Method for eliminating the induced polarization effect in transient electromagnetic responses, computer program product
By using a deep neural network model with convolutional and recurrent networks in series, the problems of low computational efficiency and non-uniqueness of results in the excitation polarization effect in transient electromagnetic responses are solved, achieving fast and accurate elimination of the excitation polarization effect, which is suitable for large-scale data processing.
Patent Information
- Application Number
- CN202611528275.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies for eliminating the induced polarization effect in transient electromagnetic responses suffer from low computational efficiency, non-unique results, and reliance on subjective judgment, making it difficult to meet the engineering needs of large-scale data processing.
A deep neural network model is used, and normalization and denormalization are performed through concatenated convolutional and recurrent network modules to achieve the mapping from induced transient electromagnetic response to pure induced transient electromagnetic response, thereby eliminating the induced polarization effect.
It achieves rapid, accurate, and automated elimination of excitation polarization effects, with high uniqueness of results, wide applicability, several orders of magnitude improvement in elimination efficiency, strong adaptability, and reduced reliance on operator experience.
Smart Images
Figure CN123029623A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of geophysical electromagnetic detection data processing technology, specifically relating to a method for eliminating the induced polarization effect in transient electromagnetic responses and a computer program product. Background Technology
[0002] The Transient Electromagnetic Method (TEM) is an important geophysical exploration method that infers subsurface electrical structure by observing the decay response of a secondary field induced by the subsurface medium. However, in actual exploration, when the medium contains polarizable minerals, an induced polarization effect occurs. This effect, superimposed on the pure electromagnetic induction effect, significantly interferes with the observed transient electromagnetic response. Specifically, the physical mechanism of the induced polarization effect is the migration and aggregation of ions or electrons in the medium under an applied electric field, macroscopically manifested as a slow charging and discharging process. This causes distortion in the measured induced electromagnetic transient response curve in the middle and late stages, such as an abnormally accelerated decay rate or even a sign reversal. If this induced electromagnetic transient response is used directly for conventional inversion without processing, false electrical structures will be obtained, severely affecting the accuracy of geological interpretation.
[0003] Currently, there are two main types of methods for eliminating induced polarization effects. The first type is the inversion method based on the complex resistivity model. This method incorporates parameters describing the induced polarization effect (such as polarizability, time constant, etc.) along with resistivity into the inversion objective, and solves for both electrical and polarization parameters simultaneously through iterative optimization. However, this method has inherent drawbacks. First, there is a strong coupling effect between resistivity and polarization parameters, leading to a high degree of ambiguity in the inversion problem. The results heavily depend on the choice of the initial model, resulting in poor reliability. Second, forward modeling involving induced polarization effects is very time-consuming, while the inversion process requires repeated calls to the forward model, resulting in extremely low overall computational efficiency and making it difficult to meet the engineering needs of large-scale data processing.
[0004] The second type of method is the signal processing method based on basis function decomposition. This method treats the measured response as a linear superposition of the pure inductive response and the induced polarization effect response. It decomposes the signal onto a set of predefined basis functions (such as power functions, exponential functions, etc.), then identifies and removes basis function components representing the induced polarization effect based on certain criteria, and finally reconstructs the pure inductive response. The main problems with this type of method are: first, the morphology of the induced polarization effect is influenced by various geoelectric parameters and is complex; any fixed set of basis functions cannot universally describe all situations. Second, treating induction and polarization as a simple linear superposition ignores their coupling relationship in the physical process, resulting in incompleteness in the model itself. Furthermore, identifying and determining which components belong to the induced polarization effect often relies on the operator's subjective experience and threshold settings, lacking an objective and unified standard.
[0005] In summary, existing technologies for handling induced polarization effects in transient electromagnetic responses are either hampered by the ambiguity and low efficiency of inversion, or limited by the subjectivity of signal decomposition methods and the incompleteness of models. Therefore, there is an urgent need in the field for a new technology capable of rapidly, accurately, and automatically recovering pure inductive responses from induced transient electromagnetic responses. Summary of the Invention
[0006] The purpose of this application is to provide a method and computer program product for eliminating the induced polarization effect in transient electromagnetic responses, aiming to solve the problems of low computational efficiency, non-unique results, and reliance on subjective judgment in existing methods for eliminating the induced polarization effect in transient electromagnetic responses.
[0007] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for eliminating excited polarization effects in transient electromagnetic responses, comprising: Acquire the induced transient electromagnetic response containing multiple time channels to be processed; The electromagnetic response containing induced transients is normalized to convert it into a normalized electromagnetic response containing induced transients; wherein the normalization process uses independent normalization parameters for each time channel. The normalized electromagnetic response with induced transient emission is input into a pre-trained deep neural network model. The deep neural network model processes the normalized electromagnetic response with induced transient emission through its sequentially connected convolutional and recurrent network modules to output a predicted normalized pure inductive transient electromagnetic response. The convolutional modules are used to extract local morphological features of the normalized electromagnetic response with induced transient emission, and the recurrent network modules are used to learn global temporal dependencies from the local morphological features. The normalized pure inductive transient electromagnetic response is denormalized to restore it to a pure inductive transient electromagnetic response free from excitation polarization effects; wherein the denormalization process uses the corresponding normalization parameter in the normalization process for each time channel.
[0008] In some embodiments, the recurrent network module is a bidirectional recurrent network module, and the convolution module is a one-dimensional convolutional neural network module.
[0009] In some embodiments, the bidirectional recurrent network module is a bidirectional long short-term memory network module; The one-dimensional convolutional neural network module includes multiple one-dimensional convolutional blocks with convolutional kernels of different sizes to extract local morphological features of the response at different time scales.
[0010] In some embodiments, the normalized parameters for each time channel include the mean and standard deviation of the induced transient electromagnetic response obtained during the pre-training of the deep neural network model at that time channel.
[0011] In some embodiments, the pre-training step of the deep neural network model includes: Construct a training sample dataset, which includes multiple sample pairs, each of which includes an induced transient electromagnetic response and a corresponding pure inductive transient electromagnetic response. The deep neural network model is trained using multiple pairs of the aforementioned samples.
[0012] In some embodiments, training the deep neural network model using multiple sample pairs specifically includes: The mean and standard deviation of the induced transient electromagnetic response and the pure induction transient electromagnetic response are calculated at each time channel to serve as the normalization parameter. Each sample pair is then normalized using the normalization parameter to obtain a normalized sample pair, which includes a normalized induced transient electromagnetic response and a corresponding normalized pure induction transient electromagnetic response. The deep neural network model is trained using the normalized sample pairs through supervised learning, so that the deep neural network model learns the mapping relationship from the normalized induced transient electromagnetic response to the normalized pure induced transient electromagnetic response.
[0013] In some embodiments, constructing a training sample dataset specifically includes: For each of the randomly generated sets of geoelectric model parameters, two forward modeling calculations are performed to obtain the induced transient electromagnetic response and the pure induced transient electromagnetic response, respectively.
[0014] In some embodiments, when training the deep neural network model, mean squared error is used as the loss function, and an adaptive moment estimation optimizer is used to update the network parameters.
[0015] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for eliminating the excited polarization effect in the transient electromagnetic response described in any one of the above.
[0016] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for eliminating the induced polarization effect in the transient electromagnetic response described above.
[0017] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method for eliminating the excited polarization effect in the transient electromagnetic response described above.
[0018] According to the specific embodiments provided in this application, the following technical effects are disclosed: First, the elimination efficiency is extremely high. The trained model only requires one forward propagation to process single-channel data, taking only milliseconds. This is several orders of magnitude more efficient than traditional inversion methods, meeting the near real-time processing needs of massive datasets. Second, the results are unique and require no manual intervention. By constructing an end-to-end mapping, this invention avoids the subjective selection of parameters, basis functions, or discrimination criteria in traditional methods, ensuring the objectivity and reproducibility of the results and reducing reliance on operator experience. Third, it has strong data adaptability. Through channel-based normalization, it effectively overcomes the negative impact of the large dynamic range of transient electromagnetic data on network training, improving the model's numerical stability and generalization ability. Furthermore, the network structure is highly targeted. Employing a hybrid architecture combining convolutional and recurrent network modules, it collaboratively extracts local morphological features and global long-range dependencies from the response curve, thereby accurately capturing and eliminating subtle anomalies caused by excitation polarization effects. Finally, this invention requires only a single-channel transient electromagnetic response as input, resulting in low data requirements and wide applicability. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a flowchart illustrating a method for eliminating the excited polarization effect in a transient electromagnetic response according to an embodiment of this application.
[0021] Figure 2 This is a schematic diagram illustrating the training and inference process of a deep neural network model in one embodiment of this application.
[0022] Figure 3 This is a schematic diagram of data processing for a deep neural network model in one embodiment of this application.
[0023] Figure 4 This is a histogram showing the distribution of the average prediction relative error of the validation set at different time points in one embodiment of this application.
[0024] Figure 5This is a comparison chart of the decay curves of the induced polarization response, the actual pure inductive response, and the predicted pure inductive response of a test sample in one embodiment of this application. Detailed Implementation
[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0026] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0027] Reference Figure 1 The technical solution of this invention mainly includes two stages: an offline model training stage and an online model inference and application stage. The core objective of the model training stage is to train a high-performance deep neural network model capable of accurately mapping from induced transient electromagnetic responses to purely induced transient electromagnetic responses using a large amount of synthetic sample data through supervised learning. The model inference and application stage corresponds to... Figure 1 The complete process shown in this stage utilizes a pre-trained model to rapidly and accurately excite and eliminate polarization effects in new, unprocessed measured electromagnetic responses containing induced transient electromagnetic signals, outputting the final pure induced transient electromagnetic response, thus providing a reliable data foundation for subsequent geological interpretation.
[0028] In one exemplary embodiment, the training method for a deep neural network model includes: Step S110: Construct a training sample dataset. The training sample dataset includes multiple sample pairs. Each sample pair includes an induced transient electromagnetic response and a corresponding pure induced transient electromagnetic response.
[0029] The construction of the training sample dataset specifically includes: for multiple sets of randomly generated geoelectric model parameters, performing two forward modeling calculations for each set of geoelectric model parameters to obtain induced transient electromagnetic responses and pure induced transient electromagnetic responses (as labels) to form a sample pair.
[0030] Specifically, the geometric and acquisition parameters of the measuring device are first set, including the coordinates of each vertex of the transmission loop, the three-dimensional coordinates of the receiving point, the amplitude of the transmitted current, the current turn-off falling edge time, and the time-channel sequence; then, a large number of geodetic model parameter samples are randomly generated, for resistivity... r and time constant tLogarithmic uniform distribution sampling is used, that is, random numbers are generated uniformly in the logarithmic domain and then the exponent is taken. For polarizability m Frequency-related factors c Using uniform distribution sampling, the number of generated samples is... N A sample set, each sample containing resistivity r i Polarizability m i Frequency correlation factor c i and time constant t i The four parameters constitute the geodetic model parameter vector P. i Then, the induced transient electromagnetic response and the purely induced transient electromagnetic response corresponding to each sample are calculated separately for each geodetic model parameter vector P. i The response values of the induced polar magnetic field component in each time channel were calculated using the Cole-Cole complex resistivity model. , ,in n.t. The time channel number is represented by the value of the pure induction transient electromagnetic response calculated by forward modeling after setting the polarizability to zero. Thus formed N A sample pair containing "induced transient electromagnetic response and pure induction transient electromagnetic response".
[0031] Step S120: Train a deep neural network model using multiple sample pairs.
[0032] Specifically, this includes: calculating the mean and standard deviation of the induced transient electromagnetic response and the pure induced transient electromagnetic response at each time channel as normalization parameters; normalizing each sample pair using the normalization parameters to obtain normalized sample pairs, each normalized sample pair consisting of a normalized induced transient electromagnetic response and a corresponding normalized pure induced transient electromagnetic response; and training a deep neural network model using the normalized sample pairs through supervised learning, so that the deep neural network model learns the mapping relationship from the normalized induced transient electromagnetic response to the normalized pure induced transient electromagnetic response.
[0033] Specifically, for each j Each time path, calculate all N The mean of each sample pair in that time slot m j and standard deviation s jThe dataset includes the mean and standard deviation of the induced transient electromagnetic response and the pure induced transient electromagnetic response on the corresponding time channel. The mean and standard deviation are used to normalize each sample on the corresponding time channel to obtain a dataset composed of normalized induced transient electromagnetic responses and normalized pure induced transient electromagnetic responses. Then, the dataset is divided into a training set and a validation set by randomly shuffling the order of the sample pairs. The deep neural network model is trained using the training set and validated using the validation set.
[0034] The deep neural network model takes the normalized excited transient electromagnetic response as input and the normalized pure induced transient electromagnetic response as output. The deep neural network model first includes a sequence input layer, which receives the normalized transient electromagnetic response in sequence form. The normalized transient electromagnetic response then enters the feature extraction module, which is composed of a one-dimensional convolutional neural network and a bidirectional long short-term memory network connected in series. The one-dimensional convolutional neural network (CNN) consists of multiple one-dimensional convolutional blocks. Each block contains a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a pooling layer. The kernel size and the number of filters increase progressively. Each one-dimensional convolutional layer uses zero-padding in the time dimension to maintain the sequence length. Local morphological features of the decay curve are extracted at different time scales using convolutional kernels of different sizes. The output of the CNN then enters a bidirectional long short-term memory (BiLSTM) network layer to learn the global temporal dependencies of the entire decay curve from both forward and backward directions. A Dropout layer with a dropout probability of 0.2 is then added to prevent overfitting. The output of the feature extraction module is flattened into a one-dimensional feature vector and then enters a fully connected regression module. This vector passes through a fully connected layer containing multiple neurons, a batch normalization layer, a ReLU activation layer, and a Dropout layer. Finally, a pooling layer is set up... n.t. A fully connected output layer with 10 neurons (the same number as the time channels) directly outputs the normalized sequence prediction value of the pure inductive transient electromagnetic response without setting an activation function.
[0035] Furthermore, when training the deep neural network model, mean squared error is used as the loss function, and an adaptive moment estimation optimizer is employed to update the network parameters. Additionally, a learning rate decay strategy and an early stopping strategy are used during training. Specifically, the learning rate decay strategy can be a piecewise constant decay strategy.
[0036] Specifically, the training process is as follows: The normalized induced transient electromagnetic responses (ETRs) from the training set are input into the aforementioned deep neural network model. The mean square error between the predicted output normalized pure induced transient electromagnetic response and the labeled normalized pure induced transient electromagnetic response is used as the loss function. An adaptive moment estimation optimizer is used for parameter updates, employing a piecewise constant decay strategy and enabling an early stopping strategy. If the verification loss does not improve after multiple iterations, training is terminated early. After training, the network structure and weight parameters, as well as the normalized parameters (mean values) for each time stage, are saved. m j and standard deviation s j This parameter is used in the subsequent prediction stage, that is, in the actual inference stage, the normalization parameter is used to normalize the actual electromagnetic response containing induced transients.
[0037] Based on the training method of the deep neural network model described above, a deep neural network model can be pre-trained to eliminate the induced polarization effect in the electromagnetic response containing induced transients that actually needs to be processed.
[0038] Based on this, in an exemplary embodiment, a method for eliminating the excited polarization effect in a transient electromagnetic response is also provided, which specifically includes: Step S210: Obtain the induced transient electromagnetic response containing multiple time channels of data to be processed.
[0039] The induced transient electromagnetic response is usually measured data obtained from field exploration, and its measurement device and time series should be consistent with the training data.
[0040] Step S220: Normalize the electromagnetic response containing induced transients to convert it into a normalized electromagnetic response containing induced transients; wherein, the normalization process uses independent normalization parameters for each time channel.
[0041] The normalized parameters for each time channel include the mean and standard deviation of the induced transient electromagnetic response obtained during the pre-training of the deep neural network model for that time channel.
[0042] Specifically, this step must use the normalized parameters, i.e., the mean, saved during the training phase for each time channel. and standard deviation The measured electromagnetic response containing induced transients... Time channel data The formula for normalization is: . For normalized That is, the normalized electromagnetic response containing induced transient response. This process transforms the measured data into the exact same distribution space as the training data, making it suitable for processing by a pre-trained deep neural network model.
[0043] Step S230: The normalized electromagnetic response with induced polarization is input into a pre-trained deep neural network model. The deep neural network model processes the normalized electromagnetic response with induced polarization through its sequentially connected convolutional and recurrent network modules to output the predicted normalized pure inductive transient electromagnetic response. The convolutional modules are used to extract the local morphological features of the normalized electromagnetic response with induced polarization, and the recurrent network modules are used to learn the global temporal dependencies from the local morphological features.
[0044] The recurrent network module is a bidirectional recurrent network module, and the convolutional module is a one-dimensional convolutional neural network module. The bidirectional recurrent network module is a bidirectional long short-term memory network module; the one-dimensional convolutional neural network module includes multiple one-dimensional convolutional blocks using convolutional kernels of different sizes to extract local morphological features of the response at different time scales.
[0045] Specifically, as described in the training method above, the deep neural network model takes the normalized excited transient electromagnetic response as input and the normalized pure induced transient electromagnetic response as output. The deep neural network model first includes a sequence input layer that receives the normalized transient electromagnetic response in sequence form. The normalized transient electromagnetic response then enters the feature extraction module, which is composed of a one-dimensional convolutional neural network and a bidirectional long short-term memory network connected in series. The one-dimensional convolutional neural network (CNN) consists of multiple one-dimensional convolutional blocks. Each block contains a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a pooling layer. The kernel size and the number of filters increase progressively. Each one-dimensional convolutional layer uses zero-padding in the time dimension to maintain the sequence length. Local morphological features of the decay curve are extracted at different time scales using convolutional kernels of different sizes. The output of the CNN then enters a bidirectional long short-term memory (BiLSTM) network layer to learn the global temporal dependencies of the entire decay curve from both forward and backward directions. A Dropout layer with a dropout probability of 0.2 is then added to prevent overfitting. The output of the feature extraction module is flattened into a one-dimensional feature vector and then enters a fully connected regression module. This vector passes through a fully connected layer containing multiple neurons, a batch normalization layer, a ReLU activation layer, and a Dropout layer. Finally, a pooling layer is set up... n.t. A fully connected output layer with 10 neurons (the same number as the time channels) directly outputs the normalized sequence prediction value of the pure inductive transient electromagnetic response without setting an activation function.
[0046] Step S240: The normalized pure inductive transient electromagnetic response is denormalized to restore it to a pure inductive transient electromagnetic response that has eliminated the excitation polarization effect; wherein, the denormalization process uses the corresponding normalization parameter in the normalization process for each time channel.
[0047] This step is the reverse of the normalization process, and it also uses the preset normalization parameters (mean). and standard deviation The first normalized pure inductive transient electromagnetic response was... Time channel data The formula for inverse normalization is: . For the reverse normalization That is, the first pure inductive transient electromagnetic response This step involves processing time-channel data. The output of the deep neural network model is then restored from a dimensionless, normalized space to a purely induced transient electromagnetic response with physical meaning and original dimensions. The resulting response sequence, which eliminates the induced polarization effect, is the purely induced transient electromagnetic response and can be directly used for subsequent resistivity inversion and geological interpretation.
[0048] It should be noted that in the above-described method for eliminating the induced polarization effect in transient electromagnetic responses, the deep neural network model is pre-trained using the training method for the deep neural network model provided in the above embodiments. For example, the pre-training steps of the deep neural network model include: constructing a training sample dataset, which includes multiple sample pairs, each sample pair including an induced transient electromagnetic response and a corresponding purely induced transient electromagnetic response; and training the deep neural network model using multiple sample pairs.
[0049] The deep neural network model is trained using multiple sample pairs. Specifically, this includes: calculating the mean and standard deviation of the induced transient electromagnetic response and the pure induced transient electromagnetic response at each time channel as normalization parameters; and normalizing each sample pair using these parameters to obtain normalized sample pairs, each consisting of a normalized induced transient electromagnetic response and a corresponding normalized pure induced transient electromagnetic response. The deep neural network model is then trained using these normalized sample pairs through supervised learning, enabling the model to learn the mapping relationship from the normalized induced transient electromagnetic response to the normalized pure induced transient electromagnetic response.
[0050] The training sample dataset is constructed by performing two forward modeling calculations for each set of randomly generated geoelectric model parameters to obtain the transient electromagnetic response including induced transient response and the pure induced transient electromagnetic response.
[0051] Furthermore, when training the deep neural network model, mean squared error is used as the loss function, and an adaptive moment estimation optimizer is used to update the network parameters. In addition, a learning rate decay strategy and an early stopping strategy are also adopted.
[0052] For specific training instructions, please refer to the detailed explanation of the training methods for deep neural network models described above.
[0053] The following implementation example, combined with a specific scenario, illustrates the training method of the deep neural network model provided above and the method for eliminating the excitation polarization effect in transient electromagnetic response.
[0054] Reference Figure 2 In a specific scenario, the training (corresponding training method) and inference process of a deep neural network model (corresponding to the method for eliminating the excitation polarization effect in transient electromagnetic response) are as follows: Step 1: Construct the training sample dataset.
[0055] For a uniform half-space model of the Earth, the resistivity range is set to 1–10000 Ω·m, the polarizability range to 0–1, the frequency factor range to 0–1, and the time constant range to 0.001–1 s. The geometric and acquisition parameters of the measuring device are set. The transmission loop is a square loop with sides of 100 m, and its four vertices are (0,0), (100,0), (100,100), and (0,100). The receiving point is located at the center of the transmission loop (50,50,0). The transmission current amplitude is 1 A, the current turn-off falling edge time is 10 μs, and the time-channel sequence consists of 41 logarithmically spaced time points, ranging from 10... -2 s to 10 2 s.
[0056] 50,000 sets of geodetic model parameter samples are randomly generated within a set range.
[0057] For resistivity r i and time constant t i Logarithmic uniform distribution sampling is used, that is, random numbers are generated uniformly in the logarithmic domain and then the exponent is taken: ; ; For polarizability m i Frequency-related factors c i Uniform distribution sampling is used: ; Where U(0,1) represents a uniformly distributed random number on the interval [0,1]. The number of generated numbers is...N A sample set of 50,000, each sample containing resistivity r i Polarizability m i Frequency correlation factor c i and time constant t i The four parameters constitute the model parameter vector P. i = [ r i , m i , c i , t i ].
[0058] A time-domain finite-difference forward modeling algorithm based on the Cole-Cole model is used for each geodetic model parameter vector P. i First, the complex resistivity at each frequency point is calculated based on the Cole-Cole complex resistivity model: ; in, r ( oh () represents the complex resistivity in the frequency domain. r 0 Zero-frequency resistivity, m Polarizability t It is a time constant. c Here, ω is the frequency correlation coefficient, and ω is the angular frequency. Substituting the complex resistivity into Maxwell's equations, the response values containing the induced polarized magnetic field component at each time channel are calculated via frequency-time domain transformation. , Obtain the electromagnetic response vector containing the induced transient response: .
[0059] At the same time, the polarization rate is set to zero. m =0), at which point the formula for complex resistivity degenerates into r ( oh )= r Substituting 0 into the forward modeling, we obtain the corresponding pure induced transient electromagnetic response vector: ; For example, =41 represents the time track number. This is formed from... N A sample pair containing both induced transient electromagnetic response and pure induction transient electromagnetic response { }, i =1,2,..., N .
[0060] Step 2: Data preprocessing and normalization.
[0061] For each of the following... j Each time path, calculate all N The mean of the samples in this time period m j and standard deviation s j : ; ; Each sample is then standardized to obtain a normalized sequence of excited transient electromagnetic responses: ; After the above preprocessing, each time channel of all response sequences follows a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in dimensions and magnitudes on network training. Finally, the normalized induced response matrix X (with dimensions of...) is obtained. N × n.t. ) and the normalized pure inductive response matrix Y (size is N × n.t. Then, the dataset is split by randomly shuffling the sample order and dividing the sample set into a training set and a validation set, where the training set contains 40,000 samples and the validation set contains 10,000 samples.
[0062] Step 3: Construct a deep neural network model.
[0063] The structure of the deep neural network model is shown in the attached figure. Figure 3 As shown, the network first sets up a sequence input layer that receives a feature vector (a single time series) of size 1×41. The input data then enters the feature extraction module, which consists of a concatenated convolutional neural network and a bidirectional long short-term memory network.
[0064] The convolutional neural network part contains three one-dimensional convolutional blocks. For the input sequence... X ∈ The operation of the first convolution block can be represented as: ; in, This represents the output of the first convolutional block. Conv1D_{7,32} represents a one-dimensional convolution with a kernel size of 7 and 32 output channels. BN represents batch normalization, ReLU is the activation function, and MaxPool is max pooling with a window size of 2 and a stride of 2. The sequence length is halved after the first pooling layer.
[0065] The operation of the second convolution block can be represented as: ; in, This represents the output of the second convolutional block. Conv1D_{5,64} represents a one-dimensional convolution with a kernel size of 5 and 64 output channels.
[0066] The operation of the third convolution block can be represented as: ; in, This represents the output of the third convolutional block, i.e., the final output of the convolutional neural network. Conv1D_{3,128} indicates a one-dimensional convolution with a kernel size of 3 and 128 output channels. This layer does not perform pooling to maintain the length of the feature sequence. Each convolutional layer uses zero-padding in the time dimension to keep the sequence length unchanged before and after the convolution operation.
[0067] Output of the convolutional neural network part The input sequence is then fed into a bidirectional long short-term memory (LSM) network layer. The BiLSTM layer models the input sequence from both the forward and reverse directions, at each time step... t The forward LSTM and backward LSTM generate the hidden states respectively, and the output at that time step is obtained by concatenating them. : ; in L The current sequence length. and These are the hidden states of the forward and reverse LSTM, respectively. This indicates a vector concatenation operation.
[0068] In this example, the number of hidden cells in the BiLSTM is set to 128, so the output dimension at each time step is 256.
[0069] The BiLSTM output is then fed into a Dropout layer with a dropout probability of 0.2. ; in, This is the final output of the BiLSTM layer. This is the output of the Dropout layer, i.e., the output of the feature extraction module. After flattening Transform into one-dimensional feature vectors Next, the fully connected regression module is entered. The fully connected regression module includes two fully connected layers: a first fully connected layer and an output fully connected layer. ; ; in, and The weights and biases of the first fully connected layer (128 neurons) are given. For a Dropout layer with a dropout probability of 0.3, This is the output of the first fully connected layer. and To output the weights and biases of the fully connected layer (41 neurons), The output of the fully connected layer is used. No activation function is applied to the output fully connected layer; it directly outputs the normalized pure inductive response predictions for 41 time channels. Y pred ∈ .
[0070] Step 4: Train the deep neural network model.
[0071] The training set is input into the deep neural network model constructed in step S3, and the mean square error between its predicted output normalized pure inductive transient electromagnetic response and the normalized pure inductive transient electromagnetic response used as the label is used as the loss function: ; in For small batch sizes, For the first The first sample Predicted values of the normalized pure inductive transient electromagnetic response over a given time period. For the first The first sample The true value of the normalized pure inductive transient electromagnetic response of a time channel.
[0072] The adaptive moment estimator (Adam) optimizer is used for parameter updates. Let the network parameters to be optimized be... In the In the next iteration, the parameter update rule is: ; ; ; ; in, = 0.9 and =0.999 are the exponential decay rates of the first and second moments, respectively. or For learning rate, d =10 -8 To prevent division by zero of extremely small constants, This represents the current iteration number. Let be the network parameter vector to be optimized. For the first Network parameters at the next iteration For the first The network parameters after the next iteration. In the first The loss function value calculated under the parameters of the next iteration. For the loss function with respect to parameters gradient, This is a first-moment estimate of the gradient. This is the first-moment estimate from the previous iteration. For the second moment estimation of the gradient, This is the second-order moment estimate from the previous iteration. This is the first-order moment estimate after bias correction. This is the second-order moment estimate after bias correction. for of Power of 1 for of Power of 1.
[0073] The learning rate employs a piecewise constant decay strategy: the initial learning rate is set to... or 0 = 0.001, and the learning rate is multiplied by a decay factor of 0.5 every 30 rounds of training, i.e., the learning rate is... r The learning rate of the round is: ; In the formula The floor rule is set to 100, the maximum number of training epochs is set to 128, and the L2 regularization coefficient is set to 1×10. -5 Meanwhile, the validation set loss is evaluated every 30 iterations, and an early stopping strategy is enabled. If the validation loss does not improve within 50 consecutive rounds, training is terminated early.
[0074] After training, save the network structure and weight parameters. and the mean of normalized parameters for each time channel. m j and standard deviation s j This is used for subsequent inference and prediction stages.
[0075] Step 5: Use the trained deep neural network model to predict the elimination of induced polarization effect.
[0076] First, obtain the measured electromagnetic response vectors containing induced transients that have the same measurement device and time sequence as the training data. Utilize the channel mean saved during the training phase. mj and standard deviation s j Normalize: ; The normalized electromagnetic response vector containing the induced transient response = [ , , ..., Inputting the pre-trained deep neural network, performing a forward propagation calculation, yields the predicted value of the normalized pure inductive transient electromagnetic response. : ; in This represents the parameters of the trained network. Defined end-to-end mapping function.
[0077] Finally, the predicted values are denormalized to restore them to their original dimensions. The denormalization process is as follows: Inverse normalization transform: ; This yields the vector of the pure induced transient electromagnetic response after eliminating the excitation effect. = [ , , ..., This data is used for subsequent resistivity inversion and geological interpretation.
[0078] Appendix Figure 4 To validate the histogram of the average prediction relative error distribution at different time points, from Figure 4 The results show that the average relative prediction error of the validation set at different time points is generally within the range of 3%-7%, with an average relative error of approximately 5%. This indicates that the designed deep neural network model performs well in learning the mapping from induced response to pure inductive response. (See attached image.) Figure 5 This paper presents a comparison (double logarithmic coordinates) of the decay curves of the induced polarization response, the actual pure inductive response, and the pure inductive response predicted by the method of this invention for several test samples. Figure 5 The results show that the predicted curve is quite close to the actual curve at each time point, verifying the feasibility and effectiveness of the method.
[0079] In summary, this invention offers significant advantages. First, it boasts extremely high efficiency; after training, processing a single channel of data requires only one network forward propagation, taking only milliseconds, representing an efficiency improvement of several orders of magnitude compared to traditional inversion methods. Second, the results are unique and require no manual intervention; the end-to-end mapping avoids the subjectivity and non-uniqueness issues of traditional methods. Third, it exhibits strong data adaptability; channel-based normalization effectively addresses the problem of vastly different data volumes. Furthermore, the network structure is highly targeted; the combination of convolutional and recurrent network modules can comprehensively extract temporal features. Finally, this method requires only a single channel input, has low data requirements, and is widely applicable.
[0080] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0081] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0082] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0083] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0084] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0085] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for eliminating excited polarization effects in transient electromagnetic responses, characterized in that, include: Acquire the induced transient electromagnetic response containing multiple time channels to be processed; The electromagnetic response containing induced transients is normalized to convert it into a normalized electromagnetic response containing induced transients; wherein the normalization process uses independent normalization parameters for each time channel. The normalized electromagnetic response with induced transient emission is input into a pre-trained deep neural network model. The deep neural network model processes the normalized electromagnetic response with induced transient emission through its sequentially connected convolutional and recurrent network modules to output a predicted normalized pure inductive transient electromagnetic response. The convolutional modules are used to extract local morphological features of the normalized electromagnetic response with induced transient emission, and the recurrent network modules are used to learn global temporal dependencies from the local morphological features. The normalized pure inductive transient electromagnetic response is denormalized to restore it to a pure inductive transient electromagnetic response free from excitation polarization effects; wherein the denormalization process uses the corresponding normalization parameter in the normalization process for each time channel.
2. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 1, characterized in that, The recurrent network module is a bidirectional recurrent network module, and the convolution module is a one-dimensional convolutional neural network module.
3. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 2, characterized in that, The bidirectional cyclic network module is a bidirectional long short-term memory network module; The one-dimensional convolutional neural network module includes multiple one-dimensional convolutional blocks with convolutional kernels of different sizes to extract local morphological features of the response at different time scales.
4. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 1, characterized in that, The normalized parameters for each time channel include the mean and standard deviation of the induced transient electromagnetic response obtained during the pre-training of the deep neural network model for that time channel.
5. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 1, characterized in that, The pre-training steps of the deep neural network model include: Construct a training sample dataset, which includes multiple sample pairs, each of which includes an induced transient electromagnetic response and a corresponding pure induced transient electromagnetic response. The deep neural network model is trained using multiple pairs of the aforementioned samples.
6. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 5, characterized in that, Training the deep neural network model using multiple sample pairs specifically includes: The mean and standard deviation of the induced transient electromagnetic response and the pure induction transient electromagnetic response are calculated at each time channel to serve as the normalization parameter. Each sample pair is then normalized using the normalization parameter to obtain a normalized sample pair, which includes a normalized induced transient electromagnetic response and a corresponding normalized pure induction transient electromagnetic response. The deep neural network model is trained using the normalized sample pairs through supervised learning, so that the deep neural network model learns the mapping relationship from the normalized induced transient electromagnetic response to the normalized pure induced transient electromagnetic response.
7. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 5, characterized in that, Constructing the training sample dataset specifically includes: For each of the randomly generated sets of geoelectric model parameters, two forward modeling calculations are performed to obtain the induced transient electromagnetic response and the pure induced transient electromagnetic response, respectively.
8. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 5, characterized in that, When training the deep neural network model, mean squared error is used as the loss function, and an adaptive moment estimation optimizer is used to update the network parameters.
9. The method for eliminating excited polarization effects in transient electromagnetic responses according to claim 8, characterized in that, When training the deep neural network model, a learning rate decay strategy and an early stopping strategy are also employed.
10. A computer program product comprising a computer program that, when executed by a processor, implements a method for eliminating the excitation polarization effect in a transient electromagnetic response as described in any one of claims 1-9.