A deep learning-based aeromagnetic anomaly detection and parameter intelligent estimation method

Abstract

The application discloses a kind of based on deep learning's aeromagnetic anomaly detection and parameter intelligent estimation method, it is related to geophysical exploration technical field. Including: through decoupling detection-inversion architecture, first, all abnormal regions are automatically located from aeromagnetic ΔT grid using target detection network, then for each abnormal region, a plurality of independent, pre-trained shape-specific inversion network is used to estimate its depth and magnetization direction parameters under different shape assumptions, finally, based on the forward residual and morphological prior knowledge fusion decision, output the best parameter combination and recommended shape of each anomaly.The present application realizes full-process automation, significantly improves the completeness and efficiency of information extraction.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, and more specifically to a deep learning-based method for detecting aeromagnetic anomalies and intelligently estimating their parameters. Background Technology

[0002] In the interpretation of aeromagnetic anomalies, accurately obtaining the spatial location of underground magnetic bodies, especially their burial depth and magnetization direction (dip angle, deflection angle), is a crucial foundation for assessing the properties of magnetic bodies and making geological inferences. Traditionally, field technicians rely primarily on manual or semi-manual graphical methods, such as the tangent method, to estimate the depth of isolated anomalies. This method requires manually selecting characteristic points such as maxima, semi-maxima, and inflection points on the magnetic anomaly profile before substituting them into empirical formulas to calculate the depth. Its limitations include: low efficiency of manual operation; results highly dependent on the operator's personal experience, resulting in poor repeatability; insufficient processing capacity for complex anomalies such as finite-length slabs and cylinders, as well as scenarios with multiple anomalies superimposed; and the inability to obtain only approximate depth values, failing to simultaneously acquire magnetization direction or shape information, making the accuracy insufficient for quantitative interpretation.

[0003] With the development of computing technology, automatic inversion methods based on optimization theory have emerged, such as the Maguire method, genetic algorithms, and simulated annealing. These methods iteratively adjust model parameters to achieve an optimal fit between the forward-modeled field and the observed field, and can simultaneously output depth and magnetization direction. However, their application is limited by problems such as strong ambiguity, heavy reliance on the initial model, susceptibility to local extrema, and high computational cost. Especially when the shape of the anomalous body is unknown, it is necessary to optimize both geometric dimensions and physical property parameters simultaneously, resulting in excessively high dimensionality of the variable space and difficulty in convergence. For complex field sources with multiple superimposed anomalous bodies, subjective anomaly separation processing is required before inversion, which is difficult to meet the needs of rapid on-site processing.

[0004] In recent years, deep learning technology has made breakthroughs in the field of image recognition, and some studies have begun to explore its application in aeromagnetic anomaly identification. For example, classification networks are used to determine the shape of anomalies, simplifying them into spheres, cylinders, and plates, or object detection networks are used to locate the center of anomalies. However, existing technical solutions are limited in function and have not yet formed a complete technical solution that integrates anomaly detection with multi-parameter (depth, tilt angle, deflection angle, shape) output.

[0005] Its main shortcomings are reflected in the following aspects: (1) Feature extraction is not fully coupled with physical mechanisms. Most networks directly use the original magnetic anomaly image as input and fail to effectively utilize multi-channel physical features such as analytical signal amplitude, gradient field, and frequency domain information. This results in limited learning ability of the model for key parameters such as depth and magnetization direction, and poor prediction reliability in scenarios with large amplitude dynamic range.

[0006] (2) Lack of prior knowledge of shape in shape discrimination. Purely data-driven classification networks are prone to confusion when faced with anomalies such as finite-length cylinders and plate-like objects with similar shapes, as their classification confidence levels are close. Existing technologies lack mechanisms to use physical shape features, such as gradient direction entropy, as auxiliary decision-making, resulting in low accuracy in recommending shapes.

[0007] (3) Adaptive processing of multiple anomalies is not achieved. Existing methods are mostly designed for single anomaly scenarios. When there are multiple interfering magnetic objects in the test area, they cannot be automatically segmented and inverted one by one, and manual intervention is still required, thus losing the advantage of automated processing.

[0008] (4) Inconsistency between inference process and training conditions. Some methods use different grid resolutions or feature normalization methods during inference than during training, which introduces resampling errors or distribution shifts, resulting in severe distortion of prediction depth and poor system practicality and robustness.

[0009] Therefore, proposing a deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0010] In view of this, the present invention provides a deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation. This method addresses the shortcomings of existing technologies, such as the low efficiency of relying on manual tangent methods for interpreting aeromagnetic anomalies, the inability to obtain magnetization direction, the limited functionality of existing deep learning methods, the lack of full utilization of physical features, the lack of prior knowledge of shape recognition, the inability to handle multiple anomalies simultaneously, and the inconsistency between inference and training conditions, which leads to distortion in depth prediction.

[0011] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation includes the following steps: S1. For three typical magnetic body models, namely sphere, finite-length cylinder and finite-length plate, construct and train independent inversion network models respectively. During training, extract multi-channel physical features of the magnetic anomaly field ΔT of each sample as network input. Each shape-specific inversion network is used to invert the depth and magnetization direction parameters under the assumption of the corresponding single shape. S2. A target detection network trained based on multiple anomaly superimposed samples is used to automatically locate the bounding boxes of all anomaly regions from the aeromagnetic ΔT grid. S3. For the input measured aeromagnetic ΔT grid data, execute the following inference process: S31. Data preprocessing: Calculate the grid spacing of the measured data and construct a multi-channel feature map; S32. Anomaly Detection and Localization: Input the multi-channel feature map into the trained target detection network to detect all abnormal bounding boxes; S33. Window cropping and feature extraction: Using each anomaly center point as the center, crop a window of fixed physical size from the original ΔT grid, and extract multi-channel physical features from the current window; S34. Parallel Inversion of Multiple Shapes: Input the processed window features into the inversion network model for each shape trained in step S1. Each network independently outputs the depth, magnetization tilt angle and magnetization deflection angle under the corresponding shape assumption. S35, Optimal Shape Fusion Decision: For the multiple sets of parameters output by each shape network, a weighted score is calculated by combining the forward fitting residual, morphological prior knowledge, and inversion confidence. The most likely shape is recommended for each anomaly, and the depth, magnetization tilt angle, and magnetization deflection angle corresponding to the shape are output as the final estimation result.

[0012] Optionally, the multi-channel physical features in S1 include: analytic signal amplitude, horizontal gradient, vertical derivative, and frequency domain amplitude spectrum; The amplitude of the analytical signal is calculated from the horizontal and vertical derivatives of the magnetic anomaly ΔT. The vertical derivative is calculated in the frequency domain. The frequency domain amplitude spectrum is obtained by taking the amplitude spectrum after performing a two-dimensional fast Fourier transform on ΔT and then performing logarithmic compression.

[0013] Optionally, each shape-specific inversion network in S1 is a three-branch convolutional neural network structure, including a depth branch, an inclination branch, and an eccentricity branch; the depth branch outputs normalized depth values ​​using a regression method, while the inclination branch and the eccentricity branch output the probability distributions of magnetization inclination and magnetization eccentricity respectively using a classification method. The input to the depth branch is a two-channel feature map composed of the analytical signal amplitude and ΔT; the input to the tilt branch is a three-channel feature map composed of ΔT, the horizontal derivative, and the vertical derivative; and the input to the skew branch is a four-channel feature map composed of ΔT, the horizontal derivative, the vertical derivative, and the frequency domain amplitude spectrum.

[0014] Optionally, the method for generating training samples in S1 is as follows: set a fixed spatial grid range, randomly generate a large number of samples within the spatial grid range, and the parameters of each sample include the center plane coordinates, burial depth, magnetization tilt angle, magnetization deflection angle and the unique geometric parameters of the corresponding shape of the magnetic body. Calculate the theoretical magnetic anomaly ΔT using the finite length forward modeling formula, and add Gaussian noise to simulate real observation.

[0015] Optionally, the target detection network in S2 is a YOLO network, and the input is a three-channel pseudo-color image. The three channels are: the robustly normalized enhancement ΔT, the robustly normalized horizontal gradient magnitude, and the robustly normalized absolute value of the vertical derivative. The multiple anomaly superposition sample is obtained by randomly generating 1 to 5 anomalies within a fixed spatial grid range and linearly superimposing them to obtain the total field. The bounding rectangle of each significant anomaly is automatically extracted as a label based on adaptive threshold segmentation and connected component analysis.

[0016] Optionally, in S32, the center of each detection box output by the target detection network is peak-corrected using the amplitude of the analyte signal within the local area to obtain the location of the abnormal center.

[0017] Optionally, the optimal shape fusion decision in S35 specifically includes: performing forward modeling calculations on the parameters output by each shape-specific inversion network to obtain the theoretical magnetic anomaly and calculating the root mean square error with the observed field; simultaneously calculating the information entropy of the magnetic anomaly gradient direction within the window as prior knowledge of the shape; combining the softmax maximum probability of the directional branches of each shape-specific inversion network as the inversion confidence, and determining the optimal shape through weighted scoring.

[0018] Optionally, S4 also includes calculating the difference between induced magnetization and observed magnetization based on the recommended shape and the inverted magnetization direction, combined with the geomagnetic field direction at the measuring point, estimating the equivalent remanence index Q value, and providing a reliability evaluation.

[0019] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation, the beneficial effects of which are: (1) Integration of anomaly detection and parameter estimation: This invention cascades the target detection network (YOLO) with multiple shape-specific parameter inversion networks to form an end-to-end intelligent interpretation process. This method can automatically identify the spatial range of multiple magnetic anomalies from the original aeromagnetic ΔT grid data and delineate the bounding box without manually presetting the number of anomalies or manually selecting the profile. At the same time, for each detected anomaly, it outputs its center burial depth, magnetization inclination angle, magnetization deflection angle and recommended shape category at once, namely sphere, finite-length cylinder or finite-length plate. Compared with the traditional method that requires separate anomaly separation, depth estimation and shape discrimination, this invention realizes full-process automation and significantly improves the completeness and efficiency of information extraction. (2) High depth estimation accuracy: Quantitative validation on synthetic datasets: For magnetic moments of 5×10¹ 4 For single anomalies with an area of ​​A·m² and a burial depth ranging from 30 to 200 m, the mean absolute error (MAE) of the depth regression branch trained in this invention on the validation set is less than 2.1 m (cylinder) to 3.5 m (sphere / plate). Under the premise that the grid resolution and training conditions are strictly consistent, the depth prediction error of typical validation samples can be controlled within 5 m. For example: The spherical anomaly with a true depth of 87.4m was predicted to be 88.1m, with an error of only +0.7m; the plate-like anomaly with a true depth of 46.9m was predicted to be 42.6m, with an error of -4.3m.

[0020] In contrast, the traditional tangent method typically has a depth estimation error of 10–30 m under ideal conditions (isolated sphere, good signal-to-noise ratio). Model approximation and human operation are the key bottlenecks restricting its accuracy, and it cannot obtain magnetization direction parameters at the same time. This invention has significant advantages in both accuracy and information content. (3) Accurate direction estimation: This invention uses a 36-category classification method for magnetization tilt angle (I) and deflection angle (D), where the tilt angle covers -90° to 90° in 5° intervals, and the deflection angle covers 0° to 360° in 10° intervals. In the synthesized verification samples, the predicted direction values ​​are highly consistent with the true values. Typical examples are as follows: Plate-shaped sample: True I = -73.1°, D = 77.1°; Predicted I = -73°, D = 77° (tilt deviation 0.1°, deflection deviation 0.1°); Sphere-shaped sample: True I = 67.1°, D = 281.1°; Predicted I = 67°, D = 281° (tilt deviation 0.1°, deflection deviation 0.1°); The above deviations are all within one classification interval (5° or 10°), indicating that the classification strategy can reliably recover the magnetization direction of the magnetic body, providing a valuable reference for remanence assessment and geological interpretation; (4) Shape discrimination introduces morphological priors, resulting in strong robustness: To address the problem of similar magnetic anomaly morphology between finite-length cylinders and plate-like bodies, and the ease with which pure data-driven classification networks can confuse them, this invention introduces physical morphological prior knowledge based on gradient direction entropy in the shape discrimination stage. Specifically: When the classification confidence scores output by the three shape-specific networks are close to each other (confidence difference less than a preset threshold, such as 0.05), the information entropy of the magnetic anomaly gradient direction within the anomaly window is automatically calculated: a high entropy value indicates that the gradient direction is chaotic (approaching spherical or point source features), while a low entropy value indicates that the gradient direction is concentrated, approaching the linear features of a plate-like body. Based on the entropy distribution, the scores of the three shapes are weighted and corrected. Experimental verification shows that this strategy reduces the probability of spherical anomalies being misclassified as plate-like bodies by about 60%, improves the overall accuracy of recommended shapes by more than 20%, and significantly enhances the decision reliability of the system in fuzzy scenarios; (5) Automatic segmentation and anti-interference capability of multiple anomalies: This invention adopts an improved YOLOv8 target detection network. The training data comes from samples generated by random combination of multiple anomalies (1 to 5) in the superimposed field. During detection, the network outputs the bounding box of each anomaly, and then uses the peak position of the analytical signal amplitude (ASA) in the local area to perform secondary fine correction on the center of the box. This strategy can reliably segment multiple magnetic objects that are adjacent to each other and provide a clipping window for each independent anomaly for subsequent object-specific network to perform parameter inversion. At the same time, the system is designed with a fallback mechanism: when the YOLO model does not detect any anomalies or the model is unavailable, it automatically switches to the full-field single-object inversion mode to perform single anomaly object parameter estimation on the whole image, thereby ensuring the robustness and availability of the system in single anomaly or edge scenes; (6) Strict consistency between inference and training conditions eliminates resampling errors: This invention strictly follows the data processing specifications of the training phase during the inference phase: ① Unified grid resolution (the grid is 128×128 during training, and the effective area after cropping is 104×104, corresponding to an actual physical width of 1000 m); ② Unified feature calculation process (the calculation parameters of ASA, gradient, and FFT amplitude spectrum are completely consistent with the training script); ③ Unified use of Z-score normalization parameters (mean and standard deviation) saved during the training phase to normalize the input features. The above consistency design avoids the distribution shift introduced by resolution mismatch, feature calculation differences, or different normalization parameters, fundamentally eliminating resampling errors. Experiments show that there is no significant difference in the depth prediction deviation of the model between synthetic data and measured data, ensuring the credibility and stability of the system in actual aeromagnetic data processing; (7) Significantly improves work efficiency: The traditional tangent method requires interpreters to manually select feature points such as maximum values, semi-maximum values, and inflection points on each survey line profile. For a medium-sized survey area containing thousands of survey points, a complete interpretation may take several days or even months. The method of this invention can complete the entire process of full-field anomaly detection and multi-parameter inversion on an ordinary computer (CPU or entry-level GPU) in only a few seconds to tens of seconds, improving the processing speed by more than two orders of magnitude, which can meet the needs of rapid on-site judgment and real-time processing; (8) Reduced reliance on expert experience: This invention transforms complex physical inversion problems into automatic mapping using deep learning models. Non-expert technicians only need to load aeromagnetic grid data and click run to obtain depth, magnetization direction, and shape suggestions for each anomaly. This significantly reduces the reliance on the operator's geophysical background knowledge and manual diagramming experience, which is conducive to promoting advanced quantitative interpretation technology to grassroots geological teams, mineral exploration units, and teaching and training scenarios, and promoting the widespread application of aeromagnetic interpretation technology. (9) Provides rich information for quantitative interpretation: In addition to depth parameters, the magnetization direction (dip angle, deflection angle) and recommended shape output by this invention can be directly used to infer the type of magnetic body (such as spheroidal ore bodies associated with intrusive rocks and vein-like ore bodies associated with faults). Combined with the estimation of the equivalent remanent magnetization index Q value, it can also provide auxiliary basis for judging the degree of remanent magnetization influence of magnetite. The above multidimensional information provides more comprehensive quantitative constraints for subsequent three-dimensional geological modeling, resource estimation and borehole deployment, significantly improving the application value of aeromagnetic data in mineral exploration and geological research. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation, provided in an embodiment of the present invention. Figure 2 This is an example image of YOLO detection results provided in an embodiment of the present invention; Figure 3 This is a diagram illustrating the inversion results of a single anomaly provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This method employs a decoupled detection-inversion architecture. First, it uses a target detection network to automatically locate all anomalous regions from the aeromagnetic ΔT grid. Then, for each anomalous region, multiple independent, pre-trained shape-specific inversion networks are used to estimate its depth and magnetization direction parameters under different shape assumptions. Finally, based on the fusion decision of forward modeling residuals and morphological prior knowledge, the optimal parameter combination and recommended shape for each anomalous region are output.

[0025] This invention discloses a deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation, comprising the following steps: S1. For three typical magnetic body models, namely sphere, finite-length cylinder and finite-length plate, construct and train independent inversion network models respectively. During training, extract multi-channel physical features of the magnetic anomaly field ΔT of each sample as network input. Each shape-specific inversion network is used to invert the depth and magnetization direction parameters under the assumption of the corresponding single shape. S2. A target detection network trained based on multiple anomaly superimposed samples is used to automatically locate the bounding boxes of all anomaly regions from the aeromagnetic ΔT grid. S3. For the input measured aeromagnetic ΔT grid data, execute the following inference process: S31. Data preprocessing: Calculate the grid spacing of the measured data and construct a multi-channel feature map; S32. Anomaly Detection and Localization: Input the multi-channel feature map into the trained target detection network to detect all abnormal bounding boxes; S33. Window cropping and feature extraction: Using each anomaly center point as the center, crop a window of fixed physical size from the original ΔT grid, and extract multi-channel physical features from the current window; S34. Parallel Inversion of Multiple Shapes: Input the processed window features into the inversion network model for each shape trained in step S1. Each network independently outputs the depth, magnetization tilt angle and magnetization deflection angle under the corresponding shape assumption. S35, Optimal Shape Fusion Decision: For the multiple sets of parameters output by each shape network, a weighted score is calculated by combining the forward fitting residual, morphological prior knowledge, and inversion confidence. The most likely shape is recommended for each anomaly, and the depth, magnetization tilt angle, and magnetization deflection angle corresponding to the shape are output as the final estimation result.

[0026] Furthermore, the multi-channel physical characteristics in S1 include: analytical signal amplitude, horizontal gradient, vertical derivative, and frequency domain amplitude spectrum; The amplitude of the analytical signal is calculated from the horizontal and vertical derivatives of the magnetic anomaly ΔT. The vertical derivative is calculated in the frequency domain. The frequency domain amplitude spectrum is obtained by taking the amplitude spectrum after performing a two-dimensional fast Fourier transform on ΔT and then performing logarithmic compression.

[0027] Specifically, for the magnetic anomaly field ΔT of each sample, the following multi-channel physical feature map is calculated to enhance the network's ability to perceive depth and orientation information: Analyze the signal amplitude:

[0028] Among them, vertical derivative Calculated in the frequency domain.

[0029] Gradient field: horizontal derivative , and vertical guide number .

[0030] Frequency domain amplitude spectrum: Perform a two-dimensional fast Fourier transform on ΔT, take the amplitude spectrum, and then perform logarithmic compression and normalization.

[0031] Data augmentation: Square root transformation of ΔT This is to suppress the dynamic range of large value anomalies.

[0032] Z-score standardization is applied to all feature maps, and the mean and standard deviation of each feature are saved as standardization parameters.

[0033] Furthermore, each shape-specific inversion network in S1 is a three-branch convolutional neural network structure, including a depth branch, an inclination branch, and an eccentricity branch; the depth branch outputs normalized depth values ​​using a regression method, while the inclination and eccentricity branches output the probability distributions of magnetization inclination and magnetization eccentricity respectively using a classification method. The input to the depth branch is a two-channel feature map composed of the analytical signal amplitude and ΔT; the input to the tilt branch is a three-channel feature map composed of ΔT, the horizontal derivative, and the vertical derivative; and the input to the skew branch is a four-channel feature map composed of ΔT, the horizontal derivative, the vertical derivative, and the frequency domain amplitude spectrum.

[0034] Specifically, design a three-branch convolutional neural network, including: Depth branch: The input is a two-channel feature map of [ASA, ΔT], and the output is a normalized depth value (between 0 and 1), which is then denormalized. Obtain the actual depth.

[0035] Inclination branch: Input is The three-channel feature map outputs a probability distribution for 36 categories, corresponding to the magnetization tilt angle I. Each category represents a 5° interval, ranging from -90° to 90°.

[0036] Angle branch: Input is The four-channel feature map outputs a probability distribution for 36 categories, corresponding to the magnetization deflection D. Each category represents a 10° interval, ranging from 0° to 360°.

[0037] During training, the depth branch uses the L1 loss function, and the two directional branches use the cross-entropy loss function. The weighted sum of these three loss functions is used as the total loss. The above steps are repeated for the three shapes: sphere, cylinder, and plate, to train three independent network models and their corresponding standardized parameters.

[0038] Furthermore, the method for generating training samples in S1 is as follows: a fixed spatial grid range is set, and a large number of samples are randomly generated within the spatial grid range. The parameters of each sample include the center plane coordinates, burial depth, magnetization tilt angle, magnetization deflection angle, and the unique geometric parameters of the corresponding shape of the magnetic body. The theoretical magnetic anomaly ΔT is calculated using the finite length forward modeling formula, and Gaussian noise is added to simulate real observations.

[0039] Specifically, a fixed spatial grid range is defined (e.g., -500m to 500m in both the X and Y directions, with a grid size of 128×128 points). Within this range, a large number of samples are randomly generated. The parameters for each sample include: the center plane coordinates (x0, y0) of the magnetic object, the burial depth z (30-200m), the magnetization tilt I (-90° to 90°), the magnetization declination D (0° to 360°), and shape-specific parameters, such as the radius and length of a cylinder, and the length, width, and thickness of a plate. The theoretical magnetic anomaly ΔT is calculated using the finite-length forward modeling formula, and Gaussian noise is added to simulate real observations.

[0040] Furthermore, the target detection network in S2 is a YOLO network, and the input is a three-channel pseudo-color image. The three channels are: the robustly normalized enhancement ΔT, the robustly normalized horizontal gradient magnitude, and the robustly normalized absolute value of the vertical derivative. The multiple anomaly superposition sample is obtained by randomly generating 1 to 5 anomalies within a fixed spatial grid range and linearly superimposing them to obtain the total field. The bounding rectangle of each significant anomaly is automatically extracted as a label based on adaptive threshold segmentation and connected component analysis.

[0041] Specifically, multiple anomaly detection samples are generated: within the same spatial grid, 1-5 anomalies are randomly generated, with random shapes, parameters, and positions. These are linearly superimposed to obtain the total field. Based on adaptive threshold segmentation and connected component analysis, the bounding rectangle of each significant anomaly is automatically extracted from the total field as a label.

[0042] Constructing a three-channel input image: To adapt to the YOLO network, the magnetic anomaly field of each multi-anomaly sample is converted into a three-channel pseudo-color image. Channel 1: Enhanced ΔT, i.e. Robust normalization (linear mapping to 0-255 after truncating 2%-98% quantiles).

[0043] Channel 2: Horizontal gradient magnitude, i.e. Robust normalization.

[0044] Channel 3: Absolute value of the vertical derivative, i.e. Robust normalization.

[0045] Training the detection network: Using the generated multi-anomaly three-channel images and corresponding bounding box labels, a single-class (anomaly) YOLOv8 object detection model is trained to automatically locate all anomalous regions in the aeromagnetic map.

[0046] Furthermore, in S32, the center of each detection box output by the target detection network is peak-corrected using the amplitude of the analyte signal within the local area to obtain the location of the abnormal center.

[0047] Furthermore, the optimal shape fusion decision in S35 specifically includes: performing forward modeling calculations on the parameters output by each shape-specific inversion network to obtain the theoretical magnetic anomaly and calculating the root mean square error with the observed field; simultaneously calculating the information entropy of the magnetic anomaly gradient direction within the window as prior knowledge of the shape; combining the softmax maximum probability of the directional branches of each shape-specific inversion network as the inversion confidence, and determining the optimal shape through weighted scoring.

[0048] Specifically, data preprocessing: Calculate the grid spacing (dx, dy) based on the grid range and number of points in the measured data. Calculate the three-channel feature map according to the method in S2.

[0049] Anomaly detection and localization: The three-channel feature map is input into the trained YOLO model to detect all anomaly bounding boxes. For the center of each detected box, peak correction is performed using the analyzed signal amplitude (ASA) within the local region to obtain a more accurate anomaly center location.

[0050] Window cropping and feature extraction: Centered on each corrected anomaly center point, a window of fixed physical size (e.g., 104×104 points, corresponding to approximately 1000m×1000m) is cropped from the original ΔT grid. For this window, multi-channel physical feature extraction and Z-score normalization are repeated, using pre-stored standard deviation and mean parameters during normalization.

[0051] Parallel inversion of multiple shapes: The processed window features are input into three trained inversion networks for spheres, cylinders and plates, respectively. Each network independently outputs its depth, magnetization tilt angle and magnetization deflection angle under the shape assumption.

[0052] Optimal shape fusion decision: For the three sets of parameters output by the three networks, forward modeling is performed to obtain the theoretical magnetic anomaly. And calculate with the observed field The root mean square error (RMSE) is calculated. Simultaneously, the information entropy of the magnetic anomaly gradient direction within the window is used as morphological prior knowledge. Finally, combining the inversion confidence (i.e., the softmax maximum probability of the directional branch, the forward fitting residual, and the morphological prior knowledge), a weighted score is applied to recommend the most probable shape for each anomaly, and the corresponding depth, magnetization tilt angle, and magnetization declination angle are output as the final estimation result.

[0053] Furthermore, S4 calculates the difference between induced magnetization and observed magnetization based on the recommended shape and the inverted magnetization direction, combined with the geomagnetic field direction at the measuring point, estimates the equivalent remanence index Q value, and provides a reliability evaluation.

[0054] Specifically, based on the recommended shape and the inverted magnetization direction ( , ), combined with the direction of the geomagnetic field at the measuring point ( , The difference between induced magnetization and observed magnetization is calculated, the equivalent remanence index Q is estimated, and a reliability evaluation is given, providing users with a qualitative reference for the impact of remanence.

[0055] Through the above method, the present invention realizes fully automated intelligent interpretation of the entire process from inputting the aeromagnetic ΔT grid to outputting the top or center depth, magnetization direction (I, D) and recommended shape of multiple magnetic anomalies.

[0056] In one specific embodiment, the following is included: This invention provides a deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation, the overall technical process of which is as follows: Figure 1 As shown, the main steps include: Step S1: Obtain grid data of total aeromagnetic magnetic field anomaly (ΔT).

[0057] Step S2: Data preprocessing and feature image generation.

[0058] Step S3: Multiple Anomaly Detection (YOLO).

[0059] Step S4: Fine correction of the anomaly center ASA.

[0060] Step S5: Local window feature extraction.

[0061] Step S6: Parallel inference of polymorphic parameters.

[0062] Step S7: Shape recommendation decision.

[0063] Step S8: Forward modeling residual and remanence assessment.

[0064] Step S9: Summarize and output the results.

[0065] The following provides further explanation of the key technical details in each step.

[0066] Data preprocessing and feature image generation (step S2): Let the input aeromagnetic ΔT grid data be matrix F with size . The grid spacing is and (Unit: m). First, calculate the following physical characteristics: Analytical Signal Amplitude (ASA):

[0067] Where the horizontal derivative , Using spatial domain central difference and vertical derivative Calculated in the frequency domain: For Perform a two-dimensional FFT and multiply by the wavenumber amplitude. The real part is obtained by inverse transformation.

[0068] Horizontal total gradient magnitude (HGM): ; Absolute value of vertical guide: ; Enhanced ΔT: This is used to suppress the dynamic range of strong anomalies and improve the visualization of weak anomalies. Subsequently, robust percentile normalization is applied to each feature channel: taking the lower quantile of 2% and the higher quantile of 98%, and cropping the channel values ​​to... Interval, linearly mapped to Generate a single-channel image. Stack the three channels of enhancement ΔT, horizontal total gradient magnitude, and vertical derivative absolute value to form a three-channel RGB feature image, which is used as input to the YOLO detection network.

[0069] Multiple anomaly target detection (step S3): This embodiment uses the YOLOv8n model, a single-stage object detection network. The dataset used to train this model consists entirely of synthetic forward modeling data, generated as follows: Randomly generate 1 to 5 magnetic bodies, randomly selecting one of a sphere, a finite number of cylinders, or a finite number of plates, and randomly setting the center coordinates within a range of 500m to -500m. , ), depth 30m~200m, dip angle deflection angle Shape-specific parameters: Spheres have no additional parameters; finite cylinders have a radius of 15m~40m and a length of 300m~800m; finite plates have a length of 300m~800m, a width of 50m~150m, and a thickness of 10m~40m. Uniform magnetic moment. .

[0070] The discrete dipole superposition method is employed, with the sphere represented by an analytical formula. The cylinder and plate are discretized into point elements, and forward modeling is performed by summing these elements. The field values ​​of multiple anomalies are linearly superimposed to obtain the total field. .

[0071] After adding 5% Gaussian noise to the total field, generate a three-channel feature image according to step S2.

[0072] Label extraction: Bounding boxes are automatically extracted directly from the superimposed total field through connected component analysis. Specifically, a threshold is applied. After binarization, OpenCV is used to extract each salient region, and the bounding rectangle is converted to YOLO normalized format.

[0073] A total of 20,000 samples were generated, and the training set and validation set were divided in an 8:2 ratio.

[0074] Training parameters: epochs=200, batch=16, initial learning rate=0.001, optimizer AdamW, cosine annealing scheduling, image size 256×256, and the trained model weights are saved as best.pt.

[0075] During inference, the three-channel image generated in step S2 is input into the YOLO model, and a confidence threshold is set. The IoU threshold for NMS is 0.3, and the model outputs the normalized coordinates of each detection box. Scaling these coordinates back to the original grid coordinate system yields the detection centers. .

[0076] ASA fine-tuning of anomaly centers (step S4): Since the center of the YOLO detection bounding box may not be the true source of the anomaly, local correction is required using the ASA peak value. Specific steps: [The rest of the text appears to be a separate, unrelated section about anomalies and ASA peak values.] Centered on the original ΔT grid, draw a... A local sub-region of a pixel. Calculate the ASA within this sub-region, find the pixel coordinates of the pixel with the maximum ASA, and convert them to global coordinates. If the correction offset exceeds 25 pixels (an empirical threshold), the original detection center is retained. This coordinate is used as the final center for subsequent window extraction.

[0077] Local window feature extraction (step S5): To maintain consistency with the training conditions of the individual parametric regression network, the resolution and size of the local fields need to be standardized. The training data for the individual model are all in... Forward modeling on a mesh, with a physical range of -500m to 500m and a mesh spacing of approximately 7.874m, and cropping 12 pixels from each side, yields the following result. The input. During inference, the following sub-steps are used to achieve complete alignment: S51: Full-field downsampling to The original field F is scaled to its original size using OpenCV's resize function. Size. The grid spacing is 7.874m at this point.

[0078] S52: Coordinate mapping of the target center in the downsampling field. Let... The corrected center pixel coordinates in the original grid are given by the following physical coordinates: ,

[0079] exist Corresponding pixel coordinates in the field: ,

[0080] S53: with Centered on, extract The window has a half-width of 52 pixels. If the cropped area exceeds the 128-pixel boundary, the excess portion is filled by edge copying. The cropped window is denoted as [insert denot here]. The size is .

[0081] S54: In Calculate multi-channel features: Call the same function as in step S2, preserving the full size.

[0082] , , : Gradients in three directions.

[0083] :right Perform a two-dimensional FFT, take the logarithmic magnitude, and normalize to [0,1], specifically: ,

[0084] Parallel inference of polymorphic parameters (step S6): Network structure: Each of the three shapes (sphere, finite cylinder, and finite plate) corresponds to an independent neural network model with an identical architecture, but its parameters are trained independently. The network structure consists of three parallel convolutional branches and a final fully connected head. Deep branching: The input is 2 channels, passing through 4 Conv2d(3×3)-BatchNorm2d-ReLU-MaxPool2d(2) units, with the number of channels successively being 16→32→64→128. Finally, it is connected to AdaptiveAvgPool2d, flattened, and then passed through a fully connected layer (128→64→1, including Dropout 0.3). After Sigmoid activation, it is linearly mapped to... rice.

[0085] I-branch (tilt angle): Input is 3 channels. The backbone network is the same as the deep branch, with 36 fully connected outputs corresponding to -90° to 85°, in 5° intervals, trained using CrossEntropyLoss.

[0086] D branch (angle): Input is 4 channels. The backbone network is the same, and the fully connected output has 36 classes, corresponding to 0° to 350°, with an interval of 10°. It is trained using CrossEntropyLoss.

[0087] Training data and standardized parameters: The training data was generated using synthetic forward modeling, with 5000 samples for each shape. The forward modeling parameters were within the same range as the YOLO training data, but each sample contained only one anomalous body. For each sample, following the procedure in step S5, the data was directly generated from... Main performance stage cutting Generate feature maps and save the mean and standard deviation of each channel as standardization parameters.

[0088] Training hyperparameters: epochs=100, batch=32, learning rate=0.001, optimizer Adam, ReduceLROnPlateau scheduler (patience=8), early stopping patience=25. The loss function is... The weights are 1:1:1.

[0089] Preprocessing during inference: During reasoning, for the data extracted in step S5 , , , , , Z-score standardization was performed based on the mean and std values ​​saved for each shape model: The standardized data are stacked into tensors with batch=1, and fed into the corresponding models to obtain 36-class logits for depth values ​​d and I, and 36-class logits for D. Softmax is applied to the logits, and the class index corresponding to the highest probability is selected. and Converted to angles: ,

[0090] Simultaneously, the maximum probability value of Softmax is recorded as the classification confidence of the model for I and D. and .

[0091] Body shape recommendation decision (step S7): For the same anomaly, the three models will output three sets of parameters and confidence scores. The recommendation decision is divided into two steps: Base score calculation: For each shape, calculate the geometric mean of the classification confidence scores:

[0092] Morphological prior weighting: only when the difference between the highest and second-highest base scores... At that time, morphological prior weights based on gradient directional entropy are introduced. Calculation method: exist Calculate the gradient direction angle for each pixel. .

[0093] Pixels with gradient magnitude greater than 20% of the maximum magnitude are considered valid pixels, and their gradient direction histograms (18 intervals) are calculated.

[0094] Calculate the normalized entropy: ; according to Assign different prior weights: Table 1. Different Prior Weights

[0095] Final score: ; Otherwise, take directly. ; The shape with the highest score is selected as the interpretation result for this anomaly and recommended, while the parameters of the other two shapes are retained for reference.

[0096] Forward modeling residual and remanence assessment (step S8): To provide more quality references, the forward fitting residuals and equivalent Q values ​​were further calculated for the recommended shape.

[0097] Orthogonal residual: in On the mesh, using the predicted depth d, dip angle I, and deflection angle D as parameters, a normalized magnetic moment is applied. Forward modeling templates (the sphere uses analytical formulas, the cylinder uses a fixed template with a radius of 25m and a length of 500m, and the plate uses a fixed template with a length of 500m, a width of 100m, and a thickness of 20m) are used to perform forward modeling to obtain the theoretical field. The equivalent magnetic moment is estimated using least squares: ; Final prediction field Root mean square of residuals: ; Equivalent Q-value estimation: The Q-value (i.e., the Königsberg ratio) is estimated using empirical formulas such as the ratio of positive to negative extreme values ​​in the observation field and the ASA center offset. This embodiment uses the following empirical formula:

[0098] in The local geomagnetic field inclination angle is calculated using the IGRF model. This is based on the Q value and signal-to-noise ratio. Provide qualitative labels: If SNR>8, then the induced magnetic field is dominant; Then there is moderate remanence; If the value is 0, it indicates that strong remanence is dominant; if SNR≤8, it indicates low signal-to-noise ratio and the remanence assessment result is unreliable.

[0099] Multiple exception handling and rollback mechanisms: If no anomalies are detected in step S3, and YOLO has no output or the model is not loaded, the system automatically reverts to the "full-field single-unit inversion" mode: downsampling the entire field to... Simply crop 12 pixels from each of the four sides to get the full field. The feature map is fed into three models for inversion, and a single anomaly result is output. When multiple anomalies are detected, steps S4 to S8 are executed for each anomaly, and finally all results are summarized into a table for output.

[0100] See Figure 2 The image shown is an example of YOLO detection results provided in an embodiment of the present invention. See Figure 3 The image shown is a diagram illustrating the inversion results of a single anomaly provided in an embodiment of the present invention. Examples of implementation results: Taking a synthetic test sample as an example, it contains three anomalies: A0 is a plate-like object (true depth 46.9m, I=-73.1°, D=77.1°), A1 is a sphere (true depth 87.4m, I=67.1°, D=281.1°), and A2 is a cylinder (true depth 135.2m, I=15.2°, D=102.3°). This system detected these three anomalies, and after ASA correction, the inversion results are as follows: Table 2 Inversion Results

[0101] The mean absolute error in depth is approximately 2.6 m, and the deviations in inclination and deflection are both within one classification interval, i.e., 5° or 10°. The processing time, from loading the file to outputting the CSV, is approximately 3.2 seconds. This demonstrates that the proposed method can quickly and accurately achieve automatic detection and intelligent parameter estimation of multiple aeromagnetic anomalies.

[0102] Table 3 Explanation of English Abbreviations and Terms

[0103] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0104] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based method for aeromagnetic anomaly detection and intelligent parameter estimation, characterized in that, Includes the following steps: S1. For three typical magnetic body models, namely sphere, finite-length cylinder and finite-length plate, construct and train independent inversion network models respectively. During training, extract multi-channel physical features of the magnetic anomaly field ΔT of each sample as network input. Each shape-specific inversion network is used to invert the depth and magnetization direction parameters under the assumption of the corresponding single shape. S2. A target detection network trained based on multiple anomaly superimposed samples is used to automatically locate the bounding boxes of all anomaly regions from the aeromagnetic ΔT grid. S3. For the input measured aeromagnetic ΔT grid data, execute the following inference process: S31. Data preprocessing: Calculate the grid spacing of the measured data and construct a multi-channel feature map; S32. Anomaly Detection and Localization: Input the multi-channel feature map into the trained target detection network to detect all abnormal bounding boxes; S33. Window cropping and feature extraction: Using each anomaly center point as the center, crop a window of fixed physical size from the original ΔT grid, and extract multi-channel physical features from the current window; S34. Parallel Inversion of Multiple Shapes: Input the processed window features into the inversion network model for each shape trained in step S1. Each network independently outputs the depth, magnetization tilt angle and magnetization deflection angle under the corresponding shape assumption. S35, Optimal Shape Fusion Decision: For the multiple sets of parameters output by each shape network, a weighted score is calculated by combining the forward fitting residual, morphological prior knowledge, and inversion confidence. The most likely shape is recommended for each anomaly, and the depth, magnetization tilt angle, and magnetization deflection angle corresponding to the shape are output as the final estimation result.

2. The method for aeromagnetic anomaly detection and intelligent parameter estimation based on deep learning according to claim 1, characterized in that, The multi-channel physical characteristics in S1 include: analytical signal amplitude, horizontal gradient, vertical derivative, and frequency domain amplitude spectrum; The amplitude of the analytical signal is calculated from the horizontal and vertical derivatives of the magnetic anomaly ΔT. The vertical derivative is calculated in the frequency domain. The frequency domain amplitude spectrum is obtained by taking the amplitude spectrum after performing a two-dimensional fast Fourier transform on ΔT and then performing logarithmic compression.

3. The method for aeromagnetic anomaly detection and intelligent parameter estimation based on deep learning according to claim 1, characterized in that, Each shape-specific inversion network in S1 is a three-branch convolutional neural network structure, including a depth branch, an inclination branch, and an eccentricity branch. The depth branch outputs normalized depth values ​​using a regression method, while the inclination and eccentricity branches output probability distributions of magnetization inclination and magnetization eccentricity respectively using a classification method. The input to the depth branch is a two-channel feature map composed of the analytical signal amplitude and ΔT; the input to the tilt branch is a three-channel feature map composed of ΔT, the horizontal derivative, and the vertical derivative; and the input to the skew branch is a four-channel feature map composed of ΔT, the horizontal derivative, the vertical derivative, and the frequency domain amplitude spectrum.

4. The method for aeromagnetic anomaly detection and intelligent parameter estimation based on deep learning according to claim 1, characterized in that, The method for generating training samples in S1 is as follows: a fixed spatial grid range is set, and a large number of samples are randomly generated within the spatial grid range. The parameters of each sample include the center plane coordinates, burial depth, magnetization tilt angle, magnetization deflection angle, and the unique geometric parameters of the corresponding shape of the magnetic body. The theoretical magnetic anomaly ΔT is calculated using the finite length forward modeling formula, and Gaussian noise is added to simulate real observation.

5. The method for aeromagnetic anomaly detection and intelligent parameter estimation based on deep learning according to claim 1, characterized in that, The target detection network in S2 is a YOLO network, and the input is a three-channel pseudo-color image. The three channels are: the robustly normalized enhancement ΔT, the robustly normalized horizontal gradient magnitude, and the robustly normalized absolute value of the vertical derivative. The multiple anomaly superposition sample is obtained by randomly generating 1 to 5 anomalies within a fixed spatial grid range and linearly superimposing them to obtain the total field. The bounding rectangle of each significant anomaly is automatically extracted as a label based on adaptive threshold segmentation and connected component analysis.

6. The method for aeromagnetic anomaly detection and intelligent parameter estimation based on deep learning according to claim 1, characterized in that, In S32, the center of each detection box output by the target detection network is peak-corrected using the amplitude of the analyte signal within the local area to obtain the location of the abnormal center.

7. The method for aeromagnetic anomaly detection and intelligent parameter estimation based on deep learning according to claim 1, characterized in that, The optimal shape fusion decision in S35 specifically includes: performing forward modeling calculations on the parameters output by each shape-specific inversion network to obtain the theoretical magnetic anomaly and calculate the root mean square error with the observed field; simultaneously calculating the information entropy of the magnetic anomaly gradient direction within the window as prior knowledge of the shape; combining the softmax maximum probability of the directional branches of each shape-specific inversion network as the inversion confidence, and determining the optimal shape through weighted scoring.

8. The method for aeromagnetic anomaly detection and intelligent parameter estimation based on deep learning according to claim 1, characterized in that, It also includes S4, which, based on the recommended shape and the inverted magnetization direction, combined with the geomagnetic field direction at the measuring point, calculates the difference between induced magnetization and observed magnetization, estimates the equivalent remanence index Q value, and provides a reliability evaluation.

Images