A method for separating magnetic anomaly regional field and local field based on deep learning network

Abstract

This invention discloses a method for separating regional and local magnetic anomaly fields based on deep learning networks. The method acquires gridded total magnetic anomaly data, inputs it into a convolutional neural network model based on an improved UNet++, and outputs separated regional and local fields. Model improvements include: adding a Convolutional Block Attention (CBAM) module at the end of nested convolutional blocks to introduce an attention mechanism for adaptive recalibration of feature importance; using a loss function with physical constraints for training to ensure the output conforms to the theoretical laws of magnetic potential fields; and employing a hybrid fusion strategy of shallow summation and deep concatenation in the decoding process. This invention improves the accuracy and physical plausibility of magnetic anomaly separation, and solves the problems of weak anomalies being easily obscured and traditional methods relying on prior assumptions.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of geophysical exploration and deep learning, specifically to a method for separating the magnetic anomaly region field from the local field based on an improved UNet++ convolutional neural network. Background Technology

[0002] In gravity and magnetic exploration, the measured gravity and magnetic fields or anomalies are superimposed responses caused by all underground geological bodies with differences in density or magnetic properties, exhibiting a complex mixture of anomalies of different scales and amplitudes. This superposition effect of multiple anomalies poses significant challenges to the inversion and interpretation of the target geological body. Therefore, how to effectively separate the anomalies caused by a specific target geological body from the superimposed gravity and magnetic fields—a process known as "anomaly decomposition"—has always been a core research direction and challenge in gravity and magnetic data processing and interpretation.

[0003] Existing potential field separation methods are numerous and can be traditionally categorized into spatial domain methods (such as moving average, polynomial fitting, trend surface analysis, and interpolation cutting) and frequency domain methods (such as matched filtering, Wiener filtering, and wavelet transform). However, most of these methods rely on manually designed filters or strong prior geological assumptions, resulting in significant limitations when facing complex geological scenarios: First, their adaptability is poor. For example, the matched filtering method requires model simplification and neglect of scale effects during derivation, leading to poor separation results and blurred boundaries under complex geological conditions; second, their multi-scale signal extraction capabilities are insufficient, making it difficult to accurately separate superimposed anomalies at different depths.

[0004] Despite continuous efforts to improve existing technologies, such as developing regularized filtering methods to enhance stability, utilizing wavelet multi-resolution analysis and attempting to select the optimal mother wavelet, and exploring new approaches based on blind source separation or low-rank decomposition, the problem of separating gravity and magnetic anomalies has not yet been fully solved. In particular, the high-precision, high-efficiency, and high-fidelity separation of weak anomalous signals in complex magnetic environments remains a significant challenge. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to overcome the limitations of existing potential field separation methods, such as reliance on manually designed filters, strong prior geological assumptions, and the existence of blurred separation boundaries, insufficient weak signal extraction capabilities, contradictions between noise processing and signal preservation, and weak generalization, so as to achieve high-precision and high-fidelity separation of magnetic anomaly region fields and local fields.

[0006] To address the aforementioned technical problems, this invention provides a method for separating the magnetic anomaly region field from the local field based on a deep learning network, comprising the following steps:

[0007] S1. Obtain the gridded total magnetic anomaly data of the area to be processed;

[0008] S2. Input the total magnetic anomaly data into the pre-trained magnetic anomaly separation model, and output the separated regional field data and local field data;

[0009] The magnetic anomaly separation model is a convolutional neural network model based on the UNet++ architecture. A CBAM (Convolutional Block Attention Module) is added to the end of each nested convolutional block in the UNet++ architecture. The CBAM includes cascaded channel attention sub-modules and spatial attention sub-modules, enabling the network to adaptively recalibrate the channel and spatial importance of feature maps, thus increasing attention to key regions. The magnetic anomaly separation model is trained using a loss function that includes physical constraints, which ensure that the model's output conforms to the fundamental physical laws of magnetic potential field theory.

[0010] Preferably, the channel attention submodule is used to weight each channel of the input feature map, which is achieved through global pooling and shared MLP (multilayer perceptron); the spatial attention submodule is used to make the network focus on the spatially richest position, which is achieved through 7×7 convolution.

[0011] Preferably, the physical constraint is an additive consistency constraint. The formula is expressed as follows:

[0012] (1)

[0013] in, It is the total field of the input model. This is the part predicted as regional anomaly. It is predicted to be a local anomaly. It is the batch size; the additive consistency constraint term is used to constrain the sum of the predicted regional anomalies and local anomalies to equal the total observed field.

[0014] Preferably, the magnetic anomaly separation model decoding process is based on a UNet++ architecture convolutional neural network model, starting from the deepest decoding node X of the encoder. i,0 The feature map obtained from upsampling is then compared with the previous decoding node X. i-1,0 The decoded node X is obtained by performing skip connections and convolutional fusion. i-1,1 Feature map; then decoded by the encoder's second-deepest decoding node X i-1,0 The feature map obtained after upsampling is compared with the decoding node X of the layer above the encoder. i-2,0 The decoding node X is obtained by fusing the fusion feature with skip connections and convolutional fusion. i-2,1 Feature map, decoding node X i-2,1Feature map, decoding node X i-2,0 The feature map is fused with convolutional decoding nodes X via skip connections. i-1,1 After upsampling again, the feature maps are fused to obtain the decoding node X. i-2,2 The feature map is then used; this process is repeated with progressive upsampling, skip connections, and convolutional fusion, finally passing through the last decoding node X of layer 0. 0,i The output feature map is then mapped to two channels through a 1×1 convolution, corresponding to region anomalies and local anomalies respectively, thus completing the feature reconstruction from the highest semantic layer to the original resolution.

[0015] Preferably, a model training step is included before step S2, the model training step including:

[0016] S01. Construct a training sample set, which includes multiple sets of sample data. Each set of sample data includes input samples and label data. The input samples are simulated or pre-processed actual total magnetic anomaly data. The label data corresponds to the regional anomaly field and local anomaly field data in the input samples.

[0017] S02. Construct the magnetic anomaly separation model described above;

[0018] S03. Train the model using the training sample set, wherein the loss function used is a weighted sum of the mean squared error loss function and the physical constraint term, expressed by the formula:

[0019] (2)

[0020] (3)

[0021] in, It is the mean squared error loss function. It is a hyperparameter that balances the two terms. The part that is actually labeled as an anomaly in the region. The part that is actually labeled as a local anomaly.

[0022] The method for simulating and generating actual total magnetic anomaly data is as follows: First, a 32×32×16 voxel grid is assigned to the forward model, and the forward propagation (forward()) method is called to simulate the theoretical magnetic anomaly (in nT) on the observation plane according to the given magnetic susceptibility spatial distribution, and finally a usable magnetic anomaly dataset is obtained.

[0023] Preferably, in step S01, the training sample set is further expanded by data augmentation techniques, which include one or more of rotating, flipping, and adding random noise to the original data.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] (1) This invention has a strong ability to enhance and extract weak nT-level magnetic anomaly signals with low signal-to-noise ratio through lightweight network design, CBAM and shallow summation strategy, and solves the problem that weak anomalies are easily submerged.

[0026] (2) An additive consistency constraint is introduced and the deep learning loss function is embedded in the form of a regularization term. This forces the sum of the regional anomalies and local anomalies predicted by the model to equal the total observed field, ensuring that the separation results conform to the basic laws of magnetism. This significantly improves the reliability of the geological interpretation of the output results, which is an advantage that pure data-driven methods do not have.

[0027] (3) By introducing a constraint term based on the Laplace operator, the regional field satisfies the Laplace equation in the source-free region, which further enhances the physical rationality of the separation results;

[0028] (4) The end-to-end training method reduces the reliance on prior assumptions. The model trained through data augmentation and deep supervision has a better generalization ability for unseen regional data.

[0029] (5) The hybrid feature fusion strategy takes into account both shallow detail preservation and deep semantic extraction, and improves the fusion efficiency of multi-scale abnormal features. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the improved UNet++ convolutional neural network structure provided in an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the CBAM structure provided in an embodiment of the present invention;

[0032] Figure 3 This is a flowchart of the method for separating the magnetic anomaly region field from the local field provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0034] Example 1: Network Construction and Inference

[0035] like Figure 1 and Figure 3 As shown, the magnetic anomaly separation method based on the improved UNet++ convolutional neural network of this invention has the following network construction steps:

[0036] A symmetrical framework based on a U-shaped encoder-decoder is adopted as the basic architecture. Small U-shaped sub-networks are embedded between each layer of the encoder and decoder to form a nested structure. This design draws on the core idea of ​​UNet++, replacing the traditional direct concatenation operation with dense skip connections. The features of each layer of the encoder are refined by convolutional layers and then densely transmitted to the corresponding nodes and deeper nodes of the decoder, realizing deep fusion and full utilization of multi-scale features from both shallow and deep layers.

[0037] like Figure 1 As shown, all convolutional blocks are... In this representation, the superscript i represents the downsampling level (from 0 to 4, with higher levels having lower feature resolution and stronger semantic information), and the subscript j represents the convolutional block number within that level; nodes with j=0 (such as...) , … ( ) is the raw downsampled feature node of the encoder (Backbone), responsible for extracting multi-scale features step by step. It is also the starting point of the network's input features; nodes where j>0 (such as...) ~ , ~ , ~ , () is a feature fusion node on a nested jump path, which simultaneously receives features from preceding nodes at the same level and upsampled features from nodes at deeper levels, achieving multi-scale feature fusion through convolution and propagating it upwards; where It is the final main output feature node of the network, and , , It will also serve as an auxiliary output node for deep supervision, participating in loss calculation during training.

[0038] Each convolutional block ends with an integrated CBAM (such as...). Figure 2 As shown, the network first uses a channel attention submodule (global pooling + shared MLP) to weight each channel, and then uses a spatial attention submodule (7×7 convolution) to focus the network on the spatially richest location, thereby recalibrating the features to enhance the discriminative ability.

[0039] A hybrid feature fusion strategy is adopted during the decoding process: at shallow decoding nodes, the feature maps from the encoder are fused by summation, which helps to retain shallow details and reduce the number of parameters; at deep decoding nodes, the feature maps are fused by splicing, which helps to integrate deep semantic information.

[0040] To address the characteristics of weak magnetic signals and low signal-to-noise ratio in magnetic exploration data, the network adopts a 2D lightweight architecture, including 3 to 4 layers of downsampling and upsampling operations, which reduces the computational burden while ensuring performance.

[0041] During inference, the gridded total magnetic anomaly data to be processed is input into the trained model, which then outputs the separated regional and local field results end-to-end. The final separation results need to be compared and analyzed with actual geological data and drilling data of the work area to verify their geological rationality.

[0042] Example 2: Data Preparation and Model Training

[0043] Model training requires constructing a "sample-label" dataset. First, a geological model is built and magnetic anomaly forward modeling is performed, saving the data as magnetic anomaly data. The method is as follows: First, a 32×32×16 voxel grid is assigned to the forward model. The `forward()` method is called to simulate the theoretical magnetic anomaly (in nT) on the observation plane based on the given spatial distribution of magnetic susceptibility, ultimately obtaining a usable magnetic anomaly dataset.

[0044] For each data point, the following steps are performed sequentially: add channel dimensions, crop to a uniform size, transpose to PyTorch format, and convert to two-dimensional gridded total magnetic anomaly data of type float32 as model input samples.

[0045] Label data is generated through high-precision forward modeling, and the anomalies it causes are calculated using a constructed geological model to ensure accuracy. Alternatively, traditional filtering separation results verified by those skilled in the art can be used as pseudo-labels, but simulated labels are generally more accurate than pseudo-labels. The structure of the network input layer is modified to receive single-channel data, and geometric transformations (rotation, flipping, adding random noise, etc.) are used to expand the training sample set to improve the model's generalization ability.

[0046] During training, mean squared error (MSE) is used as the main loss function to ensure that the model's predicted regional anomalies and local anomalies closely approximate the true labels. An additive consistency constraint is added as a physical regularization term to ensure that the sum of the two predicted components conforms to the physical superposition principle, i.e., the sum of the predicted regional anomalies and local anomalies equals the total field observations in the input. A constraint term based on the Laplace operator is also included to ensure that the regional field satisfies the Laplace equation in the source-free region. The total loss function is a weighted sum of the MSE loss and the physical constraint term, where the weights are adjusted through hyperparameters. The Adam optimizer is used for training, and the losses on the training and validation sets are monitored. The model parameters with optimal validation performance are saved.

[0047] Compared with traditional methods, the method based on the improved UNet++ in this invention automatically captures field source features through dense nesting and skip connections, eliminating the need for manual filter design and improving the separation accuracy in complex geological scenarios. The introduction of physical constraint terms makes the separation results conform to the basic laws of magnetism, significantly enhancing the reliability of geological interpretation. The end-to-end framework has no prior assumptions, resulting in stronger generalization and robustness in separating complex anomalies.

[0048] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for separating the magnetic anomaly region field from the local field based on a deep learning network, characterized in that, The method includes the following steps: S1. Obtain the gridded total magnetic anomaly data of the area to be processed; S2. Input the total magnetic anomaly data into the pre-trained magnetic anomaly separation model, and output the separated regional field data and local field data; The magnetic anomaly separation model is a convolutional neural network model based on the UNet++ architecture. A convolutional block attention module (CBAM) is added to the end of each nested convolutional block in the UNet++ architecture. The CBAM includes a cascaded channel attention submodule and a spatial attention submodule, which enables the network to adaptively recalibrate the channel and spatial importance of the feature map and improve the attention to key regions. The magnetic anomaly separation model is trained by a loss function that includes physical constraints. The physical constraints are used to make the output of the model conform to the basic physical laws of magnetic potential field theory.

2. The method for separating the magnetic anomaly region field and local field based on deep learning networks according to claim 1, characterized in that, The channel attention submodule in the Convolutional Block Attention Module (CBAM) is used to weight each channel of the input feature map, which is achieved through global pooling and a shared multilayer perceptron (MLP); the spatial attention submodule is used to focus the network on the spatially richest location, which is achieved through 7×7 convolution.

3. The method for separating the magnetic anomaly region field and local field based on deep learning networks according to claim 1, characterized in that, The physical constraint is an additive consistency constraint. The formula is expressed as follows: （1） in, It is the total field of the input model. This is the part predicted as regional anomaly. It is predicted to be a local anomaly. It is the batch size; the additive consistency constraint term is used to constrain the sum of the predicted regional anomalies and local anomalies to equal the total observed field.

4. The method for separating the magnetic anomaly region field and local field based on deep learning networks according to claim 1, characterized in that, The magnetic anomaly separation model decoding process is based on a UNet++ architecture convolutional neural network model, starting from the deepest decoding node X of the encoder. i,0 The feature map obtained from upsampling is then compared with the previous decoding node X. i-1,0 The decoded node X is obtained by performing skip connections and convolutional fusion. i-1,1 Feature map; then decoded by the encoder's second-deepest decoding node X i-1,0 The feature map obtained after upsampling is compared with the decoding node X of the layer above the encoder. i-2,0 The decoding node X is obtained by fusing the fusion feature with skip connections and convolutional fusion. i-2,1 Feature map, decoding node X i-2,1 Feature map, decoding node X i-2,0 The feature map is fused with convolutional decoding nodes X via skip connections. i-1,1 After upsampling again, the feature maps are fused to obtain the decoding node X. i-2,2 The feature map is then used; this process is repeated with progressive upsampling, skip connections, and convolutional fusion, finally passing through the last decoding node X of layer 0. 0,i The output feature map is then mapped to two channels through a 1×1 convolution, corresponding to region anomalies and local anomalies respectively, thus completing the feature reconstruction from the highest semantic layer to the original resolution.

5. The method for separating the magnetic anomaly region field and local field based on deep learning networks according to claim 1, characterized in that, Before step S2, a model training step is included, which includes: S01. Construct a training sample set, which includes multiple sets of sample data. Each set of sample data includes input samples and label data. The input samples are simulated or pre-processed actual total magnetic anomaly data. The label data corresponds to the regional anomaly field and local anomaly field data in the input samples. S02. Construct the magnetic anomaly separation model described above; S03. Train the model using the training sample set, wherein the loss function used is a weighted sum of the mean squared error loss function and the physical constraint term, expressed by the formula: （2） （3） in, It is the mean squared error loss function. It is a hyperparameter that balances the two terms. The part that is actually labeled as an anomaly in the region. The part that is actually labeled as a local anomaly.

6. The method for separating the magnetic anomaly region field and local field based on a deep learning network according to claim 5, characterized in that, The method for simulating and generating actual total magnetic anomaly data in step S01 is as follows: First, a 32×32×16 voxel grid is assigned to the forward model, and the forward propagation method is called to simulate the theoretical magnetic anomaly on the observation plane according to the given magnetic susceptibility spatial distribution, and finally a usable magnetic anomaly dataset is obtained.

7. The method for separating the magnetic anomaly region field and local field based on a deep learning network according to claim 5, characterized in that, In step S01, the training sample set is further expanded by data augmentation techniques, which include one or more of rotating, flipping, and adding random noise to the original data.

Images