An intelligent multi-modal fusion method for multi-source gravity and magnetic measurement data and related devices
By using a deep learning network model to preprocess and separate anomalies in multi-source gravity and magnetic measurement data, and then weighted and fused the data, the problem of low accuracy in multi-source gravity and magnetic data fusion was solved, achieving efficient and accurate data fusion results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINESE PEOPLES LIBERATION ARMY UNIT 61540
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-12
AI Technical Summary
Existing technologies have not effectively combined multimodal fusion technology with deep learning, making it impossible to achieve efficient and high-precision fusion of multi-source gravity and magnetic measurement data, and thus failing to fully realize the information value of multi-source gravity and magnetic data.
A deep learning network model is used to preprocess and separate anomalies in multi-source gravity and magnetic measurement data. Weights are configured based on the height characteristics of gravity and magnetic measurements for weighted fusion. The model is trained using a residual network model with the batch normalization layer removed and an adaptive learning rate optimization algorithm.
It achieves efficient and high-precision fusion of multi-source gravity and magnetic measurement data. The fusion results are more consistent with the geophysical reality, with a significant improvement in accuracy and a substantial reduction in computation time. It is suitable for processing large amounts of gravity and magnetic measurement data.
Smart Images
Figure CN122196910A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of geophysical exploration and gravity and magnetic matching navigation technology, and in particular to an intelligent multimodal fusion method and related device for multi-source gravity and magnetic measurement data. Background Technology
[0002] Gravity and magnetic measurements are core technologies for geophysical exploration and gravity-magnetic matching navigation. Depending on the measurement platform, they are mainly categorized into satellite gravity and magnetic measurements, airborne gravity and magnetic measurements, ground-based gravity and magnetic measurements, and shipborne gravity and magnetic measurements. Different observation methods observe the Earth's gravity and magnetic field from different perspectives, reflecting information from different wavebands of the Earth's gravity and magnetic field, resulting in differences in the information content of the acquired gravity and magnetic measurement data.
[0003] Geophysical theory and practice demonstrate that fusing different types of gravity and magnetic observation data can eliminate systematic differences between data, achieve information complementarity, improve the quality and reliability of solution results, and broaden the spectral range. This is crucial for constructing high-precision, high-resolution global gravity and magnetic field models and ocean gravity and magnetic background field models. Currently, the classic methods for gravity and magnetic data fusion are mainly statistical methods and equivalent source methods. The least squares collocation method is a representative of statistical methods. Although it can combine multiple observation data, it suffers from difficulties in constructing local covariance models and the high difficulty in inverting high-dimensional covariance matrices. While the equivalent source method is robust and can add constraint information, it is time-consuming to invert matrices when the data volume is large, has high requirements for computing equipment, and the theoretical requirements for determining the weights of different types of data are high.
[0004] To address the technical challenges of traditional fusion methods, deep learning technology has been introduced into the field of gravity and magnetic data processing. It can uncover complex patterns and rules within data through deep network structures. However, current technologies lack a mature solution that combines multimodal fusion with deep learning to perform anomaly separation based on the highly specific characteristics of gravity and magnetic data before weighted fusion. This makes it impossible to achieve efficient and high-precision fusion of multi-source gravity and magnetic measurement data, thus failing to fully realize the informational value of such data. Therefore, an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data is urgently needed. Summary of the Invention
[0005] The purpose of this application is to provide an intelligent multimodal fusion method and related device for multi-source gravity and magnetic measurement data, which can achieve efficient and high-precision fusion of multi-source gravity and magnetic measurement data.
[0006] To achieve the above objectives, this application provides the following solution.
[0007] In a first aspect, this application provides an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data, which includes the following steps.
[0008] Acquire multi-source gravity and magnetic measurement data; the multi-source gravity and magnetic measurement data are raw gravity and magnetic measurement data obtained by various observation methods.
[0009] The multi-source gravity and magnetic measurement data are preprocessed and reduced to the same height surface to obtain gridded gravity and magnetic data with the same distribution range, data type and resolution.
[0010] The gridded gravity and magnetic data are input into a deep learning network model for anomaly separation processing to obtain regional anomaly data and local anomaly data corresponding to the gridded gravity and magnetic data. The deep learning network model is a residual network model with the batch normalization layer removed, and the residual network model is trained by the mean square error loss function and the adaptive learning rate optimization algorithm.
[0011] Based on the height characteristics of gravity and magnetic measurements, height weights are configured for the regional anomaly data and the local anomaly data, and the regional anomaly data and the local anomaly data are fused based on the height weights. Then, the fused regional anomaly data and the local anomaly data are synthesized to obtain the fusion result corresponding to the multi-source gravity and magnetic measurement data.
[0012] Optionally, the multiple observation methods include at least two of satellite gravity and magnetic measurement, airborne gravity and magnetic measurement, ground gravity and magnetic measurement, and shipborne gravity and magnetic measurement; the preprocessing includes at least one of data denoising, data interpolation, and data format unification.
[0013] Optionally, a high weight is configured for the regional abnormal data and the local abnormal data, specifically including the following steps.
[0014] For the abnormal data in the region, the configured height weight increases with the increase of the measured height.
[0015] For the local anomaly data, the configured height weight decreases as the measured height increases.
[0016] In this case, the sum of the height weights of each type of abnormal data is 1.
[0017] Optionally, the configuration method for the height weight includes at least one of manual configuration based on the reliability of the observation data and configuration through multiple trial calculations using a check line.
[0018] Optionally, the residual network model includes at least four residual blocks, multiple convolutional layers, and multiple modified linear unit layers; at least one convolutional kernel of size 1×1 is in each convolutional layer, and the remaining convolutional kernels are 3×3 in size, and the stride of all convolutional layers is set to 1. After convolution, a zero-padding strategy is used to keep the feature map size unchanged; the last convolutional layer of the residual network model has only one convolutional kernel to ensure that the input data and output data have the same size.
[0019] Optionally, the adaptive learning rate optimization algorithm is the Adam algorithm.
[0020] The training process of the residual network model specifically includes the following steps.
[0021] a) Construct a gravity and magnetic data training set; the training set includes gravity and magnetic anomaly sample data with regional anomaly labels.
[0022] b) Input the gravity and magnetic anomaly sample data into the initial residual network model, and obtain the regional anomaly prediction value through forward propagation.
[0023] c) Calculate the loss value between the predicted anomaly value and the anomaly label of the region based on the mean squared error loss function, and optimize the weights and biases of the model through backpropagation using the Adam algorithm.
[0024] d) Repeat steps b) to c) until the loss value converges to the preset threshold, and obtain the trained residual network model.
[0025] Secondly, this application provides an intelligent multimodal fusion system for multi-source gravity and magnetic measurement data, characterized in that the intelligent multimodal fusion system for multi-source gravity and magnetic measurement data is used to implement the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data as described in the first aspect, and the intelligent multimodal fusion system for multi-source gravity and magnetic measurement data includes the following modules.
[0026] The data acquisition module is used to acquire multi-source gravity and magnetic measurement data; the multi-source gravity and magnetic measurement data are raw gravity and magnetic measurement data obtained by various observation methods.
[0027] The preprocessing module is used to preprocess the multi-source gravity and magnetic measurement data, reduce the multi-source gravity and magnetic measurement data to the same height surface, and obtain gridded gravity and magnetic data with the same distribution range, data type and resolution.
[0028] An anomaly separation module is used to input the gridded gravity and magnetic data into a deep learning network model for anomaly separation processing, to obtain regional anomaly data and local anomaly data corresponding to the gridded gravity and magnetic data; the deep learning network model is a residual network model with the batch normalization layer removed, and the residual network model is trained by the mean square error loss function and the adaptive learning rate optimization algorithm.
[0029] The fusion module is used to configure height weights for the regional anomaly data and the local anomaly data based on the height characteristics of gravity and magnetic measurements, and to fuse the regional anomaly data and the local anomaly data based on the height weights. Then, the fused regional anomaly data and local anomaly data are combined to obtain the fusion result corresponding to the multi-source gravity and magnetic measurement data.
[0030] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data as described in any one of the first aspects.
[0031] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data as described in the first aspect.
[0032] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data as described in the first aspect.
[0033] According to the specific embodiments provided in this application, this application has the following technical effects.
[0034] This application provides an intelligent multimodal fusion method and related apparatus for multi-source gravity and magnetic measurement data. It combines multimodal fusion technology with deep learning and introduces it into gravity and magnetic data processing. A deep learning network model is used to separate regional and local anomalies in the gravity and magnetic data. Weighted fusion is then performed based on the height characteristics of the gravity and magnetic measurements, fully utilizing the sensitivity of measurement data at different heights to gravity and magnetic field anomalies. This makes the fusion result more consistent with geophysical reality, significantly improving fusion accuracy. Furthermore, the deep learning network model used in this application is essentially a residual network model with the batch normalization layer removed, trained using a mean squared error loss function and an adaptive learning rate optimization algorithm. This avoids forced normalization of data features, protects the geophysical characteristics of the gravity and magnetic data, ensures the accuracy of anomaly separation, and thus guarantees the accuracy of multimodal fusion. In addition, the deep learning network model of this application can be trained in advance. For novel gravity and magnetic measurement data, only the trained deep learning network model needs to be run to directly perform anomaly separation and fusion without repeated training, which greatly saves computing time and improves the fusion efficiency of multi-source gravity and magnetic data. It can achieve efficient and high-precision fusion of multi-source gravity and magnetic measurement data, give full play to the complementary role of gravity and magnetic data from different observation methods, and is suitable for processing large amounts of gravity and magnetic measurement data. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] Figure 1 This diagram illustrates the application environment of an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data, provided in one embodiment of this application.
[0037] Figure 2 This is a flowchart illustrating an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data provided in an embodiment of this application.
[0038] Figure 3 This is a schematic diagram of the network structure of a residual network model provided in an embodiment of this application.
[0039] Figure 4 This is a schematic diagram of an anomaly separation strategy for a residual network model provided in an embodiment of this application.
[0040] Figure 5 This is a schematic diagram of the multimodal fusion technology calculation process provided in an embodiment of this application.
[0041] Figure 6This is a schematic diagram of the horizontal position of a theoretical model provided in an embodiment of this application.
[0042] Figure 7 This is a schematic diagram of the magnetization tilt angle, magnetization deflection angle, and total field strength of the experimental area provided in an embodiment of this application.
[0043] Figure 8 A schematic diagram of a theoretical model forward modeling magnetic anomalies at altitudes of 1km and 0km provided in an embodiment of this application.
[0044] Figure 9 This is a schematic diagram of the calculation results provided by the least squares configuration adaptive fusion method according to an embodiment of this application.
[0045] Figure 10 A schematic diagram showing the fusion-derived magnetic field strength anomaly at a height of 0.5 km and its difference from the theoretical value, provided as an embodiment of this application.
[0046] Figure 11 This is a schematic diagram illustrating the calculation results of a deep learning-based intelligent multimodal fusion method provided in an embodiment of this application.
[0047] Figure 12 This is a schematic diagram of the structure of an intelligent multimodal fusion system for multi-source gravity and magnetic measurement data provided in an embodiment of this application.
[0048] Figure 13 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0050] Based on the measurement platform, current gravity and magnetic measurement methods mainly include satellite gravity and magnetic measurements, airborne gravity and magnetic measurements, and ground-based and shipborne gravity and magnetic measurements. With the development of gravity and magnetic measurement methods and technologies, gravity and magnetic measurement data have become more diverse. Different gravity and magnetic measurement methods observe the Earth's gravity and magnetic field from different perspectives, reflecting information about the Earth's gravity and magnetic field in different wavebands. Therefore, the information contained in the obtained gravity and magnetic measurement data varies.
[0051] Geophysical theory and practice demonstrate that only by fusing as much different types of gravity and magnetic observation data as possible, integrating them according to a specific mathematical model into data with consistent accuracy and resolution, can systematic differences between different data types be eliminated. Furthermore, this allows for mutual complementarity among various data sets, thereby enhancing the quality and reliability of the solution results and broadening their spectral range, leading to further improvements in accuracy and spatial resolution. To construct high-precision, high-resolution global gravity and magnetic field models and oceanic gravity and magnetic background field models suitable for geophysical exploration and gravity and magnetic matching navigation, it is essential to fuse all available multi-source gravity and magnetic measurement data. Therefore, research on the fusion of gravity and magnetic field data is urgently needed to comprehensively utilize multi-source gravity and magnetic field information and effectively construct reliable magnetic fundamental data. However, current research results in this area are scarce.
[0052] Among numerous methods for fusing gravity and magnetic data, statistical methods and equivalent source methods are the most classic. Statistical methods were initially applied to geophysical fields to interpolate and extrapolate gravity and magnetic observation data. The least squares collocation method is a representative of statistical methods, its significance lying in introducing statistical theory and analytical methods into physical geodesy. Because the least squares collocation method can combine multiple types of observation data, it has a natural advantage in fusing multi-source gravity and magnetic data. This is because the solution of the least squares collocation method does not depend on the number of signal parameters to be estimated and is independent of the linear transformation of data and results. If the covariance function is chosen appropriately, the solution obtained by this method is optimal. However, the least squares collocation method faces two difficulties in practical applications: first, constructing a suitable, accurate, and concise local covariance model that meets the needs of the regional geomagnetic field; and second, effectively handling the problem of inverting high-dimensional covariance matrices. The equivalent source method is one of the commonly used and effective methods in gravity and magnetic field data processing. This method uses virtual field sources to calculate the physical field to be determined, without needing to solve integral or differential equations, making it a relatively robust method for potential field data transformation. The equivalent source method can directly process observation data at the original points, avoiding the introduction of gridding or leveling errors. It allows for joint calculation of multiple data types, such as using gradient tensors, three components, and two or three types of data from the total magnetic field strength anomaly to establish an equivalent source, fully considering the common source attributes between data. Its calculation process facilitates the addition of constraint information, such as prior geological structural information, physical law information, or spatial distribution trend information of the field source. Therefore, the equivalent source technique is suitable for multi-source gravity and magnetic data fusion processing. However, the equivalent source method also presents two challenges in application: one is the efficient matrix inversion operation when the data volume is large, which is time-consuming and requires advanced computing power; the other is the determination of weights for different types of data, which needs to be based on variance component estimation or parametric covariance, requiring advanced theoretical knowledge.
[0053] To address the challenges of traditional fusion methods, this embodiment proposes an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data. This method combines multimodal fusion technology with deep learning, performing anomaly separation based on the high-precision characteristics of gravity and magnetic data before weighted fusion. This enables efficient and high-precision fusion of multi-source gravity and magnetic measurement data, fully leveraging the complementary information from different observation methods. It solves the problems of high computational difficulty, complex weight determination, and low fusion accuracy inherent in traditional gravity and magnetic data fusion methods. By using a deep network structure with multi-layer nonlinear mapping for deep learning, hidden complex patterns and rules are uncovered in the data. Training with a large amount of multi-source gravity and magnetic detection data, statistical features are automatically learned from multi-source gravity field element and geomagnetic element data, thus truly grasping the inherent properties of multi-source gravity and magnetic detection data. Different weights are then assigned during fusion to achieve the fusion of different observation data.
[0054] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0055] The intelligent multimodal fusion method for multi-source gravity and magnetic measurement data provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on other servers. Terminal 102 can send multi-source gravity and magnetic measurement data to server 104. After receiving the multi-source gravity and magnetic measurement data, server 104 preprocesses the data, reducing it to the same height surface to obtain gridded gravity and magnetic data with the same distribution range, data type, and resolution. The gridded gravity and magnetic data is then input into a deep learning network model for anomaly separation processing, obtaining corresponding regional anomaly data and local anomaly data. Based on the height characteristics of gravity and magnetic measurements, height weights are configured for the regional and local anomaly data, and the regional and local anomaly data are fused based on these weights. Finally, the fused regional and local anomaly data are synthesized to obtain the fusion result corresponding to the multi-source gravity and magnetic measurement data. Server 104 can feed back the obtained fusion result corresponding to the multi-source gravity and magnetic measurement data to terminal 102. In addition, in some embodiments, the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data can also be implemented by the server 104 or the terminal 102 separately. For example, the terminal 102 can directly perform intelligent multimodal fusion processing on the multi-source gravity and magnetic measurement data, or the server 104 can obtain the multi-source gravity and magnetic measurement data from the data storage system and perform intelligent multimodal fusion processing on the multi-source gravity and magnetic measurement data.
[0056] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, and IoT devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.
[0057] In one exemplary embodiment, such as Figure 2 As shown, an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the following steps are included.
[0058] S1: Acquire multi-source gravity and magnetic measurement data. The multi-source gravity and magnetic measurement data refers to the raw gravity and magnetic measurement data obtained through various observation methods.
[0059] S2: Preprocess the multi-source gravity and magnetic measurement data to reduce the multi-source gravity and magnetic measurement data to the same height surface to obtain gridded gravity and magnetic data with the same distribution range, data type and resolution.
[0060] S3: The gridded gravity and magnetic data are input into a deep learning network model for anomaly separation processing to obtain regional anomaly data and local anomaly data corresponding to the gridded gravity and magnetic data. The deep learning network model is a Residual Network (ResNet) model with batch normalization (BN) layers removed, and the Residual Network model is trained using a Mean Square Error (MSE) loss function and an adaptive learning rate optimization algorithm.
[0061] S4: Based on the height characteristics of gravity and magnetic measurements, height weights are configured for the regional anomaly data and the local anomaly data, and the regional anomaly data and the local anomaly data are fused based on the height weights. Then, the fused regional anomaly data and the local anomaly data are synthesized to obtain the fusion result corresponding to the multi-source gravity and magnetic measurement data.
[0062] By implementing steps S1 to S4 above, multimodal fusion technology is combined with deep learning. Anomalies are separated based on the high characteristics of gravity and magnetic data before weighted fusion. This enables efficient and high-precision fusion of multi-source gravity and magnetic measurement data, fully leveraging the complementary information of gravity and magnetic data from different observation methods. It also solves the problems of high computational difficulty, complex weight determination, and low fusion accuracy in traditional gravity and magnetic data fusion methods.
[0063] As an optional implementation, in step S1, the multiple observation methods include at least two of satellite gravity and magnetic measurement, airborne gravity and magnetic measurement, ground gravity and magnetic measurement, and shipborne gravity and magnetic measurement.
[0064] As an optional implementation, in step S2, the preprocessing includes at least one of data denoising, data interpolation, and data format unification.
[0065] As an optional implementation, in step S3, the residual network model includes at least four residual blocks, multiple convolutional layers (conv) and multiple modified linear unit (ReLU) layers; at least one convolutional kernel of size 1×1 is in each convolutional layer, and the remaining convolutional kernels are 3×3 in size, and the stride of all convolutional layers is set to 1. After convolution, a zero-padding strategy is used to keep the feature map size unchanged; the last convolutional layer of the residual network model has only one convolutional kernel to ensure that the input data and output data have the same size.
[0066] As an optional implementation, in step S3, the adaptive learning rate optimization algorithm is the Adam algorithm.
[0067] The training process of the residual network model specifically includes the following steps.
[0068] a) Construct a gravity and magnetic data training set; the training set includes gravity and magnetic anomaly sample data with regional anomaly labels.
[0069] b) Input the gravity and magnetic anomaly sample data into the initial residual network model, and obtain the regional anomaly prediction value through forward propagation.
[0070] c) Calculate the loss value between the predicted anomaly value and the anomaly label of the region based on the mean squared error loss function, and optimize the weights and biases of the model through backpropagation using the Adam algorithm.
[0071] d) Repeat steps b) to c) until the loss value converges to the preset threshold, and obtain the trained residual network model.
[0072] As an optional implementation, step S4 involves configuring high weights for the regional abnormal data and the local abnormal data, specifically including the following steps.
[0073] For the abnormal data in the region, the configured height weight increases with the increase of the measured height.
[0074] For the local anomaly data, the configured height weight decreases as the measured height increases.
[0075] In this case, the sum of the height weights of each type of abnormal data is 1.
[0076] As an optional implementation, in step S4, the configuration method of the height weight includes at least one of manual configuration based on the confidence level of the observation data and configuration through multiple trial calculations using a check line.
[0077] To make the technical solution of this application clearer, the specific implementation process of the technical solution of this application will be explained in detail below with examples.
[0078] This embodiment provides an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data. Based on the residual network of deep learning theory, the method decomposes multi-source gravity and magnetic measurement data into regional anomaly data and local anomaly data, and assigns weights to them respectively (for regional anomalies, the weight of low-altitude data is small and the weight of high-altitude data is large; for local anomalies, the weight of low-altitude data is large and the weight of high-altitude data is small, where the weight of low-altitude data and high-altitude data depends on the measurement height), thereby achieving accurate and efficient fusion of multi-source gravity and magnetic measurement data from different observation methods.
[0079] Multimodal fusion technology is widely used in image fusion, such as the fusion of infrared and visible light images. In low-light or obstructed conditions, infrared images can effectively identify targets, and fusion with visible light images can integrate features from both types of images, achieving information complementarity, enriching image information, and improving image quality. This embodiment introduces multimodal fusion technology into the fields of geophysical exploration and gravity / magnetic matching navigation, aiming to achieve the fusion of multi-source gravity and magnetic measurement data. In the field of gravity and magnetics, multimodal refers to multi-source gravity and magnetic measurement data obtained through various observation methods (ground-based gravity and magnetic measurements, airborne gravity and magnetic measurements, oceanic gravity and magnetic measurements, and satellite gravity and magnetic measurements). The prerequisite for fusion is that the observation area, data resolution, and data type are the same, and the data needs to be pre-calculated to the same altitude surface. Data obtained from measurements at different altitudes have different sensitivities to gravity and magnetic fields. As altitude increases, the sensitivity of the measuring instrument to regional anomalies increases, while its sensitivity to local anomalies decreases. The idea behind multimodal fusion technology is to use a deep residual network based on deep learning theory to decompose gravity and magnetic data into regional anomalies and local anomalies, and assign weights to them respectively (for regional anomalies, the weight of low-altitude data is small and the weight of high-altitude data is large; for local anomalies, the weight of low-altitude data is large and the weight of high-altitude data is small), so as to achieve the fusion of different observation data.
[0080] The network structure of the residual network model based on deep learning theory is as follows: Figure 3 As shown, the structure contains 4 residual blocks, 11 convolutional layers, and 5 modified linear unit layers. Figure 3The pink boxes represent convolutional layers (conv), and the green boxes represent ReLU layers (corrected linear unit layers). The ReLU layers perform non-linear mapping between different network layers, resulting in sparse activation of neurons, improving convergence speed, and facilitating the learning of gravity and magnetism data features. The convolutional layer on the skip connection of the third residual block has a 1×1 kernel, while the kernels of the remaining convolutional layers are set to 3×3 with a stride of 1. Zero-padding is applied after each convolution to ensure the feature map size remains unchanged after convolution, reducing errors introduced by the loss of gravity and magnetism data features and data boundaries during convolution. The number of kernels in each convolutional layer is marked in the structure diagram. The last convolutional layer has only one kernel to reduce the feature map depth to 1, ensuring consistent input and output data sizes. Meanwhile, to address the issue of abnormal separation in gravity and magnetic data, the learning and training adopts an end-to-end network structure. Since the input and output data have similarities and actual geophysical characteristics, the basic residual structure used in this embodiment removes the batch normalization layer. This avoids the batch normalization layer from forcibly normalizing the data features learned by the network, which would damage the geophysical characteristics of the data and cause difficulties for end-to-end learning.
[0081] Regional anomaly data in multi-source gravity and magnetic measurement data can be expressed as follows.
[0082] (1).
[0083] in, x This is the original multi-source gravity and magnetic measurement data. This is data from an abnormal region. n For local outlier data, the gravity and magnetic data anomaly separation model based on deep residual networks can be expressed as follows.
[0084] (2).
[0085] in, The regional gravity and magnetic anomaly data obtained through forward propagation computation of a neural network; Represents the structure of a neural network; For neural network parameters, and , w As weight, b This is a bias term.
[0086] The essence of training a residual network model is to minimize the loss function using an optimization algorithm. By using the error between the feedforward output and the sample label values, backpropagation continuously optimizes the neuron weights and other parameters. This embodiment uses the mean squared error loss function, defined as follows.
[0087] (3).
[0088] in, For data volume.
[0089] The Adam algorithm is an improved learning rate adaptive optimization algorithm based on the RMSProp (Root Mean Square Prop) algorithm. This algorithm combines the classic momentum idea of the RMSProp algorithm with the L2 norm, making it suitable for optimization problems with large amounts of hyperparameters. It also has the advantages of high computational efficiency and fast convergence speed. Therefore, this embodiment uses the Adam algorithm to optimize the objective and minimize the loss function.
[0090] In this embodiment, a residual network model is used to denoise and reconstruct gravity and magnetic data, which is divided into two stages: training and prediction. In the training stage, a labeled sample dataset is used for training. Multi-layer convolution is used to extract geomagnetic data features, and then a deep learning optimization algorithm is used to adjust the network parameters, establishing a complex neural network denoising model. In the prediction phase, the raw magnetic anomaly data is input into a pre-trained ResNet neural network model to obtain predicted regional anomaly data. Figure 4 This diagram illustrates the anomaly separation strategy for a residual network model, showcasing the complete process of implementing anomaly separation using a ResNet neural network.
[0091] In this embodiment, the multimodal fusion strategy based on deep learning theory includes the following steps.
[0092] Step 1: Amplify the observation data acquired through multiple methods to the same elevation surface, forming grid data with the same distribution range, data type, and resolution. The amplified data is denoted as... .
[0093] Step 2: Using the ResNet neural network model to... Anomaly separation is performed to obtain regional anomalies. and local abnormalities .
[0094] Step 3: Assign weights The merged data is represented by the following formula.
[0095] (4).
[0096] in, This represents the merged data; Indicates the i-th weight; This indicates that the i-th region is abnormal. Represents the i-th local anomaly; weight satisfy In practical applications, weights can be assigned manually based on the reliability of the observed data, or a check line can be used to determine the weight allocation scheme through multiple trial calculations.
[0097] This embodiment uses the fusion of ship magnetic anomalies and aeromagnetic anomalies as an example to demonstrate the data processing flow of multimodal fusion technology based on deep learning theory, such as... Figure 5 As shown.
[0098] This embodiment uses magnetic data as an example to conduct experimental verification of multi-source gravity and magnetic detection data fusion technology. The theoretical model consists of six cuboids of different geometric dimensions distributed at different depths, numbered 1, 2, 3, 4, 5, and 6, respectively. Figure 6 As shown in Table 1, the model parameters are as follows. The experimental range is 110°E~115°E and 15°N~20°N. Considering the variations in magnetization tilt, magnetization declination, and total field intensity over a large scale, the IGRF (International Geomagnetic Reference Field) model is used to simulate the magnetization tilt, magnetization declination, and total field intensity in the experimental area. Figure 7 As shown.
[0099] Table 1 Spatial coordinates and magnetic parameters of the source body in the theoretical model This experiment simulated the total magnetic field anomalies at altitudes of 1 km and 0 km, such as... Figure 8 As shown.
[0100] First, based on the least squares configuration adaptive fusion method, outlier data from the forward modeling at altitudes of 0km and 1km are fused. The fused outlier value at altitude 0km is then subtracted from the theoretical value. For example... Figure 9 As shown. Figure 9 From left to right, the data represents 1km altitude data, 0km altitude data, the predicted 0km altitude data after fusion, and the difference between the fusion result and the theoretical value. From top to bottom, the data represents the calculation results for the total intensity anomaly and the vertical component anomaly of the magnetic field. In the experiment, this method was only used for fusion calculations between data of the same type, taking the total magnetic field intensity anomaly and the vertical component of the magnetic field as examples. The theoretical model and experiments demonstrate that the fusion process of this method is feasible and stable.
[0101] Secondly, based on the equivalent source method, ground and airborne geomagnetic measurement data are fused to obtain a magnetic anomaly at a height of 0.5 km, which is then subtracted from the theoretical value. Figure 10As shown, only minor boundary effects exist. It should be noted that there are some differences between the magnetic anomaly data at 0km and 1km altitudes. Direct fusion would lead to mismatch issues. Therefore, the magnetic anomaly data at 0km altitude was expanded during fusion, while the magnetic anomaly data at 1km altitude was not expanded because the transition from 0km to 0.5km is an upward extension process, which is more stable.
[0102] Finally, utilizing the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data proposed in this application, only for the same data type (taking magnetic total field strength anomaly, magnetic field vertical component anomaly, and magnetic anomaly vertical derivative as examples), the method fuses the anomaly data calculated by the theoretical model at altitudes of 0 km and 1 km through forward modeling. After fusion, the method predicts the anomaly data at an altitude of 0.5 km and subtracts it from the theoretical value. Figure 11 As shown. Figure 11 From left to right, the data shows 0km altitude, 1km altitude, the predicted 0.5km altitude after fusion, and the difference between the fusion result and the theoretical value. From top to bottom, the data shows the calculated results of the total magnetic field anomaly, the vertical component anomaly of the magnetic field, and the vertical derivative of the magnetic anomaly. Figure 11 As can be seen, the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data is feasible and stable for this experimental case.
[0103] Comparing the fusion results of the least squares configuration adaptive fusion method, the equivalent source method, and the method proposed in this application, it can be seen that the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data proposed in this application has the best effect, with a maximum difference of less than 1 nT (Nt); followed by the equivalent source method, with a maximum difference of less than 5 nT; the least squares configuration adaptive fusion method has the worst accuracy, with a maximum difference of 240 nT.
[0104] To further verify the effectiveness and feasibility of the method in this application, this embodiment acquired raw data of total magnetic field intensity anomalies from satellites (100km altitude), aircraft (1km altitude), and ground (0km altitude) within the region of 110°E~115°E and 15°N~20°N. The raw data was preprocessed, and all data were reduced to a 0.5km altitude surface to obtain gridded magnetic anomaly data with a resolution of 0.05°×0.05°. Then, a training set was constructed, and the magnetization tilt, magnetization declination, and total field intensity of the region were simulated based on the IGRF model. A theoretical model consisting of six cuboids with different depths and geometric dimensions was constructed, and forward modeling was used to generate magnetic anomaly sample data with regional anomaly labels, forming the training set. An initial residual network model is then constructed, consisting of 4 residual blocks, 11 convolutional layers, and 5 modified linear unit layers. The convolutional kernels of the convolutional layers on the skip connections of the third residual block are 1×1, while the kernels of the remaining convolutional layers are 3×3. All convolutional layers have a stride of 1, with zeros padded after each convolution. The last convolutional layer has only one kernel, and the batch normalization layer is removed. The training data is then input into the initial residual network model, and forward propagation is used to obtain the predicted regional anomalies. The loss value is calculated using the mean squared error loss function, and the Adam algorithm is used for backpropagation to optimize the model parameters. The learning rate is set to 0.001, and the number of iterations is 1000. Training stops when the loss value converges to 0.001, resulting in the trained residual network model, i.e., the deep learning network model. Then, during anomaly separation, gridded magnetic anomaly data is input into the deep learning network model to obtain regional and local anomaly data corresponding to satellite, aerial, and ground gridded data, respectively. Finally, weighted fusion was performed. Weights were assigned to regional anomaly data: satellite data 0.7, aerial data 0.2, and ground data 0.1. Weights were also assigned to local anomaly data: satellite data 0.1, aerial data 0.2, and ground data 0.7. The sum of all weights was 1. The weighted sums of the regional and local anomaly data were then performed separately. Finally, the fused regional and local anomaly data were superimposed to obtain the multi-source magnetic measurement data fusion result at a 0.5 km altitude surface.
[0105] Based on the same inventive concept, this application also provides an intelligent multimodal fusion system for multi-source gravity and magnetic measurement data to implement the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data described above. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent multimodal fusion system for multi-source gravity and magnetic measurement data provided below can be found in the limitations of the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data described above, and will not be repeated here.
[0106] In one exemplary embodiment, such as Figure 12As shown, an intelligent multimodal fusion system for multi-source gravity and magnetic measurement data is provided, which includes the following modules.
[0107] The data acquisition module is used to acquire multi-source gravity and magnetic measurement data; the multi-source gravity and magnetic measurement data are raw gravity and magnetic measurement data obtained by various observation methods.
[0108] The preprocessing module is used to preprocess the multi-source gravity and magnetic measurement data, reduce the multi-source gravity and magnetic measurement data to the same height surface, and obtain gridded gravity and magnetic data with the same distribution range, data type and resolution.
[0109] An anomaly separation module is used to input the gridded gravity and magnetic data into a deep learning network model for anomaly separation processing, to obtain regional anomaly data and local anomaly data corresponding to the gridded gravity and magnetic data; the deep learning network model is a residual network model with the batch normalization layer removed, and the residual network model is trained by the mean square error loss function and the adaptive learning rate optimization algorithm.
[0110] The fusion module is used to configure height weights for the regional anomaly data and the local anomaly data based on the height characteristics of gravity and magnetic measurements, and to fuse the regional anomaly data and the local anomaly data based on the height weights. Then, the fused regional anomaly data and local anomaly data are combined to obtain the fusion result corresponding to the multi-source gravity and magnetic measurement data.
[0111] The intelligent multimodal fusion method and related device for multi-source gravity and magnetic measurement data proposed in this application have the following advantages.
[0112] (1) The deep learning network model is trained by a large amount of multi-source gravity and magnetic measurement data, and the statistical characteristics are automatically learned from the multi-source gravity field element and geomagnetic element data, so as to truly grasp the inherent properties of multi-source gravity and magnetic detection data, and thus make the fusion results more accurate and reliable.
[0113] (2) Multimodal fusion technology is introduced into the gravity and magnetic data processing scenario. Combined with deep learning network model, regional anomalies and local anomalies of gravity and magnetic data are separated. Then, weighted fusion is performed according to the height characteristics of gravity and magnetic measurements. The sensitivity characteristics of different height measurement data to gravity and magnetic field anomalies are fully utilized, making the fusion result more consistent with the actual geophysics and greatly improving the fusion accuracy. In the experimental verification, the maximum difference between the fusion result and the theoretical value is less than 1nT, which is far better than the traditional least squares collocation method and equivalent source method.
[0114] (3) The residual network model used removes the batch normalization layer, avoids the forced normalization of data features, and protects the geophysical characteristics of gravity and magnetic data. At the same time, the convolutional layer adopts a zero-padding strategy and a specific convolutional kernel, which reduces the loss of data features and boundaries, and ensures the accuracy of anomaly separation.
[0115] (4) The deep learning network model can be trained in advance. For new gravity and magnetic measurement data, only the trained model needs to be run to separate and fuse anomalies. There is no need to repeat the training, which greatly saves the computing time and improves the fusion efficiency of multi-source gravity and magnetic data. It is suitable for processing large amounts of gravity and magnetic measurement data.
[0116] (5) The high weight configuration method is flexible. It can be configured manually according to the reliability of the observation data, or it can be configured through multiple trial calculations of the check line. It can adapt to different exploration scenarios and data quality conditions, and has strong practicality and versatility.
[0117] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 13 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores multi-source gravity and magnetic measurement data and their fusion results. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements an intelligent multimodal fusion method for multi-source gravity and magnetic measurement data.
[0118] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0119] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0120] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0121] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0122] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.
[0123] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0124] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0125] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0126] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A smart multimodal fusion method for multi-source gravity and magnetic measurement data, characterized in that, The intelligent multimodal fusion method for multi-source gravity and magnetic measurement data includes: Acquire multi-source gravity and magnetic measurement data; the multi-source gravity and magnetic measurement data are raw gravity and magnetic measurement data obtained through multiple observation methods; The multi-source gravity and magnetic measurement data are preprocessed and reduced to the same height surface to obtain gridded gravity and magnetic data with the same distribution range, data type and resolution. The gridded gravity and magnetic data are input into a deep learning network model for anomaly separation processing to obtain regional anomaly data and local anomaly data corresponding to the gridded gravity and magnetic data; the deep learning network model is a residual network model with the batch normalization layer removed, and the residual network model is trained by the mean square error loss function and the adaptive learning rate optimization algorithm. Based on the height characteristics of gravity and magnetic measurements, height weights are configured for the regional anomaly data and the local anomaly data, and the regional anomaly data and the local anomaly data are fused based on the height weights. Then, the fused regional anomaly data and the local anomaly data are synthesized to obtain the fusion result corresponding to the multi-source gravity and magnetic measurement data.
2. The intelligent multimodal fusion method for multi-source gravity and magnetic measurement data according to claim 1, characterized in that, The various observation methods include at least two of satellite gravity and magnetic measurement, airborne gravity and magnetic measurement, ground gravity and magnetic measurement, and shipborne gravity and magnetic measurement; the preprocessing includes at least one of data denoising, data interpolation, and data format unification.
3. The intelligent multimodal fusion method for multi-source gravity and magnetic measurement data according to claim 1, characterized in that, Assigning height weights to the regional anomaly data and the local anomaly data respectively, specifically including: For the abnormal data in the region, the configured height weight increases with the increase of the measured height; For the aforementioned local anomaly data, the configured height weight decreases as the measured height increases; In this case, the sum of the height weights of each type of abnormal data is 1.
4. The intelligent multimodal fusion method for multi-source gravity and magnetic measurement data according to claim 1, characterized in that, The configuration method for the height weight includes at least one of the following: manual configuration based on the reliability of the observation data, and configuration through multiple trial calculations using a check line.
5. The intelligent multimodal fusion method for multi-source gravity and magnetic measurement data according to claim 1, characterized in that, The residual network model includes at least four residual blocks, multiple convolutional layers, and multiple modified linear unit layers; at least one convolutional kernel in each convolutional layer is 1×1 in size, and the remaining convolutional kernels are 3×3 in size, and the stride of all convolutional layers is set to 1. After convolution, a zero-padding strategy is used to keep the feature map size unchanged; the last convolutional layer of the residual network model has only one convolutional kernel to ensure that the input and output data have the same size.
6. The intelligent multimodal fusion method for multi-source gravity and magnetic measurement data according to claim 1, characterized in that, The adaptive learning rate optimization algorithm is the Adam algorithm; The training process of the residual network model specifically includes: a) Construct a gravity and magnetic data training set; the training set includes gravity and magnetic anomaly sample data with regional anomaly labels; b) Input the gravity and magnetic anomaly sample data into the initial residual network model, and obtain the regional anomaly prediction value through forward propagation; c) Calculate the loss value between the predicted value of the anomaly in the region and the anomaly label in the region based on the mean squared error loss function, and optimize the weights and biases of the model through backpropagation using the Adam algorithm; d) Repeat steps b) to c) until the loss value converges to the preset threshold, and obtain the trained residual network model.
7. An intelligent multimodal fusion system for multi-source gravity and magnetic measurement data, characterized in that, The intelligent multimodal fusion system for multi-source gravity and magnetic measurement data is used to implement the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data as described in any one of claims 1-6. The intelligent multimodal fusion system for multi-source gravity and magnetic measurement data includes: The data acquisition module is used to acquire multi-source gravity and magnetic measurement data; the multi-source gravity and magnetic measurement data are raw gravity and magnetic measurement data obtained by various observation methods; The preprocessing module is used to preprocess the multi-source gravity and magnetic measurement data, reduce the multi-source gravity and magnetic measurement data to the same height surface, and obtain gridded gravity and magnetic data with the same distribution range, data type and resolution. An anomaly separation module is used to input the gridded gravity and magnetic data into a deep learning network model for anomaly separation processing, to obtain regional anomaly data and local anomaly data corresponding to the gridded gravity and magnetic data; the deep learning network model is a residual network model with the batch normalization layer removed, and the residual network model is trained by a mean squared error loss function and an adaptive learning rate optimization algorithm; The fusion module is used to configure height weights for the regional anomaly data and the local anomaly data based on the height characteristics of gravity and magnetic measurements, and to fuse the regional anomaly data and the local anomaly data based on the height weights. Then, the fused regional anomaly data and local anomaly data are combined to obtain the fusion result corresponding to the multi-source gravity and magnetic measurement data.
8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data according to any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data as described in any one of claims 1-6.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the intelligent multimodal fusion method for multi-source gravity and magnetic measurement data as described in any one of claims 1-6.