Underwater stratum data deduction method

Abstract

The invention discloses an underwater stratum data deduction method, and the method comprises the steps: obtaining multi-source data of an underwater stratum, carrying out the preprocessing of the obtained multi-source data, and constructing a multi-modal data set; obtaining a layered space adjacency graph and a multi-scale time adjacency graph, and constructing a multi-scale space-time coupling graph; fusing and extracting the multi-scale spatio-temporal features based on a multi-scale spatio-temporal convolutional neural network, embedding the physical field model into the multi-scale spatio-temporal convolutional neural network in the form of a model loss function, and constructing an underwater stratum spatio-temporal deduction model; a joint training strategy and model parameters based on the combination of self-supervised learning and supervised learning; and inputting the current underwater stratum state and the marine environment prediction data of a period of time in the future into the optimized underwater stratum space-time deduction model, and outputting stratum evolution prediction results under different time scales. According to the underwater stratum data deduction method provided by the embodiment of the invention, high-precision spatio-temporal evolution prediction of the underwater stratum can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater stratum dynamic simulation, and more particularly to an underwater stratum data deduction method, device, equipment and storage medium. BACKGROUND

[0002] Underwater stratum spatio-temporal evolution prediction is crucial for marine resource development, cross-sea facility construction and marine environment research. The seabed is rich in oil, natural gas and other resources, and the construction and operation of cross-sea cluster facilities require accurate assessment of the stability of the seabed stratum. However, the evolution of underwater stratum is influenced by various complex factors, including sedimentation, tectonic movement, marine environmental changes, etc. Traditional stratum evolution prediction methods mainly rely on seismic data and simple mathematical models, which have obvious limitations in dealing with complex stratum structures and long-term evolution.

[0003] In the prior art, some studies analyze strata by combining seismic data with simple mathematical models, but this method cannot effectively integrate multi-source data and is difficult to reflect the complex changes of strata in space and time. Although some other methods based on machine learning improve the prediction accuracy to some extent, they still have many shortcomings. First, existing methods are inefficient in data processing, and data from different sources have different scales and formats, making it difficult to directly fuse and analyze. Second, existing methods lack in-depth consideration of physical mechanisms, resulting in insufficient rationality and reliability of the prediction results. In addition, existing methods also have shortcomings in multi-scale analysis, making it difficult to comprehensively and accurately simulate the evolution process of strata at different time and spatial scales. These shortcomings limit the effectiveness of existing technology in practical applications, so a new method is needed to overcome these problems and improve the accuracy and reliability of underwater stratum spatio-temporal evolution prediction. SUMMARY

[0004] In view of at least one defect or improvement demand of the prior art, the present application provides an underwater stratum data deduction method, which can solve at least one of the problems in the background art.

[0005] To achieve the above-mentioned purpose, according to the first aspect of the present application, an underwater stratum data deduction method is provided, which comprises: obtaining multi-source data of underwater stratum, preprocessing the obtained multi-source data, and constructing a multi-modal data set; obtaining a hierarchical spatial adjacency graph and a multi-scale temporal adjacency graph based on the multi-modal data set, and constructing a multi-scale spatio-temporal coupling graph; fusing and extracting the multi-scale spatio-temporal features of the multi-scale spatio-temporal coupling graph based on a multi-scale spatio-temporal convolutional neural network, embedding a physical field model in the form of a model loss function into the multi-scale spatio-temporal convolutional neural network, and constructing an underwater stratum spatio-temporal deduction model. optimizing the underwater stratum space-time evolution model parameters based on a joint training strategy combining self-supervised learning and supervised learning; inputting the current underwater stratum state and marine environment prediction data for a future period of time into the optimized underwater stratum space-time evolution model, and outputting stratum evolution prediction results at different time scales.

[0006] Further, the underwater stratum data evolution method, the multi-source data of the underwater stratum is obtained, the obtained multi-source data is preprocessed, and a multi-modal data set is constructed, specifically including: obtaining multi-source data of the underwater stratum, including marine seismic data, logging while drilling time series data, seabed topography scanning data, seabed stratum scanning data, and marine environment monitoring data; unifying the multi-source data of the underwater stratum to the same space-time coordinate system using a space-time alignment algorithm; processing missing data based on a combination of a space-time variational autoencoder and a generative adversarial network to construct a multi-modal data set.

[0007] Further, the underwater stratum data evolution method, the multi-modal data set is used to obtain a hierarchical space adjacency graph and a multi-scale time adjacency graph, and a multi-scale space-time coupled graph is constructed, specifically including: dividing the underwater stratum space into different scale voxel units, determining node connections and edge weights based on regional geological structure characteristics, stratum attribute similarity, and distance, and constructing a hierarchical space adjacency graph; connecting nodes at the same spatial position but different periods based on time slices, adjusting time edge weights according to time intervals and marine environment parameter change rates, and constructing a multi-scale time adjacency graph; fusing the space and time adjacency graphs to obtain a multi-scale space-time coupled graph.

[0008] Further, the underwater stratum data evolution method, the physical field model is embedded in the multi-scale space-time convolutional neural network in the form of a model loss function, and an underwater stratum space-time evolution model is constructed, specifically including: based on a multi-scale space-time convolutional neural network, extracting spatial features through different size convolution kernels, and learning time variation patterns using a recurrent neural network variant combined with a multi-scale time adjacency graph; embedding the physical field model into the multi-scale space-time convolutional neural network, and adding a constraint term in the loss function to ensure that the model prediction conforms to the physical law; introducing an adaptive attention mechanism, calculating a space-time node feature importance score, dynamically adjusting the information propagation path and weight, and constructing an underwater stratum space-time evolution model.

[0009] Further, the underwater stratum data deduction method, the joint training strategy combining self-supervised learning and supervised learning optimizes the underwater stratum space-time deduction model parameters, specifically including: The joint training strategy combining self-supervised learning and supervised learning is adopted, the self-supervised learning enhances the understanding of the stratum space-time characteristics of the underwater stratum space-time deduction model through the space-time comparison learning task; The supervised learning optimizes the underwater stratum space-time deduction model parameters by using the stratum evolution history data, and adjusts the neural network weight and bias by using the stochastic gradient descent and its variant algorithm.

[0010] Further, the underwater stratum data deduction method, the construction of the hierarchical space adjacency graph further includes: The edge weight is calculated by using the improved Gaussian kernel function, and the formula is

[0011] wherein is the edge weight of the node and the node , is the distance between nodes, is the scale parameter, is the attribute similarity of the node and the node .

[0012] Further, the underwater stratum data deduction method, the embedding of the physical field model into the multi-scale space-time convolutional neural network specifically includes: In terms of deposition, the deposition rate is calculated by formula according to the flow velocity, sediment particle size and density parameters, and is added as a constraint term to the model loss function; In terms of tectonic movement, the stress field distribution is simulated by using the theory of elasticity and plate motion, and a penalty term for the difference between the stress calculated based on the physical model and the stress predicted by the model is added to the loss function.

[0013] According to the second aspect of the present application, an underwater stratum data deduction device is also provided, comprising: A data acquisition module is configured to acquire multi-source data of an underwater stratum, pre-process the acquired multi-source data, and construct a multi-modal data set; A feature extraction module is configured to acquire a hierarchical space adjacency graph and a multi-scale time adjacency graph based on the multi-modal data set, and construct a multi-scale space-time coupled graph; A model construction module is configured to fuse and extract multi-scale space-time features of the multi-scale space-time coupled graph based on a multi-scale space-time convolutional neural network, embed a physical field model into the multi-scale space-time convolutional neural network in the form of a model loss function, and construct an underwater stratum space-time deduction model; The parameter optimization module is used to optimize the parameters of the underwater stratum spatiotemporal extrapolation model based on a joint training strategy that combines self-supervised learning and supervised learning. The prediction module is used to input the current underwater strata state and the marine environment prediction data for a future period into the optimized underwater strata spatiotemporal inference model, and output the strata evolution prediction results at different time scales.

[0014] According to a third aspect of this application, an underwater stratigraphic data extrapolation device is also provided, comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of any of the methods described above.

[0015] According to a fourth aspect of this application, a storage medium is also provided, which stores a computer program executable by an underwater stratigraphic data extrapolation device, which, when run on the underwater stratigraphic data extrapolation device, causes the underwater stratigraphic data extrapolation device to perform the steps of any of the methods described above.

[0016] Overall, compared with the prior art, the above-described technical solutions conceived in this application can achieve the following beneficial effects: The underwater stratigraphic data extrapolation method provided in this application, through multi-source data fusion, multi-scale spatiotemporal map construction, physical field constraint embedding, and joint training strategies, can achieve high-precision spatiotemporal evolution prediction of underwater strata. Multi-source data fusion preprocessing constructs a multimodal dataset, providing a rich and accurate data foundation for model training; multi-scale spatiotemporal coupling map construction comprehensively expresses the spatiotemporal correlation of strata; embedding the physical field model into a multi-scale spatiotemporal convolutional neural network and introducing an adaptive attention mechanism ensures that the prediction results conform to geophysical laws and focus on key areas and time periods; a joint training strategy combining self-supervised and supervised learning is used to optimize model parameters. The final output prediction results are reliable and can provide strong technical support for marine resource development and cross-sea infrastructure construction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying 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.

[0018] Figure 1 This is a flowchart illustrating a method for extrapolating underwater stratigraphic data, provided in an embodiment of this application. Detailed Implementation

[0019] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0020] The terms "first", "second", "third", etc. in the specification and claims of the present application and the above-described drawings are used to distinguish different objects, and are not used to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0021] Figure 1 A flowchart of an underwater stratum data deduction method provided by an embodiment of the present application is shown in FIG. 1. The underwater stratum data deduction method provided by the embodiment of the present application includes the following steps. Figure 1 Obtaining multi-source data of an underwater stratum, pre-processing the obtained multi-source data, and constructing a multi-modal data set; Based on the multi-modal data set, obtaining a layered space adjacency graph and a multi-scale time adjacency graph, and constructing a multi-scale space-time coupling graph; Based on a multi-scale space-time convolutional neural network, fusing and extracting multi-scale space-time features of the multi-scale space-time coupling graph, embedding a physical field model in the form of a model loss function into the multi-scale space-time convolutional neural network, and constructing an underwater stratum space-time deduction model; Based on a joint training strategy combining self-supervised learning and supervised learning, optimizing parameters of the underwater stratum space-time deduction model; Inputting current underwater stratum state and marine environment prediction data for a future period of time into the optimized underwater stratum space-time deduction model, and outputting stratum evolution prediction results at different time scales.

[0022] ​Specifically, first, multi-source data of the underwater formation is acquired, which is widely sourced, including but not limited to marine seismic data, logging-while-drilling time series data, seafloor topography scanning data, seafloor formation scanning data, and marine environment monitoring data. Marine seismic data can provide macro information of the formation structure, and logging-while-drilling data contains physical property parameters of the formation, such as resistivity, natural gamma, etc. Seafloor topography scanning data and formation scanning data provide morphological information of the seafloor surface and the formation interior, respectively. Marine environment monitoring data covers dynamic parameters such as current velocity, temperature, salinity, etc., which have important influence on the deposition and erosion process of the formation.

[0023] The acquired multi-source data is preprocessed, and the preprocessing process includes data cleaning, correction and fusion. Data cleaning aims to remove noise and outliers, for example, wavelet transform filtering technology can be used to remove noise from seismic data, and phase correction algorithm is used to improve data resolution. The logging data is normalized to eliminate the bias caused by different measuring equipment. Data fusion unifies different sources of data into the same spatio-temporal coordinate system, and is realized by using the spatio-temporal alignment algorithm. For the problem of data missing, a completion algorithm combining variational autoencoder (VAE) and generative adversarial network (GAN) is used to generate reasonable pseudo data according to the spatio-temporal correlation of the data, so as to build a complete and accurate multi-modal data set.

[0024] Based on the preprocessed multi-modal data set, a hierarchical spatial adjacency graph and a multi-scale temporal adjacency graph are constructed, and then a multi-scale spatio-temporal coupling graph is formed. In the spatial dimension, the underwater formation space is divided into different scale voxel units, for example, larger voxel units can be divided in the macro scale to capture the large-scale structural features of the formation, and smaller voxel units can be divided in the micro scale to focus on the details of the formation. The connection relationship and edge weight between nodes are determined according to the geological structure characteristics, formation property similarity and spatial distance of different scale voxel units. The improved Gaussian kernel function can be applied to the calculation of edge weight to highlight the connection strength between nodes with high similarity and short distance.

[0025] In the time dimension, the nodes at the same spatial position in different periods are connected with a set time interval as a time slice. The determination of the time edge weight is dynamically adjusted according to the length of the time interval and the change rate of the marine environment parameters. For example, if the current velocity changes dramatically in a certain time period, it may have a large erosion or deposition effect on the formation, and the weight of the corresponding time edge will be reduced accordingly to reflect the increase of the uncertainty of the formation change in the time interval. By fusing the spatial adjacency graph and the temporal adjacency graph, a multi-scale spatio-temporal coupling graph that can comprehensively express the spatio-temporal correlation of the formation is finally constructed, providing a data basis for subsequent feature extraction and model training.

[0026] The multi-scale spatio-temporal features of the multi-scale spatio-temporal coupled graph are fused and extracted based on a multi-scale spatio-temporal convolutional neural network. In the spatial dimension, different size convolution kernels are designed to act on different scale spatial adjacency graphs to extract the spatial structure features of the strata from macro to micro. The large convolution kernel can capture the overall trend of the regional strata distribution, while the small convolution kernel can focus on the detailed changes of the local strata. In the time dimension, the recurrent neural network (RNN) variant is combined with the time adjacency graph to learn the dynamic change pattern of the strata over time, thereby realizing the deep fusion and extraction of multi-scale spatio-temporal features.

[0027] In order to embed the physical field model into the multi-scale spatio-temporal convolutional neural network, the soft constraint form is adopted to integrate it into the loss function of the model. For example, in the aspect of sedimentation, the sedimentation rate is calculated through physical formula according to parameters such as flow velocity, sediment particle size and density, and the calculated sedimentation rate is added as a constraint term to the loss function. For tectonic movement, the stress field distribution is simulated by using the theory of elasticity and plate motion, and then the difference between the stress calculated based on the physical model and the stress predicted by the model is added as a penalty term in the loss function. In this way, during the model training process, by optimizing the loss function, it can ensure that the model predicted strata evolution result conforms to the physical laws such as sediment dynamics and structural geology.

[0028] At the same time, the adaptive attention mechanism is introduced, so that the model can automatically focus on the key areas and time periods that have greater impact on strata evolution. In the analysis process, the model will calculate the importance scores of the features of different spatio-temporal nodes, and dynamically adjust the path and weight of information propagation according to the scores. For example, in the tectonically active area or rapid deposition period, the model will increase the attention weight of the features of the corresponding area and time slice, so as to improve the extraction ability and prediction accuracy of the model for complex strata evolution characteristics.

[0029] A joint training strategy combining self-supervised learning and supervised learning is adopted to optimize the parameters of the underwater strata spatio-temporal inference model. In the self-supervised learning stage, a spatio-temporal contrast learning task is designed. For example, given the states of the same strata region at different times, the model needs to judge whether these states belong to the same evolution sequence. By maximizing the similarity between positive sample pairs (i.e. the real evolution states of the same region at different times) and minimizing the similarity between negative sample pairs (i.e. randomly selected unrelated states), the model can better understand the spatio-temporal features and evolution laws of the strata.

[0030] In the supervised learning phase, known stratigraphic evolution history data is used as labels, which can be obtained from historical seismic data, drilling core analysis results, etc. The error between the model prediction result and the true label is calculated, and the weights and biases of the neural network are adjusted using stochastic gradient descent (SGD) and its variants to further optimize the model parameters. The joint training strategy fully takes advantage of self-supervised learning and supervised learning, enabling the model to effectively learn the characteristics and patterns of stratigraphic evolution even in the absence of large amounts of labeled data, and continuously improve the accuracy of the prediction.

[0031] The current underwater stratigraphic state and future marine environment prediction data for a period of time are input into the optimized underwater stratigraphic space-time inference model. Marine environment prediction data can be obtained based on global climate change models, including sea level rise rate, sea current pattern changes, etc. The model outputs stratigraphic evolution prediction results at different time scales in the future, covering changes in stratigraphic interfaces, changes in lithology distribution, fault activity, etc.

[0032] To evaluate the reliability of the prediction results, various indicators can be used for evaluation. The root mean square error can be used to measure the error of stratigraphic thickness prediction, and the intersection over union can be used to evaluate the accuracy of lithology distribution prediction. In addition, qualitative evaluation of the prediction results is based on physical rationality, checking whether it conforms to geological common sense and known physical laws. By comprehensively using these evaluation indicators, the performance of the model and the credibility of the prediction results can be verified comprehensively and objectively, providing reliable basis for practical applications such as marine resource development and cross-sea facility construction.

[0033] The underwater stratigraphic data inference method provided by the embodiments of the present application can realize high-precision space-time evolution prediction of underwater strata through multi-source data fusion, multi-scale space-time graph construction, physical field constraint embedding, and joint training strategy, etc. Multi-source data fusion preprocessing constructs a multi-modal data set, providing a rich and accurate data basis for model training; multi-scale space-time coupling graph construction fully expresses the space-time correlation of strata; the physical field model is embedded in the multi-scale space-time convolutional neural network and the adaptive attention mechanism is introduced to ensure that the prediction results conform to the geological and physical laws and focus on key areas and time periods; the joint training strategy combining self-supervised and supervised learning is used to optimize the model parameters. The final output prediction result is reliable and can provide strong technical support for marine resource development, cross-sea facility construction, etc.

[0034] Optionally, the underwater stratigraphic data inference method provided by the embodiments of the present application comprises: acquire multi-source data of the underwater formation, including marine seismic data, logging while drilling time series data, seafloor topography scanning data, seafloor formation scanning data and marine environment monitoring data; unify the multi-source data of the underwater formation to the same space-time coordinate system by using a space-time alignment algorithm; process the missing data based on a completion algorithm combined with a space-time variational autoencoder and a generative adversarial network, and construct a multi-modal data set.

[0035] Specifically, multi-source data of the underwater formation is collected, including marine seismic data, logging while drilling time series data, seafloor topography scanning data, seafloor formation scanning data and marine environment monitoring data. Marine seismic data can provide macro-structure information of the formation, and logging while drilling data contains physical property parameters of the formation, such as resistivity, natural gamma, etc. Seafloor topography scanning data and formation scanning data provide morphological information of the seafloor surface and the formation interior, respectively, while marine environment monitoring data covers dynamic parameters such as current velocity, temperature, salinity, which have important influence on the deposition and erosion process of the formation.

[0036] Next, the multi-source data is unified to the same space-time coordinate system by using a space-time alignment algorithm. The space-time alignment algorithm can process data of different sources and different scales, ensuring the consistency of all data in time and space, and providing a basis for subsequent data fusion and analysis.

[0037] Finally, for the missing parts that may exist in the data, a completion algorithm based on a combination of a space-time variational autoencoder and a generative adversarial network is used for processing. The algorithm uses the space-time correlation of the data to generate reasonable pseudo-data to fill in the missing parts, thereby constructing a complete and accurate multi-modal data set. In this way, the integrity and accuracy of the data can be ensured, and a high-quality data basis is provided for subsequent multi-scale space-time graph construction and model training.

[0038] Optionally, the underwater formation data deduction method provided by the embodiment of the application comprises: dividing the underwater formation space into different scale voxel units, determining node connection and edge weight based on regional geological structure characteristics, formation property similarity and distance, and constructing a hierarchical spatial adjacency graph; connecting nodes of the same spatial position but different periods based on time slices, adjusting time edge weight according to time interval and marine environment parameter change rate, and constructing a multi-scale time adjacency graph; fusing the spatial and time adjacency graphs to obtain a multi-scale space-time coupled graph.

[0039] Specifically, the underwater stratigraphic space is divided into voxel units of different scales, such as macroscopic, mesoscopic, and microscopic scales. For each scale, the connectivity and edge weights between nodes are determined based on the region's geological structural characteristics, stratigraphic property similarity, and spatial distance. An improved Gaussian kernel function is used to calculate edge weights, highlighting the connection strength between nodes with high similarity and close proximity. At the macroscopic scale, nodes are divided based on the macroscopic similarity of the regional stratigraphic distribution; at the mesoscopic scale, node connections are determined by combining lithology and stratigraphic structure characteristics; at the microscopic scale, adjacency relationships are constructed based on the similarity of microscopic properties such as lithology and porosity, as well as spatial distance, ultimately constructing a layered spatial adjacency graph.

[0040] A multi-scale temporal adjacency graph is constructed by connecting nodes at the same spatial location in different periods using time slices. The weights of time edges are dynamically adjusted based on the time interval and the rate of change of marine environmental parameters. For example, when marine environmental parameters change drastically within a certain time period, the weight of the corresponding time edge will be reduced to reflect the increased uncertainty of stratigraphic changes within that time interval.

[0041] By fusing spatial and temporal adjacency graphs, a multi-scale spatiotemporal coupled graph is formed. This graph can comprehensively express the variation characteristics of strata at different spatial and temporal scales. Each node contains rich stratigraphic attributes (lithology, permeability, stratigraphic thickness, etc.) and timestamp information, while the edges clearly express the spatiotemporal relationships at different scales, providing a data foundation for detailed analysis in spatial and temporal dimensions.

[0042] Optionally, the underwater stratigraphic data extrapolation method provided in this application embodiment, wherein embedding the physical field model into the multi-scale spatiotemporal convolutional neural network in the form of a model loss function to construct an underwater stratigraphic spatiotemporal extrapolation model specifically includes: Based on a multi-scale spatiotemporal convolutional neural network, spatial features are extracted by convolutional kernels of different sizes, and temporal variation patterns are learned by combining a variant of a recurrent neural network with a multi-scale temporal adjacency graph. Embedding a physics field model into a multi-scale spatiotemporal convolutional neural network, and adding a constraint term to the loss function to ensure that the model predictions conform to physical laws; An adaptive attention mechanism is introduced to calculate the importance score of spatiotemporal node features, dynamically adjust the information propagation path and weight, and construct a spatiotemporal inference model of underwater strata.

[0043] Specifically, a multi-scale spatiotemporal convolutional neural network is used to process multi-scale spatiotemporal coupled graph data. In the spatial dimension, convolutional kernels of different sizes are used to extract spatial features at each scale. Large convolutional kernels grasp the overall trend of the strata, while small convolutional kernels focus on details, capturing the spatial structural features of the strata from macroscopic to microscopic. In the temporal dimension, a variant of the recurrent neural network is combined with a multi-scale temporal adjacency graph to learn the dynamic change patterns of the strata, achieving deep fusion and extraction of multi-scale spatiotemporal features.

[0044] The deposition dynamics, physical field models such as structural geology, and the like are embedded in the neural network in the form of soft constraints, and a constraint term is added in the loss function. In terms of deposition, the deposition rate is calculated according to the flow velocity, sediment particle size density and the like, and is set as a constraint. The deposition rate is calculated through a physical formula, and is added as a constraint term to the model loss function, so as to ensure that the thickness change of the deposition layer predicted by the model conforms to the actual physical law; in terms of tectonic movement, the stress field distribution is simulated by using the theory of elasticity and plate motion, and the displacement and deformation prediction of the stratum near the fault are constrained. A penalty term based on the difference between the stress calculated by the physical model and the stress predicted by the model is added in the loss function, so that the model continuously adjusts the parameters during the learning process to meet the physical field constraints.

[0045] An adaptive attention mechanism is introduced, and the model automatically focuses on key areas and time periods that have a greater impact on stratum evolution when processing multi-scale spatio-temporal data. When analyzing the tectonically active area, the model increases the attention weight of the spatio-temporal features of the area through the attention mechanism; when studying the period of rapid deposition, the features of the corresponding time slice are highlighted. By calculating the importance score of the spatio-temporal node features, the information propagation path and weight are dynamically adjusted, and the extraction ability of the model for complex stratum evolution features is improved. The finally constructed underwater stratum spatio-temporal inference model can accurately predict stratum evolution.

[0046] Optionally, the underwater stratum data inference method provided in the embodiments of the present application, the joint training strategy combining self-supervised learning and supervised learning is used to optimize the parameters of the underwater stratum spatio-temporal inference model, and specifically includes: The joint training strategy combining self-supervised learning and supervised learning is used, the self-supervised learning enhances the understanding of the underwater stratum spatio-temporal inference model for stratum spatio-temporal features through a spatio-temporal contrast learning task; The supervised learning optimizes the parameters of the underwater stratum spatio-temporal inference model by using stratum evolution history data, and adjusts the neural network weight and bias by using a stochastic gradient descent algorithm and its variants.

[0047] Specifically, in the self-supervised learning stage, a spatio-temporal contrast learning task is designed, and data pairs of different periods of the same stratum area are input, and the model needs to determine whether they belong to the same evolution sequence. By maximizing the similarity of positive sample pairs (real evolution sequence) and minimizing the similarity of negative sample pairs (random unrelated sequence), the understanding of the model for stratum spatio-temporal features is enhanced. This process does not depend on labeled data, and can fully utilize a large amount of unlabeled data to learn spatio-temporal features.

[0048] In the supervised learning phase, known historical stratigraphic data are used as labels, derived from historical seismic data and borehole core analysis results. By calculating the error between the model's predictions and the true labels, stochastic gradient descent and its variants (such as Adagrad, Adadelta, and Adam) are used to adjust the weights and biases of the neural network. This supervised learning approach further refines the model's understanding of stratigraphic evolution and improves the accuracy of model predictions.

[0049] Throughout the joint training process, self-supervised learning and supervised learning complement each other. Self-supervised learning extracts spatiotemporal features from unlabeled data, while supervised learning uses labeled data to correct and optimize model parameters. This combination not only makes full use of abundant unlabeled data but also avoids overfitting to some extent, allowing the model to converge continuously during training and ultimately improving prediction accuracy.

[0050] Optionally, the underwater stratigraphic data extrapolation method provided in this application embodiment, which constructs a hierarchical spatial adjacency graph, further includes: The edge weights are calculated using an improved Gaussian kernel function, as shown in the formula below.

[0051] in For nodes With nodes edge weights, The distance between nodes. For scale parameters, For nodes With nodes Attribute similarity.

[0052] Specifically, in this embodiment of the application, an improved Gaussian kernel function is used to calculate edge weights when constructing the hierarchical spatial adjacency graph. The specific formula is as follows:

[0053] in, For nodes With nodes edge weights, For nodes and nodes Spatial distance between them This is a scale parameter used to control the range of influence of distance on the weights. For nodes and nodes The similarity of attributes between them reflects the degree of similarity between the two in terms of stratigraphic attributes (such as lithology, porosity, etc.).

[0054] In the calculation process, the spatial distance between nodes is first determined. which can be obtained by the Euclidean distance or other suitable spatial distance measurement methods. Attribute similarity is calculated based on the stratigraphic properties of the nodes, the more similar the properties are, the higher the value is. The scale parameter is selected according to the stratigraphic characteristics and analysis scale of the specific study area to balance the contribution of distance and attribute similarity to the weight.

[0055] The edge weight calculated by the formula can comprehensively reflect the spatial proximity and attribute similarity between nodes. In the construction of the hierarchical spatial adjacency graph, the connection between nodes with larger edge weights is closer, which helps to more accurately capture the spatial structure characteristics and attribute distribution rules of the strata in the subsequent model training and deduction process, thereby improving the prediction ability of the model for strata evolution.

[0056] Optionally, the underwater strata data deduction method provided by the embodiment of the application, the physical field model is embedded into the multi-scale spatio-temporal convolutional neural network, specifically comprising: In terms of deposition, the deposition rate is calculated by a formula according to the flow velocity, sediment particle size and density parameters, and is added as a constraint term to the model loss function; In terms of tectonic movement, the stress field distribution is simulated by using the theory of elasticity and plate movement, and a penalty term for the difference between the stress calculated based on the physical model and the stress predicted by the model is added to the loss function.

[0057] In a specific embodiment, a kind of underwater strata data deduction method provided by the embodiment of the application is adopted, and the specific execution steps are as follows: Data acquisition and preprocessing: in a target sea area of the Hong Kong-Zhuhai-Macao Bridge, past cross-period seismic data with a resolution of 10m x 10m x 5m covering different geological periods are collected; logging-while-drilling data of 5 exploration wells are selected, including resistivity, natural gamma and other parameters; detailed seabed topographic data of the region are obtained by using a multi-beam depth sounding device; stratigraphic profile data of the region are obtained by using a sub-bottom profiler; long-term marine environmental monitoring data of the region are collected, including sea current velocity, temperature, salinity, etc. Through a spatio-temporal alignment algorithm, these data are unified to the same spatio-temporal coordinate system. For seismic data, wavelet transform filtering is used to remove noise, and a phase correction algorithm is used to improve resolution; logging data are normalized. The topographic data and stratigraphic data are fused, and the water depth value of the topographic data is used to correct the water depth value of the stratigraphic data. For the problem of missing part of the stratigraphic profile data, a completion algorithm combining a spatio-temporal variational autoencoder with a generative adversarial network is used to generate reasonable pseudo data, and a complete multi-modal data set is constructed. Multi-scale spatio-temporal graph construction: In the spatial dimension, the regional strata are divided into macro (1 km x 1 km x 100 m), meso (100 m x 100 m x 10 m), and micro (10 m x 10 m x 1 m) scale voxel units. At the macro scale, based on the regional geological structure characteristics, the areas with similar sedimentary facies are divided into the same node, and the connection between nodes is based on the macro similarity of regional strata distribution; at the meso scale, the node connection is determined by combining lithology and strata structure characteristics; at the micro scale, the adjacency relationship is constructed according to the similarity of micro properties such as lithology and porosity and spatial distance. The weight of the edge is calculated using an improved Gaussian kernel function, highlighting the connection strength of nodes with high similarity and short distance. In the time dimension, 1 year is taken as the time slice, and nodes at the same spatial location in different periods are connected. The time edge weight is dynamically adjusted according to the change rate of marine environmental parameters, such as when the current speed changes more than a certain threshold in a certain time period, the corresponding time edge weight is reduced. By fusing multi-scale spatial adjacency graph and time adjacency graph, a multi-scale spatio-temporal coupled graph is formed. Multi-physical field constrained neural network modeling: A multi-scale spatio-temporal convolutional neural network is constructed. In the spatial dimension, different size convolution kernels are used to extract features from the macro, meso, and micro scale spatial adjacency graphs, capturing the spatial structure characteristics of the strata from the regional to the local. In the time dimension, the gated recurrent unit (GRU) is used to learn the dynamic change pattern of the strata over time in combination with the time adjacency graph. The sediment dynamics and tectonic geology physical field model are embedded in the neural network in the form of soft constraint, and a penalty term based on the difference between the calculated sedimentation rate and tectonic stress and the model prediction results is added to the loss function, ensuring that the model prediction conforms to the physical law. An adaptive attention mechanism is introduced to allow the model to automatically focus on key areas and time periods that have a greater impact on strata evolution, such as areas near active faults and rapid deposition periods. Model training and optimization: A joint training strategy combining self-supervised and supervised learning is adopted. The spatio-temporal contrast learning task is designed, such as given the state of the same strata region at different times, the model needs to determine whether it belongs to the same evolution sequence, by maximizing the similarity of positive samples and minimizing the similarity of negative samples, the model's understanding of strata spatio-temporal features is enhanced. Supervised learning uses the known strata evolution history data (obtained through historical seismic data and drilling core analysis results) of the region as labels, and further optimizes the model parameters by minimizing the error between the predicted results and the true labels. The Adam optimization algorithm is used to adjust the weights and biases of the neural network, allowing the model to continuously converge during training. Stratigraphic evolution deduction and result evaluation: input the trained model into the current stratigraphic state of the region and the future 10-year marine environment prediction data (sea level rise rate predicted by global climate change model, sea current mode change, etc.), the model outputs the stratigraphic evolution prediction results at different time scales, including stratigraphic interface changes, lithology distribution changes, fault activity, etc. The root mean square error is used to measure the stratigraphic thickness prediction error, and the intersection ratio is used to evaluate the accuracy of the lithology distribution prediction. At the same time, geology experts are invited to qualitatively evaluate the prediction results based on physical rationality. Compared with the actual observation data and expert evaluation results, the reliability of the method of the application in the prediction of stratigraphic evolution in the region is improved by 53% compared with the traditional single prediction method.

[0058] Optionally, the embodiment of the application further provides an underwater stratigraphic data deduction device, comprising: A data acquisition module is configured to acquire multi-source data of an underwater stratum, pre-process the acquired multi-source data, and construct a multi-modal data set. A feature extraction module is configured to acquire a hierarchical spatial adjacency graph and a multi-scale time adjacency graph based on the multi-modal data set, and construct a multi-scale spatiotemporal coupling graph. A model construction module is configured to fuse and extract multi-scale spatiotemporal features of the multi-scale spatiotemporal coupling graph based on a multi-scale spatiotemporal convolutional neural network, embed a physical field model in the multi-scale spatiotemporal convolutional neural network in the form of a model loss function, and construct an underwater stratigraphic spatiotemporal deduction model. A parameter optimization module is configured to optimize parameters of the underwater stratigraphic spatiotemporal deduction model based on a joint training strategy combining self-supervised learning and supervised learning. A prediction module is configured to input a current underwater stratigraphic state and marine environment prediction data for a period of time into the optimized underwater stratigraphic spatiotemporal deduction model, and output stratigraphic evolution prediction results at different time scales.

[0059] The application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above method. The computer-readable storage medium can include but is not limited to any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, micro-drives, and magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0060] It should be noted that, for the foregoing method embodiments, the sequences of the described actions are not necessarily required to achieve the objects of the application, and certain steps can be performed in other sequences or even concurrently. Additionally, the described embodiments are merely provided as examples, and not all of the actions described are necessarily required to achieve desired results.

[0061] In the above embodiments, the description of each embodiment is focused on different aspects, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0062] In several embodiments provided in the present application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, another division manner can be adopted. For example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some service interfaces, devices or units, and can be electrical or other forms.

[0063] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0064] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0065] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable memory. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a memory and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned memory includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0066] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by a program instructing relevant hardware, and the program can be stored in a computer readable memory, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0067] The above is only exemplary embodiments of the present disclosure, and cannot limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Those skilled in the art will easily think of embodiments of the present disclosure after considering the specification and practicing the disclosure herein. The present application is intended to cover any variations, uses or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or conventional technical means in the technical field not described in the present disclosure. The specification and examples are only considered as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

[0068] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict, they should be considered as within the scope of the present disclosure.

[0069] Those skilled in the art readily understand that the above only describes preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of inferring subsurface formation data, comprising: The method comprises the following steps: acquiring multi-source data of an underwater formation, preprocessing the acquired multi-source data, and constructing a multi-modal data set; acquiring a hierarchical spatial adjacency graph and a multi-scale time adjacency graph based on the multi-modal data set, and constructing a multi-scale spatio-temporal coupling graph; embedding a physical field model in the form of a model loss function into the multi-scale spatio-temporal convolutional neural network based on a multi-scale spatio-temporal convolutional neural network, and constructing an underwater formation spatio-temporal deduction model; optimizing the parameters of the underwater formation spatio-temporal deduction model based on a joint training strategy combining self-supervised learning and supervised learning; inputting the current underwater formation state and future marine environment prediction data into the optimized underwater formation spatio-temporal deduction model, and outputting formation evolution prediction results at different time scales.

2. The method of claim 1, wherein, The method comprises the following steps: acquiring multi-source data of an underwater formation, preprocessing the acquired multi-source data, and constructing a multi-modal data set, specifically comprising: acquiring multi-source data of an underwater formation, including marine seismic data, logging while drilling time series data, seabed topography scanning data, seabed formation scanning data, and marine environment monitoring data; unifying the multi-source data of the underwater formation to the same spatio-temporal coordinate system using a spatio-temporal alignment algorithm; 3. The method of claim 1, wherein, processing missing data based on a completion algorithm combining a spatio-temporal variational autoencoder and a generative adversarial network to construct a multi-modal data set. The method comprises the following steps: acquiring multi-source data of an underwater formation, preprocessing the acquired multi-source data, and constructing a multi-modal data set, specifically comprising: dividing the underwater formation space into different scale voxel units, determining node connections and edge weights based on regional geological structure characteristics, formation property similarity, and distance, and constructing a hierarchical spatial adjacency graph; 4. The method of claim 1, wherein, connecting nodes at the same spatial position but different periods based on time slices, adjusting time edge weights according to time intervals and marine environment parameter change rates, and constructing a multi-scale time adjacency graph; fusing the spatial and time adjacency graphs to obtain a multi-scale spatio-temporal coupling graph. The method comprises the following steps: embedding a physical field model in the form of a model loss function into the multi-scale spatio-temporal convolutional neural network based on a multi-scale spatio-temporal convolutional neural network, and constructing an underwater formation spatio-temporal deduction model, specifically comprising:

5. The method of claim 1, wherein, extracting spatial features through different size convolution kernels based on a multi-scale spatio-temporal convolutional neural network, and learning time variation patterns using a recurrent neural network variant combined with a multi-scale time adjacency graph; embedding a physical field model into a multi-scale spatio-temporal convolutional neural network, and adding a constraint term in the loss function to ensure that the model prediction conforms to the physical law; introducing an adaptive attention mechanism to calculate the importance score of spatio-temporal node features, dynamically adjusting the information propagation path and weight, and constructing an underwater formation spatio-temporal deduction model. The method comprises the following steps: adopting a joint training strategy combining self-supervised and supervised learning, self-supervised learning enhances the understanding of the spatio-temporal features of the underwater formation spatio-temporal deduction model through a spatio-temporal contrast learning task; supervised learning optimizes the parameters of the underwater formation spatio-temporal deduction model using historical formation evolution data, and adjusts the weights and biases of the neural network using a stochastic gradient descent algorithm and its variants.

6. The method of claim 3, wherein, The illustrated hierarchical spatial adjacency graph also includes: The edge weight is calculated by using the improved Gaussian kernel function, and the formula is wherein is a node is a node edge weight of a node is a distance between nodes is a scale parameter is a node attribute similarity of a node attribute similarity of a node 7. The method of claim 4, wherein, The physical field model is embedded in the multi-scale space-time convolutional neural network, and the embedding process includes: In the aspect of deposition, the deposition rate is calculated by formula according to the parameters of water flow velocity, sediment particle size and density, and the deposition rate is added as a constraint term to the model loss function; In the aspect of tectonic movement, the stress field distribution is simulated by using the theory of elasticity and plate motion, and a penalty term based on the difference between the stress calculated by the physical model and the stress predicted by the model is added to the loss function.

8. An apparatus for inferring data of a subterranean formation, comprising: The method comprises the steps of: The data acquisition module is used for acquiring multi-source data of the underwater stratum, pre-processing the acquired multi-source data, and constructing a multi-modal data set; The feature extraction module is used for acquiring a hierarchical spatial adjacency graph and a multi-scale time adjacency graph based on the multi-modal data set, and constructing a multi-scale space-time coupled graph; The model construction module is used for fusing and extracting multi-scale space-time features of the multi-scale space-time coupled graph based on a multi-scale space-time convolutional neural network, embedding a physical field model in the multi-scale space-time convolutional neural network in the form of a model loss function, and constructing an underwater stratum space-time deduction model; The parameter optimization module is used for optimizing the parameters of the underwater stratum space-time deduction model based on a joint training strategy combining self-supervised learning and supervised learning. The prediction module is used for inputting the current underwater stratum state and future marine environment prediction data into the optimized underwater stratum space-time deduction model, and outputting stratum evolution prediction results at different time scales.

9. An underwater formation data extrapolation device, characterized by, The computer program is stored in the storage unit and can be executed by the underwater stratum data deduction device, and when the computer program runs on the underwater stratum data deduction device, the underwater stratum data deduction device executes the steps of the method of any one of claims 1-7.

10. A storage medium, characterized by The computer program is stored in the storage unit and can be executed by the underwater stratum data deduction device, and when the computer program runs on the underwater stratum data deduction device, the underwater stratum data deduction device executes the steps of the method of any one of claims 1-7.

Images