Method, system and device for converting magnetic field data using transient electromagnetic method induced electromotive force

Abstract

The present invention discloses a method, system, storage medium, and computer device for converting induced electromotive force (EMF) using a transient electromagnetic method into magnetic field data. The method relates to the field of geophysical exploration and primarily includes: constructing a dataset of induced electromotive force and magnetic field data using the transient electromagnetic method; loading the generated down-step induced electromotive force data with a full waveform response; building a deep learning model based on a residual attention mechanism; training the model and performing parameter tuning based on Bayesian optimization; and using the trained model to convert the induced electromotive force into magnetic field data. Implementing the method, system, storage medium, and computer device provided by the present invention improves the accuracy and efficiency of magnetic field data conversion and effectively mines deep geoelectric structure information contained in the existing massive induced electromotive force data.

Description

Technical Field

[0001] The present invention relates to the field of geophysical exploration, and more particularly to a method, system, storage medium and computer equipment for converting magnetic field data by using transient electromagnetic method to induce electromotive force. Background Art

[0002] Transient electromagnetic (TEM) is a geophysical observation method widely used in the exploration of deep metal deposits. TEM observations are still primarily based on induced electromotive force (EMF). Compared to induced EMF, magnetic field data decreases monotonically with observation time and offers advantages such as higher signal-to-noise ratios in later stages, greater detection depths, and improved detection of deep, good conductors. my country has accumulated a vast amount of TEM induced EMF data over time. To effectively mine the deep geoelectric structure information contained in this induced EMF data, there is an urgent need to convert it into magnetic field data.

[0003] Most transient electromagnetic instruments only record the induced electromotive force data after the transmitting current is turned off, resulting in the unknown initial value of the magnetic field and the inability to convert the induced electromotive force into magnetic field data. Summary of the Invention

[0004] The purpose of the present invention is to provide a method, system, storage medium and computer equipment for converting magnetic field data from induced electromotive force using transient electromagnetic method, which can improve the accuracy and efficiency of magnetic field data conversion and effectively mine the deep geoelectric structure information contained in the existing massive induced electromotive force data.

[0005] The present invention provides a method for converting magnetic field data from induced electromotive force using a transient electromagnetic method, comprising the following steps: S1: acquiring transient electromagnetic method data, transforming the transient electromagnetic method data to obtain induced electromotive force and magnetic field strength; S2: constructing a residual deep learning model using a one-dimensional convolution residual block, a bidirectional long short-term memory network residual block, and a multi-head self-attention mechanism; S3: obtaining training data based on the induced electromotive force and magnetic field strength; training and optimizing the residual deep learning model using the training data, a loss function, and a hyperparameter selection method to obtain a trained model; S4: converting the induced electromotive force data to be converted into magnetic field data using the trained model.

[0006] Furthermore, step S1 of the method for converting induced electromotive force into magnetic field data using the transient electromagnetic method specifically includes: obtaining transient electromagnetic method data, the transient electromagnetic method data including information about the transient electromagnetic method transmitting source, receiving device, and measuring point; performing one-dimensional forward modeling on the transient electromagnetic method data, performing sine and cosine transforms on the frequency domain magnetic field to obtain induced electromotive force and magnetic field intensity, and applying a full waveform response to the induced electromotive force, as shown in the formula:

[0007]

[0008] in, represents the induced electromotive force, h(t) represents the magnetic field intensity in the time domain, Im[H(ω)] is the imaginary part of the electromagnetic field in the frequency domain, H(ω) is the electromagnetic field in the frequency domain, ω is the angular frequency of the electromagnetic field, and f w (t) is the full waveform response, f s is the impulse response excited by the step-cutoff source, t is the observation time, and dI(τ) / dτ is the derivative of the emission current waveform.

[0009] Furthermore, the residual deep learning model of the above-mentioned method of converting magnetic field data into induced electromotive force by transient electromagnetic method includes a first parallel component, a second parallel component, a multi-head attention layer, a flattening layer and a fully connected layer; the first parallel component includes a residual-connected convolutional neural network block, and the second parallel component includes a residual-connected bidirectional long short-term memory network block; the residual-connected convolutional neural network block captures local features through multiple layers of one-dimensional convolutional layers, while the residual-connected bidirectional long short-term memory network block uses bidirectional units to capture long-term temporal dependencies; the multi-head attention layer is used to learn different attention weights; the flattening layer is used to flatten the signal to one dimension; and the fully connected layer is used to output predicted data.

[0010] Furthermore, the loss function of the method for converting magnetic field data using the transient electromagnetic method induced electromotive force is the mean square error, and the hyperparameter selection method is the Bayesian optimization method.

[0011] The present invention also provides a system, which includes the following modules: a data transformation module, configured to: obtain transient electromagnetic method data, transform the transient electromagnetic method data, and obtain induced electromotive force and magnetic field strength; a model construction module, configured to: use one-dimensional convolution residual blocks, bidirectional long short-term memory network residual blocks and multi-head self-attention mechanisms to construct a residual deep learning model; a model training module, configured to: obtain training data based on the induced electromotive force and magnetic field strength; use the training data, loss function and hyperparameter selection method to train and tune the residual deep learning model to obtain a trained model; a magnetic field data conversion module, configured to: use the trained model to convert the induced electromotive force data to be converted into magnetic field data.

[0012] Furthermore, the data conversion module of the above system is specifically configured to: obtain transient electromagnetic method data, which includes transient electromagnetic method transmitter, receiving device and measurement point information; perform one-dimensional forward modeling on the transient electromagnetic method data, perform sine and cosine transform on the frequency domain magnetic field to obtain induced electromotive force and magnetic field intensity, and load the induced electromotive force with a full waveform response, as shown in the formula:

[0013]

[0014] in, represents the induced electromotive force, h(t) represents the magnetic field intensity in the time domain, Im[H(ω)] is the imaginary part of the electromagnetic field in the frequency domain, H(ω) is the electromagnetic field in the frequency domain, ω is the angular frequency of the electromagnetic field, and f w (t) is the full waveform response, f s is the impulse response excited by the step-cutoff source, t is the observation time, and dI(τ) / dτ is the derivative of the emission current waveform.

[0015] Furthermore, the residual deep learning model of the above system includes a first parallel component, a second parallel component, a multi-head attention layer, a flattening layer and a fully connected layer; the first parallel component includes a residual-connected convolutional neural network block, and the second parallel component includes a residual-connected bidirectional long short-term memory network block; the residual-connected convolutional neural network block captures local features through multiple layers of one-dimensional convolutional layers, while the residual-connected bidirectional long short-term memory network block uses bidirectional units to capture long-term temporal dependencies; the multi-head attention layer is used to learn different attention weights; the flattening layer is used to flatten the signal to one dimension; and the fully connected layer is used to output predicted data.

[0016] Furthermore, the loss function of the above system is mean square error, and the hyperparameter selection method is Bayesian optimization method.

[0017] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for converting magnetic field data by induced electromotive force using the transient electromagnetic method.

[0018] The present invention also provides a computer device, including a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for converting magnetic field data using transient electromagnetic induced electromotive force are implemented.

[0019] The method, system, storage medium, and computer device for converting magnetic field data using transient electromagnetic induced electromotive force provided by the present invention have the following beneficial effects:

[0020] The present invention constructs an induced electromotive force and magnetic field data set for the ground loop source transient electromagnetic method; loads the generated down-step induced electromotive force data with a full waveform response, accurately converting the induced electromotive force into magnetic field intensity; constructs a sample library based on a sine-cosine transform algorithm using a digital filtering method, converts the frequency domain electromagnetic field into a time domain response, and thus obtains the induced electromotive force and magnetic field;

[0021] We built a deep learning model called ResCNN-BiLSTM-AttNet based on the residual attention mechanism, proposing a novel architecture that combines residual connections, convolutional neural networks, bidirectional long short-term memory networks, and a multi-head attention mechanism to optimize the feature extraction and fusion process.

[0022] The model was trained and parameters were tuned based on Bayesian optimization. The trained model was used to convert the induced electromotive force into magnetic field data, significantly improving accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0024] Figure 1 This is a flow chart of the method for converting magnetic field data using transient electromagnetic induced electromotive force provided by the present invention;

[0025] Figure 2 It is a technical flow chart provided by the present invention;

[0026] Figure 3 This is a schematic diagram of a single-cycle "bipolar" waveform provided by the present invention;

[0027] Figure 4 It is a residual network structure diagram provided by the present invention;

[0028] Figure 5 It is a schematic diagram of the stratum structure and resistivity provided by the present invention;

[0029] Figure 6 It is the down-step and full waveform response of the induced electromotive force provided by the present invention;

[0030] Figure 7 The ResCNN-BiLSTM-AttNet model provided by the present invention;

[0031] Figure 8 are the training loss and validation loss of the model provided by the present invention;

[0032] Figure 9 It is the induced electromotive force and magnetic field response of the same abnormal body provided by the present invention;

[0033] Figure 10 It is a comparison between the predicted magnetic field response provided by the present invention and the actual magnetic field response;

[0034] Figure 11 It is a structural block diagram of the computer device provided by the present invention. DETAILED DESCRIPTION

[0035] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0036] Figure 1 A schematic diagram of a method for converting magnetic field data using transient electromagnetic method induction electromotive force is shown in this embodiment. In this embodiment, the method for converting magnetic field data using transient electromagnetic method induction electromotive force includes the following steps:

[0037] S1: Acquire transient electromagnetic method data, transform the transient electromagnetic method data, and obtain induced electromotive force and magnetic field intensity;

[0038] Specifically, step S1 of the method for converting induced electromotive force into magnetic field data using the transient electromagnetic method specifically includes: obtaining transient electromagnetic method data, which includes information about the transient electromagnetic method transmitting source, receiving device, and measuring point; performing one-dimensional forward modeling on the transient electromagnetic method data, performing sine and cosine transforms on the frequency domain magnetic field to obtain induced electromotive force and magnetic field intensity, and applying a full waveform response to the induced electromotive force, as shown in the formula:

[0039]

[0040] in, represents the induced electromotive force, h(t) represents the magnetic field intensity in the time domain, Im[H(ω)] is the imaginary part of the electromagnetic field in the frequency domain, H(ω) is the electromagnetic field in the frequency domain, ω is the angular frequency of the electromagnetic field, and f w (t) is the full waveform response, f s is the impulse response excited by the step-cutoff source, t is the observation time, and dI(τ) / dτ is the derivative of the emission current waveform;

[0041] S2: Build a residual deep learning model using one-dimensional convolutional residual blocks, bidirectional long short-term memory network residual blocks, and multi-head self-attention mechanism;

[0042] Specifically, the residual deep learning model of the method for converting magnetic field data by induced electromotive force using the transient electromagnetic method includes a first parallel component, a second parallel component, a multi-head attention layer, a flattening layer, and a fully connected layer; the first parallel component includes a residual-connected convolutional neural network block, and the second parallel component includes a residual-connected bidirectional long short-term memory network block; the residual-connected convolutional neural network block captures local features through multiple layers of one-dimensional convolutional layers, while the residual-connected bidirectional long short-term memory network block uses bidirectional units to capture long-term temporal dependencies; the multi-head attention layer is used to learn different attention weights; the flattening layer is used to flatten the signal into one dimension; and the fully connected layer is used to output predicted data;

[0043] S3: Obtain training data based on the induced electromotive force and magnetic field strength; use the training data, loss function, and hyperparameter selection method to train and optimize the residual deep learning model to obtain a trained model;

[0044] Specifically, the loss function of the method for converting the induced electromotive force into magnetic field data using the transient electromagnetic method is the mean square error, and the hyperparameter selection method is the Bayesian optimization method;

[0045] S4: Use the trained model to convert the induced electromotive force data to be converted into magnetic field data.

[0046] This embodiment provides a system, which includes a data transformation module, a model construction module, a model training module and a magnetic field data conversion module;

[0047] The data conversion module is configured to: obtain transient electromagnetic method data, transform the transient electromagnetic method data, and obtain induced electromotive force and magnetic field strength;

[0048] Specifically, the data conversion module of the above system is specifically configured to: obtain transient electromagnetic method data, which includes transient electromagnetic method emission source, receiving device and measurement point information; perform one-dimensional forward modeling on the transient electromagnetic method data, load the full waveform response on the induced electromotive force, and perform sine and cosine transformation to obtain the induced electromotive force and magnetic field intensity, as shown in the formula:

[0049]

[0050] in, represents the induced electromotive force, h(t) represents the magnetic field intensity in the time domain, Im[H(ω)] is the imaginary part of the electromagnetic field in the frequency domain, H(ω) is the electromagnetic field in the frequency domain, ω is the angular frequency of the electromagnetic field, and f w (t) is the full waveform response, f s is the impulse response excited by the step-cutoff source, t is the observation time, and dI(τ) / dτ is the derivative of the emission current waveform;

[0051] The model building module is configured as follows: using one-dimensional convolution residual blocks, bidirectional long short-term memory network residual blocks and multi-head self-attention mechanism to build a residual deep learning model;

[0052] Specifically, the residual deep learning model of the above system includes a first parallel component, a second parallel component, a multi-head attention layer, a flattening layer and a fully connected layer; the first parallel component includes a residual-connected convolutional neural network block, and the second parallel component includes a residual-connected bidirectional long short-term memory network block; the residual-connected convolutional neural network block captures local features through multiple layers of one-dimensional convolutional layers, while the residual-connected bidirectional long short-term memory network block uses bidirectional units to capture long-term temporal dependencies; the multi-head attention layer is used to learn different attention weights; the flattening layer is used to flatten the signal to one dimension; the fully connected layer is used to output predicted data;

[0053] The model training module is configured to: obtain training data based on the induced electromotive force and magnetic field strength; train and optimize the residual deep learning model using the training data, loss function, and hyperparameter selection method to obtain a trained model;

[0054] Specifically, the loss function of the above system is the mean square error, and the hyperparameter selection method is the Bayesian optimization method;

[0055] Among them, the magnetic field data conversion module is configured to: use the trained model to convert the induced electromotive force data to be converted into magnetic field data.

[0056] In some embodiments, the above-mentioned method of converting magnetic field data by inducing electromotive force using transient electromagnetic method can also be implemented in the following manner. Figure 2 The figure shows a flow chart of the method. The method for converting magnetic field data by transient electromagnetic method induced electromotive force includes the following steps:

[0057] Step 1: Dataset Construction

[0058] Based on the transient electromagnetic method emission source, receiving device and measurement point information, a one-dimensional forward model is carried out, and the induced electromotive force and magnetic field are obtained by using sine and cosine transforms. The expressions of the induced electromotive force and magnetic field strength are as follows:

[0059]

[0060]

[0061] in, represents the induced electromotive force, h(t) represents the magnetic field intensity in the time domain, and Im[H(ω)] is the imaginary part of the electromagnetic field in the frequency domain;

[0062] Step 2: Load the full waveform response to the induced electromotive force

[0063] In the ground transient electromagnetic method, a "bipolar" emission current waveform is usually used, that is, a transmission process includes positive and negative power supply; both the positive and negative power supply half-cycles include the current start-up, stabilization and shutdown stages, as well as the measurement stage (see Figure 3 ), and the corresponding measurement phase; in the "bipolar" waveform, each turn-on and turn-off of the transmitting current will induce eddy currents in the underground medium, and the eddy currents generated by the previous turn-on or turn-off will partially offset the eddy currents generated by the next turn-on or turn-off; due to the influence of this phenomenon, there will be differences in the electromagnetic response amplitudes measured in the two half-cycles of positive and negative power supply;

[0064] By calculating the impulse response excited by a pure step-off source, and then convolving it with the derivative of the emission current waveform to calculate its corresponding full waveform response f w :

[0065]

[0066] where f s is the impulse response excited by the step-cutoff source, t is the observation time, and dI(τ) / dτ is the derivative of the emission current waveform;

[0067] The complete full-waveform response provides richer information, helps to more accurately invert the electrical characteristics of the underground medium, and reduces measurement errors caused by the mutual cancellation of eddy currents. By loading and analyzing the full-waveform response of the induced electromotive force, a more comprehensive understanding of the underground structure can be achieved, improving the application of transient electromagnetic methods in geological exploration.

[0068] Step 3: Building a Residual Deep Learning Model

[0069] The ResCNN-BiLSTM-AttNet model of this example combines a one-dimensional convolutional neural network (CNN) residual block, a bidirectional long short-term memory network (BiLSTM) residual block, and a multi-head self-attention mechanism to optimize the ability to extract and fuse features from TEM response data. The following section details the key components of the model and their interactions:

[0070] Residual Networks (RNNs) are a deep learning network structure designed to address problems such as vanishing and exploding gradients that occur during the training of deep neural networks. The core idea of ​​this network structure is to introduce "skip connections" or "residual connections" that allow the network to learn a residual function, that is, the difference between the input and the desired output. These residual connections allow information to propagate more directly in the network, helping to alleviate the vanishing gradient problem and making the network easier to train.

[0071] In the residual network, each residual block consists of two main branches: the main branch and the skip connection. The main branch is usually composed of multiple convolutional layers and activation functions, which are used to learn feature representations. The skip connection directly adds the input or the input after a simple transformation to the output of the main branch to generate a residual. This structure allows the network to learn the residual function instead of directly learning the entire mapping. In this embodiment, two different residual connections are adopted (see Figure 4 ) method, when the input and output of the residual module are of the same size, the input is directly skipped and added to the output; when the input and output of the residual module are of different sizes, one-dimensional convolution or bidirectional long short time sequence is used to change the input shape before adding it to the output;

[0072] 3.1 Input data to the network flows through two parallel core components: a residually connected CNN block and a residually connected BiLSTM block. The CNN block captures local features using multiple layers of one-dimensional convolutional layers with 256 filters. In contrast, the BiLSTM block captures long-term temporal dependencies using 128 bidirectional units. These two blocks process the data independently, extracting different types of features and merging the results. This allows the model to leverage the local perception capabilities of the CNN while also capturing the time series dynamics of the LSTM.

[0073] 3.2 The core idea of ​​the multi-head attention mechanism is to introduce multiple attention heads, each of which learns different attention weights; in each head, each position of the input sequence is associated with a query, a key, and a value; then, by calculating the similarity between the query and the key, and the weighted value of the similarity, the attention distribution of each position is obtained; finally, the attention outputs of multiple heads are weightedly fused to obtain the final multi-head attention output; the weight α of each head i is α. i for:

[0074]

[0075] Then, the values ​​are weighted summed using the attention weights to get the output Z of each head i :

[0076] Z i =α i V i (5)

[0077] where Q i , K i , V i are the linear transformations of the input sequence to query, key, and value, respectively, and dk is the dimension of each key vector;

[0078] By introducing a multi-head attention mechanism, the model can better utilize the global information of the input sequence, thereby improving the model's generalization ability and performance. The merged feature vector then enters a multi-head self-attention mechanism configured with 8 heads. This layer processes different subspaces of the merged features in parallel, modeling local and global dependencies. Through the weighted integration of the outputs of different attention heads, it dynamically adjusts the focus on features, improving the quality of feature representation and prediction accuracy.

[0079] 3.3 Finally, after being processed by the multi-head attention layer, the signal is flattened to one dimension, passes through two fully connected layers with 128 units, and then outputs to a fully connected layer consisting of 30 neurons to complete the prediction of the magnetic field data;

[0080] Due to the use of residual connections, the ReLU activation function is used as the activation function. At the same time, batch normalization technology is used after each network layer to enable faster network training and fitting. In addition, the Dropout technology is introduced after the CNN block, BiLSTM block and fully connected layer of the model to prevent overfitting.

[0081] 3.4 Dropout is a regularization technique whose core idea is to randomly set the output of some neurons to zero with a certain probability p in each training iteration, thereby forming different sub-networks. This method improves the generalization ability of the model by destroying the co-adaptability between neurons. Given the input x and weight W of the neural network, the output after applying Dropout is:

[0082] y=W(xe m) (6)

[0083] Where e represents the element-by-element product, m is a mask vector with the same dimension as x, and the elements in m are 0 with probability p and 1 / (1-p) with probability 1-p, so as to ensure that the output expectation of the neural network remains unchanged, and to some extent compensate for the error caused by deleting some neurons;

[0084] Step 4: Train and tune the model

[0085] In the model training process of this embodiment, the mean square error (MSE) is selected as the loss function because it is very suitable for the sensitivity of the prediction error in the regression problem and the ability to intuitively reflect the difference between the predicted value and the actual value. The mathematical expression of MSE is:

[0086]

[0087] Where n is the number of samples, y i is the true value of the i-th sample, is the model's predicted value for the i-th sample; the Adam optimizer is used for parameter update; the Adam optimizer is chosen because of its internal momentum mechanism and adaptive learning rate, which makes the optimization process more stable and efficient;

[0088] Hyperparameter selection is performed using Bayesian optimization, which systematically explores the hyperparameter space to find the best parameter combination that minimizes validation set loss. Bayesian optimization is an efficient method for optimizing black-box functions, particularly for computationally expensive functions, such as hyperparameter optimization in deep learning. It effectively explores and exploits the search space by constructing a surrogate model (usually a Gaussian process or a tree-structured Parzen estimator) to approximate the target function and using this surrogate model to select the next evaluation point.

[0089] The basic steps of Bayesian optimization are:

[0090] 4.1 Initialization: Randomly select some hyperparameter combinations And calculate their loss on the validation set as the objective function;

[0091] 4.2 Build a surrogate model: Use the initial hyperparameters to calculate the objective function value through the Gaussian process and train a surrogate model;

[0092] 4.3 Select the next evaluation point: Use the alternative model to select the next hyperparameter combination to be evaluated; this selection process relies on an acquisition function:

[0093] α(x)=(μ n (x)-f(x))Φ(Z)+σ n (x)φ(Z) (8)

[0094] in Φ is the cumulative distribution function of the standard normal distribution, and φ is the probability density function of the standard normal distribution;

[0095] 4.4 Evaluate new points: Calculate the value of the newly selected hyperparameter combination on the objective function and add it to the training dataset;

[0096] 4.5 Update model: Update the alternative model with new data points;

[0097] 4.6 Iteration: Repeat steps 3-5 until a predetermined stopping condition is reached, such as a maximum number of iterations or a time limit;

[0098] To enhance the model's predictive capabilities on unseen data, we primarily employed Dropout technology, using early stopping and learning rate decay to monitor and maintain training efficiency. We also used real-time visualization tools, such as TensorBoard, to track training progress and performance metrics, ensuring transparency and traceability.

[0099] Apply the model on an independent test set and evaluate its performance using the mean squared error (MSE) and mean absolute error (MAE) metrics;

[0100] The absolute mean error reflects the average level of the absolute value of the difference between the predicted value and the true value. Compared with MSE, MAE is less sensitive to outliers and can provide an intuitive understanding of the prediction error. The formula of MAE is:

[0101]

[0102] The meanings of the symbols are the same as in the MSE formula;

[0103] We trained for approximately 300 epochs on a high-performance computing platform equipped with an NVIDIA A40 GPU. We selected a batch size of 512, an initial learning rate of 0.01, and decayed it to 10% of the previous rate every 30 epochs. Training lasted approximately 253 minutes. We stopped training after detecting no loss decrease after 50 epochs using the early stopping method.

[0104] Step 5: Use the trained model to convert the induced electromotive force into a magnetic field

[0105] The core of the entire method is to use the trained deep learning model to convert induced electromotive force data into magnetic field data. This step includes specific steps such as data preprocessing, model prediction, and result analysis. The following is a detailed operation process:

[0106] 5.1 Data normalization: Normalize the induced electromotive force data and scale the data to between 0 and 1 to improve the prediction accuracy of the model;

[0107] 5.2 Model loading: Load the previously trained and saved deep learning model;

[0108] 5.4 Data input: The pre-processed induced electromotive force data is used as the input of the model;

[0109] 5.5 Magnetic field data prediction: Use the model for forward propagation to obtain the corresponding magnetic field data; the model will predict the magnetic field value corresponding to each time point based on the learned mapping relationship between the induced electromotive force and the magnetic field data.

[0110] In some embodiments, the above-mentioned method of converting magnetic field data by using transient electromagnetic method to induce electromotive force can also be implemented in the following manner. In this embodiment, the method of converting magnetic field data by using transient electromagnetic method to induce electromotive force includes the following steps:

[0111] Step 1: Dataset Construction

[0112] Set up the formation model (see Figure 5 ) The maximum depth is 500m, and 5 anchor points are randomly selected from the range of 0m to 500m, with a conductivity of 10 -1 S / m to 10 -4 S / m are randomly selected and interpolated to a fixed 20 layers with thicknesses ranging from 5m to 40m. The interpolation operation ensures that the generated bottom layer model is relatively smooth and reduces frequent jumps. The above algorithm is used to randomly generate 50,000 sets of models including formation conductivity, induced electromotive force and magnetic field data (see Figure 6 ); the induced electromotive force, as the input of the neural network, is logarithmically transformed to stabilize the variance and reduce the influence of extreme values, and then normalized to make it more suitable for neural network processing; at the same time, the magnetic field response, as the output of the neural network, is also logarithmically transformed and normalized to maintain consistency in data scale; finally, the dataset is divided into training, validation, and test sets in a ratio of 8:1:1, aiming to provide sufficient training data while ensuring the evaluation of the model's generalization ability;

[0113] Step 2: Load the full waveform

[0114] When loading a full waveform response, selecting the appropriate on-time and off-time is crucial to data quality and the accuracy of the conversion results. The following is an example of the effect of selecting an on-time of 2 milliseconds and an off-time of 10 microseconds and loading the full waveform.

[0115] Step 3: Building a Residual Deep Learning Model

[0116] The ResCNN-BiLSTM-AttNet model of this embodiment (see Figure 7 ) combines a one-dimensional convolutional neural network (CNN) residual block, a bidirectional long short-term memory network residual block, and a multi-head self-attention mechanism to optimize the ability to extract and fuse features from TEM data;

[0117] Step 4: Train and tune the model

[0118] During the training process, the training loss (dark line) and validation loss (light line) of the ResCNN-BiLSTM-AttNet model both show a significant decreasing trend with the increase in the number of iterations (see Figure 8 ), which not only proves that the model has learned an effective representation of the data, but also maintains the ability to generalize to new data;

[0119] Step 5: Use the trained model to convert the induced electromotive force into a magnetic field

[0120] A layered model with a depth of 500m is constructed. An abnormally high conductor with a thickness of 10m is buried at 350m. The conductivity of the formation in the rest of the area is 0.1S / m. For this abnormally high conductor, the conductivity is selected as 100S / m and 10S / m respectively. The corresponding conductivity is calculated by forward modeling (see Figure 9 ) induced electromotive force and magnetic field response; it can be seen that when the high-conductivity anomaly is present, the induced electromotive force has a small difference, while the magnetic field response at this moment has a large difference; this proves that magnetic field data has higher sensitivity and resolution for high-conductivity anomalies, and also illustrates the necessity of converting induced electromotive force into magnetic field data;

[0121] like Figure 10 The figure shows the comparison between the predicted magnetic field response and the actual magnetic field response; by randomly selecting the induced electromotive force from a part of the test set as the input model, the predicted magnetic field response data can be obtained; the predicted magnetic field and the actual magnetic field are almost completely fitted at most times, and a few samples have only slight errors in the late stage, which proves the reliability of the deep learning model; through the trained model, a large amount of existing induced electromotive force data can be quickly converted into magnetic field data, and the converted magnetic field data can be used for data interpretation to achieve improved data accuracy; the transient electromagnetic method data conversion method based on deep learning proposed in this embodiment has obvious advantages and achieves the expected goals.

[0122] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the steps of the method for converting induced electromotive force into magnetic field data using the transient electromagnetic method. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD). The storage medium may also include a combination of the aforementioned types of memory.

[0123] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for converting magnetic field data using transient electromagnetic method induced electromotive force are implemented.

[0124] like Figure 11 As shown, the computer device may include: at least one processor 121, such as a CPU (Central Processing Unit), at least one communication interface 123, a memory 124, and at least one communication bus 122. The communication bus 122 is used to realize the connection and communication between these components. The communication interface 123 may include a display screen and a keyboard. The optional communication interface 123 may also include a standard wired interface and a wireless interface. The memory 124 may be a high-speed RAM memory (Random Access Memory) or a non-volatile memory, such as at least one disk storage. The memory 124 may optionally be at least one storage device located away from the aforementioned processor 121. The memory 124 stores application programs, and the processor 121 calls the program code stored in the memory 124 to execute any of the above-mentioned method steps. The communication bus 122 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 122 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 11The fact that only one line is used to represent the bus does not mean that there is only one bus or only one type of bus. The memory 124 may include volatile memory, such as random-access memory (RAM); the memory may also include non-volatile memory, such as flash memory, a hard disk drive (HDD), or a solid-state drive (SSD); and the memory 124 may also include a combination of the above-mentioned types of memory. The processor 121 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. The processor 121 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof. Optionally, the memory 124 is also used to store program instructions. The processor 121 can call the program instructions to implement the method of converting magnetic field data from induced electromotive force using the transient electromagnetic method of this embodiment.

[0125] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A method for converting magnetic field data by inducing electromotive force using transient electromagnetic method, characterized in that: The following steps are involved: S1: Acquire transient electromagnetic method data, transform the transient electromagnetic method data, and obtain induced electromotive force and magnetic field strength; S2: Build a residual deep learning model using one-dimensional convolutional residual blocks, bidirectional long short-term memory network residual blocks, and multi-head self-attention mechanism; S3: Obtaining training data based on the induced electromotive force and the magnetic field strength; Training and optimizing the residual deep learning model using the training data, the loss function, and the hyperparameter selection method to obtain a trained model; S4: using the trained model to convert the induced electromotive force data to be converted into magnetic field data.

2. The method for converting magnetic field data by transient electromagnetic induced electromotive force according to claim 1, characterized in that: Step S1 specifically includes: acquiring transient electromagnetic method data, the transient electromagnetic method data including transient electromagnetic method transmitting source, receiving device and measuring point information; performing one-dimensional forward modeling on the transient electromagnetic method data, performing sine and cosine transform on the frequency domain magnetic field to obtain induced electromotive force and magnetic field intensity, and applying full waveform response to the induced electromotive force, as shown in the formula: in, represents the induced electromotive force, h(t) represents the magnetic field intensity in the time domain, Im[H(ω)] is the imaginary part of the electromagnetic field in the frequency domain, H(ω) is the electromagnetic field in the frequency domain, ω is the angular frequency of the electromagnetic field, and f w (t) is the full waveform response, f s is the impulse response excited by the step-cutoff source, t is the observation time, and dI(τ) / dτ is the derivative of the emission current waveform.

3. The method for converting magnetic field data by transient electromagnetic method induced electromotive force according to claim 1, characterized in that: The residual deep learning model includes a first parallel component, a second parallel component, a multi-head attention layer, a flattening layer, and a fully connected layer; the first parallel component includes a residual-connected convolutional neural network block, and the second parallel component includes a residual-connected bidirectional long short-term memory network block; the residual-connected convolutional neural network block captures local features through multiple layers of one-dimensional convolutional layers, and the residual-connected bidirectional long short-term memory network block uses bidirectional units to capture long-term temporal dependencies; the multi-head attention layer is used to learn different attention weights; the flattening layer is used to flatten the signal into one dimension; The fully connected layer is used to output predicted data.

4. The method for converting magnetic field data by transient electromagnetic induced electromotive force according to claim 1, characterized in that: The loss function is the mean square error, and the hyperparameter selection method is the Bayesian optimization method.

5. A system, characterized in that: The system includes the following modules: The data conversion module is configured to: obtain transient electromagnetic method data, transform the transient electromagnetic method data, and obtain induced electromotive force and magnetic field strength; The model building module is configured to build a residual deep learning model using a one-dimensional convolutional residual block, a bidirectional long short-term memory network residual block, and a multi-head self-attention mechanism; The model training module is configured to: obtain training data based on the induced electromotive force and the magnetic field strength; train and tune the residual deep learning model using the training data, the loss function, and the hyperparameter selection method to obtain a trained model; The magnetic field data conversion module is configured to convert the induced electromotive force data to be converted into magnetic field data using the trained model.

6. The system according to claim 5, characterized in that The data conversion module is specifically configured to: obtain transient electromagnetic method data, the transient electromagnetic method data including transient electromagnetic method transmitting source, receiving device and measuring point information; perform one-dimensional forward modeling on the transient electromagnetic method data, perform sine and cosine transformation on the frequency domain magnetic field to obtain induced electromotive force and magnetic field intensity, and load the induced electromotive force with a full waveform response, such as the formula: in, represents the induced electromotive force, h(t) represents the magnetic field intensity in the time domain, Im[H(ω)] is the imaginary part of the electromagnetic field in the frequency domain, H(ω) is the electromagnetic field in the frequency domain, ω is the angular frequency of the electromagnetic field, and f w (t) is the full waveform response, f s is the impulse response excited by the step-cutoff source, t is the observation time, and dI(τ) / dτ is the derivative of the emission current waveform.

7. The system according to claim 5, characterized in that The residual deep learning model includes a first parallel component, a second parallel component, a multi-head attention layer, a flattening layer, and a fully connected layer; the first parallel component includes a residual-connected convolutional neural network block, and the second parallel component includes a residual-connected bidirectional long short-term memory network block; the residual-connected convolutional neural network block captures local features through multiple layers of one-dimensional convolutional layers, and the residual-connected bidirectional long short-term memory network block uses bidirectional units to capture long-term temporal dependencies; the multi-head attention layer is used to learn different attention weights; the flattening layer is used to flatten the signal into one dimension; The fully connected layer is used to output predicted data.

8. The system according to claim 5, characterized in that The loss function is the mean square error, and the hyperparameter selection method is the Bayesian optimization method.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for converting magnetic field data by induced electromotive force using a transient electromagnetic method as claimed in any one of claims 1 to 4 are realized.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method for converting magnetic field data by induced electromotive force using transient electromagnetic method as claimed in any one of claims 1 to 4 are implemented.

Images