A seismic horizon tracking method and device based on a physical gate candidate event graph

Abstract

This invention discloses a seismic horizon tracking method and apparatus based on a physically-gated candidate event graph, belonging to the field of seismic data processing methods in geophysics. It acquires seismic profiles or 3D seismic volumes and target horizon anchor points, constructs original seismic flows and attribute / physical flows, and generates events, peaks and valleys, faults, uncertainties, dip angles, and weak priors for integrated support through a dual-flow network. Physical gating is formed by peak and valley responses, local coherence, and fault support; strong peak and valley candidates are directly retained, while weak peak and valley candidates are added to the candidate event graph only when the gated support meets the criteria and uncertainty is limited. Adaptive reachable edges are established based on dip angle prediction locations and fault support. Graph dynamic programming is performed under anchor point constraints, and a reliable mask is generated based on the cost difference between the optimal and suboptimal paths, outputting a 2D horizon trajectory or a 3D horizon surface. This invention can reduce the risks of missing weak reflection candidates, fault anomaly jumps, and cross-horizontal tracking.

Description

Technical Field

[0001] This invention pertains to earthquake data processing methods within the field of geophysics. Specifically, it relates to automatic seismic horizon tracking technology, and more particularly to a seismic horizon tracking method and apparatus based on a physically gated candidate event map. Background Technology

[0002] Seismic horizon tracing is a technique for identifying the same reflection interface in seismic profiles or 3D seismic volumes. It serves as the data foundation for structural interpretation, reservoir characterization, and 3D geological modeling. Existing automatic horizon tracing methods are mainly divided into two categories:

[0003] The first category is traditional tracking methods based on the physical characteristics of earthquakes, which typically use local peak and valley detection, waveform similarity, or dip continuity as the basis for extension. When earthquake reflections are continuous, the signal-to-noise ratio is high, and there is no obvious tectonic faulting, this type of method can maintain local continuous tracking; however, in weak reflection zones, fault displacement areas, adjacent strong in-phase axes, and local low signal-to-noise ratio areas, the local peak and valley characteristics of the target event may disappear or overlap with neighboring events, resulting in missing candidate locations, abnormal path jumps, and cross-layer tracking.

[0004] The second category is intelligent tracking methods based on deep learning. These methods use convolutional networks to output stratigraphic probability maps or segmentation results, which can supplement the features of weakly reflective areas. However, if the network output is directly used as the final stratigraphic level, the network may still generate continuous high responses in locations lacking true peaks and valleys, coherence, or structural evidence, resulting in non-physical completion. Simultaneously, the network's high response to nearby strong reflective areas may attract tracking results away from the target event. Furthermore, existing methods typically perform forced interpolation to complete sections with insufficient evidence, without providing confidence markers, which can easily mislead subsequent geological interpretations.

[0005] Therefore, the technical problem to be solved in this field is: under the premise that the deep network is not allowed to directly determine the stratigraphic location, to use the network to supplement weak reflection candidates, while making the supplemented candidates subject to physical evidence such as earthquake peaks and valleys, local coherence and fault structure; and to dynamically determine the reachability relationship between candidate nodes based on local dip angle and fault support, so that the path remains stable in continuous areas and allows limited discontinuity at fault locations; at the same time, to be able to explicitly mark the credibility of tracking sections with insufficient evidence, so as to avoid the misleading effect of forced completion.

[0006] To address the aforementioned issues, this invention limits the output of the deep network to weak priors, sets strong candidate direct-through branches and weak candidate physics-gated supplementary branches, organizes the admitted candidate positions into a fault-adaptive physics candidate event graph, and performs graph dynamic programming using multi-source evidence under initial anchor point constraints. Thus, the weak priors of the deep network only affect weak candidate supplementation and node evidence; the peak and valley attributes, dip direction, fault support, and event sequence of earthquake events jointly define the candidate graph and its transition relationships. Summary of the Invention

[0007] The purpose of this invention is to provide a seismic horizon tracking method and apparatus based on a physically-gated candidate event map, to solve the problems of insufficient candidates for weak reflection regions, inapplicability of fault location continuity constraints, cross-layer tracking caused by adjacent strong in-phase axes, and forced completion of low-evidence regions. This invention also provides an electronic device and a computer-readable storage medium.

[0008] To achieve the above objectives, this invention acquires seismic profiles or 3D seismic volumes and initial anchor points of target horizons, constructs spatially aligned original seismic flows and attribute / physical flows, generates multi-task deep weak priors through a dual-flow tracing sensing prior network, forms physical gating based on peak-valley responses, local coherence, and fault support, and filters candidate event nodes using a strong candidate direct access and weak candidate gating supplementation method, calculates node evidence based on peak-valley, coherence, dip angle, fault, and gating support at the node, and establishes adaptive reachable edges based on dip angle predicted location and fault support, performs dynamic programming of evidence consistency graph under anchor point constraints, obtains optimal and suboptimal paths, and outputs horizon results and a credibility mask based on the cost difference between the two and the evidence along the path.

[0009] Step S1: Obtain the seismic profile or 3D seismic body to be tracked, perform sampling coordinate unification and inter-trace amplitude consistency processing on the seismic profile or 3D seismic body, and determine at least one initial anchor point and the tracking range of the target layer.

[0010] Step S2: Construct the original seismic flow and attribute / physical flow with one-to-one spatial location within the range to be tracked, input the trained dual-flow tracking perception prior network, and generate a depth weak prior including in-phase event prior, peak-valley-ridge prior, fault prior, uncertainty prior, dip angle prior and comprehensive support prior;

[0011] Step S3: Calculate the physical gating value based on the peak and valley candidate mask, the comprehensive peak confidence, the local coherence support, and the fault support. Multiply the physical gating value by the comprehensive support prior position by position to obtain the gated candidate support.

[0012] Step S4: Select candidate event nodes using the strong and weak candidate differential admission rule, and construct a physical candidate event graph based on all candidate event nodes, node evidence scores, and reachability edges between adjacent seismic traces;

[0013] Step S5: For two candidate event nodes on adjacent seismic traces, calculate the predicted dip position based on the time or depth location, local dip angle, and trace spacing of the previous candidate event node, and adaptively adjust the maximum allowable vertical jump based on fault support; perform dynamic programming of the evidence consistency diagram under the initial anchor point constraint to determine the optimal path and the second-best path.

[0014] Step S6: Calculate the path credibility based on the cumulative cost difference between the optimal path and the suboptimal path, the node evidence along the optimal path, the gated candidate support, the local waveform similarity, and the vertical jump amplitude; mark continuous segments with credibility below the credibility threshold as untrustworthy segments and do not perform forced interpolation completion, and output the two-dimensional seismic horizon trajectory or three-dimensional horizon surface and the corresponding credibility mask.

[0015] In step S1, the initial anchor point includes a channel number and a time sampling position or depth sampling position corresponding to the channel number;

[0016] The original earthquake record is as follows: , height and width These correspond to the number of sampling points and seismic traces in the seismic window to be processed, respectively, and the number of channels. Determined by the current profile and the number of its adjacent profiles; the attribute / physical flow record is... Its height and width are the same as Same, number of channels Determined by the number of seismic attributes used; and Align the time or depth sampling location and channel number direction point by point.

[0017] In step S2, the original seismic flow is constructed using the current seismic profile and its adjacent seismic profiles within the range to be tracked. At least two of the following properties / physical flows are constructed from the local gradient, ridge response, local energy, local coherence approximation, multi-scale difference Gaussian response, peak-valley response, dip properties, and fault support calculated from the same seismic data window. Each prior is used only as evidence for subsequent candidate screening and path optimization, and is not directly used as the final horizon.

[0018] The dual-stream tracking and sensing prior network includes a raw seismic stream encoder, an attribute / physical stream encoder, a gated fusion module, a scale-gated module, a tracking and sensing context module, a multi-scale aggregation output module, and a multi-task prediction head. The two encoders output raw seismic features and attribute features at multiple scales, respectively. The gated fusion module adjusts the participation ratio of attribute features using a gate weight jointly calculated by the two types of features, and inputs the fusion result into the tracking and sensing context module and the multi-scale aggregation output module.

[0019] The training samples of the dual-flow tracking perception prior network include spatially corresponding original seismic flows, attribute / physical flows, and in-phase event labels, peak-valley-ridge line labels, fault labels, dip angle labels, and effective annotation masks. The labels are used to constrain the in-phase event priors, peak-valley-ridge line priors, fault priors, dip angle priors, and comprehensive support priors, respectively, and uncertainty supervision or prediction residuals are used to constrain the uncertainty priors. After training, the network parameters are fixed so that each output prior maintains the same spatial coordinates as the input seismic window.

[0020] In step S3, when the comprehensive support prior is high but lacks evidence of peaks, valleys, coherence, or fault structures, the position is suppressed; when the weak reflection position has corresponding physical evidence, the position is retained, and the value of the physical gate value ranges from 0 to 1; the gated candidate support is obtained by multiplying the physical gate value by the comprehensive support prior; the physical gate only controls the degree to which the deep weak prior participates in the weak candidate supplementation and node evidence weighting, and does not change the event polarity and candidate qualification of strong peak-valley candidates that reach the first peak-valley threshold;

[0021] The node evidence score is determined by the peak and valley intensities at the candidate node. Local coherent support Tilt angle consistency Fault support The results were obtained by weighting the gated candidate supports; among them, fault supports... Evidence from participating nodes is primarily used to regulate transitions.

[0022] In step S4, strong peak-valley candidates whose comprehensive peak confidence reaches the first peak-valley threshold are directly used as candidate event nodes. Weak peak-valley candidates whose comprehensive peak confidence is between the second peak-valley threshold and the first peak-valley threshold are only used as supplementary candidate event nodes when the gating candidate support reaches the candidate support threshold and the uncertainty prior is not higher than the uncertainty threshold. The first peak-valley threshold is greater than the second peak-valley threshold. The strong peak-valley candidates enter the physical candidate event graph through the strong candidate direct branch, and the weak peak-valley candidates enter the physical candidate event graph through the physical gating supplementary branch. The uncertainty prior is used to exclude weak peak-valley candidates with high network response but insufficient prediction stability, so that the deep weak prior only supplements the candidate positions with the lowest peak-valley evidence and physical support.

[0023] In step S5, reachable edges are established only when the residual of the candidate event node relative to the dip angle prediction position does not exceed the maximum allowable vertical jump and the polarity consistency and event sequence constraints are satisfied. The optimal path and the second-best path are determined based on polarity consistency, local waveform similarity, dip angle prediction residual, fault adaptive jump, event sequence, local continuity and node evidence score.

[0024] The maximum permissible vertical jump consists of a basic jump variable and an additional jump variable positively correlated with fault support. When the fault support is below the fault threshold, candidate nodes are only connected to consecutive candidate nodes near the dip angle prediction position. When the fault support reaches the fault threshold, the connection range is allowed to be expanded within a preset upper limit, while maintaining the order of the target event relative to the upper and lower neighboring events.

[0025] The transition costs between adjacent candidate event nodes include polarity inconsistency cost, local waveform dissimilarity cost, dip angle prediction residual cost, fault adaptive jump cost, event sequence violation cost, local curvature or second-order difference continuity cost, and node evidence reward term; the cost terms are normalized and then weighted and combined.

[0026] When performing dynamic programming of the evidence consistency graph, only candidate event nodes that are located in the same path or a preset neighborhood as the initial anchor point and whose time or depth position deviation does not exceed the anchor point tolerance are set as valid starting nodes; the initial cumulative cost of other candidate event nodes is set to infinity or an anchor point deviation penalty is added. After completing the recursion step by step, the optimal path with the minimum cumulative cost and the second-minimum cumulative cost path are obtained by backtracking.

[0027] The path credibility includes at least the path discrimination degree corresponding to the cumulative cost difference between the optimal path and the second-best path; when any one of the path discrimination degree, node evidence and gating candidate support is lower than the corresponding threshold, or when the continuous vertical jump exceeds the credibility jump threshold, the corresponding path position is marked as untrustworthy.

[0028] In 3D seismic body plane tracking, physical candidate event maps are constructed along the inline and crossline directions respectively, and direction tracking results are obtained. Consistency judgment is made on the time or depth difference between the two directions at the same spatial location. The locations that meet the consistency threshold are weighted and fused, and the confidence of the locations that do not meet the consistency threshold is reduced or marked as untrusted areas.

[0029] A seismic horizon tracking device includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. The device comprises the following modules:

[0030] The data acquisition and anchoring module is used to acquire seismic profiles or 3D seismic volumes, unify sampling coordinates and inter-trace amplitudes, and determine the initial anchoring point and tracking range of the target layer.

[0031] The weak prior generation module is used to construct the original seismic flow and attribute / physical flow that correspond one-to-one with the spatial location, and to generate priors for co-phase events, peak-valley-ridge lines, faults, uncertainties, dip angles, and comprehensive support through a dual-flow tracking and sensing prior network.

[0032] The physical gating and candidate admission module is used to calculate the physical gating value and gating candidate support, and to filter candidate event nodes through strong candidate direct access branches and weak candidate gating supplementary branches;

[0033] The candidate event graph construction module is used to calculate the node evidence score, establish reachability edges between adjacent seismic traces based on dip angle prediction location, fault support, polarity consistency and event sequence, and construct a physical candidate event graph.

[0034] An anchor point constraint path optimization module is used to perform evidence consistency graph dynamic programming on the physical candidate event graph to obtain the optimal path and the suboptimal path.

[0035] The credible result output module is used to generate a credible mask based on the cumulative cost difference between the optimal and suboptimal paths and evidence along the path, and output a two-dimensional seismic horizon trajectory or a three-dimensional horizon surface.

[0036] The computer program implements a seismic horizon tracking method when executed by a processor.

[0037] Compared with the prior art, the present invention has the following technical effects:

[0038] First, a differentiated admission mechanism is adopted, with strong candidates selected directly and weak candidates gated and supplemented. Strong peak-valley candidates are selected directly without relying on network output; weak peak-valley candidates must simultaneously meet the requirements of minimum peak-valley evidence, physical gating support, and uncertainty constraints to be added to the candidate map. The network output is only used as a weak prior, effectively reducing non-physical candidates caused by high network response and ensuring that the tracking results conform to the physical laws of earthquakes.

[0039] Secondly, the reachability between candidate nodes is determined by the dip angle prediction location and fault support. The continuous reflection zone adopts strict dip angle constraints to ensure path smoothness. The fault region allows restricted vertical jumps according to the degree of fault support, avoiding the problem that the fixed jump threshold cannot simultaneously adapt to the continuous zone and the fault zone.

[0040] Third, dynamic programming of evidence consistency graphs with anchor point initialization is adopted, which combines polarity consistency, local waveform similarity, tilt angle prediction residuals, event sequence, and local continuity multi-source evidence to constrain the global path, effectively reducing the cross-layer tracking probability caused by the attraction of neighboring strong in-phase axes.

[0041] Fourth, the path discrimination is calculated by the cumulative cost difference between the optimal and suboptimal paths. The credibility of each path is generated by combining node evidence, gating support and waveform similarity, and a credibility mask is output. Sections with insufficient evidence are marked as untrustworthy and no forced interpolation is performed to complete them, so as to provide a clear credibility reference for subsequent geological interpretation.

[0042] The technical chain of this invention is as follows: raw seismic data and its attributes generate multi-task depth weak priors; seismic physical evidence gates the comprehensive support priors; strong and weak candidate differential admission forms candidate nodes; dip prediction and fault support determine reachable edges; node evidence and edge transfer costs are jointly input into the anchor point constraint graph for dynamic programming; optimal and suboptimal paths and evidence along the path jointly generate a credibility mask. The output of each step serves as the explicit input of the next step, thus forming a complete seismic data processing flow. Attached Figure Description

[0043] Figure 1 This is a flowchart of the seismic horizon tracking method in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the deep weak prior generation module structure in an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of physical gating and strong / weak candidate admission in an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the construction of the fault adaptive physical candidate event graph in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of dynamic programming path optimization of anchor point constraint evidence consistency graph in an embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram of the seismic horizon tracking device in an embodiment of the present invention. Detailed Implementation

[0049] This invention provides a seismic horizon tracking method and apparatus based on a physically-gated candidate event graph. Instead of directly interpreting the network probability graph as horizons, this method treats the network output as a weak prior constrained by seismic physical evidence, and organizes the locations of in-phase events that pass through strong-weak candidate differential admission into a sparse candidate event graph. The following description, in conjunction with the accompanying drawings, explains the input variables, intermediate processing quantities, output quantities, and their interconnections.

[0050] Example 1

[0051] like Figure 1 As shown in the figure, this embodiment proposes a seismic horizon tracking method based on a physically gated candidate event map.

[0052] Step S1: Obtain the seismic profile or 3D seismic volume to be tracked, and determine the initial anchor point and tracking range of the target layer. The seismic profile can be an inline profile or a crossline profile, with the lateral coordinates... Indicates the seismic trace number and vertical coordinate. This represents a time sampling point or a depth sampling point. The input data undergoes sampling interval unification, outlier checking, and inter-channel amplitude consistency, ensuring that subsequent attribute calculations and comparisons of adjacent channels are performed on the same coordinate and amplitude scale. Each initial anchor point is denoted as... Its output is passed to step 5 to limit the valid starting node of the dynamic programming.

[0053] Step S2: Construct dual-stream input and generate weak depth priors. Within the tracking area, the original seismic flow is constructed using the target profile and its adjacent profiles, while the attribute / physical flow is constructed using seismic attributes calculated within the same spatial window. Both types of input have the same height. and width , so that any position ( , The two types of inputs correspond to the same time or depth sampling location and the same trace number. The original seismic stream is used to preserve amplitude texture and event continuity, while the attribute / physical stream is used to provide peak and valley, coherence, dip, and fault structure information.

[0054] Overall variable transitivity: The same seismic data window is formed first. and The two are generated separately. and Then form in sequence , , to , , and Output by the multi-task prediction head Original earthquake evidence and Further formation , , and ,filter and And form a candidate event graph ;exist Up by and Obtain the optimal path And the suboptimal path, ultimately forming , and Each of the above intermediate variables is generated from the explicit output of the previous stage and used as the explicit input of the next stage.

[0055] Specifically, the target seismic profile and its adjacent seismic profiles are combined to form the original seismic flow input Xr, which is expressed as:

[0056] ;

[0057] In the formula, Indicates the original seismic flow input; " indicates a relationship; Represents the space of real numbers; This indicates the number of sampling points along the time or depth direction of the input window; This indicates the number of seismic traces contained in the input window; Indicates the original number of seismic flow channels; superscript express The three-dimensional data dimensions, where " "" indicates the combination of various dimensions.

[0058] The The target seismic profile and its adjacent profiles obtained in step S1 are then processed by coordinate unification, missing value handling, and amplitude standardization, and arranged by channel within the same spatial window; subsequently, they are input into the original seismic flow encoder. generate And as a priori network for dual-stream tracking and perception. The first input. , and The window and channel configuration is determined and passed to the subsequent feature size correspondence and multi-task prior recovery process.

[0059] Simultaneously, attribute / physical information is extracted from the target seismic profile to form the attribute / physical flow input X. a It is represented as:

[0060] ;

[0061] In the formula, Indicates attribute / physical flow input; " indicates a relationship; Represents the space of real numbers; This indicates the number of sampling points along the time or depth direction of the input window; This indicates the number of seismic traces contained in the input window; Indicates the number of attribute / physical flow channels; superscript express The three-dimensional data dimensions, where " "" indicates a combination of dimensions. by and Gradient, ridge, local energy, coherence, differential Gaussian, peak-valley, and fault properties calculated over the exact same time or depth range and channel range are arranged by channel; then input into the attribute / physical flow encoder. generate , and as The second input. and shared and This ensures the subsequent and It can fuse elements one by one in the same position.

[0062] The aforementioned dual-stream inputs are fed into a dual-stream tracking perception prior network, and multi-scale feature extraction is performed through the original seismic flow encoder and the attribute / physical flow encoder, respectively.

[0063] The original seismic flow coding process is represented as follows:

[0064] ;

[0065] In the formula, Indicates the first Original seismic flow coding features at various scales; Indicates the encoding scale index; Indicates the first Original seismic flow coding operations corresponding to each scale; Indicates the original seismic flow input; the equal sign indicates... go through After processing, the result is The aforementioned From the original seismic flow encoder right Proceed to the first Features were extracted at each scale. The resulting... Compared with the same scale Used together for calculation And participates in the fusion feature as a residual structure term. The calculation.

[0066] After obtaining the multi-scale coding features of the original seismic flow, in order to extract the attribute / physical flow features that correspond one-to-one with it in terms of spatial location and scale, the attributes / physical flow are encoded at the same scale, specifically as follows:

[0067] ;

[0068] In the formula, Indicates the first Attribute / physical stream encoding features at various scales; Indicates the encoding scale index; Indicates the first Attribute / physical flow encoding operations corresponding to each scale; Indicates attribute / physical flow input; the equals sign indicates... go through After processing, the result is The aforementioned By attribute / physical flow encoder right Proceed to the first Features were extracted at each scale. The resulting... Compared with the same scale Used together for calculation and through Modulated participation in fusion features The calculation.

[0069] Because the contribution of raw seismic information and prior attribute information varies across different geological regions, a gating fusion mechanism is introduced between features at different scales to improve the adaptive capability of dual-stream information fusion. First, gating weights are calculated based on the two types of features:

[0070] ;

[0071] In the formula, Indicates the first Network internal control weights at various scales; Indicates the feature scale index; This represents the Sigmoid activation function, whose output ranges from 0 to 1; This represents the learnable weight parameters of the gating module; and They represent the first Original seismic features and attribute characteristics at various scales; [] indicates that the two are joined together along the channel direction; This indicates the bias parameters of the gating module; the plus sign indicates adding the bias, function Map the linear combination result to The aforementioned From the same scale and After channel splicing, using the training data and Calculation; training completed and Keep it fixed. As an attribute feature adjustment coefficient used for The calculation is to determine The level of participation in the integration of the two streams.

[0072] Based on the obtained gating weights, adaptive fusion of attributes / physical features is performed:

[0073] ;

[0074] In the formula, Indicates the first Dual-stream fusion features at various scales; Indicates the first Each scale of gating weights; " indicates element-wise multiplication; Indicates the first Attributes / physical flow characteristics at each scale; Indicates the first The original seismic flow characteristics at various scales; the plus sign indicates that the gating adjustment... and Add element by element. Based on the stated gating weights, Depend on , and Element-by-element fusion is obtained; then, three context branches are input to generate... , and At the three selected scales, the corresponding fusion features are uniformly denoted as... and further input get .

[0075] To enhance the network's ability to represent long-range continuous relationships between seismic horizons, the fused features are input into the tracking and perception context module, and multi-scale structural information is obtained through convolution at different scales:

[0076] ;

[0077] In the formula, This represents the local context features output by the first context branch; This indicates a convolution operation with a kernel size of 1×1; Indicates the first Fusion features at various scales; This indicates the feature scale index currently being processed; the subscript 1 is used to identify the first context branch. Depend on Generated through a 1×1 convolution branch, used to preserve and transform local channel information; subsequently combined with... and joint action The input forms contextual features. .

[0078] After obtaining the local contextual features, to further extract reflection structure information within a moderate range, the same fused feature is input into a 3×3 dilated convolution branch with an inflation rate of 2, specifically represented as follows:

[0079] ;

[0080] In the formula, This represents the contextual features output by the second context branch; This indicates a 3×3 kernel size and a dilated convolution dilation rate. Take a 2x2 convolution operation; Indicates the dilation rate of the convolutional cavity; Indicates the first Fusion features at various scales; This indicates the feature scale index currently being processed; subscript 2 is used to identify the second context branch. Depend on Generated by a 3×3 dilated convolution branch with an inflation rate of 2, used to extract medium-range structural information; subsequently combined with... and Common input form .

[0081] Based on the structural information at a medium scale, to enhance the perception of the continuity of layers at larger scales, the same fused feature is input into a 3×3 dilated convolution branch with an inflation rate of 4, specifically represented as:

[0082] ;

[0083] In the formula, This represents the contextual features output by the third context branch; This indicates a 3×3 kernel size and a dilated convolution dilation rate. Perform a convolution operation with a factor of 4; Indicates the dilation rate of the convolutional cavity; Indicates the first Fusion features at various scales; This indicates the feature scale index currently being processed; the subscript 3 is used to identify the third context branch. Depend on Generated by a 3×3 dilated convolution branch with an inflation rate of 4, used to extract structural information over a larger area; subsequently combined with... and Common input form .

[0084] Subsequently, contextual features at different scales are fused:

[0085] ;

[0086] In the formula, This represents the fused contextual features; This indicates that splicing is performed along the direction of the feature channel; , and These represent the output features of the three context branches; the commas in the formula are used to separate the three inputs involved in the concatenation.

[0087] The output of the above three context branches Depend on , and It is obtained by splicing along the channel direction, integrating structural information from different receptive fields; then it is passed to the corresponding scale feature transformation and multi-scale aggregation process to form a supply. Scale features processed Here These are network context features and are not used as local coherence support for candidate nodes. use.

[0088] Furthermore, regarding the first Fusion features at various scales Perform feature mapping to obtain the mapping results at the corresponding scale. :

[0089] ;

[0090] In the formula, Indicates the first Mapping features at each scale; Indicates the first Feature mapping operations corresponding to each scale; Indicates the first Fusion features at various scales; Indicates scale index; " indicates a membership relation; set" express Take 1, 2, and 3 in sequence. From the Scale-based two-stream fusion and context processing results, A mapping operation with pre-defined and trained parameters for this scale; the resulting exist When =1, 2, 3, they are respectively denoted as , and and jointly used Multi-scale aggregation calculation.

[0091] The above feature mapping yields corresponding results at three scales. To unify the mapped features of different resolutions to the same spatial scale and complete the aggregation, the features at the second and third scales are upsampled and then concatenated with the features at the first scale, as specifically:

[0092] ;

[0093] In the formula, This represents the network output features after multi-scale aggregation; Indicates splicing along the channel direction; , and These represent the mapping features at three different scales; This indicates a 2x upsampling; This indicates a 4x upsampling; upsampling is used to make... and Space dimensions and Consistent; commas are used to separate the three inputs involved in the concatenation.

[0094] After completing multi-scale alignment, Depend on Upsampling and after upsampling The result is obtained by splicing; its spatial position corresponds to the original input window. Then, a multi-task prediction head is input, and it generates... , , , , and The six outputs together constitute .

[0095] Finally, deep weak priors are generated using a multi-task prediction head:

[0096] ;

[0097] In the formula, Represents a multi-task deep weak prior set; Indicated by network parameters A defined two-stream tracking and sensing prior network; and These represent the original seismic flow input and the attribute / physical flow input, respectively. This indicates a priori events of the same phase; Indicates the a priori characteristics of peaks, valleys, and ridges; Indicates fault prior; This indicates uncertain prior knowledge; Indicates the prior of the tilt angle; The curly braces indicate the comprehensive supporting priors; the curly braces represent the set consisting of the above six priors.

[0098] After processing by the multi-task prediction head, Depend on and The process sequentially involves encoding, gated fusion, context processing, multi-scale aggregation, and multi-task prediction head generation. Subsequent processing... and Used to assist in the judgment of events and peak / valley locations. Formed by fusion with earthquake fault properties , Restricting admission for weak candidates Convert to local slope , and Multiplication None of the above priors are directly used as the final layer.

[0099] Through the above steps and After two-way encoding, multi-scale gating fusion, context modeling, and aggregation output, the following results are obtained: . and Used to determine events and peak / valley locations. Used to form fault support, Used to limit unstable weak candidates, Used to predict the target location on adjacent lanes The outputs are multiplied with the physical gate to obtain gate candidate supports. Each output serves as input evidence for subsequent steps and does not directly determine the final layer location.

[0100] like Figure 2 As shown, the dual-stream tracking sensing prior network includes a raw seismic flow encoder and an attribute / physical flow encoder, with the two encoders outputting at corresponding scales. and Gating fusion module output and ; Tracking and Perception Context Module Output Multi-scale aggregation output module output Multi-task prediction head output The variables mentioned above are passed in the following order: input—encoding features—fusion features—contextual features—aggregation features—multi-task priors.

[0101] During the network training phase, windows of the same size as those in the inference phase are extracted from seismic data with both artificial horizon interpretation and fault interpretation, and then generated accordingly. and Based on the results of manual interpretation, in-phase event labels, peak-valley-ridge line labels, fault labels, and effective annotation masks are constructed; dip labels are constructed based on the local slope or structural tensor results of the interpreted horizons. Regions without reliable interpretation are excluded using effective annotation masks, preventing these regions from participating in the monitoring loss calculation.

[0102] During training, priors for in-phase events, peak-valley-ridge lines, faults, and comprehensive support are assigned corresponding classification or distance losses; priors for dip angles are assigned direction regression losses; and priors for uncertainty are assigned prediction residual supervision or heteroscedasticity supervision. The losses for each task are weighted and combined according to the following formula, and the network parameters are updated through backpropagation. :

[0103] ;

[0104] In the formula, Indicates the total loss from multiple tasks; , , , , and These represent in-phase events, peak-valley-ridge lines, faults, uncertainties, dip angles, and losses from comprehensive support tasks, respectively. , , , , and These represent the non-negative weights of the corresponding task losses; the product represents each loss multiplied by its corresponding weight, and the plus sign represents the summation of the six weighted losses.

[0105] During the network training phase, to It is calculated from six predicted priors and corresponding artificial labels or effective labeled masks. to Determined by the validation set. Total loss. Updated via backpropagation And including and The network includes learnable parameters; these parameters are fixed after training and used by the network. and Stable generation .

[0106] To ensure seamless integration between network output and subsequent physical processing, , , , , and All restored to and according to position ( , The peak and valley responses, local coherence, and fault properties correspond point by point; therefore, any threshold value, candidate node, and node evidence can be traced back to a specific sampling location within the original seismic window.

[0107] Example 2

[0108] like Figure 3 As shown, to limit the area of ​​weak prior knowledge of depth, this embodiment uses the location (t, ...) in the seismic profile. The basic processing unit is ). First, local maxima and local minima are detected on the original seismic traces. Then, the absolute amplitude, peak prominence, and local energy of the local extrema are normalized to obtain the comprehensive peak confidence score.

[0109] ;

[0110] In the formula, Indicates position The overall peak confidence level at the location; Indicates the time or depth sampling location; Indicates the earthquake track number; Indicates the normalized peak and valley absolute amplitude; Indicates the normalized peak saliency; Represents normalized local energy; , and The symbols represent the non-negative weights of the three pieces of evidence; the product represents the evidence multiplied by its corresponding weight; and the plus sign represents the summation of the three weighted pieces of evidence.

[0111] The , and All are derived from the original seismic data at the location The nearby peaks and valleys and local windows are calculated and normalized. Subsequently used to form ,enter and through and Filter separately and Peak and valley evidence at the nodes Also determined by the peak and valley intensity at that location or Obtain.

[0112] Local coherent support By position (t, Within a time or depth window centered on ), the normalized cross-correlation coefficients of the seismic waveforms of the current trace and adjacent traces are averaged and mapped to a range of 0 to 1. Seismic fault attributes are denoted as... It is obtained from at least one of coherent gradient, variance, curvature, or fault enhancement properties; fault support Based on fault prior and The result of fusion:

[0113] ;

[0114] In the formula, Indicates position Fault support after fusion; Indicates the time or depth sampling location; Indicates the earthquake track number; This represents the fault fusion coefficient, which ranges from 0 to 1. This represents the fault prior in the network output; This indicates the fault properties calculated from the raw seismic data; The sign indicates the complementary weights corresponding to the fault attributes; the plus sign indicates the summation of two weighted fault evidences.

[0115] The From , Fault properties calculated from raw seismic data are fused together at the same location. . Then enter In candidate nodes Value formation at point and continue to be used Node evidence And the cost of fault-related transfer.

[0116] After obtaining the comprehensive peak confidence, local coherence support, and fusion fault support, these are input together with the peak / valley candidate mask into the gating mapping function to calculate the physical gating value, specifically expressed as follows:

[0117] ;

[0118] In the formula, Indicates position The physical gating value at the location; Indicates the time or depth sampling location; Indicates the earthquake track number; This represents a gating function that maps multiple pieces of physical evidence to a range of 0 to 1; Indicates peak / valley candidate masks; Indicates the overall peak confidence level; Indicates local coherent support; This indicates the fused fault support; the comma is used to separate the four inputs of the gating function.

[0119] After inputting the aforementioned physical evidence into the gating function... , and Obtained from raw seismic data, Obtained by fusing network fault priors and seismic fault attributes; four common inputs form The output was then compared with... Multiply, we get .

[0120] After obtaining the physical gating value, it is multiplied point-by-point with the comprehensive support prior output by the two-stream network to generate gating candidate supports that only take effect at locations where physical evidence allows, specifically expressed as:

[0121] ;

[0122] In the formula, Indicates position Gating candidate support at the location; Indicates the time or depth sampling location; Indicates the earthquake track number; Indicates the physical gating value; The first two represent the comprehensive support priors; multiplying them together means retaining or enhancing only the deep weak priors supported by physical evidence.

[0123] The Physically Gated and network integrated support prior The value at the candidate node is obtained by multiplying point by point. For Access, its node form For After normalization Used for path credibility .

[0124] Therefore, when the network's overall support is high but the original seismic data lacks evidence of minimum peaks and valleys, coherence, or corresponding structural evidence, the network response cannot form a supplementary candidate; when the weak reflection location still retains the minimum peak and valley response and has coherence or fault structural support, allow This mechanism reinforces the candidate. It ensures that the network output can only supplement locations allowed by seismic physical evidence.

[0125] Example 3

[0126] This embodiment transforms layer tracing from a dense pixel domain search to a sparse candidate event graph search. For each layer... First, determine the strong candidate direct path set and the weak candidate supplementary set, then merge the two into a candidate event node set. In the original formula... A second peak-valley threshold is used to form the lowest peak-valley candidate mask; for the strong candidate direct pass set, the corresponding candidate admission control is set to pass; for the weak candidate supplement set, the candidate admission is determined by... and Determined jointly. The set of candidate event nodes is represented as:

[0127] ;

[0128] In the formula, Indicates the first The set of all candidate event nodes of the Dao; Indicates the earthquake track number; Indicates the first The strong candidate direct path set of the Dao; Indicates the first The weak candidate complement set of the Tao; "" indicates finding the union of two sets.

[0129] The Depend on and The merged results ensure that strong reflection candidates and physically-gated weak reflection candidates are represented by a unified node; subsequently, all seismic traces are processed... Find the set of nodes in the graph formed by the union of sets. and as a candidate edge It forms the basis for constructing and calculating node evidence.

[0130] To further clarify the composition of the candidate set, we first filter strong candidate pass sets that do not depend on depth prior admission based on the first peak-valley threshold, specifically as follows:

[0131] ;

[0132] In the formula, Indicates the first The strong candidate direct path set of the Dao; Indicates candidate event nodes; Indicates the candidate node index; This indicates that the node is located at the th Tao, time or depth location ; Represents a node The overall peak confidence level at the location; Indicates the first peak-to-valley threshold; "" indicates not less than; the content after the colon indicates the conditions that a node must meet to enter the set. Directly by and The comparison yields results that do not rely on depth-based access control; the resulting nodes are directly merged. Subsequently participated Calculation, candidate edge construction, and path optimization.

[0133] In addition to strong candidates, for weak reflection positions located between the two-level peak and valley thresholds, a supplementary set of weak candidates is further selected by combining gated candidate support and uncertainty prior, specifically expressed as follows:

[0134] ;

[0135] In the formula, Indicates the first The weak candidate complement set of the Tao; Indicates candidate event nodes, Indicates the candidate node index; This indicates that the node is located at the th Tao, time or depth location ; Indicates the overall peak confidence level; and These represent the second peak-to-valley threshold and the first peak-to-valley threshold, respectively. Indicates candidate support for gating; Indicates the candidate support threshold; This indicates uncertain prior knowledge; Indicates the uncertainty threshold; " "and" "" indicates not greater than, less than, and greater than respectively; the comma indicates that all admission conditions must be met simultaneously.

[0136] The Simultaneously receive , and and using the values ​​determined on the validation set. , , and Screening complete; weak candidates are merged. Subsequently, the same node evidence, candidate edges, and dynamic programming processing are used as for strong candidates.

[0137] The credibility of each candidate node is determined by the original seismic response, structural evidence, and gated weak priors. This is to avoid conflict with the attribute-encoded features in step S2. and fusion features Duplicate names, node peak-valley intensity is denoted as Local coherent support is denoted as The consistency of the tilt angle is denoted as Fault support is denoted as Network tilt prior In candidate nodes Converting to local slope The dip slope attribute is obtained from the structural tensor or plane wave failure operator of the original seismic data. The difference between the two is denoted as the tilt angle difference. :

[0138] ;

[0139] In the formula, Indicates candidate nodes The difference between the two local slopes; Indicates the candidate node index; Indicates prior knowledge of the tilt angle At the node The local slope obtained from the transformation; This indicates that the original seismic attributes are at the node. The local slope calculated at that point; the minus sign indicates the former minus the latter. Depend on At the node Obtained by conversion at the point, Obtained from the structural tensor or local dip properties of the original seismic data; the difference between the two Only used as a reference for calculating the inclination consistency in the next equation Input.

[0140] After obtaining the difference between the two types of local slopes, the difference is normalized according to the dip angle difference scale parameter, and then transformed into dip angle consistency in the range of 0 to 1 through exponential mapping, specifically expressed as:

[0141] ;

[0142] In the formula, Indicates candidate nodes Consistency of tilt angle at the location; Indicates the candidate node index; Represents an exponential function with the natural constant as its base; This represents the difference between the network tilt slope and the seismic attribute slope. The superscript 2 indicates a positive dip angle difference scale parameter; the negative sign indicates an increase in dip angle difference. Decrease; Fractional expression right Scale normalization is performed. Depend on and the validation set determined The calculation yielded the result; this result was then used as input evidence for the orientation of the node structure. It can also be used to construct the tilt consistency term in the transfer cost. .

[0143] Subsequently, peak-valley strength, local coherence support, dip consistency, fault support, and gated candidate support are weighted and combined to form a comprehensive evidence score for each candidate node, specifically expressed as follows:

[0144] ;

[0145] In the formula, Indicates candidate nodes The node evidence score; Indicates the candidate node index; Indicates peak and valley intensity; Indicates local coherent support; Indicates consistency of tilt angle; Indicates fault support; Indicates candidate support for gating; , , , and The numbers represent the weights of the five pieces of evidence; the product represents each piece of evidence multiplied by its corresponding weight; and the plus sign represents the summation of the five weighted pieces of evidence.

[0146] The , , , and The nodes are supported by peak-valley, coherence, dip, fault, and gating candidates, respectively. The value is obtained by taking the value; after weighting, it forms... And save the image . exist As a node reward, In terms of credibility This serves as evidence for normalization.

[0147] After obtaining the candidate nodes and their evidence scores for each trace, all candidate nodes, reachability edges between adjacent seismic traces, and node evidence scoring functions are uniformly organized into a physical candidate event graph, specifically represented as follows:

[0148]

[0149] In the formula, Represents a graph of physical candidate events; This represents the set of nodes consisting of all candidate event nodes; This represents the set of reachable edges between adjacent candidate nodes of seismic traces; Represents the node evidence scoring function; Indicates the first The set of candidate nodes for the Dao; Indicates the earthquake track number; "Indicates the total number of seismic traces" Find the union of the graphs; the three items in parentheses give the nodes, edges, and node scores of the graph, respectively.

[0150] The set of nodes, the set of reachable edges, and the node scoring function constitute the above. From all The result of merging is Depend on , and Jointly determined, From each node Composition; the three elements constitute . Then, together with the initial anchor point, we input the graph into dynamic programming to determine... , , And suboptimal paths.

[0151] To construct the adjacent path connections in the candidate event graph, the target position of the next path is first predicted based on the position of the previous candidate node and the local slope, specifically as follows:

[0152] ;

[0153] In the formula, Indicates candidate nodes The predicted time or depth location of the next seismic trace; Indicates the index of the candidate node for the next seismic trace; Represents a node Time or depth location; Represents a node The local slope obtained by prior conversion of the tilt angle; Indicates the lateral spacing between adjacent seismic traces; product This indicates the predicted vertical displacement between tracks; the plus sign indicates that the displacement will be added to... Above. The aforementioned From the previous candidate node , From , The predicted location is generated from the seismic trace coordinates. The predicted location was then compared with the candidate node. actual location Comparison, formation .

[0154] While obtaining the predicted dip angle location, the allowable vertical jump range is adaptively adjusted based on the basic jump variable and fault support, specifically expressed as follows:

[0155] ;

[0156] In the formula, Represents a node The corresponding adaptive allowable jump variable; Indicates the candidate node index; This indicates taking the smaller of the two values ​​within the parentheses; This indicates the upper limit of allowed jumps; Represents the basic jump variables under fault-free conditions; Indicates the fault support adjustment coefficient; Represents a node The fault support at the location; the plus sign indicates that fault-related additional jump variables are added to the basic jump variables. and Determined by statistical analysis of actual inter-track displacements. Determined by the verification results of the fault zone, From At the node The value at the location; the result Subsequently with The comparison is used to determine reachable edges and to normalize the jump magnitude and related costs.

[0157] Subsequently, the difference between the actual location and the predicted dip position of the next candidate node is calculated to obtain the dip prediction residual between adjacent candidate nodes, which is specifically expressed as:

[0158] ;

[0159] In the formula, Indicates candidate nodes arrive The tilt angle prediction residual; and These represent the indexes of the candidate nodes in the previous and next steps, respectively. Indicates candidate nodes The actual time or depth location; Indicates by The predicted target location; the minus sign indicates the actual location minus the predicted location. From the candidate node of the next seismic trace , From the previous equation; the two form . The reachability determination formula is then input and used for tilt angle prediction bias correlation. calculate.

[0160] Finally, the tilt angle prediction residuals are compared with the adaptive allowable jump variables; a reachable edge is established between two candidate nodes only if the residuals are within the allowable range, with the specific criteria being:

[0161] ;

[0162] In the formula, Represents a node arrive The tilt angle prediction residual; and These represent the indices of two adjacent candidate nodes; Represents a node The corresponding adaptive allowable jump variable; the superscript 2 indicates the square; " "" indicates that a node is active only if the squared residual is not greater than the squared allowable jump variable. and Establish reachable edges between them.

[0163] Based on the residual and the allowed jump variable, the condition simultaneously receives... and A directed connection will be established when the condition is met. Add to edge set And join in If the conditions are not met, no connection is established. This results in... and Images provided separately and recursive formula use.

[0164] Example 4

[0165] like Figure 5 As shown, the candidate event graph identifies selectable physical candidate nodes and reachable edges, but it is still necessary to determine the same reflection event corresponding to the initial anchor point from multiple feasible paths. Before forming the total transition cost of adjacent candidate nodes, a local second-order difference continuity cost is first defined based on the positional changes of three consecutive candidate nodes to constrain the local curvature of the path, specifically expressed as:

[0166] ;

[0167] In the formula, Indicates from candidate nodes Transfer to The cost of local second-order difference continuity; and Indicates the index of adjacent candidate nodes; , and These represent three consecutive upstream precursor nodes. Current node and target node Time or depth location; Indicates the predecessor node index; This indicates the upper limit of the permissible jump; the two brackets represent the displacements between two adjacent traces, and the square of the difference between them is used to measure the local curvature, then divided by... Normalization is performed. , and It comes from three consecutive nodes on a candidate path. From candidate edge parameters; the result As one of the six transfer costs, input It is used to suppress local curvature abrupt changes.

[0168] After obtaining the local continuity cost, it is weighted together with the costs of polarity, waveform, dip angle, fault, and event sequence, and a target node evidence reward is introduced to form the total transfer cost between adjacent candidate nodes, specifically expressed as:

[0169]

[0170] In the formula, Indicates candidate nodes Transfer to The total transfer cost; and Indicates the index of adjacent candidate nodes; , , , , and These represent polarity, waveform, dip angle, fault, event sequence, and local continuity cost, respectively. , , , , and These represent the non-negative weights of the six costs; Indicates the node evidence reward weight; Represents the target node The node evidence score; the plus sign indicates the sum of various costs, and the minus sign indicates the reduction of the total transfer cost by node evidence.

[0171] The six sub-costs are composed of candidate node polarity, local waveform, , The relative order of events and Calculated separately, with weights determined by the validation set; node evidence Scoring from graph nodes. Used for dynamic programming to accumulate costs The recursive calculation is used as the edge cost. Accumulated path cost .

[0172] Before performing the step-by-step recursion, based on the initial anchor point in step 1 ( , Initialize the cumulative cost of candidate nodes. Assume the anchor point tolerance is... Located at the anchor point Candidate nodes on =( , The initial cumulative cost is set according to the following formula:

[0173] ;

[0174] In the formula, Indicates candidate nodes At the anchor point The initial cumulative cost on; Indicates the candidate node index; Indicates the track number where the initial anchor point is located; Indicates candidate nodes Time or depth location; Indicates the time or depth position of the initial anchor point; Indicates anchor point tolerance; superscript 2 indicates square; "" indicates that when the positional deviation between the node and the anchor point does not exceed the anchor point tolerance, the initial cumulative cost is set to 0.

[0175] Satisfying anchor point tolerance and From the initial anchor point determined in step 1, From anchor point candidate nodes, The value is determined by the anchor point picking error; nodes that meet the conditions are given an initial value of 0, which serves as the valid starting state input for dynamic programming to be recursively applied to both sides. .

[0176] For candidate nodes on the anchor path that exceed the anchor tolerance, their initial cumulative cost is set to positive infinity to exclude erroneous starting nodes, specifically expressed as follows:

[0177] ;

[0178] In the formula, Indicates candidate nodes At the anchor point The initial cumulative cost on; Indicates the candidate node index; Indicates the anchor point number; and These represent the time or depth positions of the candidate node and the initial anchor point, respectively. Indicates anchor point tolerance; "Indicates positive infinity;" "" indicates that when the position deviation exceeds the anchor point tolerance, the node is not considered a valid starting node.

[0179] The formula for not satisfying the anchor point tolerance uses the same method as the previous formula. , , and Nodes that do not meet the anchor point tolerance are assigned... Thus in the subsequent It is excluded in the minimization process to avoid the path starting from an error event.

[0180] After initializing the anchor point traces, proceed recursively along the seismic trace directions. For each candidate node, select the node with the minimum sum of cumulative cost and current transfer cost among all its reachable predecessors, specifically expressed as follows:

[0181]

[0182] In the formula, Indicates arrival at candidate node The cumulative minimum cost; Indicates the index of the target candidate node; This indicates that the minimum value is taken among all reachable predecessors; Indicates the index of the predecessor candidate node; " indicates a relationship; Indicates candidate nodes The set of reachable predecessor nodes; Indicates arrival at the predecessor node The cumulative minimum cost; Indicates from Transfer to The transfer cost; the plus sign indicates that the cumulative cost of the path is added to the current transfer cost.

[0183] During the recursive process, From the set of reachable edges Sure, From the result of the previous recursion, Derived from the transfer cost formula; obtained jointly by all three. Simultaneously, the predecessor node that achieves the minimum value is recorded. After all seismic traces are recursively calculated, the cumulative cost and predecessor records are used for backtracking. And suboptimal paths.

[0184] After completing the recursion of all seismic traces and recording the optimal predecessor, the optimal in-phase event path with the minimum cumulative cost is obtained through backtracking, specifically expressed as:

[0185] ;

[0186] In the formula, This represents the optimal path with the minimum cumulative cost. The independent variable represents the value of the objective function that is minimized. Indicates a feasible path; " indicates a relationship; It represents the set of all feasible paths in the physical candidate event graph; Representing a path The cumulative transfer cost.

[0187] The From the diagram The entirety of paths that satisfy the anchor point and reachable edge constraints constitutes the entirety of paths in the middle. The following formula is used to accumulate the path edge costs; the output is then compared. . Subsequently used to determine And together with the suboptimal path, it generates and .

[0188] To clarify the cumulative cost of each feasible path in the above formula, the corresponding transition costs are summed for each of the paths, based on all reachable edges contained within the path. Specifically, this is expressed as:

[0189] ;

[0190] In the formula, Indicates a feasible path The cumulative transfer cost; Indicates a feasible path; " indicates summation; This represents a reachable edge in the path; " indicates a relationship; Represents reachable edges The corresponding transfer cost; the formula represents the cost of the path. Sum the costs of all reachable edges in the array.

[0191] Each edge belong The cost Corresponding adjacent nodes Summing all edges of a path yields... All paths Used for selection The cost difference between the best and second-best paths, after normalization, forms the following: .

[0192] Example 5

[0193] The evidence consistency graph dynamic programming outputs the path with the most consistent evidence, but in areas with weak reflection, complex faults, and strong local interference, there may still be multiple feasible paths with similar costs. Let... Represents the optimal path In the Candidate event nodes on the path; optimal path and suboptimal path In the The cumulative cost difference before the path is normalized as follows: ( ),node The evidence score, gating candidate support, and local waveform similarity after normalization are respectively ( ), ( )and ( The normalized vertical jump amplitude of adjacent lanes is as follows: ( The path reliability is defined as follows:

[0194] ;

[0195] In the formula, Indicates the first The credibility of the path of Tao; Indicates the earthquake track number; , , , and This indicates the non-negative weights of the five credibility indicators; Represents the cumulative cost difference between the normalized optimal path and the second-best path; subscript This indicates that the corresponding indicator has been normalized to 0 to 1; Represents the optimal path node Normalized node evidence; Represents a node Normalized gating candidate support; Represents a node Similarity to the normalized local waveform of the previous optimal path node; Indicates the first The vertical jump amplitude of the normalization; the plus sign indicates the accumulation of positive and credible evidence, and the minus sign indicates that the credibility is reduced by the jump amplitude.

[0196] The The cumulative cost difference between the optimal and suboptimal paths, Evidence from nodes From gating candidate support, Waveform comparison from adjacent optimal path nodes, From Inter-channel jump; five-item weighted average obtained . Then enter Determine and write it together with the layer position. .

[0197] After obtaining the path-by-path credibility, the credibility threshold and credibility jump threshold are further combined to determine whether the result of that path layer is credible; when both conditions are met simultaneously, the credibility mask is set to 1, specifically expressed as:

[0198] ;

[0199] In the formula, Indicates the first A credible mask for the Tao; Indicates the seismic trace number; a value of 1 indicates that the stratigraphic result for that trace is reliable. Indicates path reliability; Indicates the confidence threshold; Indicates the normalized vertical jump amplitude; Indicates the reliable jump threshold; "and" "" indicates not lower than and not higher than; the comma indicates that both conditions must be met simultaneously. From the previous formula, From the optimal path Inter-track displacement. and Determined by validation data; output when both conditions are met. =1. This result was subsequently used as... A trusted marker is used to preserve the corresponding layer position.

[0200] Corresponding to the aforementioned trustworthy scenarios, when at least one of the path trustworthiness condition or vertical jump condition is not met, the trustworthiness mask is set to 0, specifically as follows:

[0201] ;

[0202] In the formula, Indicates the first A credible mask for the Tao; This indicates the seismic trace number; a value of 0 indicates that the stratigraphic result for that trace is unreliable. This indicates that the condition in the previous formula is not satisfied. and At least one of the two conditions.

[0203] This formula uses the same method as the previous one. , , and Determination result; output when any condition is not met. =0. Continuous unreliable results are further compared with a low-reliability continuity length threshold to form unreliable segments that are not forced to be completed.

[0204] After obtaining the optimal path location, path confidence level, and confidence mask, these three are combined to form the final output of two-dimensional seismic horizon tracking, specifically represented as follows:

[0205] ;

[0206] In the formula, This represents the final output set of seismic horizon tracking; the superscript " " indicates that the result corresponds to the optimal path; Indicates the first The final layer, time, or depth location of the Dao; Indicates the earthquake track number; Indicates the first The credibility of the path of Tao; Indicates the first The credibility mask of the path; curly braces indicate that the final output consists of the above three results.

[0207] The From the optimal path In the The node positions of the path are obtained. Calculated based on competing paths and evidence along the route, Determined by a confidence threshold; composed of three factors This serves as the final output for two-dimensional tonal tracing. In three-dimensional applications, the values ​​in each direction... Continue inputting inline and crossline directions to fuse them in a consistent manner, forming a three-dimensional layer plane and a three-dimensional reliable mask.

[0208] In 3D seismic volume applications, for each location to be tracked, candidate node selection, reachable edge construction, and anchor point constraint path optimization are performed along both the inline and crossline directions to obtain the candidate layer locations and their reliability in both directions. Let the directional consistency threshold be... If the time or depth difference between two directions at the same spatial location is less than Then, a weighted fusion is performed based on the path credibility of the two directions; if the difference is not less than... If the credibility of that location is reduced or it is marked as untrustworthy, a three-dimensional layer plane and a three-dimensional credibility mask are formed.

[0209] In this embodiment, all parameters are determined on independent verification data and remain fixed on the data to be tracked. Among them, the input size... , , and Determined by data window and channel configuration; network parameters , and The dilation rate of the dilated convolution obtained through training. and scale mapping operations Pre-defined by network structure; task loss weights to Loss was determined through multi-task validation. The peak confidence level was weighted. , and The fault fusion coefficient is determined based on the ability to distinguish between peak and valley candidates. The first peak-valley threshold is determined based on the verification results of fault priors and seismic fault properties. Second peak and valley threshold The candidate support threshold is determined by the confidence distribution of strong and weak reflection peaks and valleys. and uncertainty threshold The tilt angle difference scale parameter is determined by the weak candidate recall rate and the false positive rate. Determined by the error distribution between network tilt and attribute tilt. Basic jump variables. Maximum jump limit and fault adjustment coefficient The fault threshold is determined by statistical analysis of actual inter-trace displacements in the continuous and fault zones. Determined by the fault support distribution of fault and non-fault samples. Anchor point tolerance. Determined by the error in manual anchor picking; node evidence weight to Transfer cost weight to and node reward weight Determined by reducing the cross-level rate and abnormal jump rate; credibility weight to Confidence threshold Trustworthy jump threshold and low confidence continuous length threshold Determining the directional consistency threshold by reducing the low-confidence false positive rate; The difference distribution between inline and crossline directional tracking is determined. The input sources, functions, and output effects of each parameter correspond to network training, candidate admission, node weighting, candidate edge construction, path recursion, credibility evaluation, and 3D directional fusion, respectively.

[0210] Example 6

[0211] like Figure 6 As shown, the present invention also provides a seismic horizon tracking device, including a data acquisition module 610, a weak prior generation module 620, a physical gating module 630, a candidate event graph construction module 640, a path optimization module 650, and a result output module 660. The data acquisition module 610 outputs seismic data processed by coordinates and amplitude, initial anchor points, and the tracking range; the weak prior generation module 620 outputs... , , , , and The physical gating module 630 takes as input the prior and peak / valley, coherence, and fault evidence, and outputs strong / weak candidate nodes and... ; Candidate event graph construction module 640 input candidate nodes, and fault support, output ; Path optimization module 650 input The system takes two paths and evidence along the paths as input, and outputs the optimal and suboptimal paths. The result output module 660 takes two paths as input and outputs the layer results and the credible mask.

[0212] The present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. The processor reads seismic data and initial anchor points, and executes the computer program according to the data transmission order of the aforementioned modules to achieve candidate admission, candidate map construction, anchor point constraint path optimization, and reliable result output. The present invention also provides a computer-readable storage medium, on which a computer program stored is executed by the processor to implement the above methods.

[0213] The above embodiments are used to illustrate the processing relationship between various variables, parameters, and output results, and do not limit the scope of protection. Without changing the technical concept of "depth weak prior subject to physical gating constraints, strong and weak candidate differentiated admission, dip angle and fault jointly limiting reachable edges, anchor point constraint graph dynamic programming, and credibility evaluation based on competitive paths", equivalent substitutions or combinations made by those skilled in the art can fall within the scope defined by the corresponding claims.

Claims

1. A seismic horizon tracking method based on a physically gated candidate event map, characterized in that, include: Step S1: Obtain the seismic profile or 3D seismic body to be tracked, perform sampling coordinate unification and inter-trace amplitude consistency processing on the seismic profile or 3D seismic body, and determine at least one initial anchor point and the tracking range of the target layer. Step S2: Construct the original seismic flow and attribute / physical flow with one-to-one spatial location within the range to be tracked, input the trained dual-flow tracking perception prior network, and generate a depth weak prior including in-phase event prior, peak-valley-ridge prior, fault prior, uncertainty prior, dip angle prior and comprehensive support prior; Step S3: Calculate the physical gating value based on the peak and valley candidate mask, the comprehensive peak confidence, the local coherence support, and the fault support. Multiply the physical gating value by the comprehensive support prior position by position to obtain the gated candidate support. Step S4: Select candidate event nodes using the strong and weak candidate differential admission rule, and construct a physical candidate event graph based on all candidate event nodes, node evidence scores, and reachability edges between adjacent seismic traces; Step S5: For two candidate event nodes on adjacent seismic traces, calculate the predicted dip position based on the time or depth location, local dip angle, and trace spacing of the previous candidate event node, and adaptively adjust the maximum allowable vertical jump based on fault support; perform dynamic programming of the evidence consistency diagram under the initial anchor point constraint to determine the optimal path and the second-best path. Step S6: Calculate the path credibility based on the cumulative cost difference between the optimal path and the suboptimal path, the node evidence along the optimal path, the gated candidate support, the local waveform similarity, and the vertical jump amplitude; mark continuous segments with credibility below the credibility threshold as untrustworthy segments and do not perform forced interpolation completion, and output the two-dimensional seismic horizon trajectory or three-dimensional horizon surface and the corresponding credibility mask. In step S3, when the comprehensive support prior is high but lacks evidence of peaks, valleys, coherence, or fault structures, the position is suppressed; when the weak reflection position has corresponding physical evidence, the position is retained, and the value of the physical gate value ranges from 0 to 1; the gated candidate support is obtained by multiplying the physical gate value by the comprehensive support prior; the physical gate value only controls the degree to which the depth weak prior participates in the weak candidate supplementation and node evidence weighting, and does not change the event polarity and candidate qualification of strong peak-valley candidates that reach the first peak-valley threshold; The node evidence score is determined by the peak and valley intensities at the candidate node. Local coherent support Tilt angle consistency Fault support The results were obtained by weighting the gated candidate supports; among them, fault supports... Participating node evidence, primarily used to regulate transitions; In step S5, reachable edges are established only when the residual of the candidate event node relative to the dip angle prediction position does not exceed the maximum allowable vertical jump and the polarity consistency and event sequence constraints are satisfied. The optimal path and the second-best path are determined based on polarity consistency, local waveform similarity, dip angle prediction residual, fault adaptive jump, event sequence, local continuity and node evidence score. The maximum permissible vertical jump consists of a basic jump variable and an additional jump variable positively correlated with fault support. When the fault support is below the fault threshold, candidate nodes are only connected to consecutive candidate nodes near the dip angle prediction position. When the fault support reaches the fault threshold, the connection range is allowed to be expanded within a preset upper limit, while maintaining the order of the target event relative to the upper and lower neighboring events. The transition costs between adjacent candidate event nodes include polarity inconsistency cost, local waveform dissimilarity cost, dip angle prediction residual cost, fault adaptive jump cost, event sequence violation cost, local curvature or second-order difference continuity cost, and node evidence reward; each cost item is weighted and combined after normalization.

2. The seismic horizon tracking method based on a physically-gated candidate event map according to claim 1, characterized in that, In step S1, the initial anchor point includes a channel number and a time sampling position or depth sampling position corresponding to the channel number; The original earthquake record is as follows: , height and width These correspond to the number of sampling points and seismic traces in the seismic window to be processed, respectively, and the number of channels. Determined by the current profile and the number of its adjacent profiles; the attribute / physical flow record is... Its height and width are the same as Same, number of channels Determined by the number of seismic attributes used; and Align the time or depth sampling location and channel number direction point by point.

3. The seismic horizon tracking method based on a physically-gated candidate event map according to claim 1, characterized in that, In step S2, the original seismic flow is constructed using the current seismic profile and its adjacent seismic profiles within the range to be tracked. At least two of the following properties / physical flows are constructed from the local gradient, ridge response, local energy, local coherence approximation, multi-scale difference Gaussian response, peak-valley response, dip properties, and fault support calculated from the same seismic data window. Each prior is used only as evidence for subsequent candidate screening and path optimization, and is not directly used as the final horizon. The dual-stream tracking and sensing prior network includes a raw seismic stream encoder, an attribute / physical stream encoder, a gated fusion module, a scale-gated module, a tracking and sensing context module, a multi-scale aggregation output module, and a multi-task prediction head. The two encoders output raw seismic features and attribute features at multiple scales, respectively. The gated fusion module adjusts the participation ratio of attribute features using a gate weight jointly calculated by the two types of features, and inputs the fusion result into the tracking and sensing context module and the multi-scale aggregation output module. The training samples of the dual-flow tracking perception prior network include spatially corresponding original seismic flows, attribute / physical flows, and in-phase event labels, peak-valley-ridge line labels, fault labels, dip angle labels, and effective annotation masks. The labels are used to constrain the in-phase event priors, peak-valley-ridge line priors, fault priors, dip angle priors, and comprehensive support priors, respectively, and uncertainty supervision or prediction residuals are used to constrain the uncertainty priors. After training, the network parameters are fixed so that each output prior maintains the same spatial coordinates as the input seismic window.

4. The seismic horizon tracking method based on a physically-gated candidate event map according to claim 3, characterized in that, In step S4, strong peak-valley candidates whose comprehensive peak confidence reaches the first peak-valley threshold are directly used as candidate event nodes. Weak peak-valley candidates whose comprehensive peak confidence is between the second peak-valley threshold and the first peak-valley threshold are only used as supplementary candidate event nodes when the gating candidate support reaches the candidate support threshold and the uncertainty prior is not higher than the uncertainty threshold. The first peak-valley threshold is greater than the second peak-valley threshold. The strong peak-valley candidates enter the physical candidate event graph through the strong candidate direct branch, and the weak peak-valley candidates enter the physical candidate event graph through the physical gating supplementary branch. The uncertainty prior is used to exclude weak peak-valley candidates with high network response but insufficient prediction stability, so that the deep weak prior only supplements the candidate positions with the lowest peak-valley evidence and physical support.

5. The seismic horizon tracking method based on a physically-gated candidate event map according to claim 1, characterized in that, When performing dynamic programming of the evidence consistency graph, only candidate event nodes that are located in the same path or a preset neighborhood as the initial anchor point and whose time or depth position deviation does not exceed the anchor point tolerance are set as valid starting nodes; the initial cumulative cost of other candidate event nodes is set to infinity or an anchor point deviation penalty is added. After completing the recursion step by step, the optimal path with the minimum cumulative cost and the second-minimum cumulative cost path are obtained by backtracking. The path credibility includes at least the path discrimination degree corresponding to the cumulative cost difference between the optimal path and the second-best path; when any one of the path discrimination degree, node evidence and gating candidate support is lower than the corresponding threshold, or when the continuous vertical jump exceeds the credibility jump threshold, the corresponding path position is marked as untrustworthy.

6. The seismic horizon tracking method based on a physically-gated candidate event map according to claim 1, characterized in that, In 3D seismic body plane tracking, physical candidate event maps are constructed along the inline and crossline directions respectively, and direction tracking results are obtained. Consistency judgment is made on the time or depth difference between the two directions at the same spatial location. The locations that meet the consistency threshold are weighted and fused, and the confidence of the locations that do not meet the consistency threshold is reduced or marked as untrusted areas.

7. A seismic horizon tracking device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the seismic horizon tracking method according to any one of claims 1-6.

Images