An airborne electromagnetic induction-polarization effect multi-parameter extraction method and system

Abstract

The application belongs to the field of geophysical exploration, and is an airborne electromagnetic induction-polarization effect multi-parameter extraction method and system, comprising a hybrid neural network model, wherein the hybrid neural network model comprises: a spatial feature extraction module, which is used for extracting spatial features from an airborne electromagnetic magnetic induction intensity response; the spatial feature extraction module adopts a convolutional neural network, an attention module is introduced into the convolutional neural network, global compression is performed on the extracted spatial features, inter-channel dependency is captured, attention weights are generated, and the attention weights are multiplied with the spatial features element by element through point multiplication; and a time feature extraction module, which is used for capturing time dependency of the airborne electromagnetic magnetic induction intensity response, and adopts a GRU network. Efficient decoupling and accurate extraction of the resistivity and time constant of the maximum phase angle model are realized. The accuracy and resolution of double-parameter inversion are effectively improved.

Description

Technical Field

[0001] This application belongs to the field of geophysical exploration, specifically a method and system for extracting multiple parameters of airborne electromagnetic induction-polarization effect. Background Technology

[0002] In the field of electromagnetic detection, the Transient Electromagnetic Method (TEM) is an important geophysical exploration method widely used in groundwater exploration, oil and gas exploration, mineral resource exploration, and shallow geological surveys. Currently, methods for interpreting TEM data and extracting parameters mainly include traditional linear methods, nonlinear methods based on intelligent optimization algorithms, and data-driven methods based on machine learning and artificial intelligence. The problem of interpreting and extracting TEM data is characterized by nonlinearity, multiple extrema, and multiple parameters. Artificial neural networks, by fitting certain nonlinear mapping relationships in the data, have great potential in solving the problem of interpreting TEM data.

[0003] Currently, most machine learning-based studies on transient electromagnetic parameter extraction do not consider the induction-polarization co-occurrence effect, relying solely on resistivity parameters for interpretation. This approach easily leads to multiple solutions, thus limiting detection resolution and hindering accurate measurement of deep metal deposits. Existing methods ignore the influence of polarization effects, resulting in incomplete inversion models, low adaptability, and increased multiple solutions. Especially for deep mineral exploration, resistivity alone offers low resolution and accuracy, making it difficult to accurately distinguish certain geological bodies and reducing the efficiency and effectiveness of mineral exploration. Summary of the Invention

[0004] This application provides a method and system for extracting multiple parameters of airborne electromagnetic induction-polarization effects, solving the problem that resistivity alone has low resolution and accuracy, making it difficult to accurately distinguish certain geological bodies and reducing the efficiency and effectiveness of mineral exploration.

[0005] The first aspect of this application provides a multi-parameter extraction system for airborne electromagnetic induction-polarization effects, including a hybrid neural network model, the hybrid neural network model comprising:

[0006] A spatial feature extraction module is used to extract spatial features from the airborne electromagnetic induction intensity response. The spatial feature extraction module adopts a convolutional neural network and introduces an attention module into the convolutional neural network to globally compress the extracted spatial features, capture the dependencies between channels, and generate attention weights. The attention weights are multiplied element-wise with the spatial features through dot product.

[0007] A time feature extraction module is used to capture the time dependence of the airborne electromagnetic induction intensity response. The time feature extraction module adopts a GRU network.

[0008] Furthermore, the spatial feature extraction module extracts local features from the data using a convolutional module consisting of 6 convolutional layers. The kernel sizes of the convolutional layers are 3×1, 3×1, 5×1, 5×1, 5×1 and 3×1, with corresponding channel numbers of 128, 128, 64, 64, 64 and 64, and a stride of 1 for each layer. After each convolutional layer, the ReLU activation function is used to map the linear output to a non-linear output. After each ReLU activation layer, a max pooling operation with a size of 1×2 and a stride of 1 is performed.

[0009] Furthermore, the attention module calculates attention weights through a fully connected layer and ReLU and Sigmoid activation functions.

[0010] Furthermore, the time feature extraction module includes a multi-layer GRU network that captures deep temporal features of the data layer by layer; the predicted values ​​of the output parameters are obtained through a fully connected layer, and the predicted values ​​are converted back to the original data space through inverse normalization and exponential restoration operations to obtain the resistivity and time constant of the maximum phase angle model.

[0011] Furthermore, the hybrid neural network model also includes a regression layer, which calculates the root mean square error between the predicted and true values ​​of the output parameters of the fully connected layer, selects the root mean square error as the loss function of the hybrid neural network model, and optimizes the hyperparameters of the hybrid neural network model by minimizing the loss function.

[0012] The second aspect of this application provides a method for extracting multiple parameters of airborne electromagnetic induction-polarization effect, including:

[0013] Constructing a sample set includes: determining the range of resistivity and polarizability of the subsurface medium Cole-Cole model based on the geological information of the actual detection area; using forward modeling to calculate the airborne electromagnetic induction intensity response as input to the sample set; converting the resistivity and polarizability of the Cole-Cole model into the resistivity and time constant of the maximum phase angle model and using them as output to the sample set.

[0014] A hybrid neural network model is constructed, comprising: a spatial feature extraction module for extracting spatial features from input data; and a temporal feature extraction module for capturing the temporal dependence of aeronautical electromagnetic data.

[0015] A hybrid neural network model is trained using a sample set;

[0016] The airborne electromagnetic induction intensity response collected in the actual detection area is input into the hybrid neural network model to obtain the resistivity and time constant of the maximum phase angle model.

[0017] Furthermore, the spatial feature extraction module employs a convolutional neural network, and an attention module is introduced into the convolutional neural network to globally compress the extracted spatial features, capture the dependencies between channels, and generate attention weights. The attention weights are then multiplied element-wise with the spatial features through dot product.

[0018] A time feature extraction module is used to capture the time dependence of airborne electromagnetic data, and the time feature extraction module adopts a GRU network.

[0019] Furthermore, the spatial feature extraction module extracts local features from the data using a convolutional module consisting of 6 convolutional layers. The kernel sizes of the convolutional layers are 3×1, 3×1, 5×1, 5×1, 5×1 and 3×1, with corresponding channel numbers of 128, 128, 64, 64, 64 and 64, and a stride of 1 for each layer. After each convolutional layer, the ReLU activation function is used to map the linear output to a non-linear output. After each ReLU activation layer, a max pooling operation with a size of 1×2 and a stride of 1 is performed.

[0020] Furthermore, the time feature extraction module includes a multi-layer GRU network that captures deep temporal features of the data layer by layer; the predicted values ​​of the output parameters are obtained through a fully connected layer, and the predicted values ​​are converted back to the original data space through inverse normalization and exponential restoration operations to obtain the resistivity and time constant of the maximum phase angle model.

[0021] Furthermore, the step of training the hybrid neural network model using a sample set includes: employing a regression layer, calculating the root mean square error between the predicted and true values ​​of the output parameters of the fully connected layer, selecting the root mean square error as the loss function of the hybrid neural network model, and optimizing the hyperparameters of the hybrid neural network model by minimizing the loss function.

[0022] Compared with existing technologies, this application offers the following advantages: By extracting spatial features and capturing the dynamic changes of time series, and introducing an attention mechanism to assign higher weights to important features, this application achieves efficient decoupling and accurate extraction of resistivity and time constant from the maximum phase angle model. This effectively improves the accuracy and resolution of two-parameter inversion and is suitable for complex geological conditions with significant polarization effects. Attached Figure Description

[0023] Figure 1 A flowchart of the multi-parameter extraction method for airborne electromagnetic induction-polarization effect provided in the embodiments of this application;

[0024] Figure 2A model diagram of the hybrid neural network model provided in the embodiments of this application;

[0025] Figure 3 A set of resistivity-depth curves provided for embodiments of this application;

[0026] Figure 4 A set of time constant-depth curves are provided for embodiments of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] See Figure 1 The flowchart shown illustrates a method for extracting multiple parameters of airborne electromagnetic induction-polarization effects. This method includes:

[0029] S1 constructs a sample set, which includes: determining the range of resistivity and polarizability of the subsurface medium Cole-Cole model based on the geological information of the actual detection area; using forward modeling to calculate the airborne electromagnetic induction intensity response as input to the sample set; converting the resistivity and polarizability of the Cole-Cole model into the resistivity and time constant of the maximum phase angle model, and using them as outputs of the sample set.

[0030] In one example, a one-dimensional three-layered geodetic model was used to construct the sample set. First, a Cole-Cole model was used, with resistivity values ​​ranging from 100 to 900. 100 every interval The polarizability was set to [0.1, 0.5], with increments of 0.1. The time constant was 0.022, and the dispersion coefficient was 0.8. The first two layers of the three-layered model were 100m deep. The resistivity and polarizability were correlated one-to-one, and the corresponding airborne electromagnetic induction intensity response was calculated as input data for the sample set, resulting in a total of 91,125 data sets.

[0031] Because the polarizability and dispersion coefficient in the Cole-Cole model are strongly correlated, multiple solutions can easily arise during data interpretation. To weaken the correlation between variables and improve interpretation accuracy, the resistivity and polarizability of the Cole-Cole model are converted into the resistivity and time constant of the maximum phase angle model, which are then used as the output data of the sample set. The formula is as follows: ,in, It is the resistivity of the maximum phase angle model. It is the resistivity of the Cole-Cole model. It is the polarizability of the Cole-Cole model. It is the dispersion coefficient of the Cole-Cole model.

[0032] The sample set is preprocessed by logarithmic transformation and normalization. The normalization method uses Min-Max normalization, and the preprocessed sample set is then divided into training and testing sets. The formula for the Min-Max normalization method is as follows: ,in, It is the raw data. It is normalized data. and These are the maximum and minimum values ​​in the sample set, respectively. and These are the normalized minimum and maximum values, which are the lower and upper bounds of the target interval, respectively.

[0033] The first 70% of the sample set was used as the training set, and the remaining 30% was used as the test set. Then, the sample set was preprocessed by logarithmic transformation and normalization. Since the airborne electromagnetic magnetic induction intensity response with polarization effect contains negative values, the normalization range was set to [-1, 0.5]. Finally, the data was flattened and converted into sequence data format for subsequent model training.

[0034] S2 constructs a hybrid neural network model, which includes: a spatial feature extraction module for extracting spatial features from input data; and a temporal feature extraction module for capturing the temporal dependence of aeronautical electromagnetic data.

[0035] S3 uses a sample set to train a hybrid neural network model;

[0036] S4 inputs the airborne electromagnetic induction intensity response collected in the actual detection area into the hybrid neural network model to obtain the resistivity and time constant of the maximum phase angle model.

[0037] In one example, the spatial feature extraction module uses a convolutional neural network, and an attention module is introduced into the convolutional neural network to globally compress the extracted spatial features, capture the dependencies between channels, and generate attention weights. The attention weights are then multiplied element-wise with the spatial features through a dot product.

[0038] A time feature extraction module is used to capture the time dependence of airborne electromagnetic data, and the time feature extraction module adopts a GRU network.

[0039] The hybrid neural network model has an input layer dimension of 1×50 and an output layer dimension of 1×6. The input data first undergoes a series of convolutional operations, with a convolutional module consisting of 6 convolutional layers sequentially extracting local features from the data. The kernel sizes of each convolutional layer are 3×1, 3×1, 5×1, 5×1, 5×1, and 3×1, corresponding to 128, 128, 64, 64, 64, and 64 channels, with a stride of 1 for each layer. Through these convolutional operations, local features at different scales can be captured, and by deepening the network layer by layer, more complex pattern information can be extracted. After each convolutional layer, the ReLU activation function is used to map the linear output to a non-linear output, further enhancing the expressive power and enabling it to learn complex non-linear relationships. Immediately following each ReLU activation function, a max pooling operation with a size of 1×2 and a stride of 1 is performed. Pooling reduces the size of the features, thereby effectively reducing computational complexity while retaining the most significant feature information and reducing redundancy.

[0040] The convolutional part incorporates an attention module, which calculates attention weights using fully connected layers and ReLU and Sigmoid activation functions. These attention weights are then multiplied by the convolutional features to enhance focus on key features. Subsequently, the convolutional features are transformed back into sequential data through a sequence defolding layer and smoothed before being input into the temporal feature extraction module. This module receives features from the attention and spatial feature extraction modules and learns long-range dependencies in the time series through a recursive structure, capturing deep temporal features layer by layer. Finally, the predicted values ​​of the output parameters from a fully connected layer are transformed back into the original data space through inverse normalization and exponential restoration operations, thus achieving accurate extraction of the target physical quantities, namely the resistivity and time constant of the maximum phase angle model.

[0041] The attention module is implemented through two fully connected layers. First, features are globally compressed, and then the two fully connected layers capture the dependencies between channels to generate attention weights. Next, the attention weights are element-wise multiplied with the original features through a dot product layer, dynamically enhancing the expressive power of important features and suppressing irrelevant features, further improving the model's sensitivity to key features. During this process, folded and defolded layers can be used to simplify the processing of sequence data, mapping the input temporal data to a format suitable for convolutional operations, and restoring the original temporal format after the convolution and attention mechanisms are completed.

[0042] After feature extraction and attention mechanism processing, the data is input into the temporal feature extraction module. This module comprises a multi-layer GRU network, capturing deep temporal features of the data layer by layer. The predicted values ​​of the output parameters are then transformed back to the original data space through inverse normalization and exponential restoration operations to obtain the resistivity and time constant of the maximum phase angle model.

[0043] Multilayer GRU networks focus on capturing short-term and long-term dependencies in time series data. Through their gating mechanism, they can effectively selectively remember and forget information, thus avoiding long-term dependency problems. The output mode of each GRU layer is set to "last," meaning only the output of the last time step is retained. This setting allows for the direct extraction of global features from the sequence data and provides efficient representations for subsequent processing.

[0044] Fully connected layers map high-dimensional features to the target output dimension through linear transformation. The role of the fully connected layers is to synthesize various features and transform them into the final predicted output, namely the resistivity and time constant of the maximum phase angle model. Regression layers are used to calculate the root mean square error (RMSE) between the predicted and true values ​​of the fully connected layer's output parameters. The RMSE is selected as the loss function of the hybrid neural network model, and the hyperparameters of the hybrid neural network model are optimized by minimizing the loss function.

[0045] The entire hybrid neural network model is trained using the Adam optimization algorithm, employing a piecewise learning rate descent strategy. The learning rate is automatically adjusted during training to accelerate convergence. Once the required number of training iterations is reached, the trained hybrid neural network model is obtained. The root mean square error (RMSE) formula is as follows: In the formula: and The first Predicted and theoretical values ​​for each sample This represents the number of samples.

[0046] The root mean square error (RMSE) is chosen as the loss function, and the hyperparameters of the hybrid neural network model are optimized by minimizing the loss function. The Adam adaptive learning rate optimization algorithm is employed to ensure convergence. In one example, the initial learning rate is set to 0.001, the learning rate decay factor is 0.1, the learning rate decay period is 20, and the maximum number of iterations is 69.

[0047] The airborne electromagnetic induction intensity response collected from the actual detection area is input into the trained hybrid neural network model for parameter extraction. This yields the resistivity and time constant of the maximum phase angle model of the subsurface medium to be extracted. The parameter extraction effect is evaluated by calculating the Mean Absolute Percentage Error (MAPE) between the predicted and theoretical parameter values. Furthermore, by comparing the error information between the parameter extraction results and known geological models, the applicability and performance of this application under specific geological conditions are further analyzed, leading to reverse optimization of the training sample set and the hybrid neural network model structure. For example, the resistivity and polarizability distribution ranges in the sample set can be adjusted based on geological information, or the training process of the hybrid neural network model can be optimized by reselecting hyperparameters such as the learning rate, number of neurons, and number of layers, thereby improving the generalization ability of the hybrid neural network model in real-world scenarios. The formula for the Mean Absolute Percentage Error is as follows: .

[0048] A trained hybrid neural network model is used to extract parameters from the actual collected data, and the mean absolute percentage error between the theoretical and predicted values ​​of resistivity and time constant for the three layers is calculated. A set of data is selected, and the theoretical and predicted resistivity-depth curves are obtained as follows: Figure 3 The time constant-depth curve shown is as follows: Figure 4 As shown, the results visually demonstrate the agreement between the parameter predictions of this application at different depths and the actual conditions. The results show that the predicted values ​​of the extracted resistivity and time constant curves have a high degree of consistency with the theoretical values.

[0049] On the other hand, see Figure 2 As shown, this application provides a multi-parameter extraction system for airborne electromagnetic induction-polarization effects, which can be explained in comparison with the method for extracting multi-parameters of airborne electromagnetic induction-polarization effects. It includes a hybrid neural network model, which comprises:

[0050] A spatial feature extraction module is used to extract spatial features from the airborne electromagnetic induction intensity response. The spatial feature extraction module adopts a convolutional neural network and introduces an attention module into the convolutional neural network to globally compress the extracted spatial features, capture the dependencies between channels, and generate attention weights. The attention weights are multiplied element-wise with the spatial features through dot product.

[0051] A time feature extraction module is used to capture the time dependence of the airborne electromagnetic induction intensity response. The time feature extraction module adopts a GRU network.

[0052] In one embodiment, the spatial feature extraction module extracts local features from the data using a convolutional module consisting of 6 convolutional layers. The kernel sizes of the convolutional layers are 3×1, 3×1, 5×1, 5×1, 5×1 and 3×1, with corresponding channel numbers of 128, 128, 64, 64, 64 and 64, and a stride of 1 for each layer. After each convolutional layer, the ReLU activation function is used to map the linear output to a non-linear output. After each ReLU activation layer, a max pooling operation with a size of 1×2 and a stride of 1 is performed.

[0053] In one embodiment, the attention module calculates attention weights through a fully connected layer and ReLU and Sigmoid activation functions.

[0054] In one embodiment, the time feature extraction module includes a multi-layer GRU network that captures deep temporal features of the data layer by layer; the predicted values ​​of the output parameters are obtained through a fully connected layer, and the predicted values ​​are converted back to the original data space through inverse normalization and exponential restoration operations to obtain the resistivity and time constant of the maximum phase angle model.

[0055] In one embodiment, the hybrid neural network model further includes a regression layer, which calculates the root mean square error between the predicted and true values ​​of the output parameters of the fully connected layer, selects the root mean square error as the loss function of the hybrid neural network model, and optimizes the hyperparameters of the hybrid neural network model by minimizing the loss function.

[0056] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for extracting multiple parameters of airborne electromagnetic induction-polarization effect, characterized in that, include: Constructing a sample set includes: determining the range of resistivity and polarizability of the subsurface medium Cole-Cole model based on the geological information of the actual detection area; using forward modeling to calculate the airborne electromagnetic induction intensity response as input to the sample set; converting the resistivity and polarizability of the Cole-Cole model into the resistivity and time constant of the maximum phase angle model and using them as output to the sample set. A hybrid neural network model is constructed, comprising: a spatial feature extraction module for extracting spatial features from input data; and a temporal feature extraction module for capturing the temporal dependence of aeronautical electromagnetic data. The spatial feature extraction module employs a convolutional neural network, and an attention module is introduced into the convolutional neural network to globally compress the extracted spatial features, capture the dependencies between channels, and generate attention weights. The attention weights are then multiplied element-wise with the spatial features through a dot product. The time feature extraction module is used to capture the time dependence of airborne electromagnetic data. The time feature extraction module includes a multi-layer GRU network, which captures the deep temporal features of the data layer by layer. The predicted values ​​of the output parameters are passed through a fully connected layer. The predicted values ​​are transformed back to the original data space through inverse normalization and exponential restoration operations to obtain the resistivity and time constant of the maximum phase angle model. A hybrid neural network model is trained using a sample set; The airborne electromagnetic induction intensity response collected in the actual detection area is input into the hybrid neural network model to obtain the resistivity and time constant of the maximum phase angle model.

2. The method for extracting multiple parameters of airborne electromagnetic induction-polarization effect according to claim 1, characterized in that, The spatial feature extraction module consists of a convolutional module composed of 6 convolutional layers to extract local features from the data. The kernel sizes of the convolutional layers are 3×1, 3×1, 5×1, 5×1, 5×1 and 3×1, with corresponding channel numbers of 128, 128, 64, 64, 64 and 64, and a stride of 1 for each layer. After each convolutional layer, the ReLU activation function is used to map the linear output to a non-linear output. After each ReLU activation layer, a max pooling operation with a size of 1×2 and a stride of 1 is performed.

3. The method for extracting multiple parameters of airborne electromagnetic induction-polarization effect according to claim 1, characterized in that, The method of training a hybrid neural network model using a sample set includes: using a regression layer, calculating the root mean square error between the predicted and true values ​​of the output parameters of the fully connected layer, selecting the root mean square error as the loss function of the hybrid neural network model, and optimizing the hyperparameters of the hybrid neural network model by minimizing the loss function.

4. A multi-parameter extraction system for airborne electromagnetic induction-polarization effect, used to implement the method described in any one of claims 1-3, characterized in that, This includes a hybrid neural network model, which comprises: A spatial feature extraction module is used to extract spatial features from the airborne electromagnetic induction intensity response. The spatial feature extraction module adopts a convolutional neural network and introduces an attention module into the convolutional neural network to globally compress the extracted spatial features, capture the dependencies between channels, and generate attention weights. The attention weights are multiplied element-wise with the spatial features through dot product. The time feature extraction module is used to capture the time dependence of the airborne electromagnetic induction intensity response. The time feature extraction module includes a multi-layer GRU network, which captures the deep temporal features of the data layer by layer. The predicted values ​​of the output parameters are passed through a fully connected layer. The predicted values ​​are converted back to the original data space through inverse normalization and exponential restoration operations to obtain the resistivity and time constant of the maximum phase angle model.

5. The multi-parameter extraction system for airborne electromagnetic induction-polarization effect according to claim 4, characterized in that, The spatial feature extraction module consists of a convolutional module composed of 6 convolutional layers to extract local features from the data. The kernel sizes of the convolutional layers are 3×1, 3×1, 5×1, 5×1, 5×1 and 3×1, with corresponding channel numbers of 128, 128, 64, 64, 64 and 64, and a stride of 1 for each layer. After each convolutional layer, the ReLU activation function is used to map the linear output to a non-linear output. After each ReLU activation layer, a max pooling operation with a size of 1×2 and a stride of 1 is performed.

6. The multi-parameter extraction system for airborne electromagnetic induction-polarization effect according to claim 4, characterized in that, The attention module calculates attention weights using a fully connected layer and ReLU and Sigmoid activation functions.

7. The multi-parameter extraction system for airborne electromagnetic induction-polarization effect according to claim 4, characterized in that, The hybrid neural network model also includes a regression layer, which calculates the root mean square error between the predicted and true values ​​of the output parameters of the fully connected layer, selects the root mean square error as the loss function of the hybrid neural network model, and optimizes the hyperparameters of the hybrid neural network model by minimizing the loss function.

Images