A method for single-fan subdivision and distribution characterization constrained by synsedimentary faults

Abstract

The application provides a single-period fan body subdivision and distribution characterization method constrained by synsedimentary faults. A fan classification body, a sand thickness body, a fold number body and a fault distance field are obtained; a single-period fan body anchor is prepared by fusing three sources of well GR cycle, stratigraphic dip and progradation dip field; a multi-channel input body is constructed, features are extracted by a double-distance field grouping gate encoder, and a gate diagram is divided into a main fault group and an adjusting fault group to modulate corresponding channels to decouple differential control; a single-period fan body sand thickness is forced to monotonically decay along a source and meet a retrogradation or progradation sequence through joint optimization of four loss terms of anchor body cross entropy, smoothing constraint, geological attenuation and period sequence; and a single-period fan body classification result is output and distribution quantification and fault staged activity characterization are performed. The application is used for fine geological modeling of multi-period fans controlled by synsedimentary faults, and serves well site deployment and reserve calculation.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, specifically to a method for subdividing and characterizing single-stage fan bodies constrained by syn-sedimentary faults, applicable to three-dimensional geological modeling of multi-stage fan-type sedimentary systems controlled by syn-sedimentary faults in rift basins. Background Technology

[0002] In oil and gas exploration in rift basins, multi-stage fan-shaped sedimentary systems controlled by syn-sedimentary faults (such as fan deltas, underwater fans, turbidite fans, and alluvial fans) have significant reservoir value. These sedimentary systems are characterized by rapid near-source accumulation, multiple superpositions, and phased activity under the influence of syn-sedimentary faults, resulting in extremely complex internal structures.

[0003] While existing technologies can classify and predict sand thickness for multi-stage fan bodies, they struggle to provide fine-grained subdivision and quantitative characterization of single-stage fan bodies. Specifically: existing deep learning networks (such as Swin-Unet) have fixed receptive field shapes, failing to adapt to the long-axis distribution characteristics of fan bodies along the source direction; graph attention networks (GAT), while capable of handling fault block topological relationships, lack the incorporation of geological priors such as fault strike and valley orientation; multi-source supervised data (well GR curves, stratigraphic dip angles, preseismic accretionary axis attitudes, etc.) have not been effectively fused to prepare anchor points for single-stage fan bodies; and explicit constraints on the sand thickness attenuation patterns and stage stacking sequences of single-stage fan bodies along the source direction are lacking. Therefore, a method is urgently needed that can comprehensively utilize multi-source data to achieve fine-grained subdivision and characterization of single-stage fan bodies under the constraint of syn-sedimentary faults. Summary of the Invention

[0004] The purpose of this invention is to provide a method for subdividing and characterizing single-stage fan bodies constrained by syn-sedimentary faults, so as to solve the problem in the prior art that it is impossible to perform fine subdivision and quantitative characterization of single-stage fan bodies in multi-stage fan-type sedimentary systems controlled by syn-sedimentary faults.

[0005] To achieve the above objectives, the technical solution of the present invention includes the following steps:

[0006] Multi-phase sector overlay and multi-source supervision data acquisition. Acquire multi-source data for the work area, including at least: sector classification data. (The code contains at least two fan body oil group categories); by The resulting fan-shaped mask I_fan = I( (∈ Fan-type); Thick Sand Body Number of superpositions The system includes: a signed distance field d1 for the main syn-sedimentary fault and a signed distance field d2 for the moderating syn-sedimentary fault; a step-code S ∈ {0, 1, ..., K}; a fault block sedimentary attribute vector {x_i} (N, ≥ 8); and a fault block adjacency matrix A (N, N). The 0th and 1st dimensions of the fault block sedimentary attribute vector represent the mean sand thickness within the fault blocks of the two fan-shaped oil groups, respectively, constituting the fractal dimension representation of the two fan-shaped oil groups in the fault block attribute space: the first fan-shaped oil group corresponds to a fault block cluster with the 0th dimension significantly greater than the 1st dimension, and the second fan-shaped oil group corresponds to a fault block cluster with the 1st dimension significantly greater than the 0th dimension. The two fan-shaped oil groups correspond to two fan deltaic oil groups in a dual-source fan delta, or to a fan-shaped oil group after merging multiple phases of a single-source fan and a secondary division, with a single-phase fan total M ≥ 3. The above multi-source data can be obtained by the Swin-Unet fusion modeling method constrained by syn-sedimentary faults and graph attention network, or by other syn-sedimentary fault identification and fan classification processes. Simultaneously, multi-source monitoring data are acquired, including: wellhead monitoring sources (well GR curve, well point stratigraphic dip) and seismic monitoring sources (fore-accretion axis attitude interpreted by seismic horizons, and optionally include the fore-accretion dip field φ(x,y,z) of the overlay modeling output of the fusion of upstream Swin-Unet and graph attention network).

[0007] Preparation of single-stage sector anchors. Single-stage sector anchors (three-source fusion monitoring tag) W_sub(D,H,W) are prepared by fusing three types of anchor sources. The coding rules are as follows: -1 indicates no well point, c=1~M correspond to the early / middle / late single-stage sectors within each sector oil group (M≥5 for dual sources, or M≥3 for single-source multi-stage sectors). First source (GR cycle): The mudstone interlayer corresponding to the local maximum GR value of the well serves as the boundary of the single-stage sector. Early / middle / late single-stage sector codes are obtained by segmenting within each sector oil group. Second source (well point formation dip angle): Along the source direction, the dip angle abrupt change serves as the well point auxiliary anchor for the single-stage inter-section. Third source (pre-accretion dip angle field φ): If φ has already been output from upstream, the contour line cluster along the source direction φ serves as the voxel-level soft anchor for the single-stage sector's distribution direction, fused with the first source GR hard anchor. After the anchor point is interpolated to a voxel along the well trajectory, a 3×3×3 structural element expansion is performed. Voxels of the same oil group within the expansion zone are assigned the early single-stage sector code by default. If multiple single-stage sector codes appear in the expansion neighborhood, the early single-stage sector code takes priority. If a voxel is assigned an early single-stage sector code by GR expansion, but the φ contour line indicates that it belongs to a late pre-accretionary sequence, the voxel is reassigned to the corresponding late code with priority given to the φ soft anchor. The reassignment operation only takes effect within the sector mask I_fan.

[0008] Construct a multi-channel three-dimensional input volume X∈R^{8×D×H×W}, specifically by: dividing d1, d2, S, ... , , At least eight types of attributes, including I_fan and well point indication, are superimposed along the channel dimension; optionally, the pre-product dip angle field φ is incorporated into X as the 9th channel;

[0009] A dual-range field grouped gated encoder is used. An encoder-decoder architecture is employed, with the encoder being either a MedNeXt block sequence or a Swin-Unet block sequence. The encoder path is enc1 → pooling → enc2. The output of enc2 is multiplied channel-by-channel by a fault activity gating map. This gating map is generated by sigmoid transformation (range [0,1]) of channels d1 and d2 in the input volume X after 1×1×1 convolution. Its output channels are divided into two groups according to their indices: the active group of primary co-depositional faults (channel indices 0~C / 2-1) and the active group of moderate co-depositional faults (channel indices C / 2~C-1), which modulate the corresponding channel groups of enc2, thus differentiating the feature modulation weights of the two faults. After decoding, a 1×1×1 classification head outputs M+1 classes of single-period sectors (background 0 + M single-period sectors 1~M). When the encoder is MedNeXt, the configuration is as follows: Patch Size 2×4×4; the number of channels in the 4 Stages is configured as 96 / 192 / 384 / 768 according to the composite scaling orthogonal configuration; each Stage contains 2 MedNeXt blocks, and the block is a residual inverted bottleneck structure (1×1 upscaling → 5×5×5 depthwise convolution UpKern → 1×1 downscaling), equipped with GroupNorm+GELU; enc2 is the output of the 2nd Stage, with 192 channels; the gated graph convolution input has 2 channels and the output has 192 channels, grouped into 0~95 and 96~191. When the encoder is Swin-Unet, the configuration is as follows: PatchSize 2×4×4; 4 Swin Stages with 96 / 192 / 384 / 768 channels; each Stage contains 2 Swin TransformerBlocks, Window Size 7×7×7, with alternating shift windows; enc2 has 192 channels, and the gating graph grouping is the same as the MedNeXt configuration; the major axis of the windows in each Swin Stage is stretched along the composite fault direction τ_synth of d1 and d2, with a stretching ratio... τ_synth is the weighted synthesis of the tangents d1 and d2.

[0010] Step (5) Joint optimization of four constraint losses

[0011] Total loss .

[0012] L_CE is the anchor cross-entropy, calculated only for positions where W_sub≠−1;

[0013] , where P_c is the probability body of the c-th type of single-stage fan within the fan mask.

[0014] , where

[0015] is the sand thickness probability centroid of the single-stage fan c along the source direction w, is the partial derivative along w, that is, the attenuation rate along the provenance direction, δ is the empirical lower limit of the long-direction attenuation rate of the single-stage fan in the work area along the provenance direction, which is obtained by fitting the sand thickness of well points in the work area along the provenance direction, with a value range of δ∈[0.05, 0.15] (unit: m / voxel); L_geo forces the sand thickness of the single-stage fan to decay monotonically along the provenance direction.

[0016] , where is the sand thickness probability centroid coordinate of the single-stage fan c along the provenance direction, c1, c2, c3 are the centroid serial numbers of the early / middle / late single-stage fans in the same fan oil group; L_seq forces the three stages of early / middle / late single-stage fans in the same fan oil group along the provenance direction to satisfy the retrogradation sequence w{early}<w{middle}<w{late} or the progradation sequence w{early}>w{middle}>w{late}, which is selected according to the sequence type of the work area: the retrogradation sequence corresponds to the lacustrine transgression background, and the progradation sequence corresponds to the lacustrine regression background; margin1 and margin2 are empirically determined by the retrogradation distance of the early / middle / late single-stage fans in the same fan oil group in the work area, with a value range of margin∈[2,5] voxels;

[0017] The loss weights λ1, λ2, λ3 are determined by grid search on the validation set of the work area, with the search ranges λ1∈[0.1, 0.3], λ2∈[0.03, 0.1], λ3∈[0.05, 0.2];

[0018] Step (6): Quantification of single-stage fan distribution and characterization of episodic activity of syndepositional faults. Connected domain analysis is performed on the predicted single-stage fans to obtain the number of lobes;

[0019] On the H-W plane, using the sand thickness as the weight, weighted PCA is performed on the voxel coordinates where P_c>0.5, to obtain the distribution orientation and the ratio of major to minor axis of each single-stage fan ;

[0020] Along the direction, divide into K intervals (bins), and count the average probability of the single-stage fan c in the k-th interval , and calculate the episodic activity index in combination with the step code S:

[0021]

[0022] Where k is the interval index, Let A_activity(k) be the average probability of a single-stage fan c within the k-th interval. A least-squares linear fit is performed on A_activity(k) as k increases to obtain the slope b and p values. A threshold θ = α / M is set, where α ∈ [0.3, 1.0]. If b > θ and p < 0.05, then the trend of "single-stage fan increasing with distance from the fault" is significant. Combined with the step code S: if the probability of an early single-stage fan is highest in the bin with a smaller S, and the probability of a late single-stage fan increases in the bin with a larger S, then the corresponding syn-depositional fault exhibits a phased activity pattern of "youth controlling early fans and old fans controlling late fans."

[0023] The multi-stage fan-type sedimentary systems include at least one of the following: fan delta, underwater fan, turbidite fan, and alluvial fan.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) By using three-source fusion to supervise the preparation of the tag body, and making full use of multi-source data such as well GR cycle, well point formation dip angle and seismic pre-accretion dip field, the voxel-level accurate calibration of single-period sector anchor points was achieved, which solved the problem of lack of fine supervision of single-period sectors in existing methods.

[0026] (2) By using a dual-distance field grouped gating encoder, the modulation weights of the main co-depositional fault and the regulating co-depositional fault on different channel groups of the network are decoupled, and the differential control effects of the two-way fault on different single-stage fan bodies can be learned separately, which enhances the network's ability to express the coupling of fault activity and deposition process.

[0027] (3) By jointly optimizing the four constraint losses, the monotonic decay of the single-period fan sand thickness along the source direction and the regression or progression of the period sequence are explicitly forced, which greatly improves the geological rationality of the prediction results.

[0028] (4) By quantifying the distribution and characterizing the phased activity, a detailed single-phase sector configuration model is provided for well location deployment and reserve calculation, which can directly serve the evaluation of favorable areas and the selection of exploration targets. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the multi-source data acquisition and preprocessing process in the method of the present invention.

[0030] Figure 2 This is a flowchart of the three-source fusion process for the preparation of a single-stage sector anchor body in the method of this invention.

[0031] Figure 3 This is a schematic diagram of the network architecture of the dual-distance field grouped gate encoder in the method of the present invention.

[0032] Figure 4 This is a flowchart illustrating the calculation process for the joint optimization of four constraint losses in an embodiment of the present invention.

[0033] Figure 5 This is a schematic diagram showing the results of the quantification of single-stage sector distribution and the characterization of phased activity in the method of the present invention, where A and B are material sources and the arrows indicate the direction of the material sources. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0035] The specific implementation of this invention follows the steps described in the invention summary, as follows:

[0036] Specific implementation of step (1): Obtain the post-stack seismic volume of the work area, with a voxel spacing of 12.5m × 12.5m × 5ms. Obtain the sector classification volume obtained by the Swin-Unet and graph attention network fusion modeling method constrained by the same sedimentary fault (i.e., the first patented method). Where D=30 (voxels in the depth direction, corresponding to approximately 150ms), H=200 (inline direction), and W=200 (crossline direction). The code contains at least two fan oil group categories (e.g., early fan oil group, mid-term fan oil group, late fan oil group). [By...] Obtain the fan-shaped mask I_fan = I( ∈ sector class), that is, when When the value is set to the fan body oil group category, I_fan=1; otherwise, I_fan=0. Simultaneously, the sand thickness is obtained. Number of superpositions The system includes: a signed distance field d1 for the main syn-sedimentary fault, a signed distance field d2 for the moderating syn-sedimentary fault, and a step code S∈{0,1,…,K} (K=4, corresponding to the Paleogene IV rift). The block sedimentary attribute vector {x_i} has a dimension N⩾8, where the 0th and 1st dimensions are the mean sand thicknesses within the fault blocks of the two fan-shaped oil groups, respectively, constituting the fractal dimension representation of the two fan-shaped oil groups in the block attribute space: the first fan-shaped oil group corresponds to a block cluster with the 0th dimension significantly greater than the 1st dimension, and the second fan-shaped oil group corresponds to a block cluster with the 1st dimension significantly greater than the 0th dimension. The above multi-source data can also be obtained from other syn-sedimentary fault identification and fan-shaped classification processes. Multi-source monitoring data includes: wellhead monitoring sources (well GR curves, wellpoint stratigraphic dip angles) and seismic monitoring sources (pre-accretion axis attitude interpreted by seismic stratigraphy, and optionally includes the pre-accretion dip angle field φ(x,y,z) output of the overlay modeling fusion of upstream Swin-Unet and graph attention network).

[0037] The specific implementation of step (2): A single-stage sector anchor body (three-source fusion monitoring tag body) W_sub(D,H,W) is prepared by fusing three types of anchor sources. The coding rule is: -1 indicates no well point, c=1~M correspond to the early / middle / late single-stage sector within each sector oil group. The total number of single-stage sector bodies is M⩾5 (when there are two sources, each sector oil group is subdivided into early / middle / late three stages, for a total of 6 stages) or M⩾3 (when there is a single source and multiple stages). First source (GR cycle): The mudstone interlayer corresponding to the local maximum value of well GR is used as the boundary of the single-stage sector body. The early / middle / late single-stage sector body code is obtained by segmenting within each sector oil group. Second source (well point formation dip angle): Along the source direction, the dip angle change point is used as the well point auxiliary anchor of the single-stage interval section. The third source (pre-accretion dip field φ): If φ has already been output from upstream, the contour cluster along the source direction φ serves as a voxel-level soft anchor for the single-phase fan distribution direction, merging with the first source GR hard anchor. After the anchor point is interpolated to a voxel along the well trajectory, a 3×3×3 structural element expansion is performed. Voxels of the same oil group within the expansion zone are assigned the early single-phase fan code by default; if multiple single-phase fan codes appear in the expansion neighborhood, the early single-phase fan code takes priority. If a voxel is assigned an early single-phase fan code by GR expansion, but the φ contour line indicates that it belongs to a late pre-accretion sequence, the voxel is reassigned to the corresponding late code with priority given to the φ soft anchor. The reassignment operation only takes effect within the fan mask I_fan.

[0038] The specific implementation of step (3): Construct a multi-channel three-dimensional input body X∈R^{8×D×H×W}, specifically: d1, d2, S, , , At least 8 types of attributes, including I_fan and well point indication, are superimposed along the channel dimension; optionally, the pre-product dip angle field φ is incorporated into X as the 9th channel, at which point the input volume size is 9×30×200×200.

[0039] The specific implementation of step (4) is as follows: An encoder-decoder architecture is adopted, and the encoder is a MedNeXt block sequence. The MedNeXt block contains a residual inversion bottleneck, a 5×5×5 UpKern kernel, and composite scaling. The encoder path is: enc1 → pooling → enc2. The output of enc2 is multiplied channel by channel of the fault activity gating map. The fault activity gating map is generated by sigmoid after 1×1×1 convolution of the two channels d1 and d2 in the input volume X (2 input channels, the output has the same number of channels as enc2), and the value range is [0,1]. The output channels of the gating map are divided into two groups according to the index: the active group of the main co-depositional fault (channel index 0~C / 2-1) and the active group of the modulating co-depositional fault (channel index C / 2~C-1), which modulate the corresponding channel group of enc2 respectively, so that the feature modulation weights of the two faults are differentiated. After decoding, the M+1 class single-period sector (background 0+M single-period sector 1~M) is output by a 1×1×1 classification head. Specifically, MedNeXt is configured as follows: Patch Size is 2×4×4; the number of channels in the 4 Stages is configured orthogonally by compound scaling as 96 / 192 / 384 / 768; each Stage contains 2 MedNeXt blocks, with a residual inversion bottleneck structure within each block: 1×1 upscaling → 5×5×5 depthwise convolution UpKern → 1×1 downscaling, equipped with Group Norm and GELU activation; enc2 is the output of the second Stage, with 192 channels; gating... Figure 1 The ×1×1 convolution has 2 input channels and 192 output channels, grouped into 0~95 and 96~191. When the encoder is Swin-Unet, the configuration is as follows: PatchSize 2×4×4; 4 Swin Stages with 96 / 192 / 384 / 768 channels; each Stage contains 2 SwinTransformer Blocks, with a Window Size of 7×7×7, and alternating shift windows; enc2 has 192 channels, and the gating graph grouping is the same as MedNeXt; the major axis direction ω of each Swin Stage window is stretched along the composite fault direction field τ_synth of d1 and d2, with a stretching ratio... , where η=1.2.

[0040] Specific implementation of step (5): Total loss function: The anchor cross-entropy loss is calculated only at positions where W_sub≠−1.

[0041] P_c is the probability volume of the c-th single-period sector within the sector mask.

[0042] ,in

[0043] is the sand thickness probability centroid of the single-stage fan body c along the source direction w, is the partial derivative along w, that is, the attenuation rate along the source direction; δ is the empirical lower limit of the long-direction attenuation rate of the single-stage fan body in the work area along the source direction, which is obtained by fitting the sand thickness of well points in the work area along the source direction, with a value range of δ∈[0.05,0.15] (unit: m / voxel); this loss forces the sand thickness of the single-stage fan body to decay monotonically along the source direction.

[0044] , where is the sand thickness probability centroid coordinate of the single-stage fan body c along the source direction, c1, c2, c3 are the centroid serial numbers of the early / middle / late single-stage fan bodies in the same fan body oil formation; L_seq forces the three early / middle / late single-stage fan bodies in the same fan body oil formation to satisfy the retrogradation sequence w{early}<w{middle}<w{late} or the progradation sequence w{early}>w{middle}>w{late} along the source direction, one of which is selected according to the sequence type of the work area: the retrogradation sequence corresponds to the lake transgression background, and the progradation sequence corresponds to the lake regression background; margin1 and margin2 are empirically determined by the retrogradation distance of the early / middle / late single-stage fan bodies in the same fan body oil formation in the work area, with a value range of margin∈[2,5] voxels. The loss weights λ1, λ2, λ3 are determined by grid search on the validation set of the work area, and the search ranges are: λ1∈[0.1,0.3], λ2∈[0.03,0.1], λ3∈[0.05,0.2]. The Adam optimizer is used for training, with an initial learning rate of 1×10 -4 , a batch size of 1, and 80 training epochs. The learning rate is attenuated to 1×10 using the cosine annealing strategy -6 , and the maximum norm of gradient clipping is 1.0.

[0045] The specific implementation of step (6): perform connected domain analysis on the predicted single-stage fan bodies to obtain the number of lobes of each single-stage fan body. On the H-W plane, H takes the sand thickness as the weight to perform weighted PCA on the voxel coordinates with P_c>0.5, so as to obtain the distribution orientation and the ratio of long axis to short axis of each single-stage fan body . Along the direction, K intervals (bins) are divided, and the average probability of the single-stage fan body c in the k-th interval is counted , and the staged activity index is calculated in combination with the step code S:

[0046] , where k is the interval index, Let A_activity(k) be the average probability of a single-stage fan c within the k-th interval. A least-squares linear fit is performed on A_activity(k) as k increases to obtain the slope b and p values. A threshold θ = α / M is set, where α ∈ [0.3, 1.0]. If b > θ and p < 0.05, then the trend of "single-stage fan increasing with distance from the fault" is considered significant. Combined with the step code S: if the probability of an early single-stage fan is highest in bins with smaller S values, and the probability of a late single-stage fan increases in bins with larger S values, then the corresponding oriented sedimentary faults are judged to exhibit a phased activity pattern of "youth controlling early fans and old fans controlling late fans." The multi-stage fan-type sedimentary system includes at least one of the following: fan delta, underwater fan, turbidite fan, and alluvial fan.

[0047] The following detailed description of the invention uses the Weizhou 12-2 oilfield in the southwestern Weizhou depression of the basin as a specific example.

[0048] This embodiment uses the Weizhou 12-2 oilfield as the implementation area, with an area of ​​approximately 25 km². 2 The target layer is the Liushagang Formation, Liushan Section 2, Paleogene. This area is controlled by several major syn-sedimentary faults, including Fault No. 1, No. 2, and No. 3, in the Weinan Depression. Fault No. 1 is the main controlling fault at the depression boundary, Fault No. 2 is the main fault of the NEE-trending fault zone within the depression, and Fault No. 3 is the southern boundary fault of the Weinan Low Uplift. Several secondary faults also exist within the work area, dividing it into multiple fault blocks in plan view, making it a typical complex fault-block reservoir. This area experienced three phases of extensional rifting during the Paleogene, and the activity of syn-sedimentary faults significantly controlled the distribution of the fan-shaped sedimentary bodies. The method of this invention was used for single-phase fan subdivision and distribution characterization. The specific parameter values ​​for each step are as follows:

[0049] Step (1): The voxel spacing is 12.5m × 12.5m × 5ms. Obtain the sector classification volume obtained by the first patented method. (4 categories: early fan-shaped oil group, mid-stage fan-shaped oil group, late-stage fan-shaped oil group, and non-fan-shaped oil group), thick sand body Number of superpositions The data includes the main syn-depositional fault distance field d1, the regulating syn-depositional fault distance field d2, and the step code S (K=4). The fault block sedimentary attribute vector {x_i} is 12-dimensional, containing 8 basic dimensions (mean sand thickness of the two fan oil groups, etc.) and extended dimensions. Multi-source monitoring data includes well GR curves, well point stratigraphic dip angles, and the attitude and dip angle field φ of the progradation axis interpreted by seismic horizons.

[0050] Step (2): Prepare single-phase sector anchor body W_sub. The total number of single-phase sectors M=6 (dual source, each sector oil group is subdivided into early / middle / late phases). GR cycle segmentation is used to obtain the phase codes; auxiliary anchors are used at dip angle abrupt changes; φ contour line clusters are used as soft anchors. The expansion structural element is 3×3×3, and in case of conflict, the φ soft anchor takes priority.

[0051] Step (3): The input volume X is 9×30×200×200 (8 basic attributes + front product tilt field φ as the 9th channel).

[0052] Step (4): The encoder uses MedNeXt, with a patch size of 2×4×4, 96 / 192 / 384 / 768 stage channels, and 192 enc2 channels. The gated graph convolution input has 2 channels and the output has 192 channels, grouped into 0~95 and 96~191. After decoding, the classification header outputs 7 classes (background + 6 single-stage sectors).

[0053] Step (5): The loss weights are set to λ1=0.2, λ2=0.05, λ3=0.1; δ=0.1m / voxel; margin1=3 vxels, margin2=3 vxels. Training uses the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 Batch size 1, number of training rounds 80.

[0054] Step (6): Connectivity analysis yields the number of leaves in each single-period sector; weighted PCA is used to calculate the distribution orientation; along... Divide the data into four intervals, calculate the periodic activity index, fit the slope b and p value, and determine the periodic activity.

[0055] Comparing the model prediction results with actual well point data, the single-stage sector subdivision results are in good agreement with the stage division described by the core wells, and the sand thickness attenuation law is consistent with geological understanding, thus verifying the effectiveness of the present invention.

[0056] In another embodiment of the present invention, the encoder in step (4) can be a Swing-Unet, specifically configured as described above. The long axis of the window of each stage of Swing is stretched along the direction of the composite fault, with a stretching ratio of , where η=1.2.

[0057] In another embodiment of the present invention, the three-source anchor fusion rule in step (2) can be specifically as follows: the GR hard anchor determines the main body of the single-phase fan body coding, the well point where the formation dip angle changes abruptly is forced to be the well point marker of the single-phase inter-section, and the contour cluster of the pre-accretion dip angle field φ is used as a voxel-level soft anchor to vote and fuse with the GR hard anchor; if a voxel is GR expanded and coded as an early single-phase fan body, but the φ contour line indicates that the location belongs to a late pre-accretion sequence, then the φ soft anchor is given priority and the voxel is reassigned to the corresponding late code. The reassignment operation only takes effect within the fan body mask I_fan.

[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for subdividing and characterizing single-stage fan bodies constrained by syn-sedimentary faults, applicable to multi-stage fan-type sedimentary systems controlled by syn-sedimentary faults in rift basins, characterized in that... Includes the following steps: Step (1) Acquisition of multi-phase sector overlay and multi-source supervision data To obtain multi-source data for the work area, at least the following should be included: Fan-shaped classification The code contains at least two fan body oil group categories; by The fan-shaped mask I_fan = I( ∈ Sector class) (i.e. When the value is of the fan body oil group type, I_fan=1; otherwise, I_fan=0); sand-thick body Number of superpositions The main co-depositional fault has a signed distance field d1, and the regulating co-depositional fault has a signed distance field d2; the step code S∈{0,1,…,K}; the fault block sedimentary attribute vector {x_i}(N,≥8) and the fault block adjacency matrix A(N,N), where the 0th and 1st dimensions of the fault block sedimentary attribute vector are the mean sand thicknesses within the fault blocks of the two fan oil groups, constituting the fractal dimension representation of the two fan oil groups in the fault block attribute space, i.e., the first fan oil group area corresponds to the fault block aggregation area where the 0th dimension of the fault block sedimentary attribute vector is significantly greater than the 1st dimension, and the second fan oil group area corresponds to the fault block aggregation area where the 1st dimension is significantly greater than the 0th dimension; the two fan oil groups correspond to the two fan delta oil groups of a dual-source fan delta, or to the fan oil group and secondary division after the merging of multiple fan bodies of a single source, with the total number of fan bodies in a single phase M≥3; the fan classification system , thick sand body Number of superpositions , d1, d2, S, {x_i}, A can be obtained by the Swin-Unet fusion modeling method constrained by syn-sedimentary faults and graph attention network, or by other syn-sedimentary fault identification and sector classification processes; Multi-source monitoring data: wellhead monitoring sources include well GR curves and wellpoint formation dip angles; seismic monitoring sources include the pre-accretion axis attitude interpreted by seismic horizons, and optionally include the pre-accretion dip field φ(x,y,z) output of the superimposed modeling output of the upstream Swin-Unet and graph attention network; the monitoring data is used in step (2) to prepare the single-stage fan-shaped anchor body W_sub; Step (2) Preparation of single-stage sector anchor body A single-stage sector anchor body (i.e., a three-source fusion monitoring tag body) W_sub(D,H,W) is prepared by fusing three types of anchor sources. The code -1 indicates no well point, and c=1~M correspond to the early / middle / late single-stage sector within each sector oil group, respectively. M≥5 (when there are two sources) or M≥3 (when there is a single source and multiple stages): First source (GR cycle): The mudstone interlayer corresponding to the local maximum value of GR in the well is the boundary of a single-stage fan body. The early / middle / late single-stage fan body code is obtained by segmenting within each fan body oil group. Second source (wellpoint formation dip angle): along the source direction, the point where the dip angle changes abruptly serves as the wellpoint auxiliary anchor for a single-period cross section; Third source (pre-integration dip field φ): If φ has been output from upstream, the contour cluster along the source direction φ serves as a voxel-level soft anchor in the single-period fan-body expansion direction, which is then integrated with the GR hard anchor. After the anchor point is interpolated to a voxel along the well trajectory, a 3×3×3 structural element expansion is performed. Voxels of the same oil group within the expansion zone are assigned an early single-stage fan body code by default. If multiple single-stage fan body codes appear in the expansion neighborhood, the early single-stage fan body code takes priority. If a voxel is assigned an early single-stage fan body code by GR expansion, but the φ contour line indicates that it belongs to a late pre-sequence layer, the voxel is reassigned to the corresponding late code with priority given to the φ soft anchor. The reassignment only takes effect within the fan body mask I_fan. Step (3) Construction of multi-channel geological attribute input body Construct a multi-channel three-dimensional input volume X∈R^{8×D×H×W}, specifically by: dividing d1, d2, S, ... , , At least eight types of attributes, including I_fan and well point indication, are superimposed along the channel dimension; optionally, the pre-product dip angle field φ is incorporated into X as the 9th channel; Step (4) Dual-distance field grouped gate encoder An encoder-decoder architecture is adopted, with the encoder being either a MedNeXt block sequence or a Swin-Unet block sequence. The MedNeXt block contains a residual inversion bottleneck, a 5×5×5 UpKern kernel, and composite scaling; the Swin-Unet block contains a shifted window attention. The encoder enc1 → pooling → enc2, and the output of enc2 is multiplied channel by channel by channel of the fault activity gating map (generated by sigmoid after 1×1×1 convolution of channels d1 and d2 in X (input 2, output with the same number of channels as enc2), with a value range of [0,1]). The output channels of the gating map are divided into two groups according to the index: the active group of the main co-depositional faults (channel index 0~C / 2-1) and the active group of the modulating co-depositional faults (channel index C / 2~C-1), which respectively modulate the corresponding channel group of enc2, so that the feature modulation weights of the two faults are differentiated. After decoding, the M+1 class single-period sector is output through the 1×1×1 classification head (background 0 + M single-period sector 1~M). When the encoder is MedNeXt, the configuration includes, but is not limited to: Patch Size 2×4×4, 4 Stages with 96 / 192 / 384 / 768 channels configured by composite scaling orthogonal, each Stage containing 2 MedNeXt blocks, with a residual inverted bottleneck structure within each block: 1×1 upscaling → 5×5×5 depthwise convolution UpKern → 1×1 downscaling, with Group Norm + GELU; enc2 is the output of the 2nd Stage, with 192 channels; the gated graph has 1×1×1 convolution with 2 channels for input and 192 channels for output, grouped into 0~95 and 96~191; When the encoder is Swing-Unet, the configuration includes, but is not limited to: Patch Size 2×4×4, 4 SwingStages with 96 / 192 / 384 / 768 channels, each Stage containing 2 Swing Transformer Blocks, WindowSize 7×7×7, and alternating shift windows; enc2 is the output of the 2nd Stage with 192 channels; gating graph grouping is the same as MedNeXt configuration; the major axis direction ω of each Swing Stage window is stretched along the composite fault direction field τ_synth of d1 and d2, with a stretching ratio... τ_synth is the weighted synthesis of the tangents d1 and d2; Step (5) Joint optimization of four constraint losses Total loss Where L_CE is the anchor cross-entropy, calculated only for the position W_sub≠−1; P_c is the probability volume of the c-th single-period sector within the sector mask; ,in Let c be the probability centroid of sand thickness along the source direction w in a single-phase fan body. for The partial derivative along w is the decay rate along the source. δ is the empirical lower limit of the longitudinal decay rate of a single-phase fan body along the source direction in the work area. It is obtained by fitting the sand thickness at the well point in the work area along the source direction, and the value range is δ∈[0.05,0.15] (unit: m / voxel). L_geo forces the sand thickness of a single-phase fan body to decay monotonically along the source direction. , wherein is the sand thickness probability centroid coordinate of the single-stage fan body c along the provenance direction, and c1, c2, c3 are the centroid serial numbers of the early / middle / late single-stage fan bodies within the same fan body oil layer; L_seq mandates that the three early / middle / late single-stage fan bodies within the same fan body oil layer along the provenance direction satisfy a retrogradation sequence w{early}<w{middle}<w{late} or a progradation sequence w{early}>w{middle}>w{late}, one of which is selected according to the sequence type of the work area: the retrogradation sequence corresponds to a lacustrine transgression background, and the progradation sequence corresponds to a lacustrine regression background; margin1 and margin2 are empirically determined based on the retrogradation distances of the early / middle / late single-stage fan bodies within the same fan body oil layer in the work area, with a value interval of margin∈[2,5] voxel; The loss weights λ1, λ2, and λ3 are determined by a grid search of the work area validation set, with the search ranges λ1∈[0.1,0.3], λ2∈[0.03,0.1], and λ3∈[0.05,0.2]. Step (6) Quantification of single-stage fan body distribution and characterization of the activity of syn-sedimentary fault stages The number of leaves is obtained by performing connected component analysis on the predicted single-period sector. HW plane with sand thickness Weighted PCA was performed on the voxel coordinates with P_c>0.5 to obtain the orientation and major-minor axis ratio of the sector distribution in each period. ; along Divide the direction into K intervals (bins), and calculate the average probability of a single-period sector c within the k-th interval. The phased activity index is calculated by combining the ladder code S: Where k is the interval index, Let A_activity(k) be the average probability of a single-stage fan c within the k-th interval. A least-squares linear fit is performed on A_activity(k) as k increases to obtain the slope b and p values. A threshold θ = α / M is set, where α ∈ [0.3, 1.0]. If b > θ and p < 0.05, then the trend of "single-stage fan increasing with distance from the fault" is significant. Combined with the step code S: if the probability of an early single-stage fan is highest in the bin with a smaller S, and the probability of a late single-stage fan increases in the bin with a larger S, then the corresponding syn-depositional fault exhibits a phased activity pattern of "youth controlling early fans and old fans controlling late fans." The multi-stage fan-type sedimentary systems include at least one of the following: fan delta, underwater fan, turbidite fan, and alluvial fan.

2. The method according to claim 1, characterized in that, The fusion rule of the three source anchors in step (2) is as follows: the GR hard anchor determines the main body of the single-phase fan body coding, the well point where the formation dip angle changes abruptly is forced to be the well point marker of the single-phase section, and the contour cluster of the preaccretion dip angle field φ is used as the voxel-level soft anchor and fused with the GR hard anchor; if a voxel is coded as an early single-phase fan body by GR expansion, but the φ contour line indicates that the location belongs to a late preaccretion sequence, then the φ soft anchor takes priority and the voxel is reassigned to the corresponding late code. The reassignment only takes effect within the fan body mask I_fan.

3. The method according to claim 1, characterized in that, In the dual-distance field grouping gating described in step (4), the gating map channels are divided into two groups, which are dominated by d1 and d2 respectively. The purpose is to decouple the modulation weights of the main co-deposition fault and the regulating co-deposition fault on different channel groups of the encoder, so that the network can learn the differentiated control effects of the two faults on different single-stage sectors.

Images