Method for constructing transient electromagnetic inversion model driven by target data set

Abstract

The present invention relates to a method for constructing a transient electromagnetic inversion model driven by a target data set, comprising: constructing an initial training set, training a first convolutional neural network by using the initial training set, and obtaining an initial transient electromagnetic inversion network model; inputting electromagnetic response data to be measured into an initial transient electromagnetic inversion model for prediction, and obtaining a resistivity value; performing forward simulation on the resistivity value, obtaining forward simulated electromagnetic response data, and constructing prediction data on the basis of the forward simulated electromagnetic response data and the resistivity value; extracting data in the initial training set similar to the electromagnetic response data to be measured, and constructing a target data set on the basis of the similar data and the prediction data; and migrating parameters in the initial transient electromagnetic inversion model to a second convolutional neural network to iteratively learn target data set pairs, and obtaining a transient electromagnetic inversion model. The present invention aims to improve the accuracy of inversion of data to be measured in a network when deep learning inversion of data to be measured is performed by means of an iterative training policy.

Description

A method for constructing transient electromagnetic inversion models driven by target datasets Technical Field

[0001] This invention relates to the field of transient electromagnetic inversion technology, and in particular to a method for constructing transient electromagnetic inversion models driven by a target dataset. Background Technology

[0002] Currently, transient electromagnetic inversion is mainly divided into two types: linear inversion and nonlinear inversion. Linear inversion methods, such as Occam inversion and Marquardt inversion, are highly dependent on the initial model, and the quality of the inversion results largely depends on whether a suitable initial model can be found. Nonlinear inversion methods, such as simulated annealing, particle swarm optimization, and Bayesian inversion, typically require enormous computational resources due to the complexity of geophysical models. With the improvement of hardware computing power, electromagnetic inversion based on deep learning has become a research hotspot. Vladimir Puzyrev (2019) first explored the potential of deep learning methods in electromagnetic inversion. In recent years, deep learning-based methods have also been extensively studied in fields such as CSEM inversion, MT inversion, and AEM inversion.

[0003] However, the aforementioned deep learning-based inversion methods are entirely data-driven, and their performance largely depends on the richness of the training sample set. When dealing with data whose distribution differs from the training sample set, the results are often unsatisfactory. To address this, many researchers have made improvements by adding physical constraints to the loss function of the deep learning network to control the network's training process. For example, Jin et al. (2019) introduced the Jacobi differential operator into the forward model of the logging-while-drilling (LWD) electromagnetic response predicted by a convolutional neural network to construct a composite loss function of the model and data mismatch function. Sun et al. (2020) constructed a forward operator Γ based on a recurrent neural network (RNN) to simulate wave propagation, achieving unsupervised deep learning seismic inversion; its training process is equivalent to an optimization of conventional deterministic inversion methods. W. Liu et al. (2022) proposed incorporating the physical laws of magnetotelluric wave propagation into a pure data-driven deep learning method (PlainDNN), adding physics-based misfit data to the loss function to guide network training, and validating this method in one-dimensional magnetotelluric inversion. Additionally, some researchers combine deep learning with traditional methods to leverage the advantages of both. For example, Asif et al. (2022) predicted the Jacobian matrix in least squares inversion using a neural network, integrating the neural network into traditional least squares inversion, and validating this method in one-dimensional transient electromagnetic inversion. Building on this, Asif et al. (2022) trained a forward modeling operator using a neural network, integrated it into least squares inversion, and ultimately achieved excellent performance in one-dimensional airborne transient electromagnetic inversion.

[0004] Due to the complexity, non-uniqueness, and high time and computational resource requirements of geophysical models, constructing large and detailed geophysical datasets and training networks is quite challenging. Furthermore, purely data-driven deep learning methods primarily learn the inversion operator L hidden between input and output within the training set; therefore, their performance is often unsatisfactory when testing on data outside the training set. In other words, data-driven deep learning inversion methods struggle to handle fine-grained details between data and models when faced with unseen data, unless large-scale training is performed on the test data with the same statistical distribution. This limits the practical application of data-driven machine learning methods in geophysical inversion.

[0005] While the improvements of the two methods mentioned above have reduced the dependence on large-scale data and improved the generalization ability of the network to some extent, they have also complicated the training process of the network and increased the consumption of computing resources during the training phase. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing transient electromagnetic inversion models based on target datasets. By repeatedly updating the model and adjusting parameters, the method gradually approximates the true inversion results, thereby overcoming the limitations of traditional data-driven methods and improving inversion accuracy and reliability.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A method for constructing transient electromagnetic inversion models driven by the target dataset includes:

[0009] Construct an initial training set, and use the initial training set to train the first convolutional neural network to obtain an initial transient electromagnetic inversion network model;

[0010] The electromagnetic response data to be measured is input into the initial transient electromagnetic inversion model for prediction to obtain the resistivity value;

[0011] A forward modeling simulation is performed on the resistivity value to obtain the electromagnetic response data of the forward modeling simulation, and prediction data is constructed based on the electromagnetic response data of the forward modeling simulation and the resistivity value;

[0012] Extract similar data from the initial training set that are similar to the electromagnetic response data to be tested, and construct a target dataset based on the similar data and the prediction data;

[0013] The parameters in the initial transient electromagnetic inversion model are transferred to a second convolutional neural network to iteratively learn the target dataset pair and obtain the transient electromagnetic inversion model.

[0014] Optionally, obtaining the initial training set includes:

[0015] The thickness of the top layer and the depth of the bottom layer of the underground geological structure are predetermined, and the structure is divided into several layers according to the method of cumulatively increasing thickness of each layer;

[0016] Several control points are determined between the top and bottom layers, and the control points are interpolated to obtain resistivity-thickness sample data between the top and bottom layers.

[0017] A transient electromagnetic one-dimensional numerical simulation was performed based on the resistivity-thickness sample data to obtain the corresponding electromagnetic response.

[0018] Based on the resistivity value and the corresponding electromagnetic response, a sample dataset is obtained, and the initial training set is divided from the sample dataset.

[0019] Optionally, interpolating the control points includes using the B-spline interpolation method.

[0020] Optionally, the convolutional neural network includes an input layer, an output layer, a convolutional layer, and a fully connected layer, wherein the input layer is used to input the electromagnetic response, the output layer is used to output the resistivity value, and the data features extracted by the convolutional layer are input to the fully connected layer.

[0021] Optionally, extracting similar data from the initial training set that are similar to the electromagnetic response data to be tested includes:

[0022] Obtain the relative average error between the electromagnetic response data in the initial training set and the electromagnetic response data to be tested; obtain the electromagnetic response data in the initial training set corresponding to the relative average error within a preset range as similar electromagnetic response data, and extract the resistivity value corresponding to the similar electromagnetic response data in the initial training set to obtain the similar data.

[0023] This invention also provides a transient electromagnetic inversion method driven by a target dataset. The transient electromagnetic inversion model is constructed using a method driven by a target dataset. The electromagnetic response data to be predicted is input into the transient electromagnetic inversion model to obtain the predicted resistivity value.

[0024] The beneficial effects of this invention are as follows: This invention aims to improve the accuracy of deep learning inversion of test data by using an iterative training strategy. In this method, by repeatedly updating the model and adjusting parameters, the network gradually approximates the true inversion result, thereby overcoming the limitations of traditional data-driven methods and improving inversion accuracy and reliability. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 is a framework diagram of the transient electromagnetic inversion model construction method based on target dataset driven by an embodiment of the present invention;

[0027] Figure 2 is a schematic diagram of sample set data generation according to an embodiment of the present invention;

[0028] Figure 3 is a diagram of the iterative inversion network framework driven by the target dataset in an embodiment of the present invention;

[0029] Figure 4 is a diagram of the network model training loss in an embodiment of the present invention;

[0030] Figure 5 is a performance evaluation diagram of the initial network model in an embodiment of the present invention;

[0031] Figure 6 is a comparison of the response error of the iterative inversion network driven by the target dataset in an embodiment of the present invention;

[0032] Figure 7 is a comparison diagram of a certain test sample point in different methods according to an embodiment of the present invention. In the figure, (a) is a response data fitting diagram and (b) is a resistivity value fitting diagram. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] To make the above-mentioned 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.

[0035] This embodiment provides a method for constructing a transient electromagnetic inversion model driven by a target dataset, including:

[0036] Construct an initial training set and use it to train the first convolutional neural network to obtain an initial transient electromagnetic inversion network model.

[0037] The electromagnetic response data to be measured is input into the initial transient electromagnetic inversion model for prediction, and the resistivity value is obtained.

[0038] Forward modeling is performed on the resistivity values ​​to obtain electromagnetic response data from the forward modeling, and prediction data is constructed based on the electromagnetic response data and resistivity values ​​from the forward modeling.

[0039] Extract similar data from the initial training set and the electromagnetic response data to be tested, and construct the target dataset based on the similar data and the prediction data;

[0040] Furthermore, similar data extracted from the initial training set and the electromagnetic response data to be tested include:

[0041] Obtain the relative average error between the electromagnetic response data in the initial training set and the electromagnetic response data to be tested; obtain the electromagnetic response data in the initial training set corresponding to the relative average error of the response within a preset range as similar electromagnetic response data, and extract the resistivity value corresponding to the similar electromagnetic response data in the initial training set to obtain similar data.

[0042] The parameters in the initial transient electromagnetic inversion model are transferred to the second convolutional neural network to iteratively learn the target dataset pair and obtain the transient electromagnetic inversion model.

[0043] Furthermore, obtaining the initial training set includes: pre-setting the top layer thickness and bottom layer boundary depth of the underground geological structure, and dividing it into several layers according to the method of cumulatively increasing the thickness of each layer; determining several control points between the top and bottom layers, interpolating the control points, and obtaining resistivity-thickness sample data between the top and bottom layers; performing transient electromagnetic one-dimensional numerical simulation based on the resistivity-thickness sample data to obtain the corresponding electromagnetic response; obtaining a sample dataset based on the resistivity value and the corresponding electromagnetic response, and dividing the initial training set from the sample dataset.

[0044] Specifically, considering that the resistivity values ​​of actual underground strata are usually continuously distributed rather than segmented, and that continuous distribution characteristics can well reflect the complexity of underground geological structures, this embodiment uses the following method to generate sample data in order to generate relatively smooth and vertically continuous underground resistivity values: First, the thickness of the top layer is set to 1m, and the boundary depth of the last layer is set to 750m. The layers are divided into 15 layers according to the cumulative increase in thickness to ensure that the random characteristics of the underground resistivity distribution are fully reflected. Five control points are determined between 1m and 750m, as shown by the dark dots in Figure 2. Two of these control points are fixed at depths of 1m and 750m, while the remaining three are randomly determined, but the depth between the control points is required to be within 50-200m. The resistivity of the control points is randomly generated within the range of 10-1000Ω·m. B-spline interpolation is used based on the five control points to interpolate, thereby making the resistivity curve smoother and finally obtaining a continuous resistivity distribution from 1m to 750m, as shown in Figure 2. A rectangular loop source device was used, and the generated resistivity-thickness sample data was substituted into a one-dimensional numerical simulation method for transient electromagnetics to calculate the corresponding electromagnetic response. The coil used was 500m in length and width, with 61 receiving channels and a sampling time of 1e. -4 ~1e -1 The emission current is 10A.

[0045] Furthermore, the convolutional neural network includes an input layer, an output layer, a convolutional layer, and a fully connected layer. The input layer is used to input the electromagnetic response, the output layer is used to output the resistivity value, and the data features extracted by the convolutional layer are input to the fully connected layer.

[0046] Specifically, a convolutional neural network model framework is adopted as shown in Figure 3. The network input is electromagnetic response data, and the output is resistivity value. The network structure mainly consists of an input layer and an output layer, four convolutional layers and two fully connected layers. The input layer has 61 neurons, representing 61 electromagnetic response values; the output layer has 15 neurons, corresponding to 15 resistivity values. The kernel sizes of the four convolutional layers are 2×1, 3×1, 3×1 and 3×1, respectively, and the number of kernels are 64, 128, 256 and 512, respectively, with a stride of 1 for each layer; the pooling layer uses max-pooling with a size of 2×1 and a stride of 2. The data features extracted by the convolutional layers are input into the fully connected layers through flattening layers. The number of neurons in the two fully connected layers are 1024 and 512, respectively. The activation function used is tanh. To quantitatively evaluate the convergence of the network training process, the following loss function is defined:

[0047] Where N represents the number of layers with high resistivity. y represents the resistivity value of the j-th layer in the network prediction model. j This represents the resistivity value corresponding to the actual geoelectric model.

[0048] The network was trained using the Adam optimizer, with the following hyperparameters: learning rate lr = 0.0001, batch size = 1024, epochs = 5000, and early stopping as the training strategy. A dataset of 40,000 samples was generated and partitioned according to a training set:validation set:test set ratio of 6.3:2.7:1.0.

[0049] Figure 4 shows the training loss function of the model network. The dashed line represents the validation set error, and the solid line represents the training set error. Employing an early stopping mechanism, it can be seen that no overfitting occurred during model training. The loss values ​​on both the training and validation sets continuously decreased during training and eventually converged. The decrease in the training set loss value indicates that the model gradually learned the patterns and rules in the training data and continuously optimized its parameters. This demonstrates that the model did not exhibit significant overfitting during training and possesses strong fitting ability.

[0050] This embodiment selected three geoelectric model data points corresponding to the minimum, average, and maximum relative average errors from the test set. Figure 5 shows a comparison of the resistivity-depth curves of the original and inverted models, as well as a comparison of the electromagnetic responses of the original and inverted models. Figure 5 uses three data points with the minimum, average, and maximum relative average errors to plot the resistivity-depth map (first row) and electromagnetic response signal map (second row) of the original and inverted geoelectric models. It can be observed from the figures that the inverted model fits the original model well. Although there are some discrepancies in some layers, the overall results are still within an acceptable range. In particular, for the model with the largest relative average error, although there are large errors in resistivity prediction at some layers, the overall trend of resistivity with depth is consistent with the original model, and its corresponding electromagnetic response error is only 9.4%. This further demonstrates the high accuracy of the inversion results. This satisfies the assumption that deep learning inversion methods are efficient and can guarantee accuracy to a certain extent.

[0051] Specifically, as shown in Figure 1, the upper left module is the initial network model training part, the upper right part is the dataset required for iterative inversion, and the lower part is the iterative training part. First, an initial training set is generated through numerical simulation, i.e., the Training data in the upper left module. Then, a convolutional neural network model is used for training to obtain the initial transient electromagnetic inversion network model. Finally, the electromagnetic response data to be tested, i.e., the Objective data in the upper right module, is used for the first prediction to obtain the corresponding resistivity value DL Predict. Next, data whose relative average error (as shown in Equation 2) between the electromagnetic response data in the Training data and the electromagnetic response data to be tested (Objective data) is extracted, i.e., the Objective data end, including the corresponding resistivity value. Then, the Objective data end, DL Predict (resistivity value), and the electromagnetic response obtained by forward simulation of DL Predict are merged to form the dataset New Data (in the lower module) required for iterative inversion. Finally, a transfer learning strategy was adopted on the initial network Model-1, i.e. the initial transient electromagnetic inversion network model. That is, while keeping the network framework unchanged, the model parameters in Model-1 were transferred to Model (the second convolutional neural network) to initialize the model parameters. Then, the model was learned on New Data, and this process was iterated multiple times to achieve iterative inversion driven by the target dataset.

[0052] Where M is the number of time channels, d j For the actual electromagnetic response of the j-th time channel, Predict the electromagnetic response of the network for the j-th time channel.

[0053] In Figure 6, the solid line represents the model response error predicted by the initial network model for 50 test data points, and the dashed line represents the model response error predicted by the iterative inversion network driven by the target dataset. It is evident from Figure 6 that after iterative inversion driven by the target dataset, the overall response error of the test data, except for a few individual data points, is reduced to a low level, indicating that the iterative inversion strategy driven by the target dataset has a certain effect. Furthermore, Figure 7 compares the results of a specific test sample point using different methods. The dashed line represents the true value, the dashed line represents the result obtained from the initial network inversion, and the dotted line represents the result after 5 iterations of the iterative inversion network driven by the target dataset. From the subplot in Figure 7(a), it can be observed that after 5 iterations, the response error of the data approaches the true response value, and the response accuracy is improved to some extent. Additionally, from Figure 7(b), it is evident that the result after 5 iterations (dashed line) is closer to the true resistivity value, which is more pronounced in the shallow layer with a depth of 100–200 m. In summary, this embodiment improves the inversion accuracy of the network for the test data through simple iterative inversion.

[0054] A deep learning-based transient electromagnetic inversion network, without any intervention, achieved a relative average error of 0.0504 and a relative average error of 0.0310 for inversion of 50 samples (with similar sample sizes for a single measurement line) that differed from the training set. This high accuracy fulfills the expectation that deep learning inversion methods offer both high efficiency and a degree of guaranteed accuracy.

[0055] The target dataset proposed in this embodiment consists of two parts:

[0056] (1) Data in the training set similar to the test data, i.e., datasets with response errors less than a specified threshold (default 5%); (2) Datasets consisting of resistivity values ​​predicted by the network and corresponding responses. A transfer learning strategy is used on the initial training model to learn from the target dataset. After five iterations of inversion, the proposed method achieves a relative average error of 0.0379 and a relative average error of 0.0105 for both the model and response, respectively. Compared to the initial prediction, both model error and response error are reduced by 3 to 5 times.

[0057] In summary, this embodiment improves the inversion accuracy of the network for the data under test through simple iterative inversion.

[0058] This embodiment also provides a transient electromagnetic inversion method driven by a target dataset. The transient electromagnetic inversion model is constructed using the transient electromagnetic inversion model construction method driven by the target dataset. The electromagnetic response data to be predicted is input into the transient electromagnetic inversion model to obtain the predicted resistivity value.

[0059] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for constructing a transient electromagnetic inversion model driven by a target dataset, characterized in that, include: Construct an initial training set, and use the initial training set to train the first convolutional neural network to obtain an initial transient electromagnetic inversion network model; The electromagnetic response data to be measured is input into the initial transient electromagnetic inversion model for prediction to obtain the resistivity value; A forward modeling simulation is performed on the resistivity value to obtain the electromagnetic response data of the forward modeling simulation, and prediction data is constructed based on the electromagnetic response data of the forward modeling simulation and the resistivity value; Extract similar data from the initial training set that are similar to the electromagnetic response data to be tested, and construct a target dataset based on the similar data and the prediction data; The parameters in the initial transient electromagnetic inversion model are transferred to a second convolutional neural network to iteratively learn the target dataset pair and obtain the transient electromagnetic inversion model.

2. The method for constructing a transient electromagnetic inversion model based on a target dataset as described in claim 1, characterized in that, Obtaining the initial training set includes: The thickness of the top layer and the depth of the bottom layer of the underground geological structure are predetermined, and the structure is divided into several layers according to the method of cumulatively increasing thickness of each layer; Several control points are determined between the top and bottom layers, and the control points are interpolated to obtain resistivity-thickness sample data between the top and bottom layers. A transient electromagnetic one-dimensional numerical simulation was performed based on the resistivity-thickness sample data to obtain the corresponding electromagnetic response. Based on the resistivity value and the corresponding electromagnetic response, a sample dataset is obtained, and the initial training set is divided from the sample dataset.

3. The method for constructing a transient electromagnetic inversion model based on a target dataset as described in claim 2, characterized in that, Interpolating the control points includes using the B-spline interpolation method.

4. The method for constructing a transient electromagnetic inversion model based on a target dataset as described in claim 2, characterized in that, A convolutional neural network includes an input layer, an output layer, a convolutional layer, and a fully connected layer. The input layer is used to input the electromagnetic response, the output layer is used to output the resistivity value, and the data features extracted by the convolutional layer are input to the fully connected layer.

5. The method for constructing a transient electromagnetic inversion model based on a target dataset as described in claim 1, characterized in that, Extracting similar data from the initial training set that are similar to the electromagnetic response data to be tested includes: Obtain the relative average error between the electromagnetic response data in the initial training set and the electromagnetic response data to be tested; obtain the electromagnetic response data in the initial training set corresponding to the relative average error within a preset range as similar electromagnetic response data, and extract the resistivity value corresponding to the similar electromagnetic response data in the initial training set to obtain the similar data.

6. A transient electromagnetic inversion method driven by a target dataset, characterized in that, Using the transient electromagnetic inversion model construction method based on target dataset as described in any one of claims 1-5, a transient electromagnetic inversion model is constructed, and the electromagnetic response data to be predicted is input into the transient electromagnetic inversion model to obtain the predicted resistivity value.

Images