposed on the features between the physical property models, so as to obtain a more accurate inversion result (multi-geophysical model 2). In order to ensure that the result of the second inversion does not significantly increase the data response residual, the embodiment fine-tunes the parameters of joint inversion network 1 based on the output of the second inversion, generates an optimized network, and a new joint inversion network 1, and uses it to perform final inversion on the data to be measured, so as to obtain an inversion result with higher precision and stability.

[0038] Further, in order to more fully extract geophysical features and improve the inversion effect, the joint inversion network 1 proposed in the embodiment selects the UT-Net network as the basic framework. UT-Net effectively improves the ability of the network in capturing complex geophysical feature space relationships and long-distance dependencies by introducing the multi-head self-attention mechanism (Multi-Head Self Attention, MHSA) and multi-head cross-attention mechanism (Multi-Head Cross Attention, MHCA) of Transformer into the U-Net framework. The UT-Net network has shown superior performance in the inversion of three-dimensional transient electromagnetic, and can fully extract long-distance spatial features and local features between electromagnetic response data. In addition, the network also has good expansibility and is suitable for inversion tasks of other three-dimensional geophysical data. Therefore, based on this network architecture, the embodiment constructs a joint inversion network driven by multi-geophysical data feature fusion and double joint strategy.

[0039] The network structure is as shown in Figure 4As shown in the figure. Since different geophysical data have different resolutions, this embodiment introduces an embedding layer to extract features from different resolution geophysical data through convolution operation to maintain the consistency of the features. In the down-sampling process, this embodiment uses two sets of convolution modules for feature extraction, including convolution, layer normalization, ReLU activation function and max pooling layer. At the same time, in order to avoid the problem of gradient explosion encountered by the network in the training process, this embodiment introduces a residual connection operation. In the feature fusion (fusion) process in the down-sampling stage, this embodiment performs two convolution operations on the connected feature maps to fully extract the fusion information of different geophysical features and realize layer-by-layer fusion. Finally, MHSA is used to fully extract the down-sampled features. In the up-sampling process, MHCA is used to process the up-sampled features and the fusion features to gradually realize the joint inversion process of different geophysical data.

[0040] This embodiment adopts a phased training strategy combined with the idea of double joint inversion to gradually optimize the model performance, such as Figure 2 . Specifically, the training process is divided into two main stages: training of joint inversion network 1 and joint inversion network 2. In the first stage, the training of joint inversion network 1 takes the geophysical response data of transient electromagnetic and magnetic methods as input and outputs the corresponding resistivity and susceptibility inversion models. In the second stage, the input and output of joint inversion network 2 are real resistivity and susceptibility models, aiming to further explore the potential feature relationship between the physical property models. In order to ensure the stability and efficiency of the above network training, this embodiment adopts a simulated annealing learning rate strategy, with an initial learning rate of 1e-4, and uses an early stopping mechanism to avoid overfitting. When the loss function of the validation set no longer decreases within 30 consecutive iterations, the training will automatically stop. In addition, the batch size is set to 64, the maximum training rounds are 800, and the optimizer uses Adam. The ratio of training set, validation set and test set is 8:1:1.

[0041] When inverting the data to be measured, first input the data to be measured into the joint inversion network 1 to generate an initial inversion model, and pass it to the joint inversion network 2 for further optimization to generate an intermediate inversion model. To further improve the inversion accuracy, this embodiment constructs a new training set based on the response data of the intermediate inversion model, and combines the idea of transfer learning to fine-tune the parameters of the joint inversion network 1, thereby obtaining an optimized new joint inversion network 1. Finally, the new joint inversion network 1 is used to invert the data to be measured, and high-precision and high-stability inversion results are obtained. During the transfer learning training process, the training parameters remain unchanged to ensure the coherence and consistency of the model optimization.

[0042] The network described above was trained on a TensorFlow 2.1 platform. The experimental computing environment is configured as follows: NVIDIA RTX 4090D (24GB) GPU, AMD EPYC 9754 (18 cores, 128 threads) CPU.

[0043] Feature fusion joint inversion results and comparison with other traditional inversion methods: To further test the efficiency of the feature fusion joint inversion method proposed in this embodiment, it is compared with three-dimensional transient electromagnetic inversion, three-dimensional magnetic inversion and U-Net multi-channel joint inversion based on deep learning. Figure 5 The loss decay curves of the above four methods in the training process are shown in the results. The loss of the feature fusion joint inversion method converges to a lower level, and compared with the deep learning-based single transient electromagnetic and magnetic inversion, it converges faster and has a lower final convergence level. The loss decay of the U-Net multi-channel joint inversion method is slower, and the final convergence level is higher than that of the single inversion method, which indicates that this method fails to effectively extract the fusion features between transient electromagnetic and magnetic data, resulting in a slow loss decay process. The above comparison results further demonstrate the effectiveness of the multi-geophysical data feature fusion joint inversion network of this embodiment.

[0044] In the three-dimensional transient electromagnetic and three-dimensional magnetic synthetic data, the loss convergence speed of the joint inversion method based on multi-geophysical data feature fusion of this embodiment is significantly faster than that of the deep learning-based three-dimensional transient electromagnetic single inversion, three-dimensional magnetic single inversion and U-Net multi-channel joint inversion, and the final convergence level is the lowest, which reflects a better inversion effect.

[0045] At the same time, this embodiment randomly selects a sample data (Model1) from the test set, and uses the above method to perform inversion and compare the effects. Figure 6 The inversion result profile of different inversion methods in the model is shown. In the X=0m profile of Model1, the single inversion effect is poor due to the influence of the background resistivity noise. After using the joint inversion method, the multi-channel joint inversion based on U-Net and the feature fusion joint inversion of this embodiment both improve the inversion accuracy, but the method of this embodiment can more accurately restore the boundary and physical properties of the anomaly body.

[0046] In the sample data (Model1) randomly selected from the test set, the single inversion effect is poor due to the influence of the background resistivity noise. Although both joint inversion methods improve the inversion accuracy, the joint inversion method based on multi-geophysical data feature fusion of this embodiment is more accurate in anomaly boundary recovery and physical property characterization.

[0047] Ablation experiment: To further verify the effectiveness of the multi-geophysical data feature fusion joint inversion method proposed in this embodiment, an ablation experiment is conducted to compare the performance of the inversion network with and without feature fusion. Table 1 shows the ablation experiment results, where fusion indicates that down-sampled feature fusion is performed, and no fusion indicates that no feature fusion is performed. This embodiment compares the accuracy indicators of the two networks in the test set, including RMSE (mean square error) and SSIM (structural similarity index). RMSE measures the difference between the predicted value and the true value, and the smaller the value, the higher the inversion accuracy and the smaller the error of the model. SSIM evaluates the structural similarity of the model, and the closer the value to 1, the more similar the inversion result is to the true data in structure and details. By combining the two indicators of RMSE and SSIM, the accuracy and structural fidelity of the inversion result can be comprehensively reflected. The results show that the network after feature fusion improves by 1-2 percentage points in both indicators, further verifying the effectiveness of the feature fusion method proposed in this embodiment.

[0048] Table 1

[0049] Joint inversion results driven by dual joint strategy: To further improve the accuracy of the joint inversion results, this embodiment uses a dual joint inversion strategy to perform a second joint inversion. Figure 7 The real geological model, feature fusion joint inversion results, and feature fusion + dual joint strategy joint inversion results are shown. As can be seen from the figure, the feature fusion joint inversion can well delineate the range of the anomaly body and its corresponding physical property values, especially in the distribution of physical properties of the anomaly body. However, due to the resolution limitation of transient electromagnetic method, especially in high-resistance areas, there is still some deviation in the identification of the boundary of the anomaly body. The boundary inversion effect of the magnetic susceptibility model is better, which is related to the strong lateral resolution of the magnetic method data. Therefore, this embodiment performs a dual joint strategy joint inversion on the inverted resistivity model and magnetic susceptibility model to extract the fused features in the physical property model. The dual joint strategy joint inversion significantly improves the identification ability of the boundary of the high-resistance anomaly body and improves the inversion accuracy of the transient electromagnetic method. In summary, the dual joint inversion strategy proposed in this embodiment has shown significant effects in improving inversion accuracy and solving the boundary identification problem.

[0050] Noise resistance test: To test the noise resistance of the network, this embodiment adds 3%, 5%, and 10% Gaussian random noise to the test data. Figure 8The inversion results under different noise levels are shown, and it can be found that even if 10% Gaussian random noise is added, the inversion performance of the network still remains small fluctuations, indicating that the network has strong anti-noise capability.

[0051] Embodiment Two Based on the same overall inventive concept, the present application also provides a multi-geophysical data feature fusion and double joint strategy driven joint inversion system. The system provided by the present application is described below, and the system described below can be referred to each other corresponding to the method described above, and the system comprises: A data acquisition module is configured to acquire geophysical response data of different types. A first joint inversion module comprises a first joint inversion network based on an encoder-decoder structure, configured to extract and fuse features from the geophysical response data to obtain a preliminary physical property model. A second joint inversion module comprises a second joint inversion network configured to learn the correlation constraints between different physical property models, and is configured to obtain an optimized physical property model by feature interaction fusion on the preliminary physical property model. A network optimization module is configured to generate simulated response data based on the optimized physical property model, and fine-tune the parameters of the first joint inversion network through transfer learning. A final inversion module is configured to invert the measured geophysical response data using the fine-tuned first joint inversion network to obtain a final inversion result.

[0052] Embodiment Three The embodiment also discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of embodiment one.

[0053] Embodiment Four The embodiment also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method of embodiment one.

[0054] Embodiment Five The embodiment also discloses a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the steps of the method of embodiment one.

[0055] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method of joint inversion driven by multi-geophysical data feature fusion and dual combination strategy, characterized in that, The method comprises the following steps: acquiring different types of geophysical response data; constructing a first joint inversion network based on an encoder-decoder structure to extract and fuse features from the geophysical response data, and obtaining a preliminary physical property model; inputting the preliminary physical property model into a second joint inversion network configured to learn the correlation constraints between different physical property models, and obtaining an optimized physical property model by feature interaction fusion on the preliminary physical property model; generating simulated response data based on the optimized physical property model, and fine-tuning the parameters of the first joint inversion network through transfer learning; using the fine-tuned first joint inversion network to perform inversion on the measured geophysical response data to obtain a final inversion result.

2. The method of claim 1, wherein, The step of constructing a first joint inversion network based on an encoder-decoder structure to extract and fuse features from the geophysical response data, and obtaining a preliminary physical property model, specifically includes the following steps: inputting different types of geophysical response data into the first joint inversion network; in the encoder stage of the first joint inversion network, extracting high-level features of each data by downsampling, and fusing the high-level features at different levels to obtain multi-modal fusion features; in the decoder stage of the first joint inversion network, reconstructing and outputting the preliminary physical property model by upsampling and introducing the multi-modal fusion features through jump connection.

3. The method of claim 2, wherein, The step of extracting high-level features of each data by downsampling in the encoder stage of the first joint inversion network, and fusing the high-level features at different levels to obtain multi-modal fusion features, specifically includes the following steps: after at least one downsampling layer of the encoder, splicing feature maps corresponding to different geophysical data modalities; performing convolution operation on the spliced feature maps to extract the multi-modal fusion features.

4. The method of claim 1, wherein, The step of inputting the preliminary physical property model into a second joint inversion network, and obtaining an optimized physical property model by feature interaction fusion on the preliminary physical property model, specifically includes the following steps: constructing a second joint inversion network, the input and output of which are both physical property models; inputting the preliminary physical property model into the second joint inversion network; learning the correlation constraints between different physical property models through the second joint inversion network, and outputting the optimized physical property model.

5. The method of claim 1, wherein, The step of generating simulated response data based on the optimized physical property model, and fine-tuning the parameters of the first joint inversion network through transfer learning, specifically includes the following steps: performing forward calculation on the optimized physical property model to generate corresponding simulated geophysical response data; constructing a loss function according to the simulated geophysical response data and the measured geophysical response data; based on the loss function, fine-tuning the parameters of the first joint inversion network through back propagation.

6. The method of claim 1, wherein, The method adopts a staged training strategy; in the first stage, the first joint inversion network is trained with geophysical response data as input and true physical property model as output; in the second stage, the second joint inversion network is trained with true physical property model as input and output.

7. A multi-geophysical data feature fusion and dual joint strategy driven joint inversion system, characterized in that, The device is used for implementing the method of any one of claims 1-6, and comprises: a data acquisition module configured to acquire geophysical response data of different types; a first joint inversion module including a first joint inversion network based on an encoder-decoder structure, configured to extract and fuse features from the geophysical response data to obtain a preliminary physical property model; a second joint inversion module including a second joint inversion network configured to learn an associated constraint between different physical property models, and configured to obtain an optimized physical property model by performing feature interaction fusion on the preliminary physical property model; a network optimization module configured to generate simulated response data based on the optimized physical property model, and to fine-tune parameters of the first joint inversion network through transfer learning; a final inversion module configured to perform inversion on measured geophysical response data using the fine-tuned first joint inversion network to obtain a final inversion result.

8. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 7. The processor executes the computer program to implement the steps of the method of any one of claims 1-6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1-6.

Images