A three-dimensional magnetotelluric data inversion method, system, device and medium

Abstract

The application provides a three-dimensional magnetotelluric data inversion method, system, computer device and medium, and belongs to the field of three-dimensional magnetotelluric inversion. The method comprises the following steps: firstly, collecting magnetotelluric data; extracting multi-scale features through multi-branch convolution and channel splicing to obtain shallow multi-scale features; compressing the channel number to extract a middle layer key trend to obtain a middle layer enhanced feature; after spatial compression, multi-scale local features are extracted and spliced to obtain a bottleneck fusion feature; the middle layer enhanced feature is fused to obtain a primary reconstruction feature; and three types of features are fused to form multi-scale electromagnetic data; finally, the Inc_Module and Inc_DeepModule lightweight modules are used to realize resistivity prediction. Through the multi-level feature fusion mechanism, the training efficiency is improved by 3 times while reducing the parameter amount by 86%, and the noise robustness is significantly enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of three-dimensional magnetotelluric inversion, and particularly relates to a three-dimensional magnetotelluric data inversion method, system, device and medium. BACKGROUND

[0002] Magnetotelluric method is widely used in mineral resource exploration, deep earth structure exploration and geodynamics research and other important scientific issues. With the continuous innovation of magnetotelluric exploration theory, electromagnetic instrument, data interpretation technology and engineering application have been rapidly developed. It is crucial to generate samples that conform to geological structures and speed up the training of three-dimensional inversion.

[0003] In recent years, deep learning technology has been widely applied in the field of geophysical exploration and has attracted great attention from many researchers. A large number of training samples in deep learning can improve the accuracy of training, but the increase of training samples also leads to a decrease in training efficiency. In one-dimensional inversion, only single line is involved and the data volume is small, the structure is simple, and the model is easy to train, so deep learning can accurately invert the results. Two-dimensional inversion involves two-dimensional plane, and the data volume is increased compared with one-dimensional, but the dimension is still lower than three-dimensional. However, three-dimensional inversion involves higher spatial dimension, and the number of model parameters increases exponentially, resulting in a doubling of training time. Therefore, the problems existing in the current three-dimensional inversion include that the generated training samples do not conform to the actual geological structure, and the increase of the number of training samples and the improvement of its complexity will lead to a significant decrease in training efficiency. In summary, the present application provides a new method for three-dimensional complex samples and improving inversion accuracy. SUMMARY

[0004] In order to solve the problem of how to reduce the model parameters and improve the inversion accuracy, the present application provides a three-dimensional magnetotelluric data inversion method, system, device and medium.

[0005] In order to achieve the above-mentioned purpose, the present application provides a three-dimensional magnetotelluric data inversion method, comprising:

[0006] Collecting magnetotelluric data.

[0007] Performing multi-scale feature extraction on the magnetotelluric data, splicing the extracted different scale features to generate shallow multi-scale features, and extracting middle key trend features of the shallow multi-scale features by compressing the channel number of the shallow multi-scale features to obtain middle enhanced features.

[0008] The middle layer enhanced features are compressed in spatial dimension, and then local features of different scales of the compressed middle layer enhanced features are extracted, and the local features of different scales are spliced to obtain bottleneck fusion features; the bottleneck fusion features and the middle layer enhanced features are fused to output primary reconstruction features; the primary reconstruction features, the bottleneck fusion features and the shallow multi-scale features are fused to obtain multi-scale electromagnetic data.

[0009] The resistivity prediction result of inversion is calculated and output by the multi-scale electromagnetic data.

[0010] Preferably, the magnetotelluric data is processed by an EfmDeepNet model to obtain the resistivity prediction result of inversion; the EfmDeepNet model is based on a U_Net network and specifically includes: replacing a convolution layer in the U_Net network with a lightweight multi-branch module Inc_Module, and replacing a pooling layer with a deep down-sampling module Inc_DeepModule; the EfmDeepNet model is connected by two layers of Inc_Module, one layer of Inc_DeepModule, two layers of Inc_Module, a convolution layer, a normalization layer, a ReLU activation function and a max pooling layer in sequence.

[0011] Preferably, the Inc_Module module is accessed by four parallel branches with the same structure, each branch is connected by a Convolution layer, a Batch Normalization layer and a ReLU layer in sequence, wherein the depth of each Convolution layer has different separable convolution kernel sizes; the outputs of the four branches are aggregated to a Depth Concatenation layer and a Drop Out layer; the Depth Concatenation layer and the Drop Out layer are responsible for feature fusion.

[0012] Preferably, the Inc_DeepModule module accesses four parallel branches, and the four branches are connected by a Convolution layer, a Batch Normalization layer and a ReLU layer in sequence, wherein one branch is first connected by an AveragePooling layer, and then connected with other branches containing a Convolution layer, a Batch Normalization layer and a ReLU layer; after processing of each branch, the branches are sequentially connected by a Depth Concatenation layer and a Drop Out layer.

[0013] Preferably, the magnetotelluric data is subjected to multi-scale feature extraction by the Inc_Module module, and different scale features extracted are spliced in channels to generate shallow multi-scale features, specifically including: different scale features of the magnetotelluric data are extracted by using depth separable convolution kernels of different scales; the different scale features extracted are subjected to normalization and ReLU activation, and feature maps after the normalization and ReLU activation are spliced in channels to obtain shallow multi-scale features.

[0014] Preferably, the Inc_DeepModule module is used to compress the spatial dimensions of the middle layer enhanced features, and then different scale local features of the compressed middle layer enhanced features are extracted and spliced to obtain bottleneck fusion features, including: the spatial dimensions of the middle layer enhanced features are compressed by using depth separable convolution, and the compressed features are subjected to depth separable convolution to obtain global statistical information features Z 1;the middle layer enhanced features are subjected to parallel convolution by using depth separable convolution of different sizes to obtain depth sensitive features Z 2、 Z 3、 Z 4;the channel number of the global statistical information features Z1 is adjusted to be consistent with the channel number of Z 2、 Z 3、 Z 4;the restored Z 1 is spliced in channels with Z 2、 Z 3、 Z 4 to obtain bottleneck fusion features.

[0015] Preferably, before the multi-scale feature extraction of the magnetotelluric data, the method further includes sample expansion of the magnetotelluric data, specifically including:

[0016] Gaussian random field method is used to randomly generate three-dimensional complex magnetotelluric sample data;

[0017] The resistivity range of real magnetotelluric data is obtained, the smoothing factor, the three-dimensional grid size and the frequency range are determined according to the resistivity range, and the sample data is subjected to forward modeling according to the smoothing factor, the three-dimensional grid size and the frequency range to obtain expanded samples.

[0018] The application further provides a three-dimensional magnetotelluric data inversion system, including:

[0019] A data acquisition module is used to acquire magnetotelluric data.

[0020] The feature extraction module is used to extract multi-scale features from magnetotelluric data. It concatenates the extracted features at different scales to generate shallow multi-scale features. By compressing the number of channels in the shallow multi-scale features, it extracts the mid-level key trend features, obtaining mid-level enhancement features. The mid-level enhancement features are then spatially compressed, and local features at different scales are extracted from the compressed mid-level enhancement features. These local features at different scales are then concatenated to obtain bottleneck fusion features. The bottleneck fusion features and mid-level enhancement features are then fused to output primary reconstruction features. Finally, the primary reconstruction features, bottleneck fusion features, and shallow multi-scale features are fused to obtain multi-scale electromagnetic data.

[0021] The prediction application module is used to output the inverted resistivity prediction results from multi-scale electromagnetic data.

[0022] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the electromagnetic inversion method.

[0023] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute any of the steps in the electromagnetic inversion method.

[0024] The three-dimensional magnetotelluric data inversion method provided by this invention has the following beneficial effects:

[0025] Multi-scale feature extraction is performed on magnetotelluric data. Basic features of the electromagnetic data are extracted by generating shallow multi-scale features, mid-level enhancement features, and bottleneck fusion features, while significantly reducing network parameters. Then, based on the fusion of bottleneck fusion features and mid-level enhancement features, primary reconstruction features are output. Furthermore, the primary reconstruction features, bottleneck fusion features, and shallow multi-scale features are fused to obtain multi-scale electromagnetic data. Finally, the resistivity prediction results are output through multi-scale electromagnetic data. This method not only improves the inversion accuracy but also increases the inversion efficiency by three times. Attached Figure Description

[0026] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of a three-dimensional magnetotelluric data inversion method according to an embodiment of the present invention;

[0028] Figure 2 Sample diagram generated by different smoothing factors of Gaussian random field of embodiments of the present application; wherein, Figure 2 The smoothing factor of (a1) of is 5, Figure 2 The smoothing factor of (a2) of is 6, Figure 2 The smoothing factor of (a3) of is 7, Figure 2 The smoothing factor of (a4) of is 8, Figure 2 The smoothing factor of (a5) of is 9, Figure 2 The smoothing factor of (a6) of is 10;

[0029] Figure 3 EfmDeepNet model structure diagram of embodiments of the present application;

[0030] Figure 4 Inc_Module module structure diagram of embodiments of the present application;

[0031] Figure 5 Inc_DeepModule module structure diagram of embodiments of the present application;

[0032] Figure 6 Theoretical sample inversion result diagram of embodiments of the present application, Figure 6 (a1-a6) of is a theoretical sample, Figure 6 (b1-b6) of is a noiseless inversion result, Figure 6 (c1-c6) of is an inversion result with 1% Gaussian noise, Figure 6 (d1-d6) of is an inversion result with 3% Gaussian noise. DETAILED DESCRIPTION

[0033] In order for those skilled in the art to better understand the technical solutions of the present application and to implement them, the present application will be described in detail below in conjunction with the drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0034] The present application provides a three-dimensional magnetotelluric data inversion method, specifically as Figure 1 shown, comprising:

[0035] S1, collecting magnetotelluric data.

[0036] Before the multi-scale feature extraction of the magnetotelluric data, the sample expansion of the magnetotelluric data is further included, specifically including: generating three-dimensional complex magnetotelluric sample data randomly by using a Gaussian random field method; obtaining a resistivity range of real magnetotelluric data, determining a smoothing factor, a three-dimensional grid size and a frequency range according to the resistivity range, and performing forward calculation on the sample data according to the smoothing factor, the three-dimensional grid size and the frequency range to obtain expanded samples.

[0037] The Gaussian random field random sample parameters are set according to the actual magnetotelluric detection resistivity range, and the random sample is generated by using the Gaussian random field method, wherein the smoothing factor is 5-10, the resistivity random range is 1-10000Ωm, and the complex sample is as shown in Figure 2 ; the three-dimensional grid size is selected in the resistivity range, and the three-dimensional grid size is , wherein the edge contains 5 air layers, the frequency range is from 16 frequency points are taken; the sample data is forward calculated by using a finite element to obtain a sample for inversion training; the training sample obtained by the forward calculation is divided into a training set, a verification set and a test set, and the training sample division ratio is 8:1:1.

[0038] S2, multi-scale feature extraction is performed on the magnetotelluric data, different scale features extracted are spliced to generate shallow multi-scale features; by compressing the channel number of the shallow multi-scale features, the middle key trend features of the shallow multi-scale features are extracted to obtain middle enhanced features. The spatial dimension of the middle enhanced features is compressed, and then different scale local features of the compressed middle enhanced features are extracted, and then different scale local features are spliced to obtain bottleneck fusion features; the bottleneck fusion features and the middle enhanced features are fused to output primary reconstruction features; the primary reconstruction features, the bottleneck fusion features and the shallow multi-scale features are fused to obtain multi-scale electromagnetic data.

[0039] The magnetotelluric data is input into the EfmDeepNet model, the Inc_Module1 module extracts multi-scale features of the magnetotelluric data through deep separable convolution of different sizes, generates shallow multi-scale feature representation after channel splicing of the extracted different scale features, inputs the shallow multi-scale features into the Inc_Module2 module, extracts middle key trend features of the shallow multi-scale features by compressing the channel number of the shallow multi-scale features, obtains middle enhanced features, compresses the spatial dimension of the middle enhanced features by using the Inc_DeepModule module, extracts features of the compressed features through deep separable convolution of different sizes, then splices the output features of each scale to obtain bottleneck fusion features, fuses the bottleneck fusion features and the middle enhanced features by using the Inc_Module3 module, and outputs primary reconstruction features; the Inc_Module4 module is used to perform up-sampling fusion features on the primary reconstruction features, the bottleneck fusion features and the shallow multi-scale features to obtain multi-scale electromagnetic data.

[0040] As shown in Figure 3 , the magnetotelluric data is processed by the EfmDeepNet model to obtain an inverted resistivity prediction result; the EfmDeepNet model is based on a U_Net network and specifically includes: replacing the convolution layer in the U_Net network with a lightweight multi-branch module Inc_Module, and replacing the pooling layer with a deep down-sampling module Inc_DeepModule; composed of two layers of Inc_Module, one layer of Inc_DeepModule, two layers of Inc_Module, a convolution layer, a normalization layer, a ReLU activation function and a max pooling layer connected in sequence.

[0041] As shown in Figure 4 , the Inc_Module module is accessed in parallel by four branches with the same structure, each branch is connected in series with a Convolution layer, a Batch Normalization layer and a ReLU layer, wherein the deep separable convolution kernel sizes of each Convolution layer are different from each other; the outputs of the four branches are aggregated to a Depth Concatenation layer and a Drop Out layer; the Depth Concatenation layer and the Drop Out layer are responsible for feature fusion.

[0042] As shown in Figure 5As shown, the Inc_DeepModule module accesses 4 parallel branches, all of which are connected by Convolution layers, Batch Normalization layers and ReLU layers in turn, one of which first passes through an AveragePooling layer and then connects with other branches containing Convolution layers, Batch Normalization layers and ReLU layers; after processing by each branch, it is in turn subjected to a Depth Concatenation layer and a Drop Out layer.

[0043] The Inc_Module module is used to extract multi-scale features from the magnetotelluric data, and the different scale features extracted are spliced in the channel to generate shallow multi-scale features; including: different scale depth separable convolution kernels are used to extract different scale features of the magnetotelluric data; the different scale features extracted are normalized and activated by ReLU, and the feature maps after normalization and ReLU activation are spliced in the channel to obtain shallow multi-scale features.

[0044] The Inc_DeepModule module is used to compress the spatial dimensions of the middle layer enhanced features, and then extract different scale local features of the compressed middle layer enhanced features, and then splice the different scale local features to obtain bottleneck fusion features; including: the spatial dimensions of the middle layer enhanced features are compressed by depth separable convolution, and the compressed features are further subjected to depth separable convolution to obtain global statistical information features Z 1;different size depth separable convolutions are used to perform parallel convolution on the middle layer enhanced features to obtain depth sensitive features Z 2、 Z 3、 Z 4;the global statistical information features Z 1 are pooled to restore the channel size to the same as Z 2、 Z 3、 Z 4;the restored Z 1 is spliced in the channel with Z 2、 Z 3、 Z 4 to obtain bottleneck fusion features.

[0045] The network model comprises an encoding and decoding structure, and the encoding and decoding layers are up-sampling and down-sampling respectively, the number of layers of the network structure comprises two encoding layers and two decoding layers, the EfmDeepNet model retains the overall encoding structure in the U-Net, can effectively extract image features and map them to a low-dimensional space to obtain feature representation, and the specific inversion network EfmDeepNet model replaces the convolutional layer and the pooling layer with the Inc_Module module and the Inc_DeepModule module in the lightweight design, the Inc_Module module comprises a convolutional layer, a normalization layer and an activation layer; the Inc_DeepModule module comprises a convolutional layer, a normalization layer, an activation layer and an average pooling layer, the width of the Inc_DeepModule module is 4, the lightweight design of the Inc_Module and the Inc_DeepModule reduces the network parameter quantity, thereby improving the training efficiency of inversion; in order to further prove the training efficiency of the network of the present application, the network training time in the prior art is compared, and the results show that the network training time of the present application is three times that of the prior art, the inversion efficiency is greatly improved, a new idea is provided for the method in the field, and the training comparison time is shown in Table 1.

[0046] Table 1: Comparison of network training time

[0047]

[0048] In the whole network design, two different lightweight modules are contained, the difference between the two modules lies in that the depths are different and the structures are also different, two different structures, each convolutional layer is followed by a normalization layer, while the convolutional layer is retained, the gradient disappearance is avoided, the lightweight module captures multi-scale features by applying various convolution kernels and pooling operations in parallel, and then the features are combined to form a more rich and comprehensive representation.

[0049] The data set is trained, the network parameters are continuously optimized to obtain the best inversion performance in the training process, the learning rate is set to a dynamic learning rate, the initial learning rate is continuously set in the training, early stopping is used to prevent overfitting of the network in the training process; the training strategy adopts the combination of Adam and L-BFGS, first, the Adam optimizer is used, and then the L-BFGS method is switched to training until the model converges, this method uses Adam as the initial optimization algorithm, can accelerate the reduction of the loss function, improves the calculation efficiency, and finally saves the final inversion model obtained by training.

[0050] S3, calculating the output resistivity prediction result of inversion through multi-scale electromagnetic data.

[0051] Output the resistivity prediction result of inversion through multi-scale electromagnetic data.

[0052] The multi-scale electromagnetic data is mapped to the resistivity physical space through convolution, and after being normalized through an activation function, a final resistivity prediction result is output.

[0053] The effect of the inversion model is evaluated by inverting the data of the test set; in actual exploration, the observation data often contains a certain degree of noise pollution, and the generalization of the inversion model is evaluated by randomly adding Gaussian noise to the test set data; 1% and 3% of Gaussian random noise are randomly added to the input data, respectively; in order to better compare the three-dimensional inversion effect, the inversion results of the section are compared, the results are sliced every 6KM, and the comparison results are as shown in Figure 6 Figure 6 (a1-a6) of the theoretical sample, Figure 6 (b1-b6) of the noiseless inversion result, Figure 6 (c1-c6) of the 1% Gaussian noise inversion result, Figure 6 (d1-d6) of the 3% Gaussian noise inversion result.

[0054] Based on the same inventive concept, the present application also provides a three-dimensional magnetotelluric data inversion system, comprising:

[0055] A data acquisition module is configured to acquire magnetotelluric data.

[0056] A feature extraction module is configured to perform multi-scale feature extraction on the magnetotelluric data, concatenate the extracted different scale features to generate shallow multi-scale features, extract middle layer key trend features of the shallow multi-scale features by compressing the channel number of the shallow multi-scale features, obtain middle layer enhanced features, perform spatial dimension compression on the middle layer enhanced features, then extract different scale local features of the compressed middle layer enhanced features, concatenate the different scale local features to obtain bottleneck fusion features, perform feature fusion on the bottleneck fusion features and the middle layer enhanced features to output primary reconstruction features, and perform feature fusion on the primary reconstruction features, the bottleneck fusion features and the shallow multi-scale features to obtain multi-scale electromagnetic data.

[0057] A prediction application module is configured to output an inversion resistivity prediction result through the multi-scale electromagnetic data.

[0058] The present application also provides a computer device, which, at the hardware level, comprises a processor, an internal bus, a network interface, a memory and a non-volatile memory, and of course can also comprise other hardware required by the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs to implement the three-dimensional magnetotelluric data inversion method provided above.

[0059] ​The application further provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the three-dimensional magnetotelluric data inversion method provided above.

[0060] The specific limitation of the three-dimensional magnetotelluric data inversion method computing system can refer to the limitation of the three-dimensional magnetotelluric data inversion method provided above, which will not be repeated here. Each module in the three-dimensional magnetotelluric data inversion system can be realized by software, hardware and a combination thereof in whole or in part. Each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operation corresponding to each module.

[0061] It should be noted that the above specific embodiments can enable those skilled in the art to more fully understand the present application, but in no way limit the present application. Therefore, although the present application has been described in detail in the specification and examples, those skilled in the art should understand that the present application can still be modified or replaced by equivalents; all technical solutions and improvements which do not deviate from the spirit and scope of the present application are covered in the protection scope of the patent of the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method of three-dimensional magnetotelluric data inversion, characterized in that, The method comprises: Collecting magnetotelluric data; Performing multi-scale feature extraction on the magnetotelluric data, splicing the extracted different scale features to generate shallow multi-scale features, and extracting middle key trend features of the shallow multi-scale features by compressing the channel number of the shallow multi-scale features to obtain middle enhanced features; Performing spatial dimension compression on the middle enhanced features, then extracting different scale local features of the compressed middle enhanced features, splicing the different scale local features to obtain bottleneck fusion features; performing feature fusion on the bottleneck fusion features and the middle enhanced features to output primary reconstruction features; performing feature fusion on the primary reconstruction features, the bottleneck fusion features and the shallow multi-scale features to obtain multi-scale electromagnetic data; Calculating and outputting the inverted resistivity prediction result through the multi-scale electromagnetic data.

2. The three-dimensional magnetotelluric data inversion method according to claim 1, characterized in that: The magnetotelluric data is processed through an EfmDeepNet model to obtain the inverted resistivity prediction result; the EfmDeepNet model is based on a U_Net network, and replaces the convolution layer in the U_Net network with a lightweight multi-branch module Inc_Module and replaces the pooling layer with a deep down-sampling module Inc_DeepModule; the EfmDeepNet model is connected in sequence by two layers of Inc_Module, one layer of Inc_DeepModule, two layers of Inc_Module, a convolution layer, a normalization layer, a ReLU activation function and a max-pooling layer.

3. The three-dimensional magnetotelluric data inversion method of claim 2, wherein, The Inc_Module module is accessed by four parallel branches with the same structure, each branch is connected in sequence by a Convolution layer, a BatchNormalization layer and a ReLU layer, wherein the depth of each Convolution layer has different separable convolution kernel sizes; the outputs of the four branches are aggregated to a Depth Concatenation layer and a Drop Out layer; the Depth Concatenation layer and the Drop Out layer are responsible for feature fusion.

4. The three-dimensional magnetotelluric data inversion method of claim 2, wherein, The Inc_DeepModule module accesses four parallel branches, and the four branches are connected in sequence by a Convolution layer, a BatchNormalization layer and a ReLU layer, wherein one branch is first connected through an Average Pooling layer, and then connected with other branches containing a Convolution layer, a BatchNormalization layer and a ReLU layer; after processing by each branch, the branches are sequentially connected through a Depth Concatenation layer and a Drop Out layer.

5. The three-dimensional magnetotelluric data inversion method of claim 2, wherein, The Inc_Module module is used for multi-scale feature extraction of the magnetotelluric data, and different scale features extracted are spliced to generate shallow multi-scale features, specifically including: different scale features of the magnetotelluric data are extracted by using different scale depth separable convolution kernels; the different scale features extracted are normalized and activated by ReLU, and the feature maps after normalization and ReLU activation are spliced in channels to obtain shallow multi-scale features.

6. The three-dimensional magnetotelluric data inversion method of claim 2, wherein, The middle layer enhanced feature is compressed in spatial dimension through the Inc_DeepModule module, local features of different scales of the compressed middle layer enhanced feature are extracted, and then local features of different scales are spliced to obtain a bottleneck fusion feature, including: compressing the spatial dimension of the middle layer enhanced feature by using a depth separable convolution, performing depth separable convolution on the compressed feature to obtain a global statistical information feature Z 1; parallel convolution of the middle layer enhanced feature by using depth separable convolutions of different sizes to obtain depth sensitive features Z 2、 Z 3、 Z 4; adjusting the channel number of the global statistical information feature Z1 to be consistent with the channel number of Z 2、 Z 3、 Z 4; splicing the restored Z 1 with Z 2、 Z 3、 Z 4 to obtain a bottleneck fusion feature.

7. The three-dimensional magnetotelluric data inversion method of claim 1, wherein, Before the multi-scale feature extraction of the magnetotelluric data, the method further includes sample expansion of the magnetotelluric data, specifically including: A three-dimensional complex magnetotelluric sample data is randomly generated by using a Gaussian random field method. The resistivity range of real magnetotelluric data is obtained, a smoothing factor, a three-dimensional grid size and a frequency range are determined according to the resistivity range, and the sample data is forward calculated according to the smoothing factor, the three-dimensional grid size and the frequency range to obtain expanded samples.

8. A three-dimensional magnetotelluric data inversion system, characterized in that, The method comprises: a data acquisition module configured to acquire magnetotelluric data; a feature extraction module configured to perform multi-scale feature extraction on the magnetotelluric data, splice different scale features extracted in channels to generate shallow multi-scale features, and extract middle layer key trend features of the shallow multi-scale features by compressing the number of channels of the shallow multi-scale features to obtain middle layer enhanced features; the middle layer enhanced features are compressed in spatial dimensions, different scale local features of the compressed middle layer enhanced features are extracted, and then different scale local features are spliced to obtain bottleneck fusion features; the bottleneck fusion features and the middle layer enhanced features are fused to output primary reconstruction features; the primary reconstruction features, the bottleneck fusion features and the shallow multi-scale features are fused to obtain multi-scale electromagnetic data; a prediction application module configured to output an inverted resistivity prediction result by using the multi-scale electromagnetic data.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-8. The processor executes the computer program to realize the method steps in any one of claims 1 to 7.

10. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to realize the method steps in any one of claims 1 to 7.

Images