Method for intelligently segmenting and identifying sedimentary microfacies based on logging curve

Abstract

The invention relates to the technical field of geophysical logging and sedimentology intelligent interpretation, and particularly discloses a method for intelligently segmenting and identifying sedimentary microfacies based on a logging curve. The method comprises the steps that a gamma ray curve and cleaning rule configuration are acquired, and a coarse segmentation segment set is generated through data cleaning, standardization and filtering processing; a section sample set is generated through interface alignment, boundary fine adjustment and section judgment in combination with a logging interpretation conclusion; feature extraction is carried out based on operation period dependence, and a segment classification result is generated through rule classification and model inference fusion processing; and finally, outputting a sedimentary microfacies type result and forming closed-loop feedback through sequence splicing, morphological rule mapping and configuration parameter self-adaptive updating. According to the method, automation and adaptive optimization of the sedimentary microfacies identification process are realized, and the identification accuracy and the engineering application efficiency are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of geophysical logging and intelligent interpretation of sedimentology, and in particular to a method for identifying sedimentary microfacies based on intelligent segmentation of logging curves. Background Technology

[0002] In identifying and characterizing sedimentary facies, lithology, grain size, sorting, clay content, vertical sequence, and the morphology and distribution of sand bodies are all important genetic indicators. These genetic indicators are the result of hydrodynamic factors in various sedimentary environments, while hydrodynamic conditions control variations in rock physical properties, such as formation spontaneous potential and natural gamma. Well logging curves are the physical responses of these physical properties along well depth, establishing accurate rock-electrical relationships in cored wells, which can then be extended to non-cored wells to deduce reservoir characteristics. Utilizing the morphology of well logging curves effectively reflects the changes of these genetic indicators in the vertical and horizontal directions, providing valuable data for identifying sedimentary facies and becoming an effective method for this purpose.

[0003] Natural gamma logging curves can take various shapes, such as box-shaped, bell-shaped, funnel-shaped, and finger-shaped. Figure 1 As shown, these patterns reflect energy changes or relative stability during sediment deposition. For example, box-shaped patterns reflect rapid accumulation under stable hydrodynamic conditions or sedimentary environments; bell-shaped patterns reflect sedimentary environments where water flow energy gradually weakens or sediment supply gradually decreases; and funnel-shaped patterns, the opposite of bell-shaped patterns, reflect sedimentary environments where hydrodynamics intensifies or sediment supply gradually increases. They exhibit significant anomalies in different sedimentary microfacies zones and reservoir belts. In particular, they are naturally present in the rock formations measured within wells, and the methods for measuring and obtaining well logging curves are simple and inexpensive. Therefore, they are widely used in the delineation and study of sedimentary facies zones.

[0004] Currently, sedimentary microfacies classification based on well logging data mainly relies on manual interpretation of the morphology of well logging curves (e.g., box, bell, and funnel shapes in well logging gamma curves represent different sedimentary environments and cycles). This method is affected by human experience and skill level, making it not only time-consuming and labor-intensive, but also subject to a certain degree of subjectivity and uncertainty in the interpretation results. Applying advanced technologies such as big data analysis and deep learning to oil and gas geological research is an exploration and attempt to address the current underutilization of big data analysis resources in the petroleum industry.

[0005] Identifying logging facies requires accurate delineation of identification units. Terrestrial sedimentary types are complex, and strata at different planar locations within the same geological time unit may exhibit significant thickness differences, making the delineation and identification of units a challenging task. The complex and variable subsurface sedimentary structures lead to differences in logging facies for the same sedimentary microfacies. While this is not difficult for experienced interpreters, it presents a significant challenge for automated computer identification.

[0006] In recent years, neural networks have emerged, capable of autonomously learning and extracting features from data, providing a new approach for identifying sedimentary microfacies from well logging. The network autonomously learns curve features rather than extracting them manually, preserving the inherent characteristics of the data to the greatest extent possible. Currently, methods for automatic segmentation and identification of sedimentary microfacies using well logging curves based on neural networks mainly involve converting the well logging curves into images for automatic segmentation and identification in target detection or image segmentation. However, automatic segmentation using neural networks requires a large amount of manual label creation, which is labor-intensive and consumes significant human and material resources. To achieve more accurate and convenient automatic segmentation and identification of sedimentary microfacies, this invention provides a method for intelligent segmentation and identification of sedimentary microfacies based on well logging curves, improving the speed of automatic segmentation and the accuracy of classification while reducing the cost of manual label creation.

[0007] Furthermore, in the field of geophysical logging and intelligent sedimentological interpretation, existing solutions for sedimentary microfacies identification typically revolve around gamma-ray logging curves, involving preprocessing, coarse segmentation, interpretation alignment, and morphological summarization, along with data cleaning, threshold setting, result mapping, and logging. These solutions suffer from limitations such as scattered and lagging cleaning rule configurations, reliance on manual experience for interface alignment leading to poor boundary stability, and a lack of a coherent process for segment determination and morphological mapping. Existing methods often employ a processing path combining manual annotation with empirical thresholds, relying on local rules or single-path models. In interpretation scenarios with significant stratigraphic variations, numerous shielded segments, and long well spans, inconsistencies across segments and chaotic version management can easily arise, making it difficult to ensure the stable generation of sedimentary microfacies type results. For the joint processing of gamma logging curves and cleaning rule configuration, interface alignment and fusion adjudication, and morphology merging, existing technologies generally have shortcomings in areas such as coverage adjudication strategy, consistency verification of curve segment category sequence and mapping table, and closed-loop update of threshold table and segment length threshold. It is difficult to form a consistent process of acquisition-alignment-judgment-recording-update in complex sedimentary environments, resulting in loose connection between sedimentary microfacies type results and segment classification results, long-term dispersion of parameters and strategies, and increased burden on production organization and data archiving. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for intelligent segmentation and identification of sedimentary microfacies based on well logging curves, comprising: The GR curve and cleaning rule configuration are obtained, and data cleaning and standardization are performed. Wavelet transform denoising is performed based on the wavelet basis order and threshold strategy table in the cleaning rule configuration, and median filtering is performed based on the median filter window width in the cleaning rule configuration to generate a coarse segment set. The system obtains a coarse segment set and well logging interpretation conclusions, and performs processing including search half-window width, matching priority, conflict resolution strategy, interface alignment and boundary fine-tuning. This includes binding interpretation interfaces within the candidate boundary neighborhood by the search half-window width, making small-range position corrections based on local window statistics, performing segment splitting and merging operations when the coarse segment crosses multiple interpretation interfaces, adjusting segment boundaries, performing statistical analysis based on the minimum segment length threshold and long segment determination threshold in the cleaning rule configuration, and performing index allocation based on the sample numbering strategy in the cleaning rule configuration to generate a segment sample set. The process includes target sampling interval, longest sequence length, short segment window width, long segment sliding stride, anomaly label mapping table and mask propagation strategy and feature extraction. For short segments, slope trend rule classification is performed based on the monotonic duration threshold, transition density threshold, main transition strength threshold and platform coverage ratio threshold in the rule version library. For long segments, LSTM-FCN network is used for inference classification based on the fusion weights in the model parameters. The classification inference process is performed to generate segment classification results. Based on the minimum hole threshold and bridging strategy in the cleaning rule configuration, sequence splicing is performed, and rule mapping is performed based on the morphology type, segment length range, stable segment coverage ratio, adjacent category transition relationship and layer limit in the micro-phase mapping table. Furthermore, segment length threshold configuration is updated based on the quantile value snapshot and anomaly percentage snapshot in the update basis bin, generating the cleaning rule configuration structure.

[0009] Furthermore, the process of generating the coarse segment set also includes: Obtain the GR curve and cleaning rule configuration. Based on the missing identification rules, anomaly type dictionary, depth unit and sampling interval standard, shielded well section list, drift verification baseline, boundary extrapolation strategy, interpolation method priority and log recording level in the cleaning rule configuration, perform missing correction and anomaly labeling to obtain the cleaning data. Depth and curve values ​​are extracted from the cleaned data. Based on the target sampling interval given by the cleaned rules, depth alignment is performed. Based on the mapping strategy table configured by the cleaned rules, interval normalization is performed to generate a standardized sequence. For the standardized sequence, wavelet denoising is performed based on the wavelet basis order and threshold strategy table in the cleaning rule configuration, and median filtering is performed based on the median filter window width and shielding segment participation strategy in the cleaning rule configuration to generate a coarse segment set structure.

[0010] Furthermore, the GR curve and cleaning rule configuration includes: The GR curve represents the relationship between well depth and gamma response. The recording unit, sampling interval, and depth reference are provided by the field recording system or historical calibration records. The cleaning rule configuration includes missing identification rules, anomaly type dictionary, depth unit and sampling interval standard, shielded well section list, drift verification baseline, boundary extrapolation strategy, interpolation method priority, and log recording level.

[0011] Furthermore, based on the minimum hole threshold and bridging strategy in the cleaning rule configuration, the sequence splicing process includes: The sedimentary microfacies type results are aggregated according to well section, stratigraphic position and morphological type. The distribution of section length, the proportion of platform coverage, the density of transition and the count of adjacent transition relationships are extracted. The length distribution of short and long sections, the proportion of stable section coverage and the proportion of shielded adjacent markers are statistically analyzed according to the section classification results. The statistical products are written into the update basis bin.

[0012] Furthermore, the snapshots of warehouse record quantile values, abnormal percentage snapshots, and cross-version comparison tables are updated. The cross-version comparison tables are used to compare the changes with the previous version configuration.

[0013] Furthermore, the process of updating the segment length threshold configuration and generating the cleaning rule configuration structure also includes: The boundary interval between short and long segments is calculated at the segment level. Combined with the segment length bandwidth corresponding to different morphologies in the sedimentary microfacies type results, a suggested group of segment length thresholds is given. If there are significant differences in bandwidth between different layers in the same well segment, a layer threshold group with layer limit conditions is generated.

[0014] Furthermore, the process of generating the cleaning rule configuration structure also includes: Check the compatibility of candidate updates with rule version library entries one by one. If there is a conflict, record the rule conflict and provide handling suggestions. At the same time, calculate and mark the impact list of potentially affected links and entries.

[0015] Furthermore, the cleaning rule configuration structure includes a threshold table, segment length threshold, bridging and overlap adjudication parameters, morphological candidate window scheme, boundary-specific rules, mask distribution strategy, and versioned metadata; the structure header contains the configuration version number, generation time, source statistical summary, and rollback strategy.

[0016] Furthermore, the process of updating the segment length threshold configuration and generating the cleaning rule configuration structure also includes: Output preparation phase: Align the configuration execution differences between this update and the previous version, write the difference alignment results into the configuration difference summary, and synchronize the summary to the operation and maintenance review channel.

[0017] Furthermore, the well logging interpretation conclusions include: Well logging interpretation conclusions refer to information such as the layer name, interface type, interface depth point or narrow range, confidence index, and interpretation version number given for the target well section. The source can be annotations by interpreters or automatic pushes from the rule base.

[0018] The key innovations of this invention include: (1) Based on the interconnected link of cleaning rule configuration structure - cleaning data - standardized sequence - coarse segment set - well logging interpretation conclusion, a sub-curve segment organization mechanism driven by interface alignment and boundary fine-tuning is constructed. Through candidate boundary neighborhood binding, cross-boundary segment splitting and dense short segment merging, sub-curve segments consistent with the stratigraphic limit are generated.

[0019] (2) Based on the segment length threshold, generate segment index and segment sample set. Construct short segment slope, turning point and platform structure features and long segment multi-scale local pattern response and sequence dependency hints on the segment input set respectively. Compile multi-path feature set and output segment classification results under fusion adjudication through trend rule classification and temporal network and convolutional network inference.

[0020] (3) Conduct coverage adjudication, bridging and overlap processing around the depth axis to generate curve segment category sequence; output sedimentary microfacies type results under the lookup of the mapping table entries; generate a statistical overview around the sedimentary microfacies type results and segment classification results and update the threshold table and segment length threshold, and write back the cleaning rule configuration structure to form a closed loop.

[0021] The following are its main beneficial effects: (1) The target of the operation is the configuration of gamma logging curves and cleaning rules. The operation link runs through the cleaning data, standardized sequence, coarse segment set to sub-curve segments. Compared with the existing interface alignment path that relies on human experience, the correspondence between candidate boundaries and interpretation interfaces is verified in the neighborhood binding and boundary fine-tuning. Cross-boundary segments and dense short segments are divided and merged to form sub-curve segments with clear structure, which is suitable for interpretation scenarios with obvious layer changes and shielding segments.

[0022] (2) The target of the operation is the segment input set and the multi-path feature set. The running link forms a complementary dual path of short segment rules and long segment model and generates segment classification results under the fusion adjudication. Compared with the existing schemes that only rely on local rules or single path models, the coverage, confidence hints and historical consistency jointly participate in the adjudication. The segment classification results form a continuous organization on the depth sequence, which is suitable for scenarios with long well spans and complex formation changes.

[0023] (3) The target of the operation is the curve segment category sequence, sedimentary microfacies type results and cleaning rule configuration structure. The operation link will classify the output and form a mappable category segment through coverage adjudication and bridging, and generate sedimentary microfacies type results through table lookup. The statistical overview drives the versioned writeback of the threshold table and segment length threshold. Compared with the existing decentralized management of parameters and strategies, it forms a closed-loop process from judgment to parameter update, which is suitable for integrated management of cross-well section production organization and data archiving.

[0024] (4) The segmentation method is simple and efficient, and has a high degree of consistency with geological features: The method of smoothing data using wavelet transform, coarse segmentation using median, and fine segmentation using well logging interpretation is simpler and more efficient than image target detection and image segmentation methods, as well as traditional manual segmentation methods. Compared with traditional smoothing methods, wavelet transform can preserve the local features of the signal while smoothing. Local anomalies or abrupt changes in GR well logging curves may correspond to important geological information, such as stratigraphic interfaces and special lithological layers. Wavelet transform can accurately locate and preserve these local features, and will not cause these key information to be lost or blurred due to smoothing, which helps to more accurately identify and analyze geological phenomena. At the same time, the median calculation is small, and when performing coarse segmentation on a large amount of GR well logging data, the results can be obtained quickly, improving processing efficiency. The initial stratification of GR curves can be completed in a short time, and the overall shape and basic features of GR curves can be well preserved, saving time for further fine processing. Well logging interpretation is based on geological theory and experience. When refining the coarsely segmented curves, it can fully consider geological information such as lithology, physical properties, and oil-bearing properties of the strata, making the segmentation results more consistent with the actual geological situation, improving the accuracy and reliability of strata division, and accurately identifying the interfaces of strata with different lithologies, as well as special geological structures such as thin interlayers in the strata.

[0025] (5) Precise Classification: Since the actual shape of GR curves is often complex and may not be a typical regular shape, the curves may exhibit characteristics of both bell and funnel shapes, or have irregular edges. For these transitional forms, it is difficult to determine a clear boundary to classify them, increasing the uncertainty of classification. Shorter curves have less obvious shapes, and image-based deep learning classification methods are relatively difficult to integrate with geological expertise, relying more on visual features of the image, which can easily lead to misclassification for inconspicuous features. This invention employs a hybrid recognition method. For short curves, classification based on slope is simple and direct, effectively handling data with obvious local features. For long curves, time-series networks can fully mine sequence information, adapting to complex patterns and long-term dependencies in long-series data. This combined approach improves the accuracy of curve recognition and more precisely determines stratigraphic interfaces and sedimentary microfacies types. Compared to single recognition methods that only focus on local or overall features, this not only improves classification accuracy but also recognition efficiency. Attached Figure Description

[0026] Figure 1 A flowchart illustrating a method for intelligent segmentation and identification of sedimentary microfacies based on well logging curves, provided for an embodiment of this application; Figure 2 A well logging curve morphology classification diagram provided for embodiments of this application; Figure 3This is a schematic diagram of the overall process of an intelligent segmentation and recognition method for sedimentary microfacies provided in an embodiment of this application; Figure 4 A specific network structure diagram of an automatic sedimentary microfacies segmentation and identification model provided in this application embodiment; Figure 5 This application provides a comparison chart of the effects of preprocessing and automatic segmentation of raw well logging curves in an embodiment of the present application. Figure 6 The final sedimentary microfacies identification effect diagram of a segment of actual well logging data provided in this application embodiment. Detailed Implementation

[0027] Example 1: Refer to Figure 1 This is a flowchart illustrating a method for intelligent segmentation and identification of sedimentary microfacies based on well logging curves provided in an embodiment of the present invention. The process may include at least steps S100-S400: S100: Obtain the GR curve and cleaning rule configuration, perform data cleaning, standardization and filtering, and generate a set of coarse segmentation segments; S200: Obtain the coarse segment set and well logging interpretation conclusions; perform segment boundary adjustment based on the search half-window width, matching priority and conflict resolution strategy in the cleaning rule configuration; perform statistical analysis based on the minimum segment length threshold and long segment determination threshold in the cleaning rule configuration; and perform index allocation based on the sample numbering strategy in the cleaning rule configuration to generate segment sample sets. S300: Based on the segment sample set and runtime dependencies pulled from the configuration center, including target sampling interval, longest sequence length, short segment window width, long segment sliding step, anomaly label mapping table and mask propagation strategy, feature extraction is performed. Based on the monotonic duration threshold, transition density threshold, main transition strength threshold and platform coverage ratio threshold in the rule version library and the fusion weight in the model parameters, classification inference is performed to generate segment classification results. S400: Obtain segment classification results, perform sequence splicing based on the minimum hole threshold and bridging strategy in the cleaning rule configuration, perform rule mapping based on the morphology type, segment length range, stable segment coverage ratio, adjacent category transition relationship and layer limitation in the micro-phase mapping table, and perform segment length threshold configuration update based on the quantile value snapshot and anomaly percentage snapshot in the update basis bin to generate the cleaning rule configuration structure.

[0028] Step S100 includes at least steps S110-S130: S110. Obtain the GR curve and cleaning rule configuration, perform missing data correction and anomaly annotation processing, and obtain the cleaned data; This step receives the GR (Gamma Ray) logging curve as the core input. The GR curve represents the correspondence between well depth and gamma response. The recording unit, sampling interval, and depth benchmark are provided by the field recording system or historical calibration records. Simultaneously, it receives the cleaning rule configuration, which is a set of extensible parameters and strategies, including missing data identification rules, anomaly type dictionary, depth unit and sampling interval standards, a list of shielded well sections, drift verification baseline, boundary extrapolation strategy, interpolation method priority, and log recording level. The cleaning rule configuration loads default entries during initial deployment, and in subsequent runs, the configuration update entries written back in step S430 overwrite entries with the same name. Specifically, the system first performs unit conversion according to the depth unit standard configured in the cleaning rules and establishes a monotonic check on the well depth field. When reverse order or duplicate depths are found, a rearrangement and merging strategy is triggered. This strategy follows the constraint of retaining the later timestamp for the same depth. If it cannot be determined, a duplicate depth mark awaiting manual verification is added to the cleaning log, and two records are temporarily stored, awaiting adjudication in the subsequent intra-segment statistical stage. Furthermore, the missing data identification rules are triggered from three sources: first, explicit missing data markers (such as null values ​​and stop-production markers written by the recording system); second, void segments caused by doubling of the sampling interval; and third, abnormal test or operating condition intervals listed in the shielded well segment directory. For explicit missing data, the system selects the effective method from neighborhood linear interpolation, spline interpolation, and intra-segment stable value filling according to the interpolation method priority in the cleaning rule configuration. For void segments, if the void length exceeds the repairable upper limit specified in the cleaning rule configuration, the entire segment is set as unrepairable and the missing unrepairable data is recorded in the anomaly field; if it does not exceed the upper limit, interpolation is performed. For shielded well segments, the shielded segment is directly recorded in the anomaly field, and the interval is written into a mask for skipping model inference in the subsequent segmentation stage. After restoring data continuity, the system enters the anomaly labeling stage. The anomaly type dictionary is pre-set with types such as pulse spike upper and lower limits, saturation platform, straight segment, negative value, anomaly, drift, and out-of-bounds. Each type of anomaly is accompanied by a trigger threshold and a neighborhood window. Specifically, pulse spikes are determined by a combination of abrupt changes in adjacent differences and neighborhood median statistics; upper and lower limit saturation is determined by detecting segments that remain close to the instrument's range boundary for an extended period; platform straight segments are identified by a combination of low variance and lower-order difference constraints; negative anomalies are marked as negative anomalies and trigger drift verification if a response less than zero still occurs after unit normalization; drift exceeding limits are determined by comparison with the drift verification baseline. For each anomaly trigger, the system simultaneously writes the anomaly start and end depth, anomaly type, anomaly intensity level, and handling suggestions. The handling suggestions are generated according to the priority relationship in the cleaning rule configuration; if multiple anomalies overlap, the one with higher priority is used for overriding. Furthermore, the system performs an index reconstruction on the cleaned records without changing the original time order, establishing a ternary mapping of deep index—original row number—cleaned row number, facilitating subsequent backtracking and result location; simultaneously, it outputs a cleaning log, including the number of triggers, a list of repaired segments, a list of unrepairable segments, and a list of masked segments.After the above processing is completed, cleaned data is generated. The cleaned data consists of fields such as depth curve value anomaly label mask flag mapping index, providing a unified and traceable data input for subsequent steps. At the end of this main step, the cleaned data is submitted to the input queue of the next step S120 through the data channel, serving as the sole source of the cleaned data in S120. At the same time, its output field name is registered as "cleaned data" in the system metadata layer, and it is noted in the cross-main step description that this product will participate in the construction of the statistical basis for configuration write-back in the subsequent S430 step.

[0029] S120. Extract depth and curve values ​​from the cleaning data, perform interval normalization and depth alignment, and generate a standardized sequence. After receiving the cleaned data output from S110, this step first extracts only the depth and curve values ​​as the main processing objects. Anomaly labels and mask flags remain in their associated state for constraint purposes and are not included in the normalization mapping calculation. Specifically, the system constructs a target depth grid according to the target sampling interval configured in the cleaned rules. If the actual sampling interval of the cleaned data deviates from the target, a depth alignment process is initiated. The alignment process employs an intra-segment ordered resampling strategy, first performing interpolation on the effective interval outside the shielded and unrepairable segments, then writing depth points within the shielded segment while keeping the mask masked for subsequent segmentation skipping. The interpolation process follows the principle of neighborhood constraints, with the neighborhood bandwidth taken from the cleaned rule configuration. If the number of effective points in the neighborhood is insufficient to trigger interpolation, a sparse neighborhood warning is recorded and the generation of that point is skipped until the next interpolable position is encountered. Further, the system enters the interval normalization stage. Interval normalization is defined as follows: for the target interpretation interval or specified well section, construct upper and lower bound statistics and map the curve values ​​to the dimensionless interval. The mapping strategy is written in the mapping strategy table configured in the cleaning rules, including two types: full well section mapping and layered mapping. When there is a list of layered well sections, the upper and lower bounds are calculated independently according to the layer boundaries to reduce the offset caused by the difference in the benchmark between formations. For sub-segments with anomaly labels of upper and lower bound saturation, the system performs boundary reestimation before normalization. The reestimation method is neighborhood unsaturated sample statistics, and the number of reestimations and the affected interval are recorded in the log. For platform straight sections, the system marks the section as low information density, and this label is maintained after normalization mapping. Further, depth alignment also includes the convergence processing of repeated depths and reverse depths. Repeated depths have been initially merged in S110, but if there are equal-depth repetitions due to interpolation landing points after resampling, the system retains the later arrival point and marks the earlier arrival point as merged and replaced. Reverse depths are written to the alignment and rearrangement identifier after the rearrangement is completed during the alignment process. After the above processing, the system obtains an information sequence that is dimensionless and carries accompanying annotations according to a unified target sampling interval. This sequence is called the standardized sequence in this invention. The standardized sequence consists of a depth-aligned raster dimensionless curve value mask flag accompanied by anomaly label index. To ensure data continuity in subsequent steps, the system simultaneously generates a traceability mapping from the cleaned data to the standardized sequence, recording the original depth interval and interpolation source information corresponding to each aligned depth point. The traceability mapping is stored in the metadata area for subsequent segment backtracking and quality auditing. The standardized sequence is written into the data channel at the end of this main step as the sole input to the standardized sequence described in S130. The output field name is explicitly stated as the standardized sequence, and the cross-main step description indicates that this product will be one of the upstream bases for the boundary fine-tuning and interface alignment actions of the coarse segment in the subsequent S210 step.

[0030] S130. Perform wavelet denoising and median filtering on the standardized sequence to generate a coarse segment set structure; This step receives the standardized sequence output from S120 as the sole input and performs two-stage smoothing and candidate segmentation according to the filtering and segmentation strategies configured in the cleaning rules. Specifically, the system first performs wavelet denoising, which is defined as: multi-scale decomposition of the standardized sequence, applying a threshold shrinkage strategy to the coefficients at each scale, and reconstructing the denoised sequence. The wavelet basis adopts the Daubechies family, with the specific order read from the cleaning rule configuration. The threshold shrinkage strategy also comes from the threshold strategy table configured in the cleaning rules, including both soft and hard thresholds. When the cleaning rule configuration enables an adaptive strategy, the system first estimates the noise level within several representative windows and then applies the threshold across the entire sequence. In terms of boundary processing, mirror extension is used in the reconstruction stage to avoid edge ringing, and the extension length is consistent with the depth grid. After wavelet denoising, the system performs median filtering on the denoised sequence. The median filtering window width is read from the cleaning rule configuration, and the window center is aligned with the depth grid points. Whether the shielded points within the window participate is determined by the shielded segment participation strategy field. If it is set to not participate, the median is calculated based on the set of valid points. If there are not enough valid points to generate, the denoised sequence value is retained and the median downgrade is recorded in the mask. After completing two levels of smoothing, the system discovers candidate boundaries according to the coarse segmentation rules. The coarse segmentation rules are defined as: triggering boundaries on the smoothed sequence based on three conditions: the magnitude of adjacent differences, the local slope change, and the length of the continuous stable segment. The adjacent difference threshold, the slope change threshold, and the shortest stable segment length are all derived from the segmentation parameter group configured in the cleaning rule. When the shielded segment penetrates the region, the segmentation algorithm forcibly inserts the segment boundary at the start and end of the shielded segment so that the interface can be fine-tuned in the subsequent S210 step in conjunction with the well logging interpretation conclusions. After candidate boundaries are generated, the system enters the boundary cleaning and intra-segment statistics stage: Boundary cleaning removes dense and isolated boundaries; dense boundaries are merged using a minimum spacing threshold constraint, and isolated boundaries are deleted using a minimum segment length threshold constraint; Intra-segment statistics calculate the start and end depths, segment length, intra-segment mean, intra-segment quantile statistics, mask coverage, and anomaly label distribution for each candidate segment, and retain source information from the traceability mapping in the statistical results to support subsequent backtracking. Furthermore, the system encapsulates all candidate segments and their statistical attributes into a coarse segment set structure. The coarse segment set structure is defined as a data set consisting of a segment index, start and end depths, segment length, intra-segment statistics, mask and anomaly overview, and source traceability index. The structure header includes a version number and a summary of the generation strategy, facilitating S210 reading and verification. To facilitate cross-master step integration, the system explicitly names the output field "coarse segment set" in the product description and marks its subsequent consumption location as the coarse segment set in S210 in the data exchange contract. This is used for interface alignment and boundary fine-tuning with the well logging interpretation conclusions loaded in S210. At the same time, the cross-master step description indicates that the segment length statistics of the coarse segment set will be read when the segment index is generated in step S220, and its intra-segment quantile statistics will be written into the sample description as intra-segment attributes when the segment sample set is generated in step S230.In summary, the technical effects of this step are as follows: Through wavelet denoising and median filtering, the standardized sequence obtains a stable shape, and the coarse segment set structure forms stable candidate boundaries and intra-segment attributes, providing directly usable upstream data for interface alignment of S210 and subsequent segment determination and sample encapsulation.

[0031] In a preferred embodiment, data preprocessing is a fundamental step in GR curve analysis, aiming to improve data quality, eliminate noise and dimensional differences, and provide clean and consistent data input for subsequent processing. This step mainly includes two key operations: data cleaning and normalization.

[0032] 1. Data Cleaning: Data cleaning mainly targets and processes missing and outlier values ​​commonly found in well logging curves. For missing values, if the number of missing values ​​is small, linear interpolation is used for imputation. The specific calculation formula is as follows: in, The values ​​of the missing data points to be filled. and These represent missing points. The values ​​of the previous and next valid adjacent data points, , , Representing data points respectively , , The corresponding measurement time or depth coordinates.

[0033] The core idea of ​​formula (1) is as follows: In the depth domain, it is assumed that the physical properties (such as the natural gamma value) between two valid data points change linearly. By calculating the rate of change between the two points, and then estimating the value proportionally based on the specific location of the missing point, this method is simple and efficient, and is suitable for segments with few missing data points and gradual data changes.

[0034] If a certain data segment is severely missing, that segment will be temporarily retained and processed in subsequent segmentation steps.

[0035] Outliers are handled using a weighted average method, which corrects them using their neighboring data points, with the weights dynamically adjusted based on their distance.

[0036] 2. Data Normalization: To eliminate the dimensional influence of data from different well sections or under different measurement conditions, the cleaned GR data is normalized using a minimum-maximum normalization method, compressing its value range to [0,1]. The specific formula is as follows: in, The original data, These are the normalized result data values. and These represent the minimum and maximum values ​​of the original GR data in the currently processed data segment, respectively.

[0037] The core idea and function of formula (2): Through normalization, all data points are transformed to the same scale. This helps to: Improve model training efficiency and stability: Avoid machine learning models (such as the LSTM-FCN network mentioned later) from having difficulty converging or encountering numerical calculation problems due to excessively large differences in the range of input feature values.

[0038] Enhanced data comparability: Enables GR curve data from different wells and measurement environments to be compared and analyzed under the same standard.

[0039] Step S200 includes at least steps S210-S230: S210. Obtain the coarse segment set and well logging interpretation conclusions, perform interface alignment and boundary fine-tuning to obtain sub-curve segments; This step receives the coarse segment set output from the preceding S130 as the core input. This coarse segment set already includes fields such as segment index, start and end depth, segment length, intra-segment statistics, mask and anomaly overview, and source tracing index during the preceding processing. Simultaneously, it receives the well logging interpretation conclusions, which refer to information such as the layer name, interface type, interface depth point or narrow range, confidence indicator, and interpretation version number given for the target well segment. The source can be annotation by interpreters or automatic push from the rule base. Specifically, the system first performs coordinate system verification, reading the depth benchmark, unit, and raster resolution recorded in the coarse segment set and the well logging interpretation conclusions. If there are differences, the coordinate mapping module is activated, performing unit conversion and depth alignment with reference to the wellbore benchmark and sampling interval registration table. Simultaneously, a coordinate mapping record is written to the metadata area, containing the mapping strategy, original benchmark and target benchmark, and the range of the segment being corrected. Further, the interface alignment process is initiated, which establishes a correspondence between the candidate boundaries in the coarse segment set and the interpretation interfaces in the well logging interpretation conclusions. To this end, the system sets a search half-window in the neighborhood of each candidate boundary. The search half-window width, matching priority, and conflict resolution strategy are all read from the configuration. When there is an interpretation interface in the neighborhood, the system binds the candidate boundary to that interface and writes it into the alignment binding table. When there are multiple interpretation interfaces in the neighborhood, the system first compares the confidence indicators, and then makes a decision based on the interface type priority (fault, interlayer, sedimentary interface, etc.). If a decision cannot be made, multiple candidates are recorded in the alignment binding table for further review. When there is no interpretation interface in the neighborhood, the candidate boundary is retained in its original position and marked as unaligned interface. After the initial binding is completed, the system enters the boundary fine-tuning stage. Boundary fine-tuning refers to making small-scale positional corrections around the bound or retained candidate boundaries based on local morphology and interpretation constraints. The specific approach is as follows: For each candidate boundary, several sampling points are taken on both sides to form a local window. The mean difference of the curves within the local window, slope changes, stable segment length, and anomaly ratio are statistically analyzed. A decision is made based on the fine-tuning threshold group to fine-tune several sampling points upwards or downwards. If the candidate boundary is bound to an interpretation interface and the interpretation interface provides a narrow range, the final value of the boundary must not exceed that range. If the bound item has multiple candidates awaiting review, the system calculates the alignment error and local window statistical difference between multiple interfaces, selects the one with the smaller error as the temporary binding, and retains the remaining candidates in the log. If the boundary crosses a shielded segment, the boundary is first projected onto the edge of the shielded segment, and then the local window statistical judgment is performed. Furthermore, the system handles intra-segment boundary issues: When a coarsely segmented segment crosses multiple interpretation interfaces, the original segment is split according to the finely adjusted boundary, generating multiple sub-segments, and the original segment index is registered as the source split. When multiple short segments appear densely in the neighborhood of the same interpretation interface and the segment spacing is lower than the minimum spacing threshold, the system performs segment merging. The merged sub-segments retain the starting depth from the first segment and the ending depth from the last segment, intra-segment statistics are recalculated, and the source tracing index retains the merged list.After completing the above processing, the system completes the attributes for each sub-segment. These attributes include: sub-segment index, start and end depths, segment length, associated explanation interface identifier, interface type, credibility identifier, mask coverage, anomaly overview, and source tracing index. After complete verification, the system generates sub-curve segments. A sub-curve segment refers to a set of curve segments that, after interface alignment and boundary fine-tuning, are consistent with the explanation interface or fall within its tolerance range. To ensure seamless integration, the system explicitly names the output field of this step as "sub-curve segment" in the data exchange description and writes a description of the next consumption position (S220) as "sub-curve segment" in the queue header. Simultaneously, the cross-main step description indicates that this output will be used for input organization work before feature construction within the S300 module.

[0040] S220. Extract segment length and amplitude statistics from the sub-curve segments, perform length threshold rule determination, and generate segment index; This step receives the sub-curve segment output by S210 as the sole input. This sub-curve segment already includes information such as the starting depth, ending depth, initial statistical values ​​within the segment, and associated interpretation interface identifiers. Specifically, the system first calculates the segment length for each sub-segment, which refers to the depth difference between the ending depth and the starting depth, using a unified unit after depth coordinate mapping. Simultaneously, amplitude statistics are performed within the sub-segment. Amplitude statistics refer to the summarization of the mean, quantiles, range, and stable segment percentage of the dimensionless curve values ​​of the GR (Gamma Ray) logging curve. If the sub-segment mask coverage is high or contains missing, irreparable markers, the system includes this sub-segment in the low-effective-sampling-segment queue and introduces the effective sampling ratio as an additional decision factor in subsequent judgments. To improve statistical robustness, the system employs a sliding-window reestimation strategy to handle sub-segments containing spikes or plateaus: when pulse spikes exist in the anomaly overview, anomaly label coverage points are first removed from the sub-segment before amplitude statistics are calculated; when plateaus exist, a low-change marker is added next to the amplitude range field for subsequent rule judgment. After completing the basic statistics, the system enters the length threshold rule determination stage. The length threshold rule determination refers to giving a segment classification conclusion for each sub-segment based on the minimum segment length threshold and the long segment determination threshold configured in the system. Specifically, if the segment length is less than the minimum segment length threshold, the sub-segment is marked as a short segment; if the segment length is greater than the long segment determination threshold, it is marked as a long segment; if the segment length is between the two thresholds, it enters the fuzzy interval, and the system continues to read the stable segment ratio and mask coverage rate from the amplitude statistics. Sub-segments with a high stable segment ratio and a low mask coverage rate are classified as long segments, and vice versa. If it is still impossible to determine, the segment classification is recorded pending determination. This mark will be added to a separate list when the S230 is bound with a number for subsequent verification by the model side or the manual side. Furthermore, the system handles the impact of cross-boundary splicing: when a sub-segment originates from the segment merging operation of S210, the intra-segment statistics may contain multiple sub-intervals with significant structural differences. The system divides the sub-intervals according to the internal stable segment threshold and calculates the weighted segment length. The weighted segment length is used to participate in the threshold rule judgment, and the adjudication result is written into the segment adjudication field and the adjudication basis is recorded. When a sub-segment is located near a shielded segment, the system records a shielded proximity identifier next to the length and amplitude statistics, which is used by S230 as a screening condition during sample structuring. In summary, the system constructs a segment index, which is an index structure that maps sub-curve segment numbers to segment conclusions (short segment, long segment, segment pending adjudication). The structure includes the sub-segment index, segment conclusion, segment length, amplitude statistics summary, effective sampling ratio, shielded proximity identifier, and adjudication basis summary. After the segment index is created, the system updates the exchange description in the data channel, indicating that the output field name of this step is segment index, and clarifies in the description that the next consumption position is the segment index of S230; at the same time, the cross-main step description indicates that the segment length and stable segment ratio in the segment index will be used as segment guidance signals in the feature construction stage of S320.

[0041] S230. Bind and number the sub-curve segments and the segment indexes to generate a segment sample set structure; This step receives the sub-curve segments output by S210 and the segment index output by S220 as parallel inputs. The sub-curve segments provide structural boundaries and interpretation constraints, while the segment index provides segment conclusions and adjudication basis. Specifically, the system first constructs a binding table, using the sub-segment index as the primary key and binding it one-to-one with the segment conclusions in the segment index. If the segment index contains segments awaiting adjudication, a pending adjudication tag is added to the binding table for that sub-segment, temporarily excluding it from subsequent short or long segment batch encapsulation. After the initial binding is completed, the system enters the numbering process, which generates a unique number for each bound sub-segment. The numbering rules are determined by the sample numbering strategy, a common approach being well number - well segment number - sub-segment sequence number - segment code. The system simultaneously writes the interpretation version number when generating the number for easy subsequent version management. If the same sub-segment is repeatedly bound under different interpretation version numbers, the system retains the number record with the later version number update, and writes the remaining records to the history area. Further, the system performs sample structuring on short and long segments respectively. Sample structuring refers to organizing sample entries around the fields required by the recognition model or rule chain. For short segments, the system reads the start and end depths, mask coverage, and anomaly overview from the sub-curve segment, and appends the segment length, amplitude statistical summary, and segment classification conclusion output by S220 to construct short segment sample entries. For long segments, the system appends the internal stable segment threshold division results and weighted segment length to the above fields to construct long segment sample entries. For sub-segments with pending adjudication, the system temporarily stores them as pending sample entries and records the supplementary information required for adjudication. Further, the system organizes batch packaging lists: short segment sample entries are added to the short segment sample list, long segment sample entries to the long segment sample list, and pending sample entries to the pending sample list. All three lists have header metadata, including generation time, numbering strategy, interpretation version number, threshold rule version number, and source tracing summary. To support the subsequent input organization of S310, the system further adds positioning and tracing indexes to the sample entries. The positioning index refers to the correspondence between the depth-aligned raster position and the original row number, while the tracing index refers to the mapping path from the sample entry back to the cleaned data and standardized sequence. Regarding data integrity, the system performs a field completeness check on each sample entry. The check rule is that required fields must exist and not be empty. Missing entries are written to a completeness anomaly list, which is also recorded in the metadata area for subsequent review. After completing all organization, the system generates a segment sample set structure. The segment sample set structure refers to a collection containing a short segment sample list, a long segment sample list, a pending sample list, and common metadata. The collection is internally partitioned, with the partition name matching the segment. The header of the collection records the version number and a data dictionary verification summary for easy reading and verification.To ensure seamless link connectivity, the system explicitly names the output field of this step as "Segment Sample Set" in the exchange description and clearly indicates in the queue that the next consumption position for "Segment Sample Set" is S310. Simultaneously, the cross-main step description indicates that the short segment sample list and the long segment sample list will respectively enter the feature construction stream of S320. The pending sample list can be incrementally synchronized into the same flow after subsequent adjudication. In summary, the technical effects of this step are: through binding and numbering mechanisms, a stable mapping is formed between sub-curve segments and segment indices; the segment sample set structure is organized in batches and with versioning, providing a clearly structured and traceable data carrier for subsequent input splitting and feature construction stages.

[0042] In a preferred embodiment, this step aims to divide the preprocessed GR curve into multiple sub-curve segments with independent morphological characteristics, providing input for subsequent classification. The automatic segmentation process includes three sub-steps: wavelet transform denoising, median coarse segmentation, and well logging interpretation fine segmentation. 1. Wavelet transform denoising: Based on the data characteristics of the GR curve and practical project experience, the Daubechies wavelet basis is selected to decompose and reconstruct the normalized curve, effectively suppressing high-frequency noise while preserving abrupt changes and detailed information in the curve, providing a high-quality data foundation for subsequent segmentation. Specifically, wavelet denoising uses mirror continuation to avoid edge ringing, and adaptive noise estimation is achieved based on the wavelet basis order and threshold strategy table in the cleaning rule configuration, ensuring the smoothness of the denoised sequence and the preservation of geological features.

[0043] 2. Median Coarse Segmentation: Median filtering, a nonlinear filtering method, is used to smooth the denoised curve. A fixed-width sliding window is set, and the median of the data within the window is calculated point by point and used to replace the original data points, thereby suppressing impulse noise while preserving edge information. Subsequently, the median of the entire curve is used as a threshold to initially identify and filter out segments without obvious morphological changes, achieving coarse segmentation of the curve. Furthermore, the coarse segmentation process combines multiple conditions such as the magnitude of adjacent differences, local slope changes, and the length of continuous stable segments to jointly trigger candidate boundaries, and forcibly inserts segment boundaries at the start and end of shielded segments to improve the robustness of the segmentation.

[0044] 3. Fine-grained segmentation based on well logging interpretation: Building upon coarse segmentation, fine-grained segmentation is achieved by aligning interfaces and fine-tuning boundaries using well logging interpretation conclusions (such as stratigraphic names, interface types, and confidence indicators). Specifically, interpretation interfaces are bound within the candidate boundary neighborhood using a search half-window width, and small-scale positional corrections are made based on local window statistics (such as curve mean difference and slope changes). When a coarse segment crosses multiple interpretation interfaces, segment splitting and merging operations are performed to generate sub-curve segments consistent with geological stratigraphy and within tolerance limits. For example, when a curve segment corresponds to multiple lithological characteristics, the curve is further segmented based on lithological change points to filter out irrelevant information, ultimately obtaining a series of sub-curve segments with clear geological significance. This process significantly improves the consistency between the segmentation results and the actual stratigraphic conditions.

[0045] Furthermore, based on the length of the sub-curve segments, this invention employs a hybrid strategy to classify them: For short curves less than 5 meters in length, classification is performed based on their slope variation characteristics. By calculating the slope between adjacent points and combining it with the curve amplitude variation trend (e.g., an initial increase followed by a decrease is considered a bell shape, and a continuous increase is considered a funnel shape), rapid and robust shape discrimination is achieved. Specifically, short segment classification is based on threshold groups such as monotonic duration threshold and inflection density threshold in the rule version library. Classification accuracy is improved through joint adjudication of multiple features including trend directionality, inflection structure, and plateau structure. Simultaneously, a conservative adjudication path is employed for high-mask samples to ensure reliability.

[0046] For long curves exceeding 5 meters in length, an LSTM-FCN time series classification network is used. The network employs a dual-path parallel architecture: the upper path uses LSTM layers to capture long-range temporal dependencies, while the lower path extracts multi-scale local patterns through convolutional layers and Squeeze and Excite layers. Finally, a weighted synthesis layer performs a fusion layer to output a class estimate. This network can learn four typical morphological patterns (bell-shaped, funnel-shaped, finger-shaped, and box-shaped) and combines a masking strategy to handle masked segments, thereby achieving high-precision classification.

[0047] By combining the above three steps, this invention achieves fully automated processing of GR curves from preprocessing to segmentation and then to classification, significantly improving the efficiency and accuracy of sedimentary microfacies identification. In addition, this step provides a unified input for subsequent fusion inference through structured encapsulation of segment sample sets (such as partitioned storage of short segment sample lists and long segment sample lists) and feature construction (such as multi-path integration of short segment slope features and long segment temporal features), further enhancing the accuracy and efficiency of the identification link.

[0048] Step S300 includes at least steps S310-S330: S310. Obtain the segment sample set, perform segment-based splitting and batch packaging processing to obtain the segment input set; This step receives the segment sample set output from preceding step S230 as its sole input. The segment sample set includes a short segment sample list, a long segment sample list, a pending sample list, and common metadata. The common metadata records the generation time, numbering strategy, interpretation version number, threshold rule version number, and source tracing summary. Specifically, the system first performs an integrity check on the segment sample set, reading the sample number, start and end depths, mask coverage, anomaly overview, segment length, amplitude statistical summary, segment conclusion, and location and tracing indexes. If any required fields are missing, the sample number is added to the completeness anomaly list and temporarily excluded from the encapsulation stream of this step. Furthermore, the system performs segment-based splitting of the segment sample set according to the segment conclusions. Segment-based splitting means sending the short segment sample list and the long segment sample list to two independent preparation queues, while the pending sample list is retained in the pending queue. When there are pending entries, the system records the decision basis and required supplementary information in the pending queue and starts asynchronous listening. If the upstream pushes supplementary information after updating the interpretation version or threshold rule version, it triggers the re-determination of the entry and incremental synchronization to the corresponding segment queue. After completing the segment-based splitting, the system enters batch encapsulation processing, which refers to batch organizing and runtime dependency injection of sample entries in the same segment queue. Specifically, the system reads the interpretation version number and threshold rule version number from the public metadata and writes them into the encapsulation header; it pulls runtime dependencies from the configuration center, including the target sampling interval, longest sequence length, short segment window width, long segment sliding step, anomaly label mapping table, and mask propagation strategy, and registers them in the encapsulation header. Furthermore, the system performs lightweight regularization on the short-segment preparation queue. Lightweight regularization refers to standardizing and organizing the fields of short-segment sample entries, retaining the start depth, end depth, segment length, amplitude statistical summary, mask coverage, anomaly overview, location index, and traceability index, and creating reserved slots for trend calculation within the entries for subsequent slope and inflection point derivation of S320. For the long-segment preparation queue, the system performs sequence regularization. Sequence regularization refers to constructing fixed-length or variable-length sequence containers according to the target sampling interval, and attaching mask bits and anomaly bits within the containers for subsequent time-series feature generation. To reduce random memory access during subsequent feature construction, the system completes batch splitting and index pre-arrangement during batch packaging. The short-segment queue uses equal-volume batch splitting, while the long-segment queue uses batch splitting by bucketing according to sequence length. The bucket boundaries are read from the configuration center, and the bucketing scheme and the sorting key (start depth or sequence length) within the bucket are recorded in the packaging header. In terms of anomaly handling, if the mask coverage of a short sample entry exceeds the configured upper limit threshold, the system marks the entry header with a high-mask sample and places it at the end of the batch; if a long sample entry has sparse neighborhood hints or the continuous masking span is too long, the system sets a downgrade calculation flag in the container to prompt the subsequent feature construction stage to adopt an alternative strategy.After completing the above processing, the system forms a segment input set. The segment input set refers to a unified input organization structure containing short segment batches and long segment bin batches. The structure header records the interpretation version number, threshold rule version number, batch splitting scheme, runtime dependencies, and exception list index. The structure internally stores a list of short segment entries and a list of long segment containers, along with their corresponding location and traceability indexes. To ensure seamless integration, the system explicitly names the output field of this step as the segment input set in the data exchange description and writes a description in the queue header indicating that the next consumption position is S320. Simultaneously, the cross-main step description specifies that this output will be directly consumed within the S300 module, and the traceability index will assist in locating the segment range during the classification output write-back in the S400 module.

[0049] S320. Extract short-segment slope features and long-segment temporal features from the segment input set, construct features, and generate a multi-path feature set; This step receives the segment input set output by S310 as the sole input. This segment input set includes short segment batches, long segment bin batches, runtime dependencies, and anomaly list indexes. Specifically, the system first performs trend analysis preparation on the short segment batches. Trend analysis preparation refers to constructing a basic quantitative description for rule determination around the short segment entries. This step includes uniform resampling, endpoint stability verification, and trend pre-screening. Uniform resampling follows the target sampling interval; endpoint stability verification determines whether an endpoint is affected by a spike by using the change magnitude threshold of the endpoint's neighborhood. If it is marked as a pulse spike, an endpoint replacement strategy is used in subsequent slope estimation; trend pre-screening quickly determines whether the short segment belongs to a low-change segment or a high-change segment based on the amplitude statistical summary and segment length information within the entry, and the pre-screening result is written to the trend label field. Further, the system proceeds to short segment slope feature extraction. Short segment slope features refer to numerical descriptions reflecting the overall rise, overall fall, local turning points, and plateau degree within the short segment. The specific approach is as follows: Two window families, a fixed window and a variable window, are constructed on the short sequence. The fixed window is used for baseline determination across the entire segment, while the variable window is used to capture local transitions. Regarding anomaly and mask handling, if the mask percentage within a window is too high, that window does not participate in derivation, and any missing windows are recorded within the entry. Under the endpoint replacement strategy, endpoint values ​​are replaced with robust neighborhood statistics, and the replacement traces are written to the entry log. The system generates feature sets in three directions around the window families: first, trend directionality, describing the overall increase / decrease trend and monotonic duration of the short segment; second, transition structure, describing the density of transition positions and the ranking of main transition intensities within the short segment; and third, platform structure, describing the coverage ratio and distribution continuity of low-change areas. These features, together with the original segment length and amplitude statistical summaries within the entry, constitute the short segment slope feature description, and the feature version number and window family configuration summary are retained at the entry level. After completing the short segment processing, the system transitions to the extraction of temporal features in long segment batches. Long segment temporal features refer to the multi-scale local patterns and sequence dependency representations required for sequence model inference. Specifically, the system performs slicing and alignment on each long-segment container. The slicing strategy uses a sliding window, with stride and window length read from runtime dependencies. The alignment strategy follows depth raster consistency. For containers with degradation calculation flags, the system automatically widens the sliding stride and reduces the window length to reduce the impact of long-span masking, and retains degradation records in the container log. Next, the system calculates a multi-scale convolutional response summary, which refers to the local pattern intensity sequence obtained by statistically analyzing the response of sequence slices using multiple local extractors with different receptive fields. Simultaneously, the system constructs sequence dependency hints, which are context pointers used for sequence model reading, describing the relative positions between slices and cross-slice continuity. To ensure readability on the model side, the system performs range normalization and mask unwrapping on each long-segment container. Range normalization maps the statistical upper and lower bounds within the container to a uniform scale; mask unwrapping converts the mask bits within the slice into slice weights for weight determination during subsequent model reading.After completing the feature derivation for short and long segments, the system performs feature construction and assembly. Feature construction and assembly refers to integrating short-segment slope features and long-segment time-series features into a unified data structure, along with necessary metadata and traceability indexes. The system generates short-segment feature vectors on short-segment entries and long-segment feature sequences on long-segment containers. Both types of features are registered through a unified multi-path feature container. The container header records the feature version number, window family configuration summary, sliding window scheme, value range standardization summary, and mask distribution strategy. To adapt to the subsequent inference operation mode, the system outputs containerized objects according to batch boundaries, and attaches alignment hints and positioning indexes to the end of the objects to ensure seamless batch continuation during S330 reading. Through the above processing, the system generates a multi-path feature set, which refers to a feature set containing batches of short-segment feature vectors and bucketed batches of long-segment feature sequences. Each entry or container in the set has a traceability index and a positioning index, facilitating the location of the source segment when writing back the classification results. To complete the link connection, the system specifies in the data exchange description that the output field name of this step is multi-path feature set, and writes in the object header that the next consumption position is S330 'multi-path feature set'. At the same time, in the cross-main step description, it is pointed out that short feature vectors will enter the rule link, and long feature sequences will enter the sequence model link. The two will complete the fusion inference within S330.

[0050] S330. Perform short-segment slope trend rule classification and long-segment LSTM-FCN inference fusion processing on the multi-path feature set to generate a segment classification result structure; This step receives the multi-path feature set output by S320 as the sole input. The multi-path feature set includes batches of short-segment feature vectors and batches of long-segment feature sequences, along with their corresponding tracing and positioning indices. Specifically, the system first performs rule classification on the batches of short-segment feature vectors. Rule classification refers to classifying short-segment feature vectors based on preset trend and structure thresholds and outputting rule categories. The thresholds are derived from a rule version library and include monotonic duration thresholds, transition density thresholds, main transition strength thresholds, and platform coverage ratio thresholds. The system sequentially reads the trend directionality, transition structure, and platform structure descriptions of the short segments, adjudicating them level by level from strong constraints to weak constraints to obtain rule categories and adjudication basis summaries. When encountering high-mask samples or samples with many missing windows, the system adds conservative adjudication paths without changing the adjudication order and records confidence prompts next to the category labels. When multiple candidates are tied in the adjudication results, the system selects the primary category according to the conflict adjudication strategy in the rule library and writes the secondary category and its basis summary into the accompanying information for reference during subsequent sequence-level merging in S410. After completing the short-segment rule classification, the system performs model inference on the long-segment feature sequences in batches. Model inference refers to calling a joint inference structure composed of a Long Short-Term Memory network and a fully convolutional network to estimate the category of the long-segment feature sequences. To comply with the convention of writing the first English abbreviation, the system clarifies in this paragraph: LSTM (Long Short-Term Memory) is used to read the contextual dependencies of the slices, and FCN (Fully Convolutional Network) is used to extract multi-scale local patterns. The two are weighted and synthesized in the fusion layer to output the category estimate. The specific process is as follows: The system loads the feature sequences within each long-segment container into the inference input buffer according to the bucketing scheme. During loading, alignment hints and mask unfolding strategies are read simultaneously, and mask bits are converted into input weights. Further, the LSTM branch sequentially traverses the slice sequence to generate a state sequence representing the context. The FCN branch computes the combined representation of multi-scale convolutional response summaries in parallel on the same slice sequence. The two branches are weighted and synthesized in the fusion layer. The fusion weights are read from the model parameters. When a degradation calculation flag is encountered, the system triggers a degradation fusion path, using restricted weights and expanding the neighborhood context length within the fusion layer. In the inference output stage, the system provides a class estimate, accompanying confidence hints, and a conflict lookup table for each long-segment container. The conflict lookup table records multiple estimated summaries of the container under different window lengths and strides, used for reference during subsequent sequence splicing and morphological merging. After completing short-segment rule classification and long-segment model inference, the system enters the fusion processing stage. Fusion processing refers to establishing a unified category space between short-segment rule categories and long-segment model categories and completing conflict resolution.The unified category space comprises two groups: basic trend categories and composite morphology categories. Basic trend categories carry short-segment rule categories, while composite morphology categories carry long-segment model categories. When short and long segments conflict in the same deep neighborhood, the system adjudicates the conflict according to the order of neighborhood coverage priority—confidence hint—historical consistency. Neighborhood coverage priority prioritizes categories with wider coverage. Confidence hints are used to select the category with higher confidence when coverage is similar. Historical consistency involves retrospectively checking the past estimates of adjacent segments within the same well using a traceability index. After adjudication, the system prepares for writing back the fusion results. This write-back preparation provides consistent segment category codes and location information for subsequent sequence splicing stages. The system generates segment category codes, primary categories, optional secondary categories, confidence hints, adjudication basis summaries, location indexes, and traceability indexes for each entry or container, and writes the rule version number and model version number into the object header for easy version management. After all processing, the system generates a segment classification result structure. This structure refers to a fused category set containing short and long segments within a unified category space. This set is organized in depth order and can be directly read by subsequent sequence-level processing. To complete the link connection, the system explicitly states in the data exchange description that the output field name for this step is "Segment Classification Result," and writes a description in the object header indicating that the next consumption position is 'Segment Classification Result' in S410. Simultaneously, the cross-main step description clarifies that this output will be directly consumed in module S400 and used as the upstream basis for morphological segment recognition during rule mapping and table lookup in S420. In summary, the technical effect of this step is: through the parallel processing and fusion of the short segment rule link and the long segment sequence model link, a unified and traceable segment classification result is formed, providing continuous, versioned, and verifiable input for subsequent sequence splicing and morphological classification.

[0051] Step S400 includes at least steps S410-S430: S410. Obtain the segment classification result, perform sequence splicing and morphological merging processing to obtain the curve segment category sequence; This step receives the segment classification results from preceding step S330 as the sole upstream input. These results are organized in depth order, and their fields include the main category, optional subcategories, adjudication basis summary, location index, traceability index, rule version number, model version number, and neighborhood coverage indication. Specifically, the system first performs a consistency check on the segment classification results, reading the depth order and coverage of adjacent entries within the same well. If there are overlapping or gaps in coverage, a splicing task sheet is generated. This task sheet records the start and end depths of the left and right blocks, the coverage difference, the source entry number, and necessary adjudication information. The system then proceeds to the sequence splicing process, which involves constructing continuous category bands around the depth axis. The system moves from the start to the end of the well segment. For adjacent entries with coverage gaps smaller than the minimum cavity threshold, a bridging strategy is used to generate transition segments. This strategy only reads the main categories and trend indicators from both sides of the neighboring area, without introducing new judgments. When coverage occurs repeatedly, the system enters the overlap adjudication sub-process. Overlap adjudication is based on the neighboring coverage range indication; entries with wider coverage are prioritized for inclusion in the main sequence. When coverage ranges are similar, the priority key in the adjudication basis summary is read. If distinction is still not made, the final assignment is made by referring to the historical consistency records of the same well, and the overlap adjudication is marked as complete in the splicing log. For connections across shielded areas, the system places the splicing breakpoint at the shielding edge based on the mask distribution identifier in the original segment and marks this breakpoint as a shielding edge breakpoint for subsequent morphological merging stages.

[0052] Furthermore, after completing the main sequence skeleton construction, the system enters the morphological merging process. Morphological merging refers to merging adjacent and category-compatible entries in the main sequence to output longer and more stable category fragments. The system first scans adjacent entries of the same category and merges combinations that meet the minimum adjacent overlap width and stability threshold. After merging, the starting depth is taken from the starting depth of the first entry in the combination, and the ending depth is taken from the ending depth of the last entry in the combination. The decision is based on the summary merged into an ordered list. If adjacent entries have the same main category but different subcategories, the system retains the consistency of the main category and writes the subcategories into a candidate subcategory set to provide a reference for the subsequent micro-phase mapping stage. If the main categories are different but there is a mergeable relationship (e.g., different intensities of the same trend), the system reads the merging relationship table and performs dimensionality reduction merging. Dimensionality reduction merging only takes effect on the pairs listed in the relationship table and is written to the relationship table in the log. Regarding boundary handling, if the length of an entry located at the beginning or end of a well segment is less than the minimum stable segment length, it is suspended and entered into the boundary candidate queue. It will be merged when it meets the merging conditions with an adjacent new entry. If the conditions are never met, it is retained as an independent short segment and a special rule is used in the subsequent mapping stage.

[0053] Understandably, after morphological merging is completed, the system generates a curve segment category sequence. This sequence refers to an ordered set of category fragments formed along the depth axis. Each fragment within the set includes the starting depth, ending depth, main category, candidate subcategories, source entry number set, location index, traceability index, and a concatenated log summary. The system specifies in the product description that the output field name is "Curve Segment Category Sequence," and in the data exchange contract, it indicates that the next consumption location is the curve segment category sequence of S420. Simultaneously, the cross-master step description indicates that this sequence will subsequently support morphological fragment extraction and rule mapping in S420, and will be used as one of the statistical bases for table entry hits during configuration updates in S430.

[0054] S420. Extract box-shaped, bell-shaped, funnel-shaped, finger-shaped and other morphological segments from the curve segment category sequence, perform rule mapping and table lookup, and generate sedimentary microfacies type results; This step receives the curve segment category sequence output by S410 as the sole input. This sequence has undergone coverage adjudication and merging, possessing sufficient depth for direct morphological recognition. Specifically, the system first performs a morphological candidate scan on this sequence. This scan involves matching the trend, turning points, and platform features of the sequence segments using a window-progression strategy. The window length is read from the configuration center and set according to layer partitioning. The system establishes candidate trigger conditions for four target morphologies: box-shaped candidates focus on rising ends and a flat middle section; bell-shaped candidates focus on a raised middle section and a smooth decline; funnel-shaped candidates focus on a low value at the top gradually increasing towards the bottom; and finger-shaped candidates focus on short, narrow, and pointed peaks. The trigger conditions are not expressed using formulas but are determined through threshold judgments based on feature description fields, specifically including the coverage ratio of stable segments, the ranking of turning point intensity, the adjacent consistency marker of the main category sequence, and the relationship between platform segment distribution density and segment length. If the trigger conditions are met within the window, the system maps the window to a morphological candidate segment. This morphological candidate segment carries a morphological type candidate, a snapshot of the trigger field, and a set of source segment numbers, and enters the candidate pool. After the candidate pool is constructed, the system performs candidate conflict resolution. Conflict resolution follows the order of coverage priority - hierarchical constraint - historical consistency. Only one candidate is retained in the same depth neighborhood. If two candidates have highly overlapping coverage and both hierarchical constraints are satisfied, the historical consistency record is read for final screening, and the unsuccessful candidate is recorded as a candidate.

[0055] Further, the system enters the rule mapping and table lookup phase. Rule mapping refers to retrieving corresponding entries from the microphase mapping table according to the feature description and hierarchical context of the morphological candidate fragments; table lookup refers to primary key matching and condition filtering based on morphological type, segment length range, stable segment coverage ratio, adjacent category transition relationship, and hierarchical constraints. The microphase mapping table is provided with versioned entries by the interpretation team during deployment. Each entry includes morphological type, hierarchical constraints, microphase candidates and priorities, remarks, and review records. The system performs a two-stage search for each morphological candidate fragment: first, it locates first-level candidates based on morphological type and segment length range, and then filters out the final entries based on hierarchical constraints and transition relationships. If multiple entries are found to be found simultaneously, they are sorted by priority, and the first one is selected as the primary found entry, with the remainder written into the candidate mapping. When there are boundary candidate fragments involved, the system searches according to boundary-specific rules, providing independent entry entries for shorter but critical fragments to avoid exclusion due to insufficient length. After completing the table lookup, the system constructs microfacies entries based on the primary hit entries. A microfacies entry refers to a record of microfacies determination for a single morphological candidate fragment. Fields include starting depth, ending depth, morphological type, microfacies type, entry number, entry version number, candidate snapshot summary, source fragment number set, and traceability index. All microfacies entries are merged in depth order into a sedimentary microfacies type result. The sedimentary microfacies type result refers to the microfacies determination set covering the target well section. The set header records the mapping table version number and a lookup log summary. The set body retains the primary hit entries and a list of candidate mappings, used as statistical basis for subsequent configuration updates.

[0056] Understandably, this step specifies the output field name as sedimentary microfacies type result in the output stage, and indicates in the data exchange contract that the next consumption position is the sedimentary microfacies type result of S430; at the same time, in the cross-master step description, it is noted that this product has a traceability relationship with the segment classification result of the previous S330. When S430 updates the threshold table and segment length threshold configuration, it will read both at the same time to complete the version reorganization of rules and parameters.

[0057] S430. The threshold table and segment length threshold configuration of the sedimentation microfacies type results and the segment classification results are updated to generate a cleaning rule configuration structure. This step receives the sedimentary microfacies type results output by S420 and the segment classification results output by S330 in parallel. Both inputs include location and tracing indices, supporting the retrieval of the preceding data generation chain from the current determination. Specifically, the system first enters the statistical aggregation stage, which refers to establishing a multi-dimensional statistical overview around microfacies and segment entries. The system aggregates sedimentary microfacies type results by well interval, stratigraphy, and morphological type, extracting segment length distribution, platform coverage ratio segments, transition density overview, and adjacent transition relationship counts. Then, based on the segment classification results, it statistically analyzes the length distribution of short and long segments, the coverage ratio overview of stable segments, and the proportion of masked neighboring markers. The statistical products are written to the update basis bin, which records quantile snapshots, anomaly ratio snapshots, and cross-version comparison tables. The cross-version comparison tables are used to compare the changes with the previous version configuration.

[0058] Further, the system enters the threshold table update process. The threshold table refers to the set of parameters that provide thresholds for the preceding segmentation, judgment, and mapping stages. Entries include minimum cavity threshold, minimum stable segment length, overlap adjudication priority key sorting, candidate trigger threshold groups, and boundary-specific rule thresholds. The system reads snapshots of quantile values ​​from the update basis bin and establishes a reference interval for each threshold. This reference interval is not expressed by a formula but is presented through robust statistical descriptions and versioned records. If the difference between the current observation and the previous configuration exceeds a preset change threshold, the system registers the new value, source, scope of impact, and rollback strategy for that threshold in the candidate update table. For segment length threshold configuration updates, the system calculates the boundary interval between short and long segment distributions at the segment level, and, combined with the segment length bandwidth corresponding to different morphologies in the sedimentary microfacies type results, provides a suggested group of segment length thresholds. If there are significant differences in bandwidth between different layers within the same well segment, the system generates a layered threshold group in the candidate update table, along with layered limitation conditions. After the candidate update table is formed, the system enters the governance verification stage, which involves assessing the compliance, conflict, and impact of the candidate updates. The system checks the compatibility of candidate updates with rule version library entries one by one. If an update conflicts with a rule entry, the system records the rule conflict in the governance log and provides a handling suggestion of keeping the old value or enabling hierarchical threshold groups. At the same time, the system calculates an impact list, which marks the links and entry numbers that may be affected, making version releases traceable.

[0059] After completing the governance verification, the system generates a cleaning rule configuration structure. This structure serves as a unified configuration carrier for S110 and subsequent steps, containing a threshold table, segment length thresholds, bridging and overlap decision parameters, morphological candidate window schemes, boundary-specific rules, masking deployment strategies, and versioned metadata. The structure header includes the configuration version number, generation time, source statistical summary, and rollback strategy. The structure body stores various parameters in partitions, with partition names consistent with the process terminology. During the output preparation phase, the system aligns the execution differences between this update and the previous configuration version. The alignment result is written to a configuration difference summary, which is synchronized to the operations and maintenance review channel for pre-delivery verification. Furthermore, the system specifies in the data exchange contract that the output field name is cleaning rule configuration, and clarifies that the next consumption position is the cleaning rule configuration of S110; in the cross-master step description, it is noted that this update has read and write associations with S200, S300, and S400, among which the bridging and overlap decision parameters will affect the splicing strategy of S410, the morphological candidate window scheme and boundary-specific rules will affect the table entry search path of S420, and the segment length threshold will affect the segment determination of S220.

[0060] In summary, the technical effects of this step are as follows: By performing versioned statistics, candidate generation, and governance verification on the sedimentary microfacies type results and segment classification results, the cleaning rule configuration structure is updated and written back to S110, forming a closed loop from data cleaning to identification output and then to parameter reconfiguration; the thresholds and strategies are centrally stored in the structured carrier, and the operation link obtains a stable and traceable parameter source.

[0061] Figure 2 A well logging curve morphology classification diagram provided for embodiments of this application, such as Figure 2 As shown, this diagram illustrates typical logging curve morphologies such as box-shaped, bell-shaped, funnel-shaped, and finger-shaped curves and their corresponding relationships in different sedimentary facies. It forms the knowledge basis for constructing the microfacies mapping table in this invention. Specifically, in step S420, after extracting morphological segments from the identified curve segment category sequence, the invention automatically searches and matches based on the mapping rules defined by this classification relationship (including morphological type, smoothness, and corresponding sedimentary microfacies types such as meandering river, braided river, meandering river delta, and lacustrine facies). This objectively and efficiently transforms the curve morphological characteristics into sedimentary microfacies type results. The introduction of this diagram ensures the professionalism of the identification process and the geological rationality of the conclusions, making the automated identification based on evidence.

[0062] like Figure 3The diagram illustrates the overall process flow of the intelligent sedimentary microfacies segmentation and identification method of this invention. It clearly demonstrates the core processing chain of this invention: starting with the input of raw data, data preprocessing is performed, corresponding to the data cleaning, standardization, and filtering steps in this invention, to generate high-quality initial data. The preprocessed data is then processed through a curve segmentation step, corresponding to the segment boundary adjustment and segment sample set construction in this invention, dividing the continuous curve into geologically significant independent units. Subsequently, through a curve morphology identification step, corresponding to feature extraction, classification inference, and rule mapping in this invention, the morphological category of each curve unit is identified. Finally, all identification results are integrated to achieve the overall goal of sedimentary microfacies identification. This process highlights the modular and sequential design philosophy of this invention, ensuring the logical rigor and operability from raw data to geological interpretation conclusions.

[0063] Figure 4 A specific network structure diagram of an automatic sedimentary microfacies segmentation and identification model provided in this application embodiment is shown below. Figure 4 As shown, this corresponds to the step of performing long-segment LSTM-FCN inference fusion processing on the multi-path feature set as described in the claims. The model receives the segmented curve segments as input and employs a dual-path parallel processing architecture: the upper path is a long sequence dependency capture pathway, using at least two LSTM layers combined with a Dropout layer to effectively extract long-range temporal context features of the curve segments; the lower path is a local morphological feature extraction pathway, using multiple convolutional layers interleaved and stacked with Squeeze and Excite layers to focus on and enhance key information in the local morphology of the curve, which is then summarized by a global pooling layer. Finally, the temporal features output from the upper path and the morphological features output from the lower path are cascaded and fused, and the output layer maps them to classification results for sedimentary microfacies types such as box-shaped, bell-shaped, funnel-shaped, and finger-shaped. This dual-path fusion structure fully leverages the respective advantages of LSTM in long sequence modeling and CNN in local feature extraction, jointly improving the accuracy and robustness of sedimentary microfacies recognition.

[0064] Figure 5 This application provides a comparison chart of the effects of preprocessing and automatic segmentation of raw well logging curves, as shown in the embodiments of this application. Figure 5As shown, the left column waveform GR_norm displays the original GR curve after standardization, which contains many spikes and noise. The right column waveform GR_filt_norm displays the result after wavelet denoising and median filtering, with a smoother curve shape. Multiple segments of different colors in the background clearly mark the effective segmentation boundaries automatically identified by the method of this invention. This effect visually demonstrates that the data cleaning, standardization, and filtering processes in steps S110 to S130 of this invention can effectively suppress noise interference and highlight the basic shape trend of the curve, laying a reliable data foundation for subsequent accurate segmentation and feature extraction. The different colored segments shown in the figure will be directly input into the subsequent classification and recognition module as segment sample sets.

[0065] Figure 6 The final sedimentary microfacies identification effect diagram of a segment of actual well logging data provided in this application embodiment is shown below. Figure 6 As shown in the figure, this diagram intuitively illustrates the correspondence between the curve segment category sequence (waveform in the figure) obtained after the sequence splicing process of step S410 of this invention and the sedimentary microfacies type results output after the rule mapping of step S420 (colored markers on the right side of the figure). As shown, the identified microfacies types (including box-shaped, bell-shaped, funnel-shaped, symmetrical, and background) are clearly and continuously marked vertically (corresponding to depth / time series), and highly consistent with the changing trends of the original curve morphology. This effect powerfully demonstrates that the complete technical solution constructed by this invention, from data preprocessing and intelligent segmentation to classification and rule mapping, can achieve high-precision and automated identification of sedimentary microfacies types. The results are clear and intuitive, possessing significant field guidance value.

Claims

1. A method for intelligent segmentation and identification of sedimentary microfacies based on well logging curves, characterized in that, include: The GR curve and cleaning rule configuration are obtained, and data cleaning and standardization are performed. Wavelet transform denoising is performed based on the wavelet basis order and threshold strategy table in the cleaning rule configuration, and median filtering is performed based on the median filter window width in the cleaning rule configuration to generate a coarse segment set. The system obtains a coarse segment set and well logging interpretation conclusions, and performs processing including search half-window width, matching priority, conflict resolution strategy, interface alignment and boundary fine-tuning. This includes binding interpretation interfaces within the candidate boundary neighborhood by the search half-window width, making small-range position corrections based on local window statistics, performing segment splitting and merging operations when the coarse segment crosses multiple interpretation interfaces, adjusting segment boundaries, performing statistical analysis based on the minimum segment length threshold and long segment determination threshold in the cleaning rule configuration, and performing index allocation based on the sample numbering strategy in the cleaning rule configuration to generate a segment sample set. The process includes target sampling interval, longest sequence length, short segment window width, long segment sliding stride, anomaly label mapping table and mask propagation strategy and feature extraction. For short segments, slope trend rule classification is performed based on the monotonic duration threshold, transition density threshold, main transition strength threshold and platform coverage ratio threshold in the rule version library. For long segments, LSTM-FCN network is used for inference classification based on the fusion weights in the model parameters. The classification inference process is performed to generate segment classification results. Based on the minimum hole threshold and bridging strategy in the cleaning rule configuration, sequence splicing is performed, and rule mapping is performed based on the morphology type, segment length range, stable segment coverage ratio, adjacent category transition relationship and layer limit in the micro-phase mapping table. Furthermore, segment length threshold configuration is updated based on the quantile value snapshot and anomaly percentage snapshot in the update basis bin, generating the cleaning rule configuration structure.

2. The method according to claim 1, characterized in that, The process of generating a coarse segment set also includes: Obtain the GR curve and cleaning rule configuration. Based on the missing identification rules, anomaly type dictionary, depth unit and sampling interval standard, shielded well section list, drift verification baseline, boundary extrapolation strategy, interpolation method priority and log recording level in the cleaning rule configuration, perform missing correction and anomaly labeling to obtain the cleaning data. Depth and curve values ​​are extracted from the cleaned data. Based on the target sampling interval given by the cleaned rules, depth alignment is performed. Based on the mapping strategy table configured by the cleaned rules, interval normalization is performed to generate a standardized sequence. For the standardized sequence, wavelet denoising is performed based on the wavelet basis order and threshold strategy table in the cleaning rule configuration, and median filtering is performed based on the median filter window width and shielding segment participation strategy in the cleaning rule configuration to generate a coarse segment set structure.

3. The method according to claim 2, characterized in that, GR curve and cleaning rule configuration includes: The GR curve represents the relationship between well depth and gamma response. The recording unit, sampling interval, and depth reference are provided by the field recording system or historical calibration records. The cleaning rule configuration includes missing identification rules, anomaly type dictionary, depth unit and sampling interval standard, shielded well section list, drift verification baseline, boundary extrapolation strategy, interpolation method priority, and log recording level.

4. The method according to claim 1, characterized in that, The sequence splicing process, based on the minimum hole threshold and bridging strategy in the cleaning rule configuration, includes: The sedimentary microfacies type results are aggregated according to well section, stratigraphic position and morphological type. The distribution of section length, the proportion of platform coverage, the density of transition and the count of adjacent transition relationships are extracted. The length distribution of short and long sections, the proportion of stable section coverage and the proportion of shielded adjacent markers are statistically analyzed according to the section classification results. The statistical products are written into the update basis bin.

5. The method according to claim 4, characterized in that, include: The update is based on the warehouse record quantile snapshot, the abnormal percentage snapshot, and the cross-version comparison table. The cross-version comparison table is used to compare the changes with the previous version configuration.

6. The method according to claim 1, characterized in that, The process of updating the segment length threshold configuration and generating the cleaning rule configuration structure also includes: The boundary interval between short and long segments is calculated at the segment level. Combined with the segment length bandwidth corresponding to different morphologies in the sedimentary microfacies type results, a suggested group of segment length thresholds is given. If there are significant differences in bandwidth between different layers in the same well segment, a layer threshold group with layer limit conditions is generated.

7. The method according to claim 1, characterized in that, The process of generating the cleaning rule configuration structure also includes: Check the compatibility of candidate updates with rule version library entries one by one. If there is a conflict, record the rule conflict and provide handling suggestions. At the same time, calculate and mark the impact list of potentially affected links and entries.

8. The method according to claim 4, characterized in that, The cleaning rule configuration structure includes a threshold table, segment length threshold, bridging and overlap decision parameters, morphological candidate window scheme, boundary-specific rules, mask distribution strategy, and versioned metadata; the structure header contains the configuration version number, generation time, source statistics summary, and rollback strategy.

9. The method according to claim 1, characterized in that, The process of updating the segment length threshold configuration and generating the cleaning rule configuration structure also includes: Output preparation phase: Align the configuration execution differences between this update and the previous version, write the difference alignment results into the configuration difference summary, and synchronize the summary to the operation and maintenance review channel.

10. The method according to claim 1, characterized in that, Well logging interpretation conclusions include: Well logging interpretation conclusions refer to information such as the layer name, interface type, interface depth point or narrow range, confidence index, and interpretation version number given for the target well section. The source can be annotations by interpreters or automatic pushes from the rule base.

Images