A ground transient electromagnetic inversion method based on CNN-lightgbm
By combining CNN and LightGBM methods, the problems of computational time consumption and limited generalization ability in ground transient electromagnetic inversion are solved, and efficient and reliable underground structure inversion is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH OF CHINA
- Filing Date
- 2023-08-16
- Publication Date
- 2026-07-31
AI Technical Summary
Existing ground transient electromagnetic inversion methods rely on accurate prior information about underground structures, which easily leads to multiple local minima, are computationally time-consuming, have limited generalization ability, and are difficult to efficiently process large amounts of data.
By combining CNN-LightGBM with ground transient electromagnetic response data, multi-scale features are extracted through CNN and regression is performed using LightGBM to establish a mapping relationship between data and resistivity, thereby improving the reliability and efficiency of the inversion.
It achieves fast and reliable underground structure inversion, improves inversion efficiency by several orders of magnitude, avoids local minima and overfitting, and enhances the model's generalization ability.
Smart Images
Figure CN117471560B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of geophysical technology, specifically relating to a ground transient electromagnetic inversion method based on CNN-LightGBM. Background Technology
[0002] Ground-based transient electromagnetic (TEM) is a branch of transient electromagnetic methods that involves ground-based electromagnetic detection. TEM has wide applications in mineral resource exploration, engineering surveys, groundwater resource investigations, and environmental geological surveys. The key to obtaining subsurface information from ground-based TEM data is inversion.
[0003] Ground transient electromagnetic inversion is a highly nonlinear problem. The inventors have found that current inversion methods can generally be divided into two main categories: deterministic and probabilistic. Deterministic inversion obtains optimal model parameters by minimizing the objective function (the difference between measured and forward modeling data). This method requires relatively accurate prior information about the subsurface structure and often encounters multiple local minima. Probabilistic inversion, on the other hand, transforms the inverse problem into solving for the probability distribution of the inversion parameters. This method is particularly suitable for highly nonlinear problems; however, probabilistic methods are computationally very time-consuming, and this time consumption increases exponentially with the number of model parameters. Inverting large amounts of ground transient electromagnetic data can take several days, thus limiting the process of understanding subsurface structures. Machine learning, as a global search algorithm, provides an alternative for ground transient electromagnetic inversion. Convolutional neural networks (CNNs) are the most common type of deep neural network and one of the most powerful developments in machine learning in recent decades. Nevertheless, the currently used fully connected CNN architectures are susceptible to limitations such as slow training speed, limited generalization ability, and susceptibility to overfitting. Unlike neural networks, Lightweight Gradient Boosting Machine Learning (LightGBM) is an ensemble learning algorithm that combines multiple weak learners into a single strong learner to improve prediction accuracy. In conclusion, there is an urgent need to improve ground-based transient electromagnetic inversion methods to enhance efficiency and reliability. Summary of the Invention
[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a ground transient electromagnetic response (TEM) inversion method based on CNN-LightGBM. This invention first inputs ground transient electromagnetic response data into a CNN to extract multi-scale features of the ground transient electromagnetic response. Then, it uses LightGBM to regress the extracted multi-scale features of the ground transient electromagnetic response to improve the reliability of the inversion.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:
[0006] A ground transient electromagnetic inversion method based on CNN-LightGBM, which specifically includes the following steps:
[0007] Step 1: Generate a sample dataset using a ground-based transient electromagnetic forward modeling program, including:
[0008] The parameters of the forward modeling procedure are determined based on the acquisition parameters obtained from ground transient electromagnetic method detection.
[0009] Randomly select the resistivity of each underground model layer within the set resistivity range;
[0010] Based on the determined parameters and the resistivity of each subsurface model, forward modeling was used to generate transient electromagnetic response data of the ground.
[0011] Preprocessing of ground transient electromagnetic response data and resistivity of each subsurface model;
[0012] The preprocessed ground transient electromagnetic response data is used as the input to CNN-LightGBM, and the resistivity of each subsurface model after preprocessing is used as the label of the input data of CNN-LightGBM.
[0013] The sample dataset is divided into a training set, a validation set, and a test set;
[0014] Step 2: Construct CNN-LightGBM and determine its optimal parameters;
[0015] The multi-scale features of the transient electromagnetic response of the ground extracted by CNN are flattened into a one-dimensional vector, and then input into LightGBM to establish the mapping relationship between the transient electromagnetic response data of the ground and the resistivity of each subsurface model. The parameters to be optimized in CNN include the number of convolutional and fully connected layers, the size of the convolutional kernel, and the learning rate. The parameters to be optimized in LightGBM include the depth of the decision tree, the number of leaves, the learning rate, and the regularization parameter.
[0016] Step 3: Train CNN-LightGBM using optimal parameters;
[0017] CNN-LightGBM is trained using the training and validation sets from step 1. During training, the optimal parameters are used to obtain the best inversion performance. Finally, the trained CNN-LightGBM is saved.
[0018] Step 4: Validate the inversion effect of CNN-LightGBM using the test set;
[0019] The test set is data that CNN-LightGBM did not use during training. It is used to evaluate the generalization ability of CNN-LightGBM. The inversion result is evaluated based on the error between the inversion result of CNN-LightGBM on the test set and the label.
[0020] Step 5: Use the trained CNN-LightGBM to invert the measured ground transient electromagnetic data;
[0021] After preprocessing the measured transient electromagnetic data of the ground, the data is input into the trained CNN-LightGBM to obtain the resistivity of each subsurface model.
[0022] Preferably, the resistivity range set in step 1 is 0.1-15000 Ω·m, which includes the resistivity of most materials in nature; the preprocessing method for the ground transient electromagnetic response data and the resistivity of each subsurface model is as follows: Subtract the mean μ from the transient electromagnetic response data x and divide by the standard deviation σ. The labels are the logarithmic transformation values of the resistivity ρ for each subsurface model: The training set, validation set, and test set comprise 90%, 5%, and 5% of the sample dataset, respectively.
[0023] Preferably, in step 2, the optimal parameters of CNN-LightGBM are determined using random search or Bayesian optimization; leaky ReLU is used as the activation function of the CNN; and a dropout operation is performed before the fully connected layers of the CNN to randomly disable a portion of neurons, thereby preventing overfitting and improving robustness and generalization ability.
[0024] Preferably, in step 3, an early stopping strategy is employed, and learning rate decay is used to prevent CNN-LightGBM from overfitting and to reduce its training time; the loss function of CNN-LightGBM in step 3 is:
[0025]
[0026] Where N represents the batch size. y and y represent the resistivity of the label and the inverted subsurface model, respectively, M is the number of network output parameters, and a regularization term with weight λ is introduced to improve the model's generalization ability. This represents the sum of squares of the trainable weights W in CNN-LightGBM.
[0027] Preferably, in step 4, the root mean square error (RMSE) is used to verify the inversion effect of CNN-LightGBM, defined as:
[0028] The beneficial effects of this invention are:
[0029] This invention inputs ground transient electromagnetic response (TEMP) data into a CNN to extract multi-scale features of the TEMP. Then, it uses LightGBM to regress the extracted key features to improve the reliability of the inversion. CNNs can learn complex abstract features, while LightGBM can better capture the nonlinear relationships and interactions between data. Combining the two improves the model's inversion capability. It overcomes the limitations of currently used fully connected CNN architectures, which are prone to slow training speed, limited generalization ability, and overfitting. Compared to traditional methods, the inversion is independent of the initial model selection, thus avoiding the local minima problem common in traditional methods. Furthermore, it significantly accelerates the inversion efficiency; this invention can invert 10,000 samples per second, an efficiency several orders of magnitude faster than traditional methods. Attached Figure Description
[0030] Figure 1 This is a flowchart of a ground transient electromagnetic inversion method based on CNN-LightGBM according to an embodiment of the present invention;
[0031] Figure 2 This is a schematic diagram of the architecture of the CNN-LightGBM inversion method of this invention;
[0032] Figure 3 This is a comparison between the inversion results of the CNN-LightGBM inversion method of this invention and the CNN inversion results. Detailed Implementation
[0033] The following description, in conjunction with the accompanying drawings, further illustrates the CNN-LightGBM-based ground transient electromagnetic response inversion method proposed in this invention. As introduced in the background section, existing techniques require relatively accurate prior information on subsurface structures, often encountering multiple local minima or being computationally very time-consuming, thus limiting the process of understanding subsurface structures. This invention inputs ground transient electromagnetic response data into a CNN to extract multi-scale features of the ground transient electromagnetic response, and then uses LightGBM to regress the extracted key features to improve the reliability of the inversion.
[0034] This embodiment provides a ground transient electromagnetic inversion method based on CNN-LightGBM, such as Figure 1 As shown, it includes the following steps:
[0035] Step 1: Generate a sample dataset using a ground-based transient electromagnetic forward modeling program.
[0036] First, the parameters of the forward modeling program are determined based on the acquisition parameters from the ground transient electromagnetic method (TEM). These parameters include the size of the transmitting coil, the observation time range, and the magnitude of the transmitting current. The resistivity of each subsurface model is randomly selected within the resistivity range of 0.1–15000 Ω·m. Then, according to the determined parameters of the forward modeling program and the resistivity of each subsurface model, the TEM generates ground transient electromagnetic response data. The input to CNN-LightGBM is the transient electromagnetic response data x minus the mean μ divided by the standard deviation σ. The labels are the logarithmic transformation values of the resistivity ρ for each subsurface model: Finally, 90% of the sample dataset was divided into a training set, 5% into a validation set, and 5% into a test set.
[0037] Step 2: Construct CNN-LightGBM and determine its optimal parameters;
[0038] Figure 2 A schematic diagram of the architecture of the CNN-LightGBM inversion method of this invention is given. The input of CNN-LightGBM is transient electromagnetic response data at 40 time sampling points, and the output is the resistivity of a 9-layer underground model. The optimal parameters of CNN-LightGBM were determined using random search or Bayesian optimization. The parameters to be optimized for CNN include the number of convolutional and fully connected layers, the size of the convolutional kernels, and the learning rate. The parameters to be optimized for LightGBM include the depth of the decision tree, the number of leaves, the learning rate, and the regularization parameter. Path 1 of CNN-LightGBM is a CNN, which includes three convolutional layers with 16, 32, and 64 kernels respectively, and three fully connected layers with 128, 128, and 9 neurons respectively. The kernel size is 5, and the learning rate is 0.001. Each convolutional layer is followed by a max pooling layer, using leaky ReLU as the activation function of the CNN, which has a non-zero gradient across the entire domain, thus avoiding the neuron death problem. A dropout operation is performed before the fully connected layers of the CNN to avoid overfitting. Path 2 of CNN-LightGBM is a LightGBM with a decision tree depth of 9, 192 leaves, a learning rate of 0.04, and a regularization parameter of 0.2. The multi-scale features of the ground transient electromagnetic response extracted by CNN are flattened into a one-dimensional vector, and then input into LightGBM to establish a mapping relationship between the transient electromagnetic response data and the resistivity of each subsurface model.
[0039] Step 3: Train CNN-LightGBM using optimal parameters;
[0040] The training set and validation set generated in step 1 are used to train CNN-LightGBM. During training, optimal parameters are used to obtain the best inversion performance. An early stopping strategy and learning rate decay are employed to prevent overfitting and reduce training time. The loss function of CNN-LightGBM is:
[0041]
[0042] Where N represents the batch size. y and y represent the resistivity of the label and the inverted subsurface model, respectively, M is the number of network output parameters, and a regularization term with weight λ is introduced to improve the model's generalization ability. This represents the sum of squares of the trainable weights W of CNN-LightGBM. Finally, save the trained CNN-LightGBM.
[0043] Step 4: Validate the inversion effect of CNN-LightGBM using the test set;
[0044] The test set consists of data not used during the training of CNN-LightGBM and can be used to evaluate its generalization ability. Noise is added to the test set data to further test the robustness of CNN-LightGBM. Figure 3 This figure compares the inversion results of the CNN-LightGBM inversion method of this invention with the CNN inversion results. In the figure, the black line represents the true resistivity of the underground model, the gray dashed line represents the CNN inversion result, and the black dashed line represents the CNN-LightGBM inversion result. The root mean square error (RMSE) is used to verify the inversion effect of CNN-LightGBM, defined as: A smaller root mean square error (RMSE) indicates a better inversion result. It can be seen that the RMSE of the CNN-LightGBM inversion result is always smaller than that of the CNN inversion result. This demonstrates the superiority of the CNN-LightGBM inversion method of this invention.
[0045] Step 5: Use the trained CNN-LightGBM to invert the measured ground transient electromagnetic data;
[0046] After preprocessing the measured transient electromagnetic data of the ground, the resistivity of the underground structure in the measured area can be obtained by using the trained CNN-LightGBM inversion.
[0047] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A ground transient electromagnetic inversion method based on CNN-LightGBM, characterized in that, The method specifically includes the following steps: Step 1: Generate a sample dataset using a ground-based transient electromagnetic forward modeling program, including: The parameters of the forward modeling procedure are determined based on the acquisition parameters obtained from ground transient electromagnetic method detection. Randomly select the resistivity of each underground model layer within the set resistivity range; Based on the determined parameters and the resistivity of each subsurface model, forward modeling was used to generate transient electromagnetic response data of the ground. Preprocessing of ground transient electromagnetic response data and resistivity of each subsurface model; The preprocessed ground transient electromagnetic response data is used as the input to CNN-LightGBM, and the resistivity of each subsurface model after preprocessing is used as the label of the input data of CNN-LightGBM. The sample dataset is divided into a training set, a validation set, and a test set; Step 2: Construct CNN-LightGBM and determine its optimal parameters; The multi-scale features of the transient electromagnetic response of the ground extracted by CNN are flattened into a one-dimensional vector, and then input into LightGBM to establish the mapping relationship between the transient electromagnetic response data of the ground and the resistivity of each subsurface model. The parameters to be optimized in CNN include the number of convolutional and fully connected layers, the size of the convolutional kernel, and the learning rate. The parameters to be optimized in LightGBM include the depth of the decision tree, the number of leaves, the learning rate, and the regularization parameter. Step 3: Train CNN-LightGBM using optimal parameters; CNN-LightGBM is trained using the training and validation sets from step 1. During training, the optimal parameters are used to obtain the best inversion performance. Finally, the trained CNN-LightGBM is saved. Step 4: Validate the inversion effect of CNN-LightGBM using the test set; The test set is data that CNN-LightGBM did not use during training. It is used to evaluate the generalization ability of CNN-LightGBM. The inversion result is evaluated based on the error between the inversion result of CNN-LightGBM on the test set and the label. Step 5: Use the trained CNN-LightGBM to invert the measured ground transient electromagnetic data; After preprocessing the measured transient electromagnetic data of the ground, the data is input into the trained CNN-LightGBM to obtain the resistivity of each subsurface model.
2. The method according to claim 1, characterized in that, The resistivity range set in step 1 is 0.1-15000 Ω·m. The preprocessing method for the ground transient electromagnetic response data and the resistivity of each subsurface model is as follows: Subtract the mean μ from the transient electromagnetic response data x and divide by the standard deviation σ. The label represents the logarithmic transformation value of the resistivity ρ for each subsurface model layer. The training set, validation set, and test set comprise 90%, 5%, and 5% of the sample dataset, respectively.
3. The method according to claim 1, characterized in that, In step 2, the optimal parameters of CNN-LightGBM are determined using random search or Bayesian optimization; leakyReLU is used as the activation function of the CNN; and a dropout operation is performed before the fully connected layers of the CNN to randomly disable a portion of neurons to prevent overfitting and improve robustness and generalization ability.
4. The method according to claim 1, characterized in that, In step 3, an early stopping strategy is employed, and learning rate decay is used to prevent CNN-LightGBM from overfitting and to reduce its training time. The loss function of CNN-LightGBM in step 3 is: Where N represents the batch size. and Let represent the resistivity of the subsurface model obtained from the label and inversion, respectively. M is the number of network output parameters. A regularization term with weight λ is introduced to improve the generalization ability of the model. This represents the sum of squares of the trainable weights W in CNN-LightGBM.
5. The method according to claim 1, characterized in that, The step 4 uses the root mean square error RMSE to verify the inversion effect of the CNN-LightGBM, which is defined as: .