Construction method and inversion method of mourhue surface depth gravity inversion model

Abstract

The invention provides a construction method and an inversion method of a mourhue surface depth gravity inversion model, and relates to the technical field of artificial intelligence, and the construction method of the mourhue surface depth gravity inversion model comprises the steps: obtaining a mourhue surface data set of a to-be-detected region; a corresponding gravity anomaly data set is obtained according to the mourhue surface data set; training and optimizing an initial model according to the gravity anomaly data set to obtain a final mourhos surface depth gravity inversion model; wherein the mourhue surface depth gravity inversion model is used for obtaining corresponding mourhue surface fluctuation data, and the initial model is obtained by adding an attention network in a U-Net basic network; according to the method, the accuracy and efficiency of inversion are improved, and a powerful tool and method are provided for research of a complex geological environment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a Moho depth gravity inversion model construction method and an inversion method. Background Art

[0002] The high sensitivity of gravity to density distribution makes it an important multidisciplinary geophysical method, which is widely used in global plate tectonic research, resource exploration and engineering exploration. By inverting gravity anomaly data, the depth model of the inhomogeneous density interface inside the earth can be determined, including the crust interface, Moho surface and Gutenberg surface. However, gravity inversion problems often show multiple solutions and are prone to divergence during the calculation process.

[0003] Nowadays, deep learning, as a branch of machine learning, provides a new strategy for solving the thorny problems in gravity inversion. However, the density interface inversion method based on deep learning is different from the traditional gravity inversion. It relies on a data-driven approach to map the observation data into the geophysical model. This method focuses on independently learning the complex mapping relationship between input variables and outputs by constructing a deep nonlinear network, thereby reducing the limitations of geophysical theory.

[0004] However, the current gravity inversion method based on deep learning can only roughly simulate the overall trend of the density interface, but has poor ability to recover details, resulting in low accuracy of the final inversion results and lack of generalization ability for complex terrain. Summary of the invention

[0005] The problem solved by the present invention is how to improve the accuracy of Moho surface depth gravity inversion results.

[0006] In order to solve the above problems, the present invention provides a method for constructing a Moho surface depth gravity inversion model, comprising:

[0007] Obtain the Moho dataset of the area to be measured;

[0008] Acquiring a corresponding gravity anomaly dataset according to the Moho dataset;

[0009] The initial model is trained and tuned according to the gravity anomaly data set to obtain the final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain the corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

[0010] Optionally, the gravity anomaly data set includes gravity anomaly data and corresponding label data; the training and tuning of the initial model according to the gravity anomaly data set to obtain the final Moho depth gravity inversion model includes:

[0011] Predicting the gravity anomaly data by using the initial model to obtain a corresponding temporary prediction result;

[0012] The initial model is tuned by using the temporary prediction result and the corresponding label data, and the tuned initial model is used as the Moho surface depth gravity inversion model.

[0013] Optionally, the initial model further includes an encoder and a decoder; and predicting the gravity anomaly data by using the initial model to obtain a corresponding temporary prediction result includes:

[0014] Encoding the gravity anomaly data by the encoder to obtain first characteristic data;

[0015] Processing the first feature data through the decoder to obtain target feature data; and concatenating the output data of the attention network with the target feature data to fuse feature data; decoding the fused feature data to obtain a temporary prediction result;

[0016] The output data of the attention network is obtained based on the output data during the processing of the encoder and the decoder.

[0017] Optionally, the encoder includes an output module and multiple upsampling modules, the attention network includes multiple attention gates, and the multiple attention gates correspond to the multiple upsampling modules one by one; the first feature data is processed by the decoder to obtain target feature data; and the output data of the attention network is spliced ​​with the target feature data to fuse feature data; the fused feature data is decoded to obtain a temporary prediction result, including:

[0018] Selecting a first upsampling module, wherein the first upsampling module is any of the upsampling modules;

[0019] Processing the first feature data through the first upsampling module to obtain first target feature data; fusing the first target feature data with the corresponding temporary data through the corresponding attention gate to obtain first temporary fused data; splicing the first temporary fused data and the first target feature data through the first upsampling module to obtain first fused feature data; processing the first fused feature data through the first upsampling module to obtain first temporary feature data;

[0020] Processing the first temporary feature data through the next first upsampling module to obtain second target feature data; fusing the second target feature data and the corresponding temporary data through the corresponding attention gate to obtain second temporary fused data; splicing the second temporary fused data and the second target feature data through the next first upsampling module to obtain second fused feature data; processing the second fused feature data through the next first upsampling module to obtain second temporary feature data, repeating the iteration until all the upsampling modules are traversed to obtain the processed feature data, and processing the processed feature data through the output module to obtain the temporary prediction result;

[0021] The temporary data is the output data of the corresponding submodule.

[0022] Optionally, obtaining a Moho dataset includes:

[0023] Acquire basic data of the area to be measured, wherein the basic data includes area range and coordinate data;

[0024] Obtaining a plurality of rectangular areas according to the area range and the coordinate data, and assigning values ​​to the rectangular areas respectively, to obtain an interface where a plurality of cuboids exist;

[0025] Superimposing and filtering all interfaces existing in the cuboids to obtain multiple Moho surface data;

[0026] The Moho surface data set is obtained according to all the Moho surface data.

[0027] Optionally, the superposition and filtering of interfaces existing on all the cuboids to obtain a plurality of Moho surface data includes:

[0028] By using formula 1, all interfaces existing in the cuboid are superimposed to obtain multiple temporary undulating height interfaces;

[0029] Smoothing each of the temporary undulating height interfaces to obtain a plurality of Moho surface data;

[0030] Wherein, the formula 1 is:

[0031]

[0032] Among them, H(m, n) is the temporary fluctuation height interface, m is the horizontal coordinate of the temporary fluctuation height interface, n is the vertical coordinate of the temporary fluctuation height interface, N is the total number of layers of the temporary fluctuation height interface, n is the number of cuboids existing in the temporary fluctuation height interface, Δh represents the height assignment of the temporary fluctuation height interface, h0 represents the corresponding initial height value of the temporary fluctuation height interface, i is the number of layers of the current temporary fluctuation height interface, and j is the jth cuboid existing in the temporary fluctuation height interface.

[0033] Optionally, the optimizing the initial model by using the temporary prediction result and the corresponding label data, and using the optimized initial model as the Moho depth gravity inversion model, comprises:

[0034] Perform loss calculation based on the temporary prediction result and the corresponding label data to obtain a loss function output;

[0035] Adjusting the model parameters of the initial model according to the output of the loss function until a stop condition is met, and using the initial model after parameter adjustment as the Moho depth gravity inversion model;

[0036] The stopping condition includes a preset training round, a preset loss condition or an early stopping mechanism.

[0037] The method for constructing a Moho depth gravity inversion model of the present invention has the following beneficial effects:

[0038] First, the Moho dataset of the area to be tested is obtained through means such as seismic wave detection. These datasets reflect the interface depth information between the crust and the mantle, providing a basis for subsequent model construction, and based on the obtained Moho dataset, the corresponding gravity anomaly dataset is calculated and obtained. The gravity anomaly data reflects the unevenness of the crust structure and is closely related to the undulating characteristics of the Moho surface. Finally, based on the obtained gravity anomaly dataset, the Moho depth gravity inversion model is finally obtained through training and tuning of the initial model. The initial model here is based on the U-Net basic network, and an attention network is added. The purpose of introducing the attention mechanism is to use the attention gate technology to connect the features between different layers in the encoding process in series to achieve the improvement of model performance.

[0039] Specifically, the focus is on the relationship between gravity anomaly and Moho surface fluctuation in areas with weak Moho surface fluctuation. By establishing a clear mapping relationship between gravity anomaly and underground interface depth, there is a high correlation between gravity anomaly changes and underground interface depth changes. For example, if the output features of the encoder block in the U-Net base network are directly connected to the features generated after the upsampling process, the features in the area with unclear fluctuations may have a negative impact on the final output of the model, and even reduce the prediction accuracy of the gravity inversion algorithm based on deep learning. Therefore, an attention mechanism is introduced in the jump connection module of the U-Net network. This design effectively suppresses the interference of the fuzzy area of ​​the mapping relationship between gravity anomaly data and interface fluctuations. By assigning different weights to features with and without obvious corresponding relationships, the AU-Net network can focus on learning the regression law of gravity inversion. The attention mechanism reduces the influence of features with no obvious corresponding relationship and potential random interference noise in the sample data set, enhances the effective data selectivity of the algorithm in the fitting process, and ultimately improves the calculation accuracy of the gravity inversion method.

[0040] Therefore, the present invention uses a method based on the U-Net basic network and an additional attention network, so that the initial model can accurately extract and model important features in the inversion task of gravity anomaly data and Moho surface fluctuation data, thereby supporting the construction of an efficient Moho surface depth gravity inversion model. This improvement not only improves the accuracy and efficiency of the inversion, but also provides a powerful tool and method for the study of complex geological environments.

[0041] In another aspect, the present invention provides a device for constructing a Moho surface depth gravity inversion model, comprising:

[0042] An acquisition unit, used to acquire a Moho surface dataset;

[0043] A processing unit, used for acquiring a corresponding gravity anomaly dataset according to the Moho dataset;

[0044] The processing unit is also used to train and tune the initial model according to the gravity anomaly data set to obtain a final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

[0045] The advantages of the Moho surface depth gravity inversion model construction device and the Moho surface depth gravity inversion model construction method described in the present invention are the same as those of the prior art, which will not be repeated here.

[0046] In another aspect, the present invention provides a Moho depth gravity inversion method, comprising:

[0047] Acquire Moho surface data of the area to be measured, and acquire corresponding gravity anomaly data according to the Moho surface data set;

[0048] The gravity anomaly data is input into the Moho depth gravity inversion model obtained by the Moho depth gravity inversion model construction method to obtain the Moho surface fluctuation data of the area to be measured.

[0049] The Moho depth gravity inversion method described in the present invention and the Moho depth gravity inversion model construction method described in the present invention have the same advantages over the prior art, which will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a flow chart of a method for constructing a Moho depth gravity inversion model according to an embodiment of the present invention;

[0051] Figure 2 is a schematic structural diagram of an initial model of an embodiment of the present invention;

[0052] Figure 3 Schematic diagram of the structure of the attention network of an embodiment of the present invention;

[0053] Figure 4 Schematic diagram of Moho surface recovery results of different types of gravity inversion methods according to an embodiment of the present invention;

[0054] Figure 5 A schematic structural diagram of a Moho depth gravity inversion model building device according to an embodiment of the present invention;

[0055] Figure 6 A schematic diagram of a flow chart of a Moho depth gravity inversion method according to an embodiment of the present invention;

[0056] Figure 7 The figure is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be interpreted as being limited to the embodiments described herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not intended to limit the scope of protection of the present invention.

[0058] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0059] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to"; the term "based on" means "based at least in part on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present invention are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0060] It should be noted that the modifications of "one" and "plurality" mentioned in the present invention are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0061] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only used for illustrative purposes, and are not used to limit the scope of these messages or information.

[0062] like Figure 1 As shown, a method for constructing a Moho depth gravity inversion model provided by an embodiment of the present invention includes:

[0063] S100, obtaining a Moho dataset of the area to be measured.

[0064] Specifically, obtaining the Moho dataset of the area to be measured is a key step in gravity inversion research, which involves multiple links, including data collection, processing and analysis. For example, relevant geological data can be collected by means of geological exploration, drilling and ground measurement. At the same time, gravity instruments can be used to conduct high-density gravity measurements in the area to be measured to record changes in gravity values, and seismic exploration technology can be used to collect seismic wave propagation speed and reflection data to infer the depth of the Moho surface. Finally, the acquired data is cleaned to ensure the accuracy of the data. In addition, data from different sources can be integrated into a unified data set. Geographic Information System (GIS) technology can be used to spatially align and integrate gravity data, seismic data and geological data.

[0065] For example: select and define the area to be measured, ensure that the area is representative and has sufficient geological complexity. First, use a high-precision gravimeter (such as an ultra-gravimeter) to perform high-density gravity measurements at selected locations. The model, accuracy (for example, 0.01mGal) and measurement interval of the gravimeter need to be recorded in detail. By measuring the change in ground gravity, collect values ​​and ensure repeatability and consistency under different conditions. Then use seismic wave exploration technology, set up the source and arrange receivers to record the propagation time and reflection data of seismic waves. The equipment here may include a source and a seismic trace to obtain data at a high resolution (such as a sampling rate ≥ 1000Hz), and calculate the wave velocity of different depth layers to infer the depth and change characteristics of the Moho surface. Finally, drill at key locations to extract underground rock and soil samples to obtain physical parameters such as density and porosity. Analyze the data results and compare them with the gravity measurement results to improve the accuracy of the data.

[0066] Preprocess the collected data, such as removing noise and unqualified data. The processing methods used should include average analysis, outlier removal, etc. to ensure the rigor and scientificity of the data.

[0067] Use Geographic Information System (GIS) technology to spatially align different types of data (gravity data, seismic data, geological profiles, etc.) to form a unified Moho dataset. Ensure that all data are in the same coordinate system and perform format conversion, such as storing the data in a GIS-friendly format (such as .shp or .geodatabase).

[0068] It should be noted that all data collection and processing processes adopted scientific methods, using modern instruments and standard operating procedures (SOPs). The reliability of the data can be verified by comparing it with existing geological data and repeated testing to further ensure its technicality.

[0069] S200: Acquire a corresponding gravity anomaly dataset according to the Moho dataset.

[0070] Specifically, the Parker gravity forward model with corrected direction can be used to obtain the desired gravity anomaly dataset based on the Moho dataset.

[0071] In a specific example, the Parker gravity forward model formula after the corrected direction is as follows:

[0072]

[0073] Among them, Δ g represents gravity anomaly, F[] represents Fourier transform operator; G represents gravitational constant, Δρ represents density difference on both sides of inhomogeneous medium interface, represents the wave number; z0 represents the average depth of the interface, and the vertical axis of the coordinate system is vertically downward as the positive direction; represents the distance vector from the inhomogeneous medium interface to the observation plane, then Indicates the depth change of the interface. This method is based on the derivation of the law of universal gravitation. The process of calculating gravity anomaly through the density difference on both sides of the inhomogeneous medium interface is stable and convergent, where z0, Δ ρ , and Δ g Obtained from the Moho dataset, e is a natural constant, also known as the Euler number.

[0074] S300, training and tuning the initial model according to the gravity anomaly data set to obtain a final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

[0075] Specifically, relevant gravity anomaly datasets are collected and organized to ensure the quality and accuracy of the data as input for model training. The gravity anomaly dataset includes gravity anomaly data and corresponding label data. The data is preprocessed, including data cleaning, normalization, and feature extraction, to meet the subsequent deep learning model input requirements.

[0076] The initial model is built based on the U-Net architecture. U-Net is a convolutional neural network commonly used for image segmentation because its encoder-decoder structure is suitable for processing spatial features. An attention network is added to U-Net to enhance the model's ability to focus on important features.

[0077] The process of improving model performance by concatenating features between different layers in the encoding process through the attention gate technology mainly focuses on feature fusion and optimization of deep neural networks. In areas with weak Moho surface fluctuations, the relationship between gravity anomalies and Moho surface fluctuations is particularly important, so careful feature extraction and selection are required. Specifically, the encoder part of U-Net extracts features of input data through multi-layer convolution. However, in areas with less obvious fluctuations, directly connecting the encoder output features with the features generated by the upsampling process may cause irrelevant features to interfere with the model's learning process. In the feature map, especially in areas with weak fluctuations, a large amount of redundant information may be contained, which may cause the model's learning to degrade and affect the accuracy of gravity inversion. The attention gate technology is introduced to use a special mechanism to weight different areas in the feature map. Specifically, the attention mechanism can help the model identify features that are more relevant to gravity anomalies, thereby improving the effectiveness of feature selection. In feature fusion, the attention gate dynamically assigns weights according to the context, focusing more attention on important features and reducing dependence on features in areas with less obvious fluctuations.

[0078] That is, by establishing a clear mapping relationship between gravity anomalies and the depth of underground interfaces, the model can better understand the high correlation between changes in gravity anomalies and changes in the depth of underground structures. In addition, during the training process, the model can learn which features are directly related to the fluctuations of the underground interface, and thus effectively use these features for inversion during prediction.

[0079] Using the attention gate technology, the features of different levels of the encoder and decoder are more reasonably integrated during the upsampling process. By selectively filtering and enhancing important features, redundant features in areas with unclear fluctuations are prevented from having a negative impact on the model output.

[0080] Through this optimization mechanism, the prediction accuracy of the final model can be significantly improved, ensuring that the results of the gravity inversion algorithm are more accurate and reliable.

[0081] Therefore, by introducing the attention gate technology, the features between different layers in the encoder can be effectively connected in series, and the importance of the features can be dynamically adjusted to reduce the impact of redundant features related to areas with weak Moho surface undulations, thereby improving model performance. This process not only helps to establish a clear mapping relationship between gravity anomalies and underground interfaces, but also improves the prediction accuracy of the gravity inversion algorithm based on deep learning. Ultimately, the model can more accurately reflect underground structural information and provide strong support for geological exploration and resource development.

[0082] And select a suitable loss function (such as mean square error) to measure the difference between the model prediction results and the actual gravity anomaly data. For example: use the back propagation algorithm to train the model, and optimize the model parameters through a large number of iterations to gradually reduce the loss function value. Adjust the model's hyperparameters (such as learning rate, batch size, training rounds, etc.) to find the best training settings. Finally, use an independent validation set to evaluate the trained model to check the generalization ability of the model. Use a variety of performance indicators (such as accuracy, recall, F1-score, etc.) to comprehensively evaluate the model effect and ensure the reliability of the model.

[0083] The trained model is used to perform gravity inversion of the Moho depth and output the corresponding Moho depth data. The Moho surface fluctuation data is extracted from the model output to show the changing characteristics of the underground structure.

[0084] In summary, by adding an attention mechanism, the model can more effectively identify and utilize key features, thereby improving the accuracy and reliability of gravity inversion. Appropriate hyperparameter tuning and enhanced network structure help improve the model's generalization ability for unknown data and better handle different types of gravity anomaly data. The deep learning model can fully mine the information in the gravity anomaly data set, thereby achieving accurate inversion of underground structures and improving the efficiency of data use. Through the deep learning method, the Moho depth is inverted and the fluctuation data is extracted, which reduces manual intervention and improves processing speed and efficiency.

[0085] Moreover, this model and process are not only suitable for the processing of gravity data, but can also be extended to the analysis and modeling of other geological exploration data, and has significant scientific research and engineering application value.

[0086] Through the above steps and processes, the Moho surface depth gravity inversion model finally constructed effectively realizes the extraction and analysis of underground structure characteristics, which is of great significance to geological research and related fields.

[0087] In this embodiment, first, a Moho dataset of the area to be measured is obtained by means of seismic wave detection or the like. This Moho dataset reflects the interface depth information between the crust and the mantle, and provides a basis for subsequent model construction. Based on the obtained Moho dataset, the corresponding gravity anomaly dataset is calculated. Among them, the gravity anomaly data reflects the unevenness of the crust structure, which is closely related to the undulating characteristics of the Moho surface.

[0088] Based on the acquired gravity anomaly dataset, the initial model is trained and tuned to finally obtain the gravity inversion model of the Moho depth. Among them, the initial model uses the U-Net basic network and introduces the attention network. The purpose of introducing the attention mechanism is to use the attention gate technology to connect the features of different layers in the encoding process to improve the model performance. That is, special attention is paid to the relationship between the gravity anomaly and the Moho surface fluctuation in the area with weak Moho surface fluctuation. By establishing a clear mapping relationship between gravity anomaly and underground interface depth, the strong correlation between the two is emphasized. For example: if the output features of the encoder block in the U-Net basic network are directly connected with the features generated after the upsampling process, the features in the area with unclear fluctuations may have a negative impact on the final output of the model, and even reduce the prediction accuracy of the gravity inversion algorithm based on deep learning. And by introducing the attention mechanism into the jump connection module of U-Net, the interference of the fuzzy mapping relationship area is effectively suppressed. At the same time, the allocation of different weights makes the AU-Net network focus on the regression law of gravity inversion. Moreover, the attention mechanism can reduce the impact of features without obvious corresponding relationships and random interference noise, and enhance the effective data selectivity of the algorithm in the fitting process, ultimately improving the computational accuracy of the gravity inversion method.

[0089] Therefore, the present invention uses a method based on the U-Net basic network and an additional attention network, so that the initial model can accurately extract and model important features in the inversion task of gravity anomaly data and Moho surface fluctuation data, thereby supporting the construction of an efficient Moho surface depth gravity inversion model. This improvement not only improves the accuracy and efficiency of the inversion, but also provides a powerful tool and method for the study of complex geological environments.

[0090] Optionally, the gravity anomaly data set includes gravity anomaly data and corresponding label data; the training and tuning of the initial model according to the gravity anomaly data set to obtain the final Moho depth gravity inversion model includes:

[0091] Predicting the gravity anomaly data by using the initial model to obtain a corresponding temporary prediction result;

[0092] The initial model is tuned by using the temporary prediction result and the corresponding label data, and the tuned initial model is used as the Moho surface depth gravity inversion model.

[0093] Specifically, in this embodiment, for the U-Net network based on the attention mechanism, that is, the initial model, a "double-layer inner and outer loop" training mechanism is used to achieve effective training and verification of the network. The inner loop focuses on the training and accuracy verification of the network, aiming to preliminarily evaluate the computing performance of the neural network on a given sample data set. Specific training process:

[0094] First, obtain the gravity anomaly data set of the area to be measured, including gravity anomaly data and corresponding label data (actual Moho surface undulation data); and divide the gravity anomaly data set into a training set, a test set and a validation set;

[0095] The gravity anomaly data in the training set is input into the U-Net network based on the attention mechanism to obtain the predicted Moho surface fluctuation data (temporary prediction result); then, the corresponding label data is used to calculate the loss between the actual Moho surface fluctuation data and the predicted Moho surface fluctuation data, and the network parameters are updated to reduce the error.

[0096] The above training is continued in multiple training cycles until the preset stopping conditions are met, which include reaching the maximum number of training rounds, convergence of the loss function, and triggering the early stopping mechanism of the network.

[0097] Through the verification of the outer loop, the purpose of the outer loop is to verify the generalization ability of the model and ensure the accuracy and applicability of the U-Net network based on the attention mechanism in the gravity inversion algorithm. That is, the test set is used to verify the accuracy of the U-Net network trained by the inner layer.

[0098] If after the above training, if the accuracy of the network meets the expected requirements, the network parameters with the minimum verification loss will be saved, and the gravity inversion algorithm based on the network will be output accordingly.

[0099] If the accuracy of the network does not meet the expected requirements, the test set is reset as training data, and the network training is continued, and the above training process is repeated until the accuracy of the network meets the predetermined requirements.

[0100] The inner and outer double-layer loop training mechanism is to train the model with a small amount of training data, and judge the accuracy of the model after the training is completed. If the model accuracy does not meet the standard, the training process of the model is repeated by using the data in the test set or the validation set as the training data. That is, through the inner and outer double-layer loop training mechanism, the U-Net network based on the attention mechanism can be more effectively optimized, thereby improving the performance and practicality of the gravity inversion algorithm.

[0101] Optionally, the initial model further includes an encoder and a decoder; and predicting the gravity anomaly data by using the initial model to obtain a corresponding temporary prediction result includes:

[0102] Encoding the gravity anomaly data by the encoder to obtain first characteristic data;

[0103] Processing the first feature data through the decoder to obtain target feature data; and concatenating the output data of the attention network with the target feature data to fuse feature data; decoding the fused feature data to obtain a temporary prediction result;

[0104] The output data of the attention network is obtained based on the output data during the processing of the encoder and the decoder.

[0105] Optionally, the encoder includes a plurality of submodules, and the processing of the gravity anomaly data by the decoder to obtain the first feature data includes:

[0106] Selecting a first submodule, wherein the first submodule is any of the submodules;

[0107] The gravity anomaly data is processed by the first submodule to obtain first temporary data; and the first temporary data is processed by the next submodule to obtain second temporary data, and the above process is repeated until all the submodules complete the processing process to obtain the first characteristic data.

[0108] Optionally, the encoder includes an output module and multiple upsampling modules, the attention network includes multiple attention gates, and the multiple attention gates correspond to the multiple upsampling modules one by one; the first feature data is processed by the decoder to obtain target feature data; and the output data of the attention network is spliced ​​with the target feature data to fuse feature data; the fused feature data is decoded to obtain a temporary prediction result, including:

[0109] Selecting a first upsampling module, wherein the first upsampling module is any of the upsampling modules;

[0110] Processing the first feature data through the first upsampling module to obtain first target feature data; fusing the first target feature data with the corresponding temporary data through the corresponding attention gate to obtain first temporary fused data; splicing the first temporary fused data and the first target feature data through the first upsampling module to obtain first fused feature data; processing the first fused feature data through the first upsampling module to obtain first temporary feature data;

[0111] Processing the first temporary feature data through the next first upsampling module to obtain second target feature data; fusing the second target feature data and the corresponding temporary data through the corresponding attention gate to obtain second temporary fused data; splicing the second temporary fused data and the second target feature data through the next first upsampling module to obtain second fused feature data; processing the second fused feature data through the next first upsampling module to obtain second temporary feature data, repeating the iteration until all the upsampling modules are traversed to obtain the processed feature data, and processing the processed feature data through the output module to obtain the temporary prediction result;

[0112] The temporary data is the output data of the corresponding submodule.

[0113] In some specific embodiments, Figure 2 As shown in the figure, a U-Net network based on the attention mechanism, i.e., the initial model, is constructed. Specifically, the attention mechanism (attention gate) is introduced to strengthen the selection and merging of feature extractions at different levels, thereby improving the learning effect of the deep network.

[0114] Construction of encoder: We can use VGG16 as the backbone network of the encoder to encode the data mapping relationship features of the gravity inversion algorithm. The encoder consists of five submodules, each of which includes two 3×3 convolutions, a batch normalization module, and a ReLU activation function.

[0115] The input data is single-channel gravity anomaly data with a resolution of 64×64, representing 4096 observation stations evenly distributed on the ground. When processing the input data, it is first normalized to the interval (-1,1); after two convolutional layers, the number of channels is expanded to 64, and as the network depth increases, the number of channels is 64, 128, 256, 512 and 1024 respectively. The resolution of each layer in the encoder decreases successively: the resolution of layers 1 to 5 is 64×64, 32×32, 16×16, 8×8 and 4×4, where the number of channels in each layer is twice that of the previous layer.

[0116] Decoder construction: The decoder is responsible for reconstructing the data mapping relationship characteristics of the gravity inversion algorithm. Its structure consists of 4 upsampling modules and 1 output module.

[0117] Each upsampling module includes a bilinear interpolation, a skip connection structure (SC), two 3×3 convolutions and a ReLU activation function. The skip connection structure performs channel superposition of the features processed by bilinear interpolation with the output of the corresponding downsampling module in the encoder, that is, the process data output from the submodule of each layer in the encoder (that is, each submodule obtains the corresponding process data after convolution twice on the corresponding input data) is superimposed with the output of the upsampling module in the corresponding encoder through the attention gate (AG) in the skip connection structure (SC);

[0118] After each upsampling, the number of channels of the feature map is reduced by half and the size is doubled. Finally, the data mapping relationship in the gravity inversion algorithm is obtained through the output module using 1×1 convolution and Sigmoid activation function.

[0119] Introducing the attention mechanism: Inserting an attention gate (AG) at the end of the skip connection structure to implement the attention mechanism, thereby eliminating irrelevant and noisy responses that may be generated in the skip connection between the encoder and the decoder.

[0120] The output of the encoder learns the correlation between channels through the skip connection module and filters out important channel features.

[0121] In the decoder part, the attention gate AG is used to combine the low-level feature map with the output of the corresponding jump connection module to obtain features that are more relevant to the gravity anomaly data and the Moho surface fluctuation data. At the same time, the output of the attention gate AG is spliced ​​with the upsampled low-level features (such as the first temporary fusion data and the second temporary fusion data), which improves the learning ability of the deep learning network for areas with inconspicuous fluctuation changes and promotes the search for data mapping relationships between input and output in gravity inversion.

[0122] Therefore, the U-Net network model based on the attention mechanism can effectively improve the accuracy and performance of the gravity inversion algorithm.

[0123] The attention gate AG is composed of the gating signal and the attention module.

[0124] In some preferred embodiments, an attention mechanism is implemented in the U-Net base network, and an attention gate can be inserted at the end of the skip connection structure to increase the learning weight of features that are strongly correlated with the gravity inversion algorithm and reduce the learning weight of features that are weakly correlated with the gravity inversion algorithm. Figure 3 As shown, a schematic diagram of the structure of the attention gate.

[0125] Specifically, by adding an attention network to the U-Net basic network:

[0126] Feature data extraction: The attention network first obtains the weight matrix by mean pooling the input data x (process data matrix) and g (the output feature data matrix of the decoder, such as the first temporary fusion data and the second temporary fusion data), as shown in the following formula:

[0127]

[0128] Among them, W x is the weight matrix of the process data of the encoder output including the current channel information, W g is the weight matrix of the decoder output feature data including the current channel information; x represents the process data matrix (feature data matrix) output by the encoder, g represents the output feature data matrix of the decoder; H and W are the height and width of the corresponding feature data matrix respectively. Here, i is the row number of the current matrix, and j is the column number of the current matrix.

[0129] The feature weight is updated by g Multiply W by g x The dot multiplication is performed with x, that is, the features obtained by the encoder are fully connected (i.e., superimposed) with the features obtained by the decoder, further reducing the amount of parameter calculation. The results of the fully connected layer (superimposed results) are summed through the ReLU activation function, and then the channels are stacked through convolution to obtain the intermediate matrix q of the attention mechanism.att , the calculation process is shown as follows:

[0130]

[0131] Among them, the intermediate matrix q att In the above equation, σ1 represents the ReLU activation function; Φ, W x and W g represents the convolution operation and T is the transpose.

[0132] Data update process, intermediate matrix q att The parameters Φ and W of the convolution operation are calculated through the Sigmod activation function. x and W g It can be obtained by back propagation of feature weights. The update expression for the weight of each data in the feature data matrix output by the encoder and decoder is as follows:

[0133]

[0134] Among them, α represents the feature weight matrix; σ2 represents the Sigmod activation function; q att The intermediate matrix representing the attention mechanism; Denotes the convolution operation Φ, W x and W g Specific parameters; Output represents the final result of each feature data.

[0135] Optionally, obtaining a Moho dataset includes:

[0136] Acquire basic data of the area to be measured, wherein the basic data includes area range and coordinate data;

[0137] Obtaining a plurality of rectangular areas according to the area range and the coordinate data, and assigning values ​​to the rectangular areas respectively, to obtain an interface where a plurality of cuboids exist;

[0138] Superimposing and filtering all interfaces existing in the cuboids to obtain multiple Moho surface data;

[0139] The Moho surface data set is obtained according to all the Moho surface data.

[0140] Optionally, the superposition and filtering of interfaces existing on all the cuboids to obtain a plurality of Moho surface data includes:

[0141] By using formula 1, all interfaces existing in the cuboid are superimposed to obtain multiple temporary undulating height interfaces;

[0142] Smoothing each of the temporary undulating height interfaces to obtain a plurality of Moho surface data;

[0143] Wherein, the formula 1 is:

[0144]

[0145] Among them, H(m, n) is the temporary fluctuation height interface, m is the horizontal coordinate of the temporary fluctuation height interface, n is the vertical coordinate of the temporary fluctuation height interface, N is the total number of layers of the temporary fluctuation height interface, n is the number of cuboids existing in the temporary fluctuation height interface, Δh represents the height assignment of the temporary fluctuation height interface, h0 represents the corresponding initial height value of the temporary fluctuation height interface, i is the number of layers of the current temporary fluctuation height interface, and j is the jth cuboid existing in the temporary fluctuation height interface.

[0146] In some embodiments, a Moho surface data set is generated by first acquiring basic data of the area to be measured, wherein the basic data includes area range and coordinate data (including diagonal point coordinates), and generating multiple rectangular areas of random size and random position through the area range and diagonal point coordinates, and the formula is as follows:

[0147]

[0148] Among them, for the position relationship of a generated rectangular area, a represents the length of the rectangular area, b represents the width of the rectangular area; L represents the length of the area to be measured, and W represents the width of the area to be measured; (x rect .,y rect. ) represents the coordinates of the lower left corner of the generated rectangular area.

[0149] Assign values ​​to the rectangular areas, set a height Δh for each randomly generated rectangular area, set an initial height value h0 for the area to be tested, and obtain an interface where multiple cuboids exist.

[0150] Layer superposition, randomly extract multiple layers of interfaces through formula 1 and superimpose them to generate a temporary undulating height interface.

[0151] Smoothing correction: Since the generated interface should conform to the actual geological conditions as much as possible, the geometric interface with steep changes in the region should be smoothed and filtered. Therefore, the mean filter is selected to smooth the superposition interface. The specific formula is as follows:

[0152]

[0153] Among them, P(m,n) is the interface result (Moho surface data) after smooth filtering, m is the horizontal coordinate of the interface after smooth filtering, and n is the vertical coordinate of the interface after smooth filtering; (2a+1) 2is the range of the filter window; H(x,y) is the simulated interface (temporary fluctuation height interface) in the filter window after the number of superimposed layers, x is the horizontal coordinate of the temporary fluctuation height interface, y is the vertical coordinate of the temporary fluctuation height interface, and a is the size of the filter window.

[0154] Among them, data normalization is performed during the processing. Based on the maximum and minimum normalization method, the gravity anomaly data and the corresponding label data are standardized, as shown in the following formula:

[0155]

[0156] in, is the result after data normalization; z i is the i-th data in the sample data set (gravity anomaly data and corresponding label data); z mi n is the minimum value of the data in the sample data set; z max is the maximum value of the data in the sample data set. Through data normalization, the dimensions can be unified, and the depth data values ​​and gravity anomaly data values ​​of different categories can be constrained within the range of (-1,1).

[0157] Optionally, the optimizing the initial model by using the temporary prediction result and the corresponding label data, and using the optimized initial model as the Moho depth gravity inversion model, comprises:

[0158] Perform loss calculation based on the temporary prediction result and the corresponding label data to obtain a loss function output;

[0159] Adjusting the model parameters of the initial model according to the output of the loss function until a stop condition is met, and using the initial model after parameter adjustment as the Moho depth gravity inversion model;

[0160] The stopping condition includes a preset training round, a preset loss condition or an early stopping mechanism.

[0161] In some embodiments, the loss function of the U-Net neural network model based on the attention mechanism selects the Huber loss function (the Huber loss function is a loss function for regression problems, which combines the advantages of the mean square error (MSE) and the absolute error (MAE) and can effectively handle outliers), as shown in the following formula:

[0162]

[0163] Among them, h prerepresents the temporary prediction result, h represents the label data; θ represents the parameters of the corresponding neural network, such as weights and biases. Based on the Huber loss function (the Huber loss function is a loss function used for regression problems. It combines the advantages of mean square error (MSE) and absolute error (MAE) and can effectively handle outliers), we added a smoothing loss term (expressed as L smooth ) to account for the assumed smoothness of the subsurface density interface. In addition, to alleviate the overfitting problem during training, we introduce a regularization norm L2 to constrain the network parameters, L(θ, h pre , h) is the final optimized loss function, L Huber (h pre , h) is the Huber loss function; L L2 (θ) is the regularization norm, which is a fixed value.

[0164] It should be noted that the Huber loss function combines the advantages of the mean absolute error (MAE, referred to as L1 loss) and the mean square error (MSE, referred to as L2 loss). Compared with the L2 loss, it is more robust to outliers and is differentiable at the center point. The specific definition of the Huber loss function is as follows:

[0165]

[0166] Among them, h pre represents the prediction result of the U-Net network, h represents the actual depth and distribution of the underground interface; δ is a non-negative parameter used to control the range between the quadratic term and the linear term of the loss function. pre When |≥δ, the Huber loss function uses the L1 (norm) loss function. When |hh pre |<δ, the Huber loss function can be regarded as the L2 (norm) loss function.

[0167] Among them, the L1 loss function is also called the minimum absolute deviation (LAD) and the absolute value loss function (LAE). In general, it converts the target value x i and the estimated value f(x i ). The L1 loss function can be expressed as:

[0168]

[0169] L2 loss function, L2 norm loss function, is also called least square error (LSE). It converts the target value x i and the estimated value f(x i ) is minimized. This loss is generally used in regression problems, and outliers have a greater impact on this loss.

[0170]

[0171] When testing the accuracy of the gravity inversion algorithm based on the attention mechanism U-Net network, the root mean square error (RMSE) indicator between the actual interface fluctuation value and the predicted interface fluctuation value (model output) is used for evaluation. The calculation formula of the RMSE indicator is as follows:

[0172]

[0173] Among them, RMSE is the root mean square error value; N is the number of point data of the interface undulation value; i is the point number of the depth data in the interface undulation model; y i represents the depth data value of the i-th point in the real interface undulation model (real interface undulation value); x i represents the gravity anomaly data value (model input) of the i-th point in the deep learning network, then f(x i ) is the depth data value of the i-th point in the predicted interface undulation model output by the deep learning network (model output).

[0174] In some specific embodiments, the Moho surface depth gravity inversion model (U-Net network with attention mechanism) is experimentally verified, and the specific implementation method is as follows:

[0175] The target area was used as the field test area, and the Moho depth gravity inversion model was used to invert and restore the measured gravity anomaly data. Similarly, as a comparative verification, the classic Parker-Oldenburg gravity inversion method and the gravity inversion method based on the U-Net network were selected for comparative experiments.

[0176] Due to the lack of Moho surface data, the Parker gravity forward model formula based on the corrected direction was used for the final data transformation. The data was converted into unified gravity anomaly data (in mGal) for effect comparison, and the root mean square error (RMSE) was selected as the evaluation indicator.

[0177] like Figure 4 As shown in the figure, the Moho surface recovery results (Moho surface fluctuation data) of different types of gravity inversion methods. Figure 4 (a) is the distribution map of gravity anomaly data related to the measured Moho depth model in the target area; Figure 4 (b) is the Moho depth distribution map of the area restored based on the classic Parker-Oldenburg gravity inversion method; Figure 4 (c) is the Moho depth distribution map obtained based on U-Net network gravity inversion; Figure 4(d) Moho depth distribution map obtained by the U-Net network gravity inversion algorithm based on the attention mechanism. The error results of each method are shown in Table 1:

[0178] Table 1 Comparison results of gravity anomalies (distribution of measured gravity anomaly data) between different gravity inversion methods and original data

[0179]

[0180] pass Figure 4 As can be seen from Table 1, the gravity inversion error of the inversion method based on the Moho surface depth gravity inversion model is reduced in practical applications compared with the direct use of the Parker-Oldenburg gravity inversion method. In addition, compared with the gravity inversion algorithm adopted by the unimproved U-Net network, the gravity inversion process of the inversion method based on the Moho surface depth gravity inversion model achieved higher accuracy after completing all six training cycles. The predicted value of gravity anomaly caused by the Moho surface model restored by the U-Net network gravity inversion algorithm based on the attention mechanism is very close to the input Bouguer gravity anomaly data value, and the algorithm performs well in depicting the inversion details of the underground interface.

[0181] In summary, the U-Net network gravity inversion algorithm based on the attention mechanism can accurately calculate the Moho surface undulation data. The overall trend of the algorithm's calculation results is closely consistent with the input data, which proves the effectiveness of the proposed inversion method based on the Moho surface depth gravity inversion model. In a specific example, the research results further show that the U-Net network gravity inversion algorithm based on the attention mechanism meets the requirements of modern gravity inversion methods for actual calculation accuracy.

[0182] like Figure 5 As shown, an embodiment of the present invention provides a Moho depth gravity inversion model construction device 500, comprising:

[0183] An acquisition unit 510 is used to acquire a Moho surface dataset;

[0184] A processing unit 510 is used to obtain a corresponding gravity anomaly dataset according to the Moho dataset;

[0185] The processing unit is also used to train and tune the initial model according to the gravity anomaly data set to obtain a final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

[0186] The Moho surface depth gravity inversion model construction device of this embodiment is used to implement the Moho surface depth gravity inversion model construction method as described above. Its advantages over the prior art are the same as the advantages of the Moho surface depth gravity inversion model construction method as described above over the prior art, and will not be repeated here.

[0187] like Figure 6 As shown, an embodiment of the present invention provides a Moho depth gravity inversion method, comprising:

[0188] Acquire Moho surface data of the area to be measured, and acquire corresponding gravity anomaly data according to the Moho surface data set;

[0189] The gravity anomaly data is input into the Moho depth gravity inversion model obtained by the Moho depth gravity inversion model construction method to obtain the Moho surface fluctuation data of the area to be measured.

[0190] The advantages of the Moho depth gravity inversion method of this embodiment over the prior art are the same as the advantages of the above-mentioned Moho depth gravity inversion model construction method over the prior art, which will not be repeated here.

[0191] An embodiment of the present invention provides a Moho depth gravity inversion device, comprising:

[0192] An acquisition module is used to acquire Moho surface data of the area to be measured, and acquire corresponding gravity anomaly data according to the Moho surface data set;

[0193] The processing module is used to input the gravity anomaly data into the Moho surface depth gravity inversion model obtained by the Moho surface depth gravity inversion model construction method to obtain the Moho surface fluctuation data of the area to be measured.

[0194] The advantages of the Moho depth gravity inversion device of this embodiment compared with the prior art are the same as the advantages of the above-mentioned Moho depth gravity inversion model construction method compared with the prior art, which will not be repeated here.

[0195] like Figure 7 As shown, an electronic device 700 provided by an embodiment of the present invention includes a memory 710 and a processor 720; the memory 710 is used to store a computer program; the processor 720 is used to implement the Moho surface depth gravity inversion model construction method as described above when executing the computer program.

[0196] In other words, an electronic device 700 includes a memory 710 and a processor 720 coupled to the memory 710; the memory 710 is configured to store a computer program; and the processor 720 is configured to perform the following operations when executing the computer program:

[0197] Obtain the Moho dataset of the area to be measured;

[0198] Acquiring a corresponding gravity anomaly dataset according to the Moho dataset;

[0199] The initial model is trained and tuned according to the gravity anomaly data set to obtain the final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain the corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

[0200] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for constructing a Moho depth gravity inversion model as described above is implemented.

[0201] In other words, a non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor performs the following operations:

[0202] Obtain the Moho dataset of the area to be measured;

[0203] Acquiring a corresponding gravity anomaly dataset according to the Moho dataset;

[0204] The initial model is trained and tuned according to the gravity anomaly data set to obtain the final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain the corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

[0205] An electronic device 700 that can be used as a server or client of the present invention will now be described, which is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device 700 is intended to represent various forms of digital electronic computer equipment, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device 700 can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples, and are not intended to limit the implementation of the present invention described herein and / or required.

[0206] The electronic device 700 includes a computing unit, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for the operation of the device can also be stored. The computing unit, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0207] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc. In the present application, the unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment of the present invention. In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0208] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A method for constructing a Moho depth gravity inversion model, characterized in that: include: Obtain the Moho dataset of the area to be measured; Acquiring a corresponding gravity anomaly dataset according to the Moho dataset; The initial model is trained and tuned according to the gravity anomaly data set to obtain the final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain the corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

2. The method for constructing a Moho depth gravity inversion model according to claim 1, characterized in that: The gravity anomaly data set includes gravity anomaly data and corresponding label data; The initial model is trained and tuned according to the gravity anomaly data set to obtain the final Moho depth gravity inversion model, including: Predicting the gravity anomaly data by using the initial model to obtain a corresponding temporary prediction result; The initial model is tuned by using the temporary prediction result and the corresponding label data, and the tuned initial model is used as the Moho surface depth gravity inversion model.

3. The method for constructing a Moho depth gravity inversion model according to claim 2, characterized in that: The initial model also includes an encoder and a decoder; the prediction of the gravity anomaly data by the initial model to obtain a corresponding temporary prediction result includes: Encoding the gravity anomaly data by the encoder to obtain first characteristic data; Processing the first feature data through the decoder to obtain target feature data; and concatenating the output data of the attention network with the target feature data to fuse feature data; decoding the fused feature data to obtain a temporary prediction result; The output data of the attention network is obtained based on the output data during the processing of the encoder and the decoder.

4. The method for constructing a Moho depth gravity inversion model according to claim 3, characterized in that: The encoder includes a plurality of submodules, and the decoder processes the gravity anomaly data to obtain first feature data, including: Selecting a first submodule, wherein the first submodule is any of the submodules; The gravity anomaly data is processed by the first submodule to obtain first temporary data; and the first temporary data is processed by the next submodule to obtain second temporary data, and the above process is repeated until all the submodules complete the processing process to obtain the first characteristic data.

5. The method for constructing a Moho depth gravity inversion model according to claim 4, characterized in that: The encoder includes an output module and multiple upsampling modules, the attention network includes multiple attention gates, and the multiple attention gates correspond to the multiple upsampling modules one by one; the first feature data is processed by the decoder to obtain target feature data; and the output data of the attention network is spliced ​​with the target feature data to fuse the feature data; Decoding the fused feature data to obtain a temporary prediction result includes: Selecting a first upsampling module, wherein the first upsampling module is any of the upsampling modules; Processing the first feature data through the first upsampling module to obtain first target feature data; fusing the first target feature data with the corresponding temporary data through the corresponding attention gate to obtain first temporary fused data; splicing the first temporary fused data and the first target feature data through the first upsampling module to obtain first fused feature data; processing the first fused feature data through the first upsampling module to obtain first temporary feature data; Processing the first temporary feature data through the next first upsampling module to obtain second target feature data; fusing the second target feature data and the corresponding temporary data through the corresponding attention gate to obtain second temporary fused data; splicing the second temporary fused data and the second target feature data through the next first upsampling module to obtain second fused feature data; processing the second fused feature data through the next first upsampling module to obtain second temporary feature data, repeating the iteration until all the upsampling modules are traversed to obtain the processed feature data, and processing the processed feature data through the output module to obtain the temporary prediction result; The temporary data is the output data of the corresponding submodule.

6. The method for constructing a Moho depth gravity inversion model according to claim 1, characterized in that: The obtaining of the Moho surface dataset comprises: Acquire basic data of the area to be measured, wherein the basic data includes area range and coordinate data; Obtaining a plurality of rectangular areas according to the area range and the coordinate data, and assigning values ​​to each of the rectangular areas respectively, to obtain an interface where a plurality of cuboids exist; Superimposing and filtering all interfaces existing in the cuboids to obtain multiple Moho surface data; The Moho surface data set is obtained according to all the Moho surface data.

7. The method for constructing a Moho depth gravity inversion model according to claim 6, characterized in that: The interfaces of all the cuboids are superimposed and filtered to obtain multiple Moho surface data, including: By using formula 1, all interfaces existing in the cuboid are superimposed to obtain multiple temporary undulating height interfaces; Smoothing each of the temporary undulating height interfaces to obtain a plurality of Moho surface data; Wherein, the formula 1 is: Among them, H(m, n) is the temporary fluctuation height interface, m is the horizontal coordinate of the temporary fluctuation height interface, n is the vertical coordinate of the temporary fluctuation height interface, N is the total number of layers of the temporary fluctuation height interface, n is the number of cuboids existing in the temporary fluctuation height interface, Δh represents the height assignment of the temporary fluctuation height interface, h0 represents the corresponding initial height value of the temporary fluctuation height interface, i is the number of layers of the current temporary fluctuation height interface, and j is the jth cuboid existing in the temporary fluctuation height interface.

8. The method for constructing a Moho depth gravity inversion model according to claim 2, characterized in that: The step of optimizing the initial model by using the temporary prediction result and the corresponding label data, and using the optimized initial model as the Moho depth gravity inversion model, comprises: Perform loss calculation based on the temporary prediction result and the corresponding label data to obtain a loss function output; Adjusting the model parameters of the initial model according to the output of the loss function until a stop condition is met, and using the initial model after parameter adjustment as the Moho depth gravity inversion model; The stopping condition includes a preset training round, a preset loss condition or an early stopping mechanism.

9. A device for constructing a Moho surface depth gravity inversion model, characterized in that: include: An acquisition unit, used to acquire a Moho surface dataset; A processing unit, used for acquiring a corresponding gravity anomaly dataset according to the Moho dataset; The processing unit is also used to train and tune the initial model according to the gravity anomaly data set to obtain a final Moho surface depth gravity inversion model; wherein the Moho surface depth gravity inversion model is used to obtain corresponding Moho surface undulation data, and the initial model is obtained by adding an attention network to the U-Net basic network.

10. A Moho depth gravity inversion method, characterized in that: include: Acquire Moho surface data of the area to be measured, and acquire corresponding gravity anomaly data according to the Moho surface data; The gravity anomaly data is input into the Moho depth gravity inversion model obtained by the Moho depth gravity inversion model construction method according to any one of claims 1 to 8 to obtain the Moho surface fluctuation data of the area to be measured.

Images