Transient electromagnetic method self-supervised inversion method based on deep learning

Abstract

The invention relates to a transient electromagnetic method self-supervised inversion method based on deep learning, and the method comprises the steps: inputting to-be-inverted electromagnetic response data into a pre-trained inversion network for inversion, and obtaining an inversion resistivity value; inputting the inversion resistivity value into a pre-trained forward modeling network for forward modeling to obtain forward modeling electromagnetic response data; updating parameters of an inversion network by minimizing loss between the output forward modeling electromagnetic response data and the input electromagnetic response data to be inverted; and performing inversion on the electromagnetic response data to be inverted again by using the inversion network with updated parameters until a preset number of iterations are reached, and finally outputting a resistivity value obtained by inversion. According to the method provided by the invention, the accuracy and the high efficiency of the inversion process are ensured while the dependence on the labeled data and the initial model is reduced, and a new thought is provided for the development of the transient electromagnetic inversion technology.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic exploration, and particularly to a self-supervised inversion method for transient electromagnetic method based on deep learning. Background Art

[0002] As an important electromagnetic exploration technology, the inversion process of the transient electromagnetic method (TEM) plays a crucial role in TEM data processing and interpretation. Currently, the inversion methods of TEM are mainly divided into linear inversion and non-linear inversion. Linear inversion has a large dependence on the initial model and is prone to falling into local optimal solutions; while non-linear inversion usually requires a large amount of computing resources. With the continuous improvement of the performance of computing devices, deep learning methods have gradually been applied to the fast inversion imaging of electrical methods and achieved remarkable progress.

[0003] Currently, in the electrical method inversion based on deep learning, the data-driven end-to-end inversion has been widely used in the field of electromagnetic inversion due to its advantages such as easy construction of the network model and simple training process. This data-driven end-to-end inversion method gives full play to the powerful non-linear mapping ability of deep learning, directly maps the electromagnetic response data to the resistivity model, and thus effectively trains an inversion network with high efficiency. Compared with traditional inversion methods, this end-to-end deep learning inversion can achieve fast inversion through data learning without manually designing the inversion process.

[0004] With the proposal of the physics-informed neural network, many geophysicists have begun to try to incorporate physical knowledge and prior information into deep learning models as constraint conditions to further reduce the dependence on large-scale data sets and improve the accuracy and generalization ability of the network. The core idea of PINN is to introduce physical laws (such as electromagnetic field equations, etc.) as regularization terms, and let the network follow physical constraints during the training process, so as to avoid the overfitting problem that may occur in traditional deep learning methods and still maintain high accuracy in the case of small data sets.

[0005] With the continuous maturity of deep learning technology, self-supervised learning has gradually become an important research direction in the field of deep learning, especially when facing a large amount of unlabeled data, its potential has been widely applied. The self-supervised full waveform inversion framework based on the autoencoder network drives the update of the network by using the loss function of seismic response instead of the traditional wave velocity model loss, thus realizing the fast training of deep learning inversion in the absence of labeled data. This method not only reduces the dependence on labeled data, but also provides an effective solution for processing large-scale and unlabeled geophysical data.

[0006] The data-driven end-to-end inversion has certain advantages over traditional inversion in terms of computational speed and accuracy. However, it has a strong dependence on high-quality training sets, which limits its generalization ability. There are often significant deviations between the labels generated by numerical simulation and the actual geological model, which directly affects the inversion accuracy of the network. The performance of the network is not only limited by the quality of the training data but also restricted by the diversity and representativeness of the data. Therefore, in the absence of a sufficiently diverse and high-quality data set, the effect of end-to-end inversion may not meet expectations.

[0007] For the physics-informed deep learning method, since the loss function needs to involve computationally time-consuming forward simulations, it significantly increases the time cost of network training. Although this method successfully introduces physical constraints into the network model, it fails to effectively transmit the gradient of the electromagnetic response loss. This to some extent limits the further improvement of the model in terms of training efficiency and performance, especially when dealing with complex geological models and large-scale data sets, where high computational overhead and training difficulty still need to be overcome.

[0008] In response to this, some scholars have combined the data-driven end-to-end inversion with traditional inversion methods, introducing physical information into the deep learning network through the physical drive in traditional inversion as a supervision signal to guide the training of the network. This method makes full use of the physical constraints in traditional inversion to ensure that the network follows geophysical laws during the training process. However, this combination method relies on traditional inversion methods as the source of physical information, and traditional inversion usually involves a large amount of forward calculations, resulting in a relatively slow inversion process. Therefore, although this method can effectively improve the inversion accuracy of the model, its slow computational speed inevitably increases the time cost of deep learning network training, especially when dealing with large-scale data sets and complex geological models.

[0009] In the field of electrical prospecting, the application of self-supervised learning is still relatively limited. The main reason is that the current seismic full-waveform forward platform based on PyTorch is relatively mature and can be effectively combined with the self-supervised inversion framework of deep learning. However, the current forward calculation platform in electromagnetic prospecting, especially in the transient electromagnetic (TEM) method, is still in a blank stage, which makes it difficult to effectively apply it to the self-supervised inversion framework of deep learning. Although relevant forward functions have been written through PyTorch and self-supervised geophysical joint inversion has been achieved, due to the high complexity of actual forward calculations, there is a large deviation between the calculated response and the actual response, so it has not been widely applied in field-measured data. In addition, existing self-supervised deep learning inversion methods usually require an initial model as input to better drive network updates, and this approach increases the dependence on the initial model to a certain extent. Summary of the Invention

[0010] The object of the present invention is a self - supervised inversion method for transient electromagnetic method based on deep learning to solve the problems existing in the above - mentioned prior art, which is different from the traditional end - to - end inversion method. By directly inputting the measured electromagnetic response data, the loss between the measured electromagnetic response and the electromagnetic response corresponding to the resistivity model output by the network is minimized to optimize the network model, and finally the corresponding resistivity model is output. This method avoids the dependence on numerical simulation training data, and has stronger generalization ability when inversing the measured data.

[0011] To achieve the above object, the present invention provides the following solutions:

[0012] A self - supervised inversion method for transient electromagnetic method based on deep learning, comprising:

[0013] Input the electromagnetic response data to be inverted into a pre - trained inversion network for inversion to obtain the inverted resistivity value;

[0014] Input the inverted resistivity value into a pre - trained forward - modeling network for forward modeling to obtain the forward - modeled electromagnetic response data;

[0015] Update the parameters of the inversion network by minimizing the loss between the output forward - modeled electromagnetic response data and the input electromagnetic response data to be inverted;

[0016] Use the inversion network with updated parameters to re - invert the electromagnetic response data to be inverted until after a preset number of iterations, and finally output the inverted resistivity value.

[0017] Optionally, the inversion network includes: an encoder part, a decoder part, and a smooth constraint part;

[0018] Inputting the electromagnetic response data to be inverted into a pre - trained inversion network for inversion includes:

[0019] Input the electromagnetic response data to be inverted into the encoder part, and perform feature extraction on the input electromagnetic response through three convolutional layers, where the convolutional layer includes a 1D convolutional layer + a pooling layer + a Tanh activation function;

[0020] In the decoder part, restore and re - process the features obtained by the encoder part through three feature up - sampling layers, where the feature up - sampling layer includes one 1D up - sampling layer and two 1D convolutional layers;

[0021] In the smooth constraint part, further smooth - constraint the output of the decoder part through a convolutional layer with a convolution kernel of 1×1, and finally output the inverted resistivity value.

[0022] Optionally, inputting the inverted resistivity value into a pre - trained forward - modeling network for forward modeling includes:

[0023] Extract features from the inverted resistivity values through two convolutional layers to capture local electromagnetic response information; wherein, the convolutional layer includes a 1D convolutional layer + a pooling layer + a Tanh activation function;

[0024] Integrate the features extracted by the convolutional layer based on three fully connected layers to extract the forward electromagnetic response data represented in the resistivity values.

[0025] Optionally, the pre-trained inversion network includes:

[0026] Generate resistivity values and corresponding electromagnetic response data through numerical simulation methods, and construct a dataset based on the resistivity values and corresponding electromagnetic response data;

[0027] Preprocess the dataset;

[0028] Use the preprocessed dataset to pre-train the inversion network; in the pre-training of the inversion network, the electromagnetic response and resistivity value in the dataset are used as the input and label of the inversion network respectively; after the inversion network outputs the predicted resistivity, calculate the error between it and the label through a loss function and drive parameter update to complete the pre-training of the inversion network.

[0029] Optionally, the pre-trained forward network includes:

[0030] Train the forward network based on the preprocessed dataset; in the pre-training of the forward network, the resistivity value and electromagnetic response in the dataset are used as the input and label of the forward network respectively; after the forward network outputs the predicted electromagnetic response, calculate the error between it and the label through a loss function and drive parameter update to complete the pre-training of the forward network.

[0031] Optionally, generating resistivity values and corresponding electromagnetic response data through numerical simulation methods includes:

[0032] Set the basic parameters of the underground formation; including: the thickness of the top layer and the boundary depth of the last layer;

[0033] Divide the formation into several layers according to the method of equal thickness for each layer;

[0034] Set several control points within the preset formation depth range, where one control point is set at each of the top and bottom representing the formation;

[0035] Randomly generate resistivity values for each control point and perform interpolation using the B-spline interpolation method;

[0036] Generate resistivity-thickness sample data;

[0037] Using a central loop device, substitute the generated resistivity-thickness sample data into the transient electromagnetic one-dimensional numerical simulation method to calculate the corresponding electromagnetic response data.

[0038] Optionally, inputting the electromagnetic response data to be inverted into the pre-trained inversion network for inversion further includes:

[0039] Based on the self-supervised inversion method, conduct an inversion test on the electromagnetic response data to be inverted; wherein, the self-supervised inversion method is based on the pre-trained inversion network; input the electromagnetic response data to be inverted into the pre-trained inversion network, the network outputs the initially predicted resistivity value, and then based on the fast forward training strategy, perform a secondary update on the pre-trained forward network, that is, use the forward numerical simulation method to solve the true electromagnetic response corresponding to the predicted resistivity value, and use it and the predicted resistivity value as new training samples to perform a secondary training update on the forward network; recalculate the corresponding electromagnetic response based on the updated forward network, and then calculate the error between it and the input electromagnetic response through the loss function to drive the update of the inversion network parameters; use the inversion network with updated parameters to re-invert the electromagnetic response data to be inverted until after reaching the preset number of iterations, and finally output the inverted resistivity value.

[0040] The beneficial effects of the present invention are:

[0041] By performing a secondary training on the forward network, using the true forward numerical simulation results to solve the electromagnetic response, and using the resistivity model output by the inversion network as new training samples, it is ensured that the inversion network can gradually update the resistivity model under the correct electromagnetic response error;

[0042] By adding a convolutional smoothing layer, a resistivity inversion result closer to the actual situation is achieved (as shown in the appendix Figure 7 ). By adding a relevant convolutional layer to the last layer of the network model and avoiding the use of a fully connected layer, a resistivity inversion result that is smoother and closer to the actual situation is finally achieved.

[0043] Compared with the traditional and mature Ocaam inversion method, this method shows higher accuracy in the inversion of measured data, specifically reflected in that the electromagnetic response of the inverted resistivity model is closer to the measured electromagnetic response. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0045] Figure 1 Comparison diagram of the self-supervised inversion method and the end-to-end inversion method based on deep learning according to the embodiments of the present invention;

[0046] Figure 2 Flow chart of training the self-supervised inversion network according to the embodiments of the present invention;

[0047] Figure 3 Structural diagram of the fast forward and inversion networks according to the embodiments of the present invention;

[0048] Figure 4 Example diagram of the resistivity model of a part of the data set according to the embodiments of the present invention;

[0049] Figure 5 Comparison diagram of test results of different inversion methods on a random electrical model according to the embodiments of the present invention; among them, (a)-(c) are comparison diagrams of resistivity model inversion; figures (d)-(e) are respectively the corresponding electromagnetic response fitting diagrams;

[0050] Figure 6 Comparison diagram of inversion results under different forward network training strategies according to the embodiments of the present invention; among them, (a) and (c) are respectively the self-supervised inversion resistivity result diagrams with and without secondary training of the forward network; (b) and (d) are respectively the corresponding average relative error diagrams of the electromagnetic response;

[0051] Figure 7 Comparison diagram of results of different structural inversion networks according to the embodiments of the present invention; among them, (a) is the resistivity diagram inverted by the network structure of this embodiment, and (b) is the resistivity inverted by the traditional convolutional + fully connected layer network structure;

[0052] Figure 8 Resistivity result diagram and electromagnetic response fitting comparison diagram corresponding to two inversion results of the measured data at the measurement point P1 according to the embodiments of the present invention; among them, (a) is the comparison diagram of the inversion resistivity at the measurement point P1 by the Occam inversion method and the self-supervised inversion method, and (b) is the comparison diagram of the electromagnetic response corresponding to the inversion resistivity at the measurement point P1 by the Occam inversion method and the self-supervised inversion method and the measured electromagnetic response; Detailed implementation manners

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] As Figure 1 shown, this embodiment proposes a self-supervised inversion method for transient electromagnetic method based on deep learning, including:

[0056] Input the electromagnetic response data to be inverted into a pre-trained inversion network for inversion to obtain the inverted resistivity value;

[0057] Input the inverted resistivity value into a pre-trained forward network for forward modeling to obtain the forward electromagnetic response data;

[0058] Update the parameters of the inversion network by minimizing the loss between the output forward electromagnetic response data and the input electromagnetic response data to be inverted;

[0059] Use the inversion network with updated parameters to re-invert the electromagnetic response data to be inverted until after a preset number of iterations, and finally output the inverted resistivity value.

[0060] As shown in the Figure 2 accompanying drawings, the process of the self-supervised inversion method proposed in this embodiment is as follows: Input the electromagnetic response data (electromagnetic response) into the pre-trained inversion network (pre-trained inversion network), and then invert to obtain the underground resistivity model (predicted resistivity value); then calculate the forward electromagnetic response of the predicted resistivity value through the forward part. In the forward part, first solve the true electromagnetic response solution through numerical simulation methods, and then use the true electromagnetic response solution and the predicted resistivity value l as a new data set to perform secondary training on the pre-trained forward network. After that, calculate the response (output response) of the resistivity model through the updated forward network; at the same time, calculate the error loss between the output response and the electromagnetic response to update the parameters of the pre-trained inversion network. By iterating the above process until the preset number of iterations is reached, and output the final inversion result.

[0061] Furthermore, the inversion network includes: an encoder part, a decoder part, and a smoothing constraint part;

[0062] Inputting the electromagnetic response data to be inverted into the pre-trained inversion network for inversion includes:

[0063] Input the electromagnetic response data to be inverted into the encoder part, and extract features of the input electromagnetic response through three convolutional layers. Among them, the convolutional layer includes a 1D convolutional layer + a pooling layer + a Tanh activation function;

[0064] In the decoder part, the features obtained from the encoder part are restored and reprocessed through three layers of feature upsampling layers. Among them, the feature upsampling layer includes one 1D upsampling layer and two 1D convolutional layers;

[0065] In the smooth constraint part, a convolutional layer with a convolution kernel of 1×1 is used to further smooth-constrain the output of the decoder part, and finally the inverted resistivity value is output.

[0066] Further, inputting the inverted resistivity value into the pre-trained forward modeling network for forward modeling includes:

[0067] Extracting features from the inverted resistivity value through two convolutional layers to capture local electromagnetic response information; among them, the convolutional layer includes a 1D convolutional layer + pooling layer + Tanh activation function;

[0068] Integrating the features extracted by the convolutional layer based on three fully connected layers to extract the forward electromagnetic response data represented in the resistivity value.

[0069] Specifically, in this embodiment, the fast forward modeling network: A 1D transient electromagnetic fast forward modeling network is built through a structure combining a convolutional layer and a fully connected layer. By combining the local feature extraction ability of the convolutional layer and the global information integration ability of the fully connected layer, the solution accuracy of the electromagnetic response and the computational efficiency of the network are improved. As shown in the attached Figure 3 In the forward modeling network part, input the resistivity model, and then extract features from the resistivity model through two convolutional layers to capture local electromagnetic response information, where the convolutional layer is composed of a 1D convolutional layer + pooling layer + Tanh activation function. Finally, the features extracted are integrated based on three fully connected layers, so as to effectively extract the electromagnetic response represented in the resistivity model and output it.

[0070] Fast inversion network: The fast inversion network is built based on the fully convolutional layer of the autoencoder structure. As shown in the attached Figure 3 In the inversion network part, in the encoder part of the network, input the electromagnetic response, and then extract features from the input electromagnetic response through three convolutional layers, where the convolutional layer is composed of a 1D convolutional layer + pooling layer + Tanh activation function; while the decoder part restores and reprocesses the features obtained from the encoder through three layers of feature upsampling layers to avoid using fully connected layers, where the feature upsampling layer consists of one 1D upsampling layer and two 1D convolutional layers. In addition, at the end of the decoder network, a convolutional layer with a convolution kernel of 1×1 is used to further smooth-constrain the result, and finally the resistivity model is output. This design can not only retain the smoothness of the resistivity model, but also help to be consistent with the actual geological model, thereby improving the accuracy and stability of the inversion result.

[0071] Further, the pre-trained inversion network includes:

[0072] Generate resistivity values and corresponding electromagnetic response data through numerical simulation methods, and construct a data set based on the resistivity values and corresponding electromagnetic response data;

[0073] Preprocess the data set;

[0074] Use the preprocessed data set to pre-train the inversion network; in the pre-training of the inversion network, the electromagnetic response and resistivity value in the data set are used as the input and label of the inversion network respectively; after the inversion network outputs the predicted resistivity, calculate the error between it and the label through the loss function and drive parameter update to complete the pre-training of the inversion network.

[0075] Furthermore, the pre-training of the forward network includes:

[0076] Train the forward network based on the preprocessed data set; in the pre-training of the forward network, the resistivity value and electromagnetic response in the data set are used as the input and label of the forward network respectively; after the forward network outputs the predicted electromagnetic response, calculate the error between it and the label through the loss function and drive parameter update to complete the pre-training of the forward network.

[0077] Furthermore, generating resistivity values and corresponding electromagnetic response data through numerical simulation methods includes:

[0078] Set the basic parameters of the underground formation; including: the thickness of the top layer and the boundary depth of the last layer;

[0079] Divide the formation into several layers according to the method of equal thickness for each layer;

[0080] Set several control points within the preset formation depth range, where one control point is set at the top and bottom representing the formation;

[0081] Randomly generate resistivity values for each control point and perform interpolation using the B-spline interpolation method;

[0082] Generate resistivity-thickness sample data;

[0083] Adopt a central loop device, substitute the generated resistivity-thickness sample data into the transient electromagnetic one-dimensional numerical simulation method, and calculate the corresponding electromagnetic response data.

[0084] Furthermore, inputting the electromagnetic response data to be inverted into the pre-trained inversion network for inversion also includes:

[0085] Based on the self-supervised inversion method, perform inversion tests on the electromagnetic response data to be inverted. Among them, the self-supervised inversion method is based on a pre-trained inversion network. Input the electromagnetic response data to be inverted into the pre-trained inversion network, and the network outputs the initially predicted resistivity value. Then, based on the fast forward training strategy, perform a secondary update on the pre-trained forward network, that is, use the forward numerical simulation method to solve the true electromagnetic response corresponding to the predicted resistivity value, and use it and the predicted resistivity value as new training samples to perform a secondary training update on the forward network. Recalculate the corresponding electromagnetic response based on the updated forward network, and then calculate the error between it and the input electromagnetic response through the loss function to drive the update of the inversion network parameters. Use the inversion network with updated parameters to re-invert the electromagnetic response data to be inverted until after a preset number of iterations, and finally output the inverted resistivity value.

[0086] Specifically, in this embodiment, the generation of the dataset used by the pre-trained network includes:

[0087] The quality of the dataset is very important for the performance of the pre-trained forward and inversion networks. Considering that the resistivity values of the actual underground strata are usually continuously distributed rather than segmented, and the continuous distribution characteristics can well reflect the complexity of the underground geological structure. Therefore, in order to generate a relatively smooth and longitudinally continuous underground resistivity model, the following method is used to generate the sample set data: First, set the top layer thickness to 1m, the boundary depth of the last layer to 100m, and divide it into 40 layers according to the method of equal thickness per layer to ensure that the random characteristics of the underground resistivity distribution are fully reflected. Determine 5 control points between 1 and 100m, where the depths of 2 control points are fixed at 0m and 100m, and the depths of the remaining 3 control points are randomly determined, but it is required that the depth between the control points is greater than 10m. The resistivity of the control points is randomly generated within 1 to 1000 Ω·m, and interpolation is performed using the B-spline interpolation method based on the 5 control points, so that the resistivity curve is smoother, and finally the resistivity values continuously distributed from 1m to 100m are obtained. Finally, using the central loop device, substitute the generated resistivity-thickness sample data into the transient electromagnetic one-dimensional numerical simulation method to calculate the corresponding electromagnetic response. Among them, the radius of the circular coil is 1.25m, the number of receiving time channels is 61, the sampling time is 1e -5 ~1e -2 s, and the magnitude of the transmitted current is 1A. Finally, 200 training sets are generated for pre-training the forward and inversion networks (part of the data is as Figure 4 shown).

[0088] To avoid the problem that the network training convergence speed is too slow due to the large scale amplitude of the electromagnetic response data, preprocessing operations of normalization are performed on the input and output data.

[0089] Specifically, in this embodiment, the performance test of the self-supervised inversion network:

[0090] During the pre-training process of the forward and inverse networks, the Adam optimizer is adopted, the learning rate is set to 1e-3, the number of training iterations (epochs) is 500, the batch size is 64, and the mean absolute error (MAE) is used as the loss function. After the forward and inverse networks are pre-trained, the measured data is inversely tested based on the self-supervised inversion method. In this process, the self-supervised inversion network is based on the pre-trained inverse network.

[0091] In addition, in order to verify the efficiency of the self-supervised inversion method, a data-driven deep learning inversion network is also trained. Except for the different number of training sets, the network model and training parameters of the data-driven inversion network are the same as those of the self-supervised inversion method. The data-driven inversion network is trained with 20,000 data sets, and the ratio of the training set to the validation set is 9:1.

[0092] The above networks are trained based on the TensorFlow 2.1 platform. The computing environment configuration of the experiment is: an Intel(R) Core(TM) i5-10300H CPU@2.50GHz processor with 8GB of memory.

[0093] To verify the advantages of the self-supervised inversion method proposed in this patent, three geoelectric models are regenerated for testing, and the results are compared with the results of the data-driven inversion method and the iteratively reweighted least squares (IRLS) inversion results of the linear inversion method (as shown in the appendix Figure 5 ). It can be seen from the results that compared with the inversion results of the data-driven inversion method, the inversion results of the self-supervised inversion method have been significantly improved in terms of the resistivity model fitting degree (× line) and the electromagnetic response data (square) fitting degree. At the same time, compared with the IRLS inversion results, the results obtained by the self-supervised network inversion also show better improvement, further proving the effectiveness and advantages of the self-supervised inversion method in resistivity model inversion. In addition, in the three examples, the calculation times of the self-supervised inversion method and the traditional IRLS inversion method are 40 seconds and 61 seconds respectively. It can be seen from this that the self-supervised inversion method is superior to the traditional IRLS inversion method in terms of inversion speed. Although the calculation time of the data-driven inversion method is about 1s, the generation of the training set in the data-driven inversion method is extremely time-consuming, about 0.5h, and this part of the time should also be considered in the inversion calculation time.

[0094] Specifically, in this embodiment, the fast forward network training strategy:

[0095] As shown in the appendix Figure 6, the self - supervised inversion was performed using the unupdated forward - modeling network and the forward - modeling network updated by secondary training respectively, and the accuracy of the finally output resistivity model was evaluated (the dotted line is the true resistivity model, and the solid line is the resistivity result map of the self - supervised inversion). At the same time, the deviation between the electromagnetic response generated based on the forward - modeling network and the electromagnetic response generated by the numerical simulation method was also compared to judge the improvement of the secondary training of the forward - modeling network on the inversion result (the dots are the electromagnetic response error curves obtained by numerically simulating the inversion result, and the plus signs are the electromagnetic response error curves obtained by using the fast forward modeling). It can be clearly observed that the inversion accuracy has been significantly improved by updating the forward - modeling network. The updated forward - modeling network has significantly improved the fitting degree of the resistivity model and effectively reduced the deviation between the generated electromagnetic response and the numerical simulation result. This shows that the secondary training of the forward - modeling network not only improves the inversion accuracy of the resistivity model but also enhances the reliability and consistency of the inversion result.

[0096] In addition, when performing secondary training on the forward - modeling network, since additional numerical simulations are required, additional computational consumption will be introduced. However, through experimental tests, it can be found that in the self - supervised inversion network, the computational times for the forward - modeling network with or without secondary update are 46 seconds and 16 seconds respectively. Although the time consumption has doubled, the inversion accuracy has been significantly improved. Considering the significant improvement in inversion accuracy, this additional computational time consumption is acceptable.

[0097] Specifically, in this embodiment, comparison of the fast inversion network structures:

[0098] In this part, the results of the inversion network of this patent and the traditional inversion network (the decoder part and the convolutional output layer of the fully convolutional network in the auto - encoder network are all replaced by fully - connected layers, and other layers remain unchanged) were compared. The results are as shown in the appendix Figure 7 . It can be found that obvious jumps occurred between different layers of the resistivity model (dotted line) inverted by the fully - connected layer, which is inconsistent with the true model, while the resistivity model (solid line) inverted by the network of this patent is still relatively smooth between adjacent layers. In summary, replacing the fully - connected layer with a convolutional layer can perform good smooth constraints on the resistivity model and obtain a continuous resistivity model.

[0099] Specifically, in this embodiment, measured data:

[0100] The transient electromagnetic dataset from the Umatilla Indian Reservation Geothermal Resources Assessment was used to verify the efficiency of this method. This dataset was collected by the U.S. Geological Survey in the areas of interest for geothermal potential, and it contains 23 ground transient electromagnetic data. TEM measurements were carried out along 4 profiles using a central loop device with 100x100m or 40x40m.

[0101] Inversion was performed on the measuring point P1 among them, and the results were compared with the traditional Ocaam inversion method (provided by Jared R. Peacock and Jeffrey D. Pepin) (as shown in the appendix Figure 8 ). The inversion results show that the resistivity models obtained by the two methods are basically the same, but in the comparison of the electromagnetic response curves, the self-supervised inversion method (triangle) shows a higher precision fitting degree.

[0102] In the synthetic data, in this embodiment (self-supervised inversion), compared with the traditional DL data-driven inversion method and the iterative reweighted least squares (IRLS) inversion method of the traditional linear inversion method, significant improvements are shown in both the resistivity model fitting degree and the electromagnetic response data fitting degree, as shown in the appendix Figure 5 . In addition, this embodiment also has a significant advantage in terms of calculation time. The time consumed by this embodiment, the traditional data-driven inversion method, and the IRLS method are 40 seconds, 1 second (but the dataset generation takes 0.5h and also needs to be considered), and 61 seconds respectively.

[0103] By retraining the forward network, using the real forward numerical simulation results to solve the electromagnetic response, and taking the resistivity model output by the inversion network as a new training sample, it is ensured that the inversion network can gradually update the resistivity model under the correct electromagnetic response error (as shown in the appendix Figure 6 ).

[0104] By adding a convolutional smoothing layer, a resistivity inversion result closer to the real situation is achieved (as shown in the appendix Figure 7 ). By adding a relevant convolutional layer to the last layer of the network model and avoiding the use of a fully connected layer, a more smooth and real-situation-approaching resistivity inversion result is finally achieved.

[0105] Compared with the traditional and mature Ocaam inversion method, this embodiment shows higher precision in the inversion of measured data, specifically reflected in that the electromagnetic response of the resistivity model obtained by inversion is closer to the measured electromagnetic response, as shown in the appendix Figure 8 .

[0106] As shown in the appendix Figure 1As shown in the figure, it is a comparison chart of the self-supervised inversion method and the end-to-end inversion method based on deep learning. The upper part is the process of the traditional data-driven deep learning inversion method, and the lower part is the process of this method - the self-supervised inversion method.

[0107] As shown in the appendix Figure 2 As shown in the figure, it is a flowchart of the self-supervised inversion network training. First, the electromagnetic response data is input into the pre-trained inversion network, and then the resistivity model predicted by the network is forward modeled. Then, the inversion network parameters are updated by minimizing the loss between the output electromagnetic response and the input electromagnetic response. After a certain number of iterations, the finally output resistivity model obtained by inversion is output.

[0108] As shown in the appendix Figure 3 As shown in the figure, it is a structural diagram of the fast forward modeling and inversion network. The upper side is the structural diagram of the fast forward modeling network, which is constructed based on convolutional layers and fully connected layers. The lower side is the structural diagram of the fast inversion network, which is based on the encoder-decoder structure and adopts the structure of fully convolutional layers, especially using one-dimensional convolution in the output layer.

[0109] As shown in the appendix Figure 4 As shown in the figure, it is an example diagram of the resistivity model of some datasets.

[0110] As shown in the appendix Figure 5 As shown in the figure, it is a comparison chart of the test results of different inversion methods on a random geoelectric model. It includes data-driven inversion, self-supervised inversion, and IRLS inversion. Figure 5 (a) - (c) of [reference] are comparison charts of resistivity model inversion; Figure 5 (d) - (e) of [reference] are respectively the corresponding electromagnetic response fitting charts.

[0111] As shown in the appendix Figure 6 As shown in the figure, it is a comparison chart of inversion results under different forward modeling network training strategies. Figure 6 (a) and (c) of [reference] are respectively the self-supervised inversion resistivity result charts with or without secondary training of the forward modeling network; Figure 6 (b) and (d) of [reference] are respectively the corresponding average relative error charts of electromagnetic responses.

[0112] As shown in the appendix Figure 7 As shown in the figure, it is a comparison chart of inversion results of different structural inversion networks. Figure 7 (a) of [reference] is the resistivity chart inverted by the network structure of this paper, Figure 7 (b) of [reference] is the resistivity inverted by the traditional convolutional + fully connected layer network structure.

[0113] As shown in the appendix Figure 8 As shown in the figure, it is a resistivity result chart and a comparison chart of electromagnetic response fitting corresponding to two inversion results of the measured data at measuring point P1.

[0114] This embodiment proposes a transient electromagnetic inversion method based on self-supervised deep learning. First, based on a data-driven end-to-end network architecture, this embodiment designs a network model dedicated to one-dimensional forward and inverse transient electromagnetic. This architecture includes two key components: a forward network and an inverse network. The main task of the forward network is to effectively integrate the transient electromagnetic forward process into the self-supervised inversion framework to ensure that the inversion model can generate accurate electromagnetic responses while enabling efficient backpropagation of gradients. The inverse network is responsible for ensuring that the resistivity model in the initial inversion result conforms to a reasonable geological morphology, thereby providing accurate calculation results for the forward network and further improving the accuracy and reliability of the inversion.

[0115] On this basis, this embodiment introduces the concept of transfer learning and constructs a self-supervised framework based on the inversion network model. Different from the traditional training method based on the loss of the resistivity model, this embodiment proposes to drive network optimization through electromagnetic response loss, enabling the network to directly learn the mapping relationship from electromagnetic responses to resistivity models. Specifically, the calculation of the forward response is implemented through the forward network to ensure that the calculation results can effectively support the backpropagation of gradients, thereby maintaining optimization efficiency and accuracy throughout the training process. In addition, to ensure the high authenticity of the electromagnetic responses calculated by the forward network, the forward network is further trained. The true forward numerical simulation results are used to solve the electromagnetic responses, and the resistivity models output by the inverse network are used as new training samples. Through this self-supervised learning process, the network can optimize the resistivity model through continuous iteration and finally obtain an accurate geoelectric model.

[0116] In summary, the method proposed in this embodiment reduces the dependence on labeled data and initial models while ensuring the accuracy and efficiency of the inversion process, providing new ideas for the development of transient electromagnetic inversion technology.

[0117] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A self-supervised inversion method for transient electromagnetic method based on deep learning, characterized in that: include: Input the electromagnetic response data to be inverted into the pre-trained inversion network for inversion to obtain the inverted resistivity value; Inputting the inverted resistivity value into a pre-trained forward modeling network for forward modeling to obtain forward modeling electromagnetic response data; Updating the parameters of the inversion network by minimizing the loss between the output forward electromagnetic response data and the input electromagnetic response data to be inverted; The electromagnetic response data to be inverted is inverted again using the inversion network with updated parameters until a preset number of iterations is reached, and the resistivity value obtained by inversion is finally output.

2. The transient electromagnetic method self-supervised inversion method based on deep learning according to claim 1 is characterized in that: The inversion network includes: an encoder part, a decoder part and a smooth constraint part; Inputting the electromagnetic response data to be inverted into the pre-trained inversion network for inversion includes: The electromagnetic response data to be inverted is input into the encoder part, and the features of the input electromagnetic response are extracted through three convolution layers, wherein the convolution layer includes a 1D convolution layer + a pooling layer + a Tanh activation function; In the decoder part, the features obtained in the encoder part are restored and reprocessed through three layers of feature upsampling layers, where the feature upsampling layer includes a 1D upsampling layer and two 1D convolutional layers; In the smoothness constraint part, a convolution layer with a convolution kernel of 1×1 is used to further smooth the output of the decoder part, and finally the inverted resistivity value is output.

3. The transient electromagnetic method self-supervised inversion method based on deep learning according to claim 1 is characterized in that: Inputting the inverted resistivity value into the pre-trained forward modeling network for forward modeling includes: Extracting features from the inverted resistivity value through two convolution layers to capture local electromagnetic response information; wherein the convolution layer includes a 1D convolution layer + a pooling layer + a Tanh activation function; The features extracted by the convolutional layer are integrated based on three fully connected layers to extract the forward electromagnetic response data represented by the resistivity value.

4. The transient electromagnetic method self-supervised inversion method based on deep learning according to claim 1, characterized in that: The pre-trained inversion network includes: generating resistivity values ​​and corresponding electromagnetic response data by numerical simulation methods, and constructing a data set based on the resistivity values ​​and the corresponding electromagnetic response data; Preprocessing the data set; The inversion network is pre-trained using the preprocessed data set; in the inversion network pre-training, the electromagnetic response and resistivity values ​​in the data set are used as the input and label of the inversion network respectively; after the inversion network outputs the predicted resistivity, the error between it and the label is calculated through the loss function and the parameter update is driven to complete the inversion network pre-training.

5. The transient electromagnetic method self-supervised inversion method based on deep learning according to claim 4 is characterized in that: The pre-trained forward network includes: The forward modeling network is trained based on the preprocessed data set; in the forward modeling network pre-training, the resistivity value and electromagnetic response in the data set are used as the input and label of the forward modeling network respectively; after the forward modeling network outputs the predicted electromagnetic response, the error between it and the label is calculated through the loss function and the parameter update is driven to complete the forward modeling network pre-training.

6. The transient electromagnetic method self-supervised inversion method based on deep learning according to claim 4 is characterized in that: The numerical simulation methods used to generate resistivity values ​​and corresponding electromagnetic response data include: Set basic parameters of underground strata, including top layer thickness and last layer boundary depth; Divide the stratum into several layers according to the method of equal thickness of each layer; A plurality of control points are set within a preset stratum depth range, wherein one control point is set at each of the top and bottom of the stratum; The resistivity value is randomly generated for each control point and interpolated using the B-spline interpolation method; Generate resistivity-thickness sample data; Using the central loop device, the generated resistivity-thickness sample data is substituted into the transient electromagnetic one-dimensional numerical simulation method to calculate the corresponding electromagnetic response data.

7. The transient electromagnetic method self-supervised inversion method based on deep learning according to claim 2 is characterized in that: Inputting the electromagnetic response data to be inverted into the pre-trained inversion network for inversion also includes: Based on the self-supervised inversion method, an inversion test is performed on the electromagnetic response data to be inverted; wherein, the self-supervised inversion method is performed on a pre-trained inversion network; the electromagnetic response data to be inverted is input into the pre-trained inversion network, and the network outputs a preliminary predicted resistivity value, and then the pre-trained forward network is updated twice based on a fast forward training strategy, that is, a forward numerical simulation method is used to solve the real electromagnetic response corresponding to the predicted resistivity value, and the real electromagnetic response and the predicted resistivity value are used as new training samples to perform a second training update on the forward network; based on the updated forward network, the corresponding electromagnetic response is recalculated, and then the error between the real electromagnetic response and the input electromagnetic response is calculated through a loss function to drive the update of the inversion network parameters; using the inversion network with updated parameters, the electromagnetic response data to be inverted is inverted again until a preset number of iterations is reached, and finally the inverted resistivity value is output.

Images