A method for predicting seismic facies and geological time and fault distance by prior construction constraint

Abstract

The application provides a method for predicting seismic facies relative geologic time and fault throw by constructing constraints in advance, comprising the following steps: constructing a synthetic seismic data volume with labels of relative geologic time, fault position and fault throw, calculating unconformity surface attributes and extracting sparse horizon data; constructing a multi-scale attention joint prediction network containing four parallel encoders, an adaptive feature fusion module and two parallel decoders; inputting seismic data, fault position, unconformity surface attributes and sparse horizon data into corresponding encoders to extract multi-source structural prior features and perform adaptive weighted fusion; outputting prediction results of relative geologic time and fault throw through parallel output of double decoders; increasing consistency constraint loss between prediction drop of relative geologic time and prediction value of fault throw in the fault neighborhood during training, and giving a higher weight; calculating a fault probability volume, an unconformity surface attribute volume and a sparse horizon surface for field three-dimensional seismic data, inputting the trained model, and outputting a relative geologic time volume and a fault throw volume.

Description

Technical Field

[0001] This invention relates to the fields of intelligent interpretation of seismic data and geophysical artificial intelligence technology, and more specifically, to a method for predicting the relative geological age and fault displacement of earthquakes based on prior tectonic constraints. Background Technology

[0002] Relative geological time is an important geological attribute characterizing the chronological order of stratigraphic deposition and sequence stratigraphic relationships, playing a crucial role in stratigraphic tracing, sequence stratigraphy analysis, stratigraphic slice construction, and tectonic reconstruction. Fault displacement is a key parameter describing the relative displacement or fault strength between the two sides of a fault, and is of great significance in quantitative fault interpretation, fault sealing assessment, and tectonic evolution analysis.

[0003] In existing technologies, the construction of relative geochronology and the interpretation of fault displacement typically employ separate processes: relative geochronology is often obtained through stratigraphic tracing, structural smoothing, or dip-constrained tracing, while fault displacement is often obtained through fault plane interpretation and manual measurement or semi-automatic estimation. This approach has the following drawbacks: relative geochronology and fault displacement are usually processed separately, failing to fully utilize their structural correspondence within the fault's vicinity; manual interpretation is labor-intensive and inefficient, and the interpretation results are heavily influenced by the interpreter's experience; while purely data-driven methods improve automation, the lack of prior structural constraints can easily lead to geologically inaccurate predictions.

[0004] With the development of deep learning technology in seismic interpretation, existing methods can model fault identification, horizon tracking, or relative geological age prediction separately. However, existing schemes are usually designed for single tasks and lack a joint interpretation scheme that integrates prior information such as fault location, unconformity properties, and sparse horizons into the input and explicitly incorporates the relative geological age difference-fault displacement consistency constraint into the output. Therefore, there is an urgent need for a joint prediction method for relative geological age and fault displacement to simultaneously improve interpretation accuracy, structural consistency, and automation. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a method for predicting relative geological age and fault displacement based on prior structural constraints. Specifically, it utilizes synthetic seismic data to construct training samples, introduces prior knowledge of fault location, unconformity surfaces, and sparse stratigraphic positions, and incorporates geological knowledge priors into the relationship between relative geological age and fault displacement at the output end. Finally, it employs a multi-scale attention neural network to jointly predict the relative geological age and fault displacement using an intelligent interpretation method. This addresses the problems of step-by-step interpretation of relative geological age and fault displacement, insufficient utilization of structural priors, low interpretation efficiency, and poor geological consistency in complex fault zones found in existing technologies.

[0006] This invention constructs a labeled synthetic seismic data volume, introduces prior structural information such as fault location, unconformity surface attributes, and sparse horizons at the input end, and introduces consistency constraint loss of relative geological age prediction difference—fault displacement prediction value within the fault neighborhood at the output end, thereby achieving joint intelligent prediction of relative geological age and fault displacement.

[0007] This invention is achieved through the following technical solution: a method for predicting the relative geological age and fault displacement of earthquakes based on prior tectonic constraints, specifically including the following steps: Step S1: Construct labeled synthetic earthquake training samples and extract tectonic prior information, including constructing a synthetic earthquake data volume with relative geological age labels, fault location labels and fault displacement labels, calculating unconformity surface properties and extracting sparse stratigraphic data, and performing standardized preprocessing on the input data. Step S2: Construct a joint prediction neural network model based on a multi-scale attention mechanism. The neural network model includes four parallel encoders, an adaptive feature fusion module, and two parallel decoders. Step S3: Establish prior structural constraints at the input end. Input the synthetic seismic data, fault location labels, unconformity surface attributes, and sparse stratigraphic data into the corresponding parallel encoders to extract multi-source structural prior features. Step S4: Perform adaptive feature fusion by inputting multi-branch, multi-scale features into the fusion module and performing weighted fusion through attention weights to generate a fused feature representation; Step S5: Perform dual-task parallel decoding prediction and output the relative geological age prediction result and the fault displacement prediction result respectively; Step S6: Establish prior constraints for the output end. In each iteration of model training, add consistency constraint loss between the predicted relative geological age drop and the predicted fault displacement within the fault neighborhood to the supervised loss, and assign weights to them. Step S7: Perform iterative training and convergence determination of the model, and backpropagate and update parameters based on the total loss function until the model converges; Step S8: Calculate the fault probability volume and unconformity surface attribute volume from the on-site 3D seismic data and obtain the sparse bedding plane. Input the trained model and output the relative geological time volume and fault displacement volume to achieve joint intelligent interpretation.

[0008] As a preferred option, step S1 specifically includes the following steps: Step S11: Construct a synthetic seismic data volume using geological structure modeling and seismic forward modeling methods. The synthetic seismic data volume includes multiple stratigraphic groups, multiple fault structural units, and unconformity surface structures, as well as spatially corresponding relative geological age labels, fault location labels, and fault displacement labels. The fault displacement labels are calculated by the difference between the corresponding layers on both sides of the fault in the relative geological age domain. This difference can be defined along the fault normal direction, the vertical direction, or a combination thereof. Step S12: Calculate the unconformity attribute volume based on the synthetic seismic data volume. The unconformity attribute volume is obtained by attribute algorithms including reflection discontinuity, structural abrupt response, and sequence boundary enhancement. Extract sparse stratigraphic data based on the relative geological age tag volume. The sparse stratigraphic data is extracted by selecting key stratigraphic intervals according to preset stratigraphic intervals, performing sparse sampling according to preset sampling ratios, and extracting representative stratigraphic intervals according to preset geological stratigraphic positions. The sparse stratigraphic data is represented as sparse stratigraphic points, stratigraphic lines, or stratigraphic surfaces. Step S13: Standardization preprocessing and spatial alignment; Standardization preprocessing and spatial alignment processing are performed on synthetic seismic data, unconformity surface attribute volumes and sparse layer data. Standardization preprocessing includes at least one of amplitude normalization, numerical range scaling, outlier suppression and missing value imputation to form model training input data.

[0009] As a preferred option, step S2 specifically includes the following steps: Step S21: Construct the overall structure of the joint prediction network; Construct a neural network model for joint prediction of relative geological age and fault displacement. The neural network model includes four parallel encoders, an adaptive feature fusion module and two parallel decoders. The neural network model adopts a "multi-branch encoding-fusion-dual decoding" structure, which enables the fusion of prior information from different sources in the feature layer and realizes joint prediction of dual tasks at the output end. Step S22: Construct four parallel encoders; the four parallel encoders correspond to the seismic data encoder, fault location encoder, unconformity surface attribute encoder and sparse layer encoder, respectively. Each encoder extracts feature representations at at least two scales. The extracted features include shallow texture features, middle-layer structural features and deep semantic features. Step S23: Construct two parallel decoders; the two parallel decoders are a relative geological age decoder and a fault displacement decoder. The relative geological age decoder outputs a relative geological age prediction volume, and the fault displacement decoder outputs a fault displacement prediction volume. The output spatial sampling of both is consistent with the input data.

[0010] As a preferred option, step S3 specifically includes the following steps: Step S31: Seismic data branch encoding; Input the synthetic seismic data obtained in step S1 into the seismic data encoder to extract multi-scale seismic features related to reflection phase axis morphology, amplitude texture, and fracture response.

[0011] Step S32: Prior data branch encoding. Input the fault location label volume into the fault location encoder to extract prior features of fault geometric location; input the unconformity surface attribute volume into the unconformity surface attribute encoder to extract prior features of sequence boundary and discontinuity boundary; input the sparse stratigraphic data into the sparse stratigraphic encoder to extract prior features of stratigraphic control. Step S33: Multi-branch feature scale alignment; Spatial scale alignment and feature registration are performed on the multi-scale features output by each branch encoder to form a multi-source prior construction feature set at the input end.

[0012] As a preferred option, step S4 specifically includes the following steps: Step S41: Input fusion module; input the multi-source prior structural feature set formed in step S3 into the adaptive feature fusion module; the fusion module is used to aggregate multi-branch information from seismic data, fault location, unconformity surface attributes and sparse horizons; Step S42: Attention-weighted fusion; The adaptive feature fusion module uses a multi-scale attention mechanism to weight and label features from different sources and at different scales, generating a unified fused feature representation; The multi-scale attention mechanism is channel attention, spatial attention, or a combination of both, and the feature weights of each branch can be adaptively updated with training iterations.

[0013] As a preferred option, step S5 specifically includes the following steps: Step S51: Input the fused feature representation obtained in step S4 into the relative geological age decoder, and output the relative geological age prediction volume. The relative geological age prediction volume can be a continuous value prediction result or a normalized ordinal value prediction result; Step S52: Input the fused feature representation obtained in step S4 into the fault displacement decoder and output the fault displacement prediction volume. The fault displacement prediction volume is used to characterize the relative displacement difference or equivalent displacement between the two strata on the fault sides. As a preferred option, step S6 specifically includes the following steps: Step S61: Calculate the supervised loss; in each iteration of model training, calculate: the first supervised loss L between the relative geological time prediction body and the relative geological time label body. rgt The second supervisory loss L between the fault displacement prediction body and the fault displacement label body throw ; First supervisory loss L rgt Defined as L ssim For structural similarity index loss, , Where x represents the relative geological age predicted by the network, and y represents the relative geological age label; μ x σ and μᵧ represent the mean values ​​of x and y within the local window, respectively; x ² and σᵧ² represent the variances of x and y within the local window, respectively; σ x ᵧ represents the covariance of x and y within the local window; C1 and C2 are stability constants; L mae For the average absolute error loss,

[0014] Among them, L mae denoted as mean absolute error loss, where yᵢ and ŷᵢ are the label value and predicted value at the i-th pixel / voxel, respectively; L normal For normal vector loss, , In the normal vector loss L_normal, n̂ᵢ represents the normal vector calculated from the prediction result at the i-th pixel / voxel, nᵢ represents the normal vector calculated from the label, and the symbol "·" represents the vector dot product; First supervisory loss L throw Defined as

[0015] Where fᵢ represents the discontinuity label value at the i-th pixel / voxel, f̂ᵢ represents the discontinuity prediction value at the i-th pixel / voxel; |·| represents absolute value operation; Step S62: Calculate the fault neighborhood consistency constraint loss; within the fault neighborhood indicated by the fault location label, extract the relative geological age prediction values ​​of the corresponding positions on both sides of the fault, and calculate the difference between the prediction values ​​on both sides as the age difference; measure the difference between the age difference and the fault displacement prediction value at the same position, and construct the mean square error consistency constraint loss. Step S63: Construct the total loss function; assign a weight higher than at least one supervision loss in step S61 to the consistency constraint loss in step S62 to construct the total loss function, thereby establishing prior constraints for the output; the total loss function is expressed as: L total = α·L rgt + β·L throw + γ·L cons Among them, L rgt For relative geological age monitoring loss, L throw For monitoring the loss of fault displacement, L consThe consistency constraint loss is denoted by α, β, and γ, which are weighting coefficients, and γ is greater than α or β. Among them, L cons Defined as , N f This represents the number of fault sampling points; , These represent the values ​​of the predicted relative geological time body at the corresponding positions along both sides of the fault at the i-th fault sampling point; This indicates the fault displacement at that point relative to the predicted geological age. This represents the predicted dislocation value at that point; |·| represents the absolute value operation.

[0016] Furthermore, the corresponding positions on both sides of the fault are determined by any of the following methods: (1) matching corresponding points within a preset window along the fault normal direction; (2) matching corresponding points in the fault neighborhood using the nearest neighbor method; (3) matching the corresponding positions of the strata on both sides of the fault using a local search method. As a preferred option, step S7 specifically includes the following steps: Step S71: Iterative Training and Parameter Update. Backpropagation and parameter updates are performed based on the total loss function constructed in step S63. Iterative training is used to optimize the joint prediction neural network model; at least one of the following strategies is employed during training: learning rate decay, gradient pruning, or early stopping. Step S72: Convergence determination; As the number of iterations increases, convergence determination is performed based on at least one of the following: changes in the total loss function, validation data error, or preset training rounds; When the preset convergence condition is met, training is stopped and the model parameters after training are output.

[0017] As a preferred option, step S8 specifically includes the following steps: Step S81: Construct field data input; calculate fault probability volume and unconformity surface attribute volume for field 3D seismic data volume, and obtain sparse layer planes; the fault probability volume is obtained by a fault identification method based on deep learning or a fault detection method based on seismic attributes; the sparse layer planes are obtained by manual interpretation, semi-automatic picking or automatic tracking methods; Step S82: Field data model inference; Input the field 3D seismic data volume, fault probability volume, unconformity surface attribute volume, and sparse bedding plane into the trained joint prediction neural network model in the same data organization form as in the training stage, and perform forward inference; Step S83: Output the joint interpretation results; output the relative geological age and fault displacement of the field data to achieve integrated joint intelligent interpretation of relative geological age and fault displacement.

[0018] By employing the above technical solutions, this invention has the following beneficial effects compared to existing technologies: 1. Strong joint prediction capability: Simultaneously predicts relative geological age and fault displacement within the same network framework, reducing the propagation of errors from step-by-step interpretation; 2. Full utilization of prior information: Introduce prior information such as fault location, unconformity surface attributes, and sparse stratigraphic positions at the input end to enhance the model's ability to express features of complex tectonic regions; 3. Higher consistency with geological structure: The consistency constraint between the predicted value of the displacement on both sides of the fault and the fault displacement is introduced at the output end, so that the prediction results are more consistent with the geological structure. 4. Reduce reliance on manual intervention: By combining supervised training with in-situ data inference, the reliance on large-scale manually labeled samples is reduced; 5. Wide range of applications: It can be used for scenarios such as relative geochronology construction in fault-developed areas, fault displacement estimation, stratigraphic slice interpretation, and structural analysis.

[0019] Additional aspects and advantages of the invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0020] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is the overall flowchart of the present invention; Figure 2 (a) is the original seismic profile of the Opunake work area in New Zealand that was entered; Figure 2 (b) is the result of superimposing the fault probability volume extracted in step S81 onto the original seismic profile; Figure 2 (c) shows the relative geological age prediction results output by the model; Figure 2 (d) is the display of the fault displacement prediction results output synchronously by the model superimposed on the original seismic profile. Detailed Implementation

[0021] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0023] The following is combined Figures 1 to 2 The method for predicting seismic relative geological age and fault displacement under prior structural constraints according to embodiments of the present invention is described in detail. In the present invention, "relative geological age" refers to a continuous or discrete sequence value characterizing the relative depositional order of strata; "fault displacement" refers to the relative displacement difference or equivalent displacement characterizing value of strata on both sides of a fault; "unconformity surface attribute" refers to the attribute response characterizing the discontinuity of stratigraphic interfaces, abrupt changes in reflective structures, or sequence boundaries; "sparse stratigraphic data" includes any one or a combination of stratigraphic points, stratigraphic lines, and stratigraphic surfaces.

[0024] like Figure 1 As shown, this invention proposes a method for predicting the relative geological age and fault displacement of earthquakes based on prior tectonic constraints, specifically including the following steps: Step S1: Construct labeled synthetic earthquake training samples and extract prior structural information, including constructing synthetic earthquake data volumes with relative geological age labels, fault location labels, and fault displacement labels; calculating unconformity surface properties and extracting sparse stratigraphic data; and performing standardization preprocessing on the input data; specifically including the following steps: Step S11: A synthetic seismic data volume is constructed using geological structure modeling and seismic forward modeling methods. The synthetic seismic data volume includes multiple stratigraphic groups, multiple fault structural units, and unconformity structures to simulate complex underground structures. It also includes spatially corresponding relative geological age tags, fault location tags, and fault displacement tags; (1) Relative geological age tags are used to characterize the relative depositional sequence of each voxel; (2) Fault location tags are used to mark the fault location or fault neighborhood; (3) Fault displacement tags are used to characterize the displacement difference or equivalent displacement between the two sides of the fault in the relative geological age domain. The fault displacement tags are calculated by the difference between the corresponding strata on both sides of the fault in the relative geological age domain. This difference can be defined along the fault normal direction, the vertical direction, or a combination thereof. Step S12: Calculate the unconformity attribute volume based on the synthetic seismic data volume. The unconformity attribute volume is obtained by attribute algorithms including reflection discontinuity, structural abrupt response, and sequence boundary enhancement, and is used to characterize the stratigraphic interface change features. Extract sparse stratigraphic data based on the relative geological age tag volume. The sparse stratigraphic data is extracted by selecting key stratigraphic positions according to a preset stratigraphic interval, performing sparse sampling according to a preset sampling ratio, and extracting representative stratigraphic positions according to preset geological stratigraphic positions. The sparse stratigraphic data is represented as sparse stratigraphic positions, stratigraphic lines, or stratigraphic surfaces. Step S13: Standardization Preprocessing and Spatial Alignment; Standardization preprocessing and spatial alignment are performed on the synthetic seismic data, unconformity surface attribute volumes, and sparse layer data. Standardization preprocessing includes at least one of amplitude normalization, numerical range scaling, outlier suppression, and missing value imputation to form the model training input data. In this step, all input data must maintain a unified spatial coordinate system and sampling grid to ensure the spatial correspondence during subsequent multi-branch coding and feature fusion.

[0025] Step S2: Construct a joint prediction neural network model based on a multi-scale attention mechanism. The neural network model includes four parallel encoders, an adaptive feature fusion module, and two parallel decoders; specifically, it includes the following steps: Step S21: Construct the overall structure of the joint prediction network; Construct a neural network model for joint prediction of relative geological age and fault displacement. The neural network model includes four parallel encoders, an adaptive feature fusion module and two parallel decoders. The neural network model adopts a "multi-branch encoding-fusion-dual decoding" structure, which enables the fusion of prior information from different sources in the feature layer and realizes joint prediction of dual tasks at the output end. Step S22: Construct four parallel encoders; the four parallel encoders correspond to the seismic data encoder, fault location encoder, unconformity surface attribute encoder and sparse layer encoder, respectively. Each encoder extracts feature representations at at least two scales. The extracted features include shallow texture features, middle-layer structural features and deep semantic features to enhance the model's ability to express geological information at different scales. Step S23: Construct two parallel decoders; the two parallel decoders are a relative geological age decoder and a fault displacement decoder, and they share the fusion feature representation from the fusion module; the relative geological age decoder outputs the relative geological age prediction volume, and the fault displacement decoder outputs the fault displacement prediction volume, and the output spatial sampling of both is consistent with the input data.

[0026] Step S3: Establish prior structural constraints at the input end. Input the synthetic seismic data, fault location labels, unconformity surface attributes, and sparse stratigraphic data into the corresponding parallel encoders to extract multi-source structural prior features. This includes the following steps: Step S31: Seismic data branch encoding; Input the synthetic seismic data obtained in step S1 into the seismic data encoder to extract multi-scale seismic features related to reflection phase axis morphology, amplitude texture, and fracture response.

[0027] Step S32: Prior data branch encoding. Input the fault location label volume into the fault location encoder to extract prior features of fault geometric location; input the unconformity surface attribute volume into the unconformity surface attribute encoder to extract prior features of sequence boundary and discontinuity boundary; input the sparse stratigraphic data into the sparse stratigraphic encoder to extract prior features of stratigraphic control. Step S33: Multi-branch feature scale alignment; Spatial scale alignment and feature registration are performed on the multi-scale features output by each branch encoder to form a multi-source prior construction feature set at the input end; This step ensures that the features of different branches can be co-expressed at the voxel level or local window level in the subsequent fusion stage.

[0028] Step S4: Perform adaptive feature fusion. Input multi-branch, multi-scale features into the fusion module, and perform weighted fusion using attention weights to generate a fused feature representation; specifically, this includes the following steps: Step S41: Input fusion module; input the multi-source prior structural feature set formed in step S3 into the adaptive feature fusion module; the fusion module is used to aggregate multi-branch information from seismic data, fault location, unconformity surface attributes and sparse horizons; Step S42: Attention-weighted fusion; The adaptive feature fusion module uses a multi-scale attention mechanism to weight and label features from different sources and at different scales to generate a unified fused feature representation; The multi-scale attention mechanism is channel attention, spatial attention, or a combination of both, and the feature weights of each branch can be adaptively updated with training iterations to improve the expressive ability of constructing sensitive features.

[0029] Step S5: Perform parallel decoding prediction with dual tasks, and output the relative geological age prediction result and the fault displacement prediction result respectively; specifically, it includes the following steps: Step S51: Input the fused feature representation obtained in step S4 into the relative geological age decoder, and output the relative geological age prediction volume. The relative geological age prediction volume can be a continuous value prediction result or a normalized ordinal value prediction result; Step S52: Input the fused feature representation obtained in step S4 into the fault displacement decoder and output the fault displacement prediction volume. The fault displacement prediction volume is used to characterize the relative displacement difference or equivalent displacement between the two strata on the fault sides. Step S6: Establish prior constraints for the output. In each iteration of model training, add a consistency constraint loss between the predicted relative geological age drop and the predicted fault displacement within the fault neighborhood to the supervised loss, and assign it a higher weight. Specifically, this includes the following steps: Step S61: Calculate the supervised loss; in each iteration of model training, calculate: the first supervised loss L between the relative geological time prediction body and the relative geological time label body. rgt The second supervisory loss L between the fault displacement prediction body and the fault displacement label bodythrow ; First supervisory loss L rgt Defined as L ssim For structural similarity index loss, , Where x represents the relative geological time predicted by the network (RGT predicted value), and y represents the relative geological time label (RGT true value); μ x σ and μᵧ represent the mean values ​​of x and y within the local window, respectively; x ² and σᵧ² represent the variances of x and y within the local window, respectively; σ x ᵧ represents the covariance of x and y within the local window; C1 and C2 are stability constants used to avoid zero denominators and improve numerical stability. L mae For the average absolute error loss, , Among them, L mae denoted as the mean absolute error loss, where yᵢ and ŷᵢ are the label value and predicted value at the i-th pixel / voxel, respectively (ŷ can be x); L normal For normal vector loss, , In the normal vector loss L_normal, n̂ᵢ represents the normal vector calculated from the prediction result at the i-th pixel / voxel, nᵢ represents the normal vector calculated from the label, and the symbol "·" represents the vector dot product; First supervisory loss L throw Defined as , Where fᵢ represents the breakpoint label value (true value) at the i-th pixel / voxel, f̂ᵢ represents the breakpoint prediction value at the i-th pixel / voxel; |·| represents absolute value operation; Step S62: Calculate the fault neighborhood consistency constraint loss; within the fault neighborhood indicated by the fault location label, extract the relative geological age prediction values ​​of the corresponding positions on both sides of the fault, and calculate the difference between the prediction values ​​on both sides as the age difference; the corresponding positions on both sides of the fault are determined by any of the following methods: (1) matching corresponding points within a preset window along the fault normal direction; (2) matching corresponding points in the fault neighborhood using the nearest neighbor method; (3) matching the corresponding positions of the strata on both sides of the fault using a local search method. The age difference is measured with the fault displacement prediction value at the same position to construct the mean square error consistency constraint loss, so as to constrain the relative geological age prediction result and the fault displacement prediction result to satisfy the structural correspondence in the fault neighborhood; Step S63: Construct the total loss function; assign a weight higher than at least one supervision loss in step S61 to the consistency constraint loss in step S62 to construct the total loss function, thereby establishing prior constraints for the output; in some implementations, the total loss function is expressed as: L total = α·L rgt + β·L throw + γ·L cons Among them, L rgt For relative geological age monitoring loss, L throw For monitoring the loss of fault displacement, L cons The consistency constraint loss is denoted by α, β, and γ, which are weighting coefficients, and γ is greater than α or β. Among them, L cons Defined as , N f This represents the number of fault sampling points; , These represent the values ​​of the predicted relative geological time body at the i-th fault sampling point along both sides of the fault (which can be defined as the hanging wall and footwall, or the two sides of the fault plane normal); - The number () indicates the fault displacement at that point relative to the predicted geological age. This represents the predicted dislocation value at that point; |·| represents the absolute value operation.

[0030] Step S7: Perform iterative model training and convergence determination, backpropagating and updating parameters based on the total loss function until the model converges; specifically including the following steps: Step S71: Iterative Training and Parameter Update. Backpropagation and parameter update are performed based on the total loss function constructed in step S63. Iterative training is used to optimize the joint prediction neural network model. During training, at least one of the following strategies is employed: learning rate decay, gradient pruning, or early stopping, to improve training stability. Step S72: Convergence determination; As the number of iterations increases, convergence determination is performed based on at least one of the following: changes in the total loss function, validation data error, or preset training rounds; When the preset convergence condition is met, training is stopped and the model parameters after training are output.

[0031] Step S8: Calculate the fault probability volume and unconformity attribute volume from the on-site 3D seismic data, and obtain the sparse bedding plane. Input the data into the trained model, and output the relative geological time volume and fault displacement volume to achieve joint intelligent interpretation. Specifically, this includes the following steps: Step S81: Construct field data input; calculate fault probability volume and unconformity surface attribute volume for field 3D seismic data volume, and obtain sparse layer planes; the fault probability volume is obtained by a fault identification method based on deep learning or a fault detection method based on seismic attributes; the sparse layer planes are obtained by manual interpretation, semi-automatic picking or automatic tracking methods; Step S82: Field data model inference; Input the field 3D seismic data volume, fault probability volume, unconformity surface attribute volume, and sparse bedding plane into the trained joint prediction neural network model in the same data organization form as in the training stage, and perform forward inference; Step S83: Output the joint interpretation results; output the relative geological time volume and fault displacement volume of the field data to realize the integrated joint intelligent interpretation of relative geological time and fault displacement; the output relative geological time volume can be used for stratigraphic tracing, sequence analysis and stratigraphic slice construction; the output fault displacement volume can be used for quantitative fault characterization, fault activity analysis and structural interpretation assistance.

[0032] In this embodiment, a two-dimensional profile of a real three-dimensional seismic data is selected, and a joint intelligent interpretation of relative geological age and fault displacement is performed according to the process described in step S8. Figure 2 (a) shows the original seismic profile of the Opunake seismic field in New Zealand. The area has multiple faults, and the stratigraphic phase axis is affected by the faults, showing obvious faulting characteristics. Figure 2 (b) shows the result of the fault probability volume extracted in step S81 superimposed on the original seismic profile. This result is used as a priori structural feature input into the fault location encoder of the neural network, which clearly defines the spatial geometric distribution of the fault.

[0033] According to steps S82 to S83, the above-mentioned actual earthquake data, fault probability volume and other structural prior information are input into the joint prediction neural network model trained according to steps S1 to S7 in a unified format, and forward inference is performed. Figure 2 (c) shows the relative geological age prediction results output by the model. Different colors in the figure represent the relative depositional sequence of the strata. It can be seen that the predicted relative geological age body not only continuously reflects the macroscopic sequence evolution law, but also accurately presents reasonable abrupt age jumps on both sides of the fault.

[0034] Figure 2 (d) shows the display of the fault displacement prediction results output synchronously by the model superimposed on the original seismic profile. (Comparison) Figure 2 (a) Figure 2 (b) and Figure 2 (c) It can be seen that, thanks to the explicit introduction of the consistency constraint loss between relative geological age difference and fault displacement during the training phase (step S6) of this invention, the predicted fault displacement is not only consistent with the prior fault location in terms of spatial location ( Figure 2 (b) is strictly aligned, and its magnitude is consistent with the relative geological timescale ( Figure 2 (c) The fault displacement amplitudes on both sides of the fault maintained a very high degree of consistency. The results of this embodiment show that the method of the present invention effectively overcomes the structural mismatch problem caused by traditional separation interpretation, and significantly improves the accuracy and geological rationality of the joint interpretation of relative geological age and fault displacement in complex fault zones.

[0035] In the description of this specification, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0036] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints, characterized in that, Specifically, the following steps are included: Step S1: Construct labeled synthetic earthquake training samples and extract tectonic prior information, including constructing a synthetic earthquake data volume with relative geological age labels, fault location labels and fault displacement labels, calculating unconformity surface properties and extracting sparse stratigraphic data, and performing standardized preprocessing on the input data. Step S2: Construct a joint prediction neural network model based on a multi-scale attention mechanism. The neural network model includes four parallel encoders, an adaptive feature fusion module, and two parallel decoders. Step S3: Establish prior structural constraints at the input end. Input the synthetic seismic data, fault location labels, unconformity surface attributes, and sparse stratigraphic data into the corresponding parallel encoders to extract multi-source structural prior features. Step S4: Perform adaptive feature fusion, input multi-branch multi-scale features into the fusion module, perform weighted fusion through attention weights, and generate fused feature representation; Step S5: Perform dual-task parallel decoding prediction and output the relative geological age prediction result and the fault displacement prediction result respectively; Step S6: Establish prior constraints for the output end. In each iteration of model training, add consistency constraint loss between the predicted relative geological age drop and the predicted fault displacement within the fault neighborhood to the supervised loss, and assign weights to them. Step S7: Perform iterative training and convergence determination of the model, and perform backpropagation and parameter update based on the total loss function until the model converges; Step S8: Calculate the fault probability volume and unconformity surface attribute volume from the on-site 3D seismic data and obtain the sparse bedding plane. Input the trained model and output the relative geological time volume and fault displacement volume to achieve joint intelligent interpretation.

2. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Construct a synthetic seismic data volume using geological structure modeling and seismic forward modeling methods. The synthetic seismic data volume includes multiple stratigraphic groups, multiple fault structural units, and unconformity surface structures, as well as spatially corresponding relative geological age labels, fault location labels, and fault displacement labels. The fault displacement labels are calculated by the difference between the corresponding layers on both sides of the fault in the relative geological age domain. This difference can be defined along the fault normal direction, the vertical direction, or a combination thereof. Step S12: Calculate the unconformity attribute volume based on the synthetic seismic data volume. The unconformity attribute volume is obtained by attribute algorithms including reflection discontinuity, structural abrupt response, and sequence boundary enhancement. Extract sparse stratigraphic data based on the relative geological age tag volume. The sparse stratigraphic data is extracted by selecting key stratigraphic intervals according to preset stratigraphic intervals, performing sparse sampling according to preset sampling ratios, and extracting representative stratigraphic intervals according to preset geological stratigraphic positions. The sparse stratigraphic data is represented as sparse stratigraphic points, stratigraphic lines, or stratigraphic surfaces. Step S13: Standardization preprocessing and spatial alignment; Standardization preprocessing and spatial alignment processing are performed on synthetic seismic data, unconformity surface attribute volumes and sparse layer data. Standardization preprocessing includes at least one of amplitude normalization, numerical range scaling, outlier suppression and missing value imputation to form model training input data.

3. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S21: Construct the overall structure of the joint prediction network; Construct a neural network model for joint prediction of relative geological age and fault displacement. The neural network model includes four parallel encoders, an adaptive feature fusion module and two parallel decoders. The neural network model adopts a "multi-branch encoding-fusion-dual decoding" structure, which enables the fusion of prior information from different sources in the feature layer and realizes joint prediction of dual tasks at the output end. Step S22: Construct four parallel encoders; the four parallel encoders correspond to the seismic data encoder, fault location encoder, unconformity surface attribute encoder and sparse layer encoder, respectively. Each encoder extracts feature representations at at least two scales. The extracted features include shallow texture features, middle-layer structural features and deep semantic features. Step S23: Construct two parallel decoders; the two parallel decoders are a relative geological age decoder and a fault displacement decoder. The relative geological age decoder outputs a relative geological age prediction volume, and the fault displacement decoder outputs a fault displacement prediction volume. The output spatial sampling of both is consistent with the input data.

4. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 1, characterized in that, Step S3 specifically includes the following steps: Step S31: Seismic data branch encoding; Input the synthetic seismic data obtained in step S1 into the seismic data encoder to extract multi-scale seismic features related to reflection phase axis morphology, amplitude texture, and fracture response. Step S32: Prior data branch encoding. Input the fault location label volume into the fault location encoder to extract prior features of fault geometric location; input the unconformity surface attribute volume into the unconformity surface attribute encoder to extract prior features of sequence boundary and discontinuity boundary; input the sparse stratigraphic data into the sparse stratigraphic encoder to extract prior features of stratigraphic control. Step S33: Multi-branch feature scale alignment; Spatial scale alignment and feature registration are performed on the multi-scale features output by each branch encoder to form a multi-source prior construction feature set at the input end.

5. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 1, characterized in that, Step S4 specifically includes the following steps: Step S41: Input fusion module; input the multi-source prior structural feature set formed in step S3 into the adaptive feature fusion module; the fusion module is used to aggregate multi-branch information from seismic data, fault location, unconformity surface attributes and sparse horizons; Step S42: Attention-weighted fusion; The adaptive feature fusion module uses a multi-scale attention mechanism to weight and label features from different sources and at different scales, generating a unified fused feature representation; The multi-scale attention mechanism is channel attention, spatial attention, or a combination of both, and the feature weights of each branch can be adaptively updated with training iterations.

6. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 1, characterized in that, Step S5 specifically includes the following steps: Step S51: Input the fused feature representation obtained in step S4 into the relative geological age decoder, and output the relative geological age prediction volume. The relative geological age prediction volume can be a continuous value prediction result or a normalized ordinal value prediction result; Step S52: Input the fused feature representation obtained in step S4 into the fault displacement decoder and output the fault displacement prediction volume. The fault displacement prediction volume is used to characterize the relative displacement difference or equivalent displacement between the two strata of the fault.

7. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 1, characterized in that, Step S6 specifically includes the following steps: Step S61: Calculate the supervised loss; in each iteration of model training, calculate: the first supervised loss L between the relative geological time prediction body and the relative geological time label body. rgt The second supervisory loss L between the fault displacement prediction body and the fault displacement label body throw ; First supervisory loss L rgt Defined as L ssim For structural similarity index loss, ， Where x represents the relative geological age predicted by the network, and y represents the relative geological age label; μ x σ and μᵧ represent the mean values ​​of x and y within the local window, respectively; x ² and σᵧ² represent the variances of x and y within the local window, respectively; σ x ᵧ represents the covariance of x and y within the local window; C1 and C2 are stability constants; L mae For the average absolute error loss, , Among them, L mae denoted as mean absolute error loss, where yᵢ and ŷᵢ are the label value and predicted value at the i-th pixel / voxel, respectively; L normal For normal vector loss, , In the normal vector loss L_normal, n̂ᵢ represents the normal vector calculated from the prediction result at the i-th pixel / voxel, nᵢ represents the normal vector calculated from the label, and the symbol "·" represents the vector dot product; First supervisory loss L throw Defined as , Where fᵢ represents the discontinuity label value at the i-th pixel / voxel, f̂ᵢ represents the discontinuity prediction value at the i-th pixel / voxel; |·| represents absolute value operation; Step S62: Calculate the fault neighborhood consistency constraint loss; within the fault neighborhood indicated by the fault location label, extract the relative geological age prediction values ​​of the corresponding positions on both sides of the fault, and calculate the difference between the prediction values ​​on both sides as the age difference; measure the difference between the age difference and the fault displacement prediction value at the same position, and construct the mean square error consistency constraint loss. Step S63: Construct the total loss function; assign a weight higher than at least one supervision loss in step S61 to the consistency constraint loss in step S62 to construct the total loss function, thereby establishing prior constraints for the output; the total loss function is expressed as: L total = α·L rgt + β·L throw + γ·L cons Among them, L rgt For relative geological age monitoring loss, L throw For monitoring the loss of fault displacement, L cons The consistency constraint loss is denoted by α, β, and γ, which are weighting coefficients, and γ is greater than α or β. Among them, L cons Defined as , N f This represents the number of fault sampling points; , These represent the values ​​of the predicted relative geological time body at the corresponding positions along both sides of the fault at the i-th fault sampling point; This indicates the fault displacement at that point relative to the predicted geological age. This represents the predicted dislocation value at that point; |·| represents the absolute value operation.

8. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 7, characterized in that, The corresponding positions on both sides of the fault are determined by any of the following methods: (1) matching corresponding points within a preset window along the fault normal direction; (2) matching corresponding points in the fault neighborhood using the nearest neighbor method; (3) matching corresponding positions of the strata on both sides of the fault using a local search method.

9. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 7, characterized in that... Step S7 specifically includes the following steps: Step S71: Iterative Training and Parameter Update. Backpropagation and parameter updates are performed based on the total loss function constructed in step S63. Iterative training is used to optimize the joint prediction neural network model; at least one of the following strategies is employed during training: learning rate decay, gradient pruning, or early stopping. Step S72: Convergence determination; As the number of iterations increases, convergence determination is performed based on at least one of the following: changes in the total loss function, validation data error, or preset training rounds; When the preset convergence condition is met, training is stopped and the model parameters after training are output.

10. The method for predicting relative geological age and fault displacement of earthquakes based on prior structural constraints according to claim 1, characterized in that, Step S8 specifically includes the following steps: Step S81: Construct field data input; calculate fault probability volume and unconformity surface attribute volume for field 3D seismic data volume, and obtain sparse layer planes; the fault probability volume is obtained by a fault identification method based on deep learning or a fault detection method based on seismic attributes; the sparse layer planes are obtained by manual interpretation, semi-automatic picking or automatic tracking methods; Step S82: Field data model inference; Input the field 3D seismic data volume, fault probability volume, unconformity surface attribute volume, and sparse bedding plane into the trained joint prediction neural network model in the same data organization form as in the training stage, and perform forward inference; Step S83: Output the joint interpretation results; output the relative geological age and fault displacement of the field data to achieve integrated joint intelligent interpretation of relative geological age and fault displacement.

Images