Three-dimensional magnetotelluric inversion method, device and equipment

Abstract

The invention discloses a three-dimensional magnetotelluric inversion method, a three-dimensional magnetotelluric inversion device and three-dimensional magnetotelluric inversion equipment, and relates to the technical field of three-dimensional magnetotelluric inversion processing. A network model suitable for three-dimensional magnetotelluric inversion is constructed based on a deep learning technology, so that the input of the model is apparent resistivity and phase; the output is a three-dimensional resistivity distribution parameter of the to-be-measured area; a data-driven and physical-driven combined weighted loss function is designed and introduced into the network model, so that three-dimensional forward modeling operation of magnetotelluric is integrated into forward propagation of the network model, the network model can follow magnetotelluric physical constraints in the training process, and the inversion precision of the anomalous body is integrally improved.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional magnetotelluric inversion processing, and in particular to a three-dimensional magnetotelluric inversion method, device and equipment. Background Art

[0002] Magnetotelluric inversion is a method that uses electric field observation data on the earth to infer the underground medium model. In this process, partial differential equations are used to describe the propagation behavior of electromagnetic fields underground and match them with the observation data to infer the underground conductivity or dielectric constant distribution. In reality, most partial differential equations cannot calculate explicit solutions, so numerical methods become a powerful tool for solving these partial differential equations. The study of partial differential equations can be divided into two aspects: direct problems and inverse problems. There are many mature methods for solving direct problems of partial differential equations, but solving inverse problems of partial differential equations is a major problem. On the one hand, due to the complexity and extensiveness of the model, it is difficult to establish a unified regularization framework. On the other hand, regularization methods can hardly get good results for highly ill-posed problems.

[0003] In recent years, deep learning has been widely used in the field of inverse problems of partial differential equations. Compared with traditional numerical methods, deep learning-based methods for solving inverse problems of partial differential equations have many advantages, such as being more robust to noise, less dependent on initial and boundary conditions, and parallelizable training. The application of deep learning in the field of inverse problems of partial differential equations has huge value and room for exploration, and is expected to change the way traditional numerical algorithms solve inverse problems of partial differential equations.

[0004] The deep learning-based method currently used is to use a deep learning network to directly calculate the numerical relationship within the electric field observation data in the detection area to infer the underground medium structure of the detection area. However, it usually ignores the connection between the deep learning network and the numerical relationship, resulting in a large difference between the inferred underground medium structure and the actual structure, affecting the accuracy of three-dimensional magnetotelluric inversion. Summary of the invention

[0005] The embodiments of the present invention provide a three-dimensional magnetotelluric inversion method, device and equipment, which can solve the problem in the prior art that the current methods usually ignore the connection between deep learning networks and numerical relationships, resulting in a large difference between the inferred underground medium structure and the actual structure.

[0006] The embodiment of the present invention provides a three-dimensional magnetotelluric inversion method, comprising the following steps:

[0007] Obtain the magnetotelluric data to be inverted, and perform forward modeling on the magnetotelluric data to obtain multiple response data;

[0008] Use 3D U-Net as the magnetotelluric inversion network model;

[0009] A data-driven loss function is constructed, and the data error representing the physically driven three-dimensional forward response is added to the data-driven loss function to construct a data-driven and physically driven joint weighted loss function; the cross entropy loss function in the magnetotelluric inversion network model is replaced with a data-driven and physically driven joint weighted loss function to obtain an improved magnetotelluric inversion network model;

[0010] The improved magnetotelluric inversion network model is trained using multiple response data. The joint weighted loss function driven by data and physics is used for constraint during the training to obtain the trained magnetotelluric inversion network model.

[0011] The response data to be measured is input into the trained magnetotelluric inversion network model to obtain the three-dimensional resistivity distribution parameters of the area to be measured to realize three-dimensional magnetotelluric inversion.

[0012] Preferably, the obtaining of multiple response data includes:

[0013] According to the detection area and magnetotelluric data, a theoretical resistivity model of the detection area is generated by random simulation; the anomaly body in the detection area is forward modeled in the theoretical resistivity model to obtain the response data of the theoretical resistivity model including apparent resistivity and phase;

[0014] In the response data containing apparent resistivity and phase, each response data contains the apparent resistivity and phase in the xy and yx directions, and also contains the impedance tensor Z xx , Z xy , Z yx , Z yy to form a data set of response data.

[0015] Preferably, the step of constructing a data-driven and physically-driven joint weighted loss function comprises:

[0016] The data error representing the three-dimensional forward response of physical drive is added to the loss objective function in data-driven, and the evolution process of the model in model training is kept consistent with the evolution of the physical environment to construct a joint weighted loss function of data-driven and physical drive.

[0017] Preferably, the joint weighted loss function of the data-driven and physical-driven loss is:

[0018]

[0019] Among them: α and β are the weight coefficients of the two errors respectively; represents the depth weighting function; Φ m(d, m, θ) represents the data-driven loss function; Φ d (d, θ) represents the data error of the physically driven three-dimensional forward response.

[0020] Preferably, the acquisition of the three-dimensional resistivity distribution parameters of the area to be measured includes:

[0021] Inputting the response data to be tested into the trained magnetotelluric inversion network model, and using the encoder of the magnetotelluric inversion network model to extract features of the response data to obtain local feature data;

[0022] The decoder is used to extract local features to obtain deep feature data, and the local feature data is superimposed with the feature elements in the deep feature data to obtain final feature data, and the three-dimensional resistivity distribution parameters of the area to be measured are identified from the final feature data.

[0023] Preferably, after identifying the three-dimensional resistivity distribution parameters of the area to be measured, the position, shape and resistivity value data of the abnormal body in the area to be measured are simulated according to the three-dimensional resistivity distribution parameters to form a three-dimensional magnetotelluric inversion result.

[0024] The embodiment of the present invention further provides a three-dimensional magnetotelluric inversion device, comprising:

[0025] The numerical module is used to obtain the magnetotelluric data to be inverted and to forward model the magnetotelluric data to obtain multiple response data;

[0026] Model module, used to use 3D U-Net network as magnetotelluric inversion network model;

[0027] A data-driven loss function is constructed, and the data error representing the physically driven three-dimensional forward response is added to the data-driven loss function to construct a data-driven and physically driven joint weighted loss function; the cross entropy loss function in the magnetotelluric inversion network model is replaced with a data-driven and physically driven joint weighted loss function to obtain an improved magnetotelluric inversion network model;

[0028] The improved magnetotelluric inversion network model is trained using multiple response data. The joint weighted loss function driven by data and physics is used for constraint during the training to obtain the trained magnetotelluric inversion network model.

[0029] The inversion module is used to input the response data to be measured into the trained magnetotelluric inversion network model to obtain the three-dimensional resistivity distribution parameters of the area to be measured to realize three-dimensional magnetotelluric inversion.

[0030] An embodiment of the present invention further provides an electronic device, including a memory and a processor;

[0031] The memory is used to store computer programs;

[0032] The processor is used to implement the steps of the three-dimensional magnetotelluric inversion method as described above when executing the computer program stored in the memory.

[0033] The embodiments of the present invention provide a three-dimensional magnetotelluric inversion method, device and apparatus. Compared with the prior art, the beneficial effects thereof are as follows:

[0034] The present invention is based on a 3D U-Net network framework, constructs a network model suitable for three-dimensional magnetotelluric inversion, designs a weighted loss function that combines data-driven and physical-driven, and introduces it into the network model, so as to integrate the three-dimensional forward modeling operation of magnetotelluric into the forward propagation of the network model, so that the network model can follow the magnetotelluric physical constraints during the training process, and strengthens the correlation between the electric field observation data in the detection area and the network model. Furthermore, the network model for three-dimensional magnetotelluric inversion formed after training can more accurately deduce the underground medium structure of the detection area when inverting the detection area, thereby improving the accuracy of three-dimensional magnetotelluric inversion as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the structure of a deep learning inversion network of a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of an overall inversion process of a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of an abnormal body used to generate training samples in a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0038] Figure 4 A schematic diagram of an example of a method for enhancing data using horizontal and vertical flipping in a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0039] Figure 5 A schematic diagram showing comparison of inversion results of a single rectangular block anomaly model of a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0040] Figure 6 A schematic diagram showing comparison of inversion results of a single step block anomaly model of a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0041] Figure 7 A schematic diagram of the inversion results of a combined rectangular block anomaly model of a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0042] Figure 8 A schematic diagram of the inversion results of a combined step block anomaly model of a three-dimensional magnetotelluric inversion method provided by an embodiment of the present invention;

[0043] Fig. 9 A schematic diagram of the inversion results of slices at different depths of an abnormal body in a model of a three-dimensional magnetotelluric inversion method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0045] See also Figure 2 The embodiment of the present invention provides a three-dimensional magnetotelluric inversion method, comprising the following steps:

[0046] Step 1: Create a training data set, generate a theoretical resistivity model using random simulation, obtain model response data (apparent resistivity and phase) through three-dimensional vector finite element forward numerical calculation, and obtain a network training sample set (input: apparent resistivity and phase data, output: 3D resistivity model).

[0047] Step 2: Use domain risk minimization to perform data enhancement during training to improve the generalization ability of network training.

[0048] Step 3: Based on the 3D U-Net network framework, a deep learning network structure suitable for 3D magnetotelluric inversion is constructed to realize the nonlinear mapping relationship from input (observation data) to output (3D resistivity model), such as Figure 1 shown.

[0049] Step 4: Design a data-driven and physically-driven joint weighted loss function, integrate the forward modeling of three-dimensional magnetotelluric into the forward propagation of the network, so that the network can follow the magnetotelluric physical constraints during training; at the same time, introduce a depth weighting function to offset the natural attenuation of the electrostatic field, so that the network has greater investment power in deep areas, thereby improving the inversion accuracy and resolution of deep anomalies.

[0050] Step 5: During network training, the optimization problem is solved by minimizing the loss function. When the loss function is lower than the set threshold or reaches the maximum number of iterations, the network stops training.

[0051] Step 6: Analyze the inversion results of the theoretical model data, adjust the neural network hyperparameter settings, and obtain the final training parameters.

[0052] Step 7: Use synthetic data to verify the inversion effect.

[0053] In step 2:

[0054] Based on the data enhancement method of three-dimensional horizontal image flipping, the number of grids remains unchanged and the original data set is expanded twice.

[0055] In step three:

[0056] The constructed network consists of an encoder stage and a decoder stage. In the encoder stage, there are three encoders, each of which contains two convolutional layers. Convolution operations are performed on the input data to generate local feature data. Batch normalization layers and activation functions are used in the feature data to improve the nonlinear fitting ability of the network. Maximum pooling layers, downsampling layers, and decoding layers are set.

[0057] In the decoder stage, it is composed of three upsampling modules; the upsampling module uses transposed convolution to expand the feature data size, and a convolution layer is connected after the transposed convolution to further decode the feature information of the data; then the resulting feature data of the upsampling module is added to the feature data of the encoder stage, so that the image contains features at different levels, making the network more robust; in addition, in the process of downsampling and upsampling, different degrees of Dropout layers are added to prevent overfitting; in the output part, the original classification output layer is replaced with the regression output layer.

[0058] In step four:

[0059] Data-driven methods transform real models m With prediction model The model error between is used as the loss objective function and can be expressed as:

[0060]

[0061] Where: N represents the number of model parameters; T represents the number of samples involved in training; m i，j Indicates the true resistivity value.

[0062] Once the network is well trained, the feed-forward network H can be used as the inversion operator F -1 , predict m from a new set of measurements d.

[0063] Effectively speed up the forward modeling progress and introduce the forward modeling data error as follows:

[0064]

[0065] Where: F represents the three-dimensional forward operator; H represents the feedforward neural network with parameter θ; d represents the forward response data.

[0066] The depth weighting function is used to improve the inversion accuracy and resolution of the anomaly:

[0067]

[0068] Among them: parameter λ is a constant related to the depth of the anomaly, and parameter δ is a constant used to control the depth weight distribution.

[0069] The joint weighted loss function of data-driven and physics-driven is defined as:

[0070]

[0071] Among them: α and β are the weight coefficients of the two errors.

[0072] In step five:

[0073] The minimization loss function set in network training is:

[0074] Loss = argminΦ(d,m,θ)

[0075] In step six:

[0076] After continuous optimization, the parameters of the inversion network model are set as follows:

[0077] Set MaxEpochs to 100, MinibatchSize to 64, optimizer to Adam, adopt dynamic learning rate adjustment scheme, set initial learning rate to 0.01, and multiply the learning rate by 0.5 every 10 epochs. At the same time, use apparent resistivity (ρ xy , ρ yx ) and phase (φ xy ,φ yx ), and the impedance tensor (Z xx , Z xy , Z yx , Z yy ) as the network input for analysis can provide more discriminant information for model inversion, especially in the abnormal body area, and can achieve better inversion accuracy.

[0078] Specifically:

[0079] In making the training data set, the determined study area is discretized into 32×32×16 three-dimensional grids, and each grid is evenly divided at intervals of 300 meters. Then, different types of anomalies are pre-set based on the actual situation, and the anomalies are randomly embedded in different positions of the study area, thereby generating various types of three-dimensional resistivity anomaly model data; Figure 3 As shown in the figure, according to the random simulation method, partial schematic diagrams of different types of anomalies and related parameters are generated; low-resistance anomalies and high-resistance anomalies are set in the data set, the values ​​of low-resistance anomalies are [100Ωm, 200Ωm, 400Ωm, 600Ωm, 800Ωm], and the values ​​of high-resistance anomalies are [2000Ωm, 3000Ωm, 4000Ωm, 5000Ωm, 10000Ωm], and the resistivity of the background area is unified as 1000Ωm; the finite element method is used in the forward modeling, in which the 5 layers above the surface are set as air layers, and the frequency value range is 10 -3 ~10 3 Hz logarithmic domain, with a frequency interval of 16, and a total of 10,000 basic sample data sets of different types are generated; each sample contains the apparent resistivity and phase in the xy and yx directions, the Z of the impedance tensor xx , Z xy , Z yx , Z yy Four subfields differ in size, horizontal distribution, burial depth and physical properties according to the random method, thus forming a diverse data set.

[0080] In the first step, we obtain the diversified data set results, and improve the generalization ability through the neighborhood risk minimization (VRM) data enhancement method, and satisfy the forward mapping relationship after the data set and label conversion. Theoretically, the forward mapping relationship still exists when the response and resistivity model are synchronously offset and the relative position remains unchanged. To verify the feasibility of this method, we first flip the model data horizontally and vertically, and then perform three-dimensional forward calculation on the flipped model. The model response data before and after the flip are compared. If the difference is within a reasonable error range, it means that the corresponding relationship of the flipped model in the forward mapping is still valid. Figure 4As shown in the figure, a1 is the resistivity model before flipping, and a2~a5 are its response data; b1~b5 are horizontally flipped model data; c1~c5 are horizontally flipped forward response data; d1~d5 are vertically flipped model data; e1~e5 are forward response data after vertical flipping; the experimental results show that the forward mapping relationship of this method still exists; therefore, the data enhancement method of horizontally flipped images is adopted to expand the data set to twice the original without changing the number of original grids; by horizontally flipping the simple geometric resistivity model, we generate a new resistivity model, significantly expand the training data, and effectively alleviate the difficulty of making a three-dimensional resistivity model data set; in summary, this data enhancement method not only increases the number of samples, but also enriches the types of anomalies, enabling the model to learn more diverse situations, thereby improving its generalization ability.

[0081] Based on the 3D U-Net network framework, a deep learning network structure suitable for three-dimensional magnetotelluric inversion is constructed to realize the nonlinear mapping relationship from input (observation data) to output (3D resistivity model).

[0082] The basic idea of ​​the deep learning inversion method is to train a neural network model to approximate the inversion operator, which can be expressed as:

[0083] m=F -1 (d)≈H(d,θ)

[0084] Where: d and m are three-dimensional matrices of N×M×Z, d is the forward response data, m is the resistivity model parameter, F -1 is the inversion operator, and H is a feedforward neural network with parameters.

[0085] The main goal of this work is to realize a neural network structure suitable for 3D magnetotelluric inversion. The design of the network has a decisive influence on the model inversion.

[0086] The designed deep learning network structure is as follows Figure 1As shown in the figure, the structure is divided into an encoding stage and a decoding stage; in the encoding stage, it contains three encoders, each of which has two convolutional layers for extracting local feature data, followed by batch normalization and ReLU activation, and then down-sampling through a 2×2×2 maximum pooling layer, and finally using a 1×1×1 convolution to enter the decoding stage; the decoding stage contains three upsampling modules, which expand the feature data size through transposed convolution, and then connect a 3×3×3 convolutional layer to decode the feature information; the upsampling result is added to the feature of the encoding stage to retain features at different levels; in addition, a random Dropout layer of 0.1-0.3 is used to prevent overfitting; in the output part, the classification output layer is replaced by a regression output layer to better handle the MT inversion problem, and the generated MT model image has more detailed quantitative information, which is helpful for subsequent analysis.

[0087] The conventional data-driven method is to compare the true model m with the predicted model The model error between is used as the loss objective function and can be expressed as:

[0088]

[0089] Where: N is the number of model parameters, T is the number of samples involved in training, m i，j is the true resistivity value. Once the network is well trained, the feedforward network H can be used as the inversion operator F -1 , predicting from a new set of measurements d m .

[0090] Although data-driven deep learning methods are powerful in solving MT inversion problems, their performance is highly dependent on training data sets that are consistent with the distribution characteristics of the detection area to ensure good generalization ability. To this end, the data error of the 3D forward response is added to the loss function, and a 3D inversion scheme combining data-driven and physics-driven joint weighted loss functions is designed. The 3D forward operation of magnetotelluric is integrated into the neural network training to optimize the inversion effect. The data error can be expressed as:

[0091]

[0092] Where: F represents the three-dimensional forward operator, H represents the feedforward neural network with parameter θ, and d is the forward response data.

[0093] In addition, there is a specific local spatial correspondence between the resistivity model and the response data. When the anomaly is at different horizontal positions, the apparent resistivity data and phase responses will be different, and the greater the depth, the more significant the impact. To address this phenomenon, a depth weighting function is used to offset the natural attenuation of the electrostatic field, thereby enhancing the performance of the network in deep areas and improving the inversion accuracy and resolution of deep anomalies. It can be defined as:

[0094]

[0095] Among them: parameter λ is a constant related to the depth of the anomaly, set to 8, parameter δ is a constant used to control the depth weight distribution, set to 1. The loss function is defined as:

[0096]

[0097] Among them: α and β are the weight coefficients of the two errors.

[0098] In neural network training, the minimization loss function is defined as:

[0099] Loss = argminΦ(d,m,θ)

[0100] In order to achieve better inversion effect, after continuously optimizing the network parameters, we set MaxEpochs to 100, MinibatchSize to 64, and the optimizer to Adam. We adopted a dynamic learning rate adjustment scheme, set the initial learning rate to 0.01, and multiply the learning rate by a decrease factor of 0.5 every 10 epochs.

[0101] In order to better reflect the inversion effect and anti-noise ability, two types of anomaly models are randomly selected from the test set, and 3% and 5% Gaussian random white noise are added to the input data respectively; the inversion results of a single anomaly model are shown in Figure 5 and Figure 6 Show; Figure 5 (a1) and (a2) are actual models; (a2) and (b2) are 0% noise inversion results; (a3) ​​and (b3) are 3% noise inversion results; (a4) and (b4) are 5% noise inversion results; Figure 6 (a1) and (a2) are the actual models; (a2) and (b2) are the inversion results with 0% noise; (a3) ​​and (b3) are the inversion results with 3% noise; (a4) and (b4) are the inversion results with 5% noise. The results show that the network inversion results are basically consistent with the real model. Whether it is a rectangular block anomaly or an inclined step block anomaly, the model can accurately describe the depth and shape of the anomaly, and the resistivity value is also relatively accurate. Especially for the inclined step block anomaly, there are only a few defects on the boundary. When the input data contains 3% or 5% noise, although the overall inversion effect of the model is good and can roughly reflect the basic shape and position of the anomaly, the resistivity value has a certain deviation. This leads to blurred resolution of the anomaly boundary, such as Figure 5 (a4,b3,b4) and Figure 6 (b3,b4).

[0102] The inversion results of the combined anomaly model are as follows: Figure 7 and Figure 8As shown, Figure 7 Figure a1 shows a 3D resistivity model consisting of two rectangular blocks with a resistivity value of 10,000. Regardless of whether there is noise in the inversion results, the model can provide accurate results with high resolution and can clearly show the shape and position. Figure 7 (a1) and (a2) are actual models; (a2) and (b2) are inversion results with 0% noise; (a3) ​​and (b3) are inversion results with 3% noise; (a4) and (b4) are inversion results with 5% noise; Figure 8 a1 in the figure is a three-dimensional resistivity model composed of anomalies of inclined step blocks with resistivity values ​​of 4000 and 10000 respectively; although the inversion results have different degrees of defects, the three inversion results still have high accuracy; Figure 8 Among them, (a1) and (a2) are actual models; (a2) and (b2) are 0% noise inversion results; (a3) ​​and (b3) are 3% noise inversion results; (a4) and (b4) are 5% noise inversion results.

[0103] In order to clearly observe the model inversion results at different depths, Figure 8 The resistivity model of a3 in the figure is sliced ​​at different depth positions (z direction); Fig. 9 (a1-a8) show the real model slices at different depths; (b1-b8) are the inversion results of the model with 0% noise data; (c1-c8) are the inversion results of the model with 3% noise data; (d1-d8) are the inversion results of the model with 5% noise data; when the network input is clean data (0% noise), Fig. 9 The sub-figures (b3, b4) show that although the boundaries of some depth anomalies are not clear enough, other areas are consistent with the true model; when the input data contains 3% or 5% noise, the noise pollution has a greater impact on the model inversion results, such as Fig. 9 The sub-figures in (c2-c5, d2-d8) show that false anomalies and severe deformation of the anomaly shapes occur; however, despite the presence of noise, the network is still able to extract weak signals from the contaminated data and accurately describe the location and basic shape of the anomaly.

[0104] In summary, the deep learning inversion method can accurately describe the location, shape and resistivity value of underground anomalies; this advantage is due to the powerful nonlinear fitting ability of neural networks and the guidance of the introduced weighted loss function based on physical drive and data drive; the neural network fits the inversion operator by continuously mapping the input observation data distribution to a new nonlinear space, similar to the pattern recognition method. When there are a large number of three-dimensional resistivity models in the input data, the network can well learn the characteristic relationship between these data and accurately invert; as the amount of data gradually increases and more data is input into the network training, the model versatility and accuracy are gradually improved; although the neural network may be time-consuming in the training process, once the network is effectively trained, it has extremely high inversion efficiency and is very convenient and efficient in some large data set inversion applications; therefore, the inversion method based on deep learning has good stability, which makes deep learning inversion applicable in real scenarios.

[0105] The present invention adopts a random simulation method to generate a theoretical resistivity model; during training, domain risk minimization is used to perform data enhancement to improve the generalization ability of network training; based on a 3D U-Net network framework, a network structure suitable for three-dimensional magnetotelluric inversion is constructed; a data-driven and physically-driven joint weighted loss function is designed so that the network can follow magnetotelluric physical constraints during training, improve the inversion accuracy and resolution of deep anomalies, and solve the optimization problem by minimizing the loss function; the neural network hyperparameter settings are adjusted through inversion experiments on theoretical model data to obtain final training parameters; synthetic data and noise experiments are used to test the inversion effect; the results show that the inversion effect of the present invention has strong noise resistance and high resolution, and has practical application value in deep learning inversion methods.

[0106] The present invention realizes end-to-end imaging of 3D structures from network input (apparent resistivity and phase) to output (resistivity model), introduces a weighted loss function based on data-driven and physical-driven joint, integrates the three-dimensional forward modeling operation of magnetotelluric into the forward propagation of the network, so that the network can follow the magnetotelluric physical constraints during training, and further adjusts the network parameters through data enhancement methods and numerical experiments. This method can effectively improve the inversion resolution and anomaly boundary information, and has practical application value in magnetotelluric inversion.

[0107] The embodiment of the present invention further provides a three-dimensional magnetotelluric inversion device, comprising:

[0108] The numerical module is used to obtain the magnetotelluric data to be inverted and to forward model the magnetotelluric data to obtain multiple response data.

[0109] Model module, used to use 3D U-Net network as magnetotelluric inversion network model.

[0110] A data-driven loss function is constructed, and the data error representing the physically driven three-dimensional forward response is added to the data-driven loss function to construct a data-driven and physically driven joint weighted loss function; the cross entropy loss function in the magnetotelluric inversion network model is replaced by the data-driven and physically driven joint weighted loss function to obtain an improved magnetotelluric inversion network model.

[0111] The improved magnetotelluric inversion network model is trained using multiple response data. During the training, a joint weighted loss function driven by data and physics is applied to constrain the model, and the trained magnetotelluric inversion network model is obtained.

[0112] The inversion module is used to input the response data to be measured into the trained magnetotelluric inversion network model to obtain the three-dimensional resistivity distribution parameters of the area to be measured to realize three-dimensional magnetotelluric inversion.

[0113] An embodiment of the present invention further provides an electronic device, including a memory and a processor.

[0114] The memory is used to store computer programs.

[0115] When the processor is used to execute the computer program stored in the memory, the steps of the above three-dimensional magnetotelluric inversion method are implemented.

[0116] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A three-dimensional magnetotelluric inversion method, characterized in that: The following steps are involved: Obtain the magnetotelluric data to be inverted, and perform forward modeling on the magnetotelluric data to obtain multiple response data; Use 3D U-Net as the magnetotelluric inversion network model; A data-driven loss function is constructed, and the data error representing the physically driven three-dimensional forward response is added to the data-driven loss function to construct a data-driven and physically driven joint weighted loss function; the cross entropy loss function in the magnetotelluric inversion network model is replaced with a data-driven and physically driven joint weighted loss function to obtain an improved magnetotelluric inversion network model; The improved magnetotelluric inversion network model is trained using multiple response data. The joint weighted loss function driven by data and physics is used for constraint during the training to obtain the trained magnetotelluric inversion network model. The response data to be measured is input into the trained magnetotelluric inversion network model to obtain the three-dimensional resistivity distribution parameters of the area to be measured to realize three-dimensional magnetotelluric inversion.

2. A three-dimensional magnetotelluric inversion method according to claim 1, characterized in that: The obtaining of multiple response data includes: According to the detection area and magnetotelluric data, a theoretical resistivity model of the detection area is generated by random simulation; the anomaly body in the detection area is forward modeled in the theoretical resistivity model to obtain the response data of the theoretical resistivity model including apparent resistivity and phase; In the response data containing apparent resistivity and phase, each response data contains the apparent resistivity and phase in the xy and yx directions, and also contains the impedance tensor Z xx , Z xy , Z yx , Z yy to form a data set of response data.

3. A three-dimensional magnetotelluric inversion method according to claim 1, characterized in that: The construction of the data-driven and physics-driven joint weighted loss function includes: The data error representing the three-dimensional forward response of physical drive is added to the loss objective function in data-driven, and the evolution process of the model in model training is kept consistent with the evolution of the physical environment to construct a joint weighted loss function of data-driven and physical drive.

4. A three-dimensional magnetotelluric inversion method according to claim 3, characterized in that: The joint weighted loss function of data-driven and physics-driven is: in: α and β are the weight coefficients of the two errors respectively; represents the depth weighting function; Φ m (d, m, θ) represents the data-driven loss function; Φ d (d, θ) represents the data error of the physically driven three-dimensional forward response.

5. A three-dimensional magnetotelluric inversion method according to claim 1, characterized in that: The acquisition of the three-dimensional resistivity distribution parameters of the area to be measured includes: Inputting the response data to be tested into the trained magnetotelluric inversion network model, and using the encoder of the magnetotelluric inversion network model to extract features of the response data to obtain local feature data; The decoder is used to extract local features to obtain deep feature data, and the local feature data is superimposed with the feature elements in the deep feature data to obtain final feature data, and the three-dimensional resistivity distribution parameters of the area to be measured are identified from the final feature data.

6. A three-dimensional magnetotelluric inversion method according to claim 5, characterized in that: After the three-dimensional resistivity distribution parameters of the area to be measured are identified, the position, shape and resistivity value data of the abnormal body in the area to be measured are simulated according to the three-dimensional resistivity distribution parameters to form a three-dimensional magnetotelluric inversion result.

7. A three-dimensional magnetotelluric inversion device, characterized in that: include: The numerical module is used to obtain the magnetotelluric data to be inverted and to forward model the magnetotelluric data to obtain multiple response data; Model module, used to use 3D U-Net network as magnetotelluric inversion network model; A data-driven loss function is constructed, and the data error representing the physically driven three-dimensional forward response is added to the data-driven loss function to construct a data-driven and physically driven joint weighted loss function; the cross entropy loss function in the magnetotelluric inversion network model is replaced with a data-driven and physically driven joint weighted loss function to obtain an improved magnetotelluric inversion network model; The improved magnetotelluric inversion network model is trained using multiple response data. The joint weighted loss function driven by data and physics is used for constraint during the training to obtain the trained magnetotelluric inversion network model. The inversion module is used to input the response data to be measured into the trained magnetotelluric inversion network model to obtain the three-dimensional resistivity distribution parameters of the area to be measured to realize three-dimensional magnetotelluric inversion.

8. An electronic device, characterized in that: include: Memory and processor; The memory is used to store computer programs; The processor is used to implement the steps of a three-dimensional magnetotelluric inversion method as described in any one of claims 1 to 6 when executing the computer program stored in the memory.

Images