An oil content prediction method, device, equipment and storage medium

Abstract

The application discloses an oil content prediction method, device, equipment and storage medium. The method comprises the following steps: acquiring a first object containing a well completion section and a corresponding depth interval, multi-curve logging time series data and an adjacent relationship containing an initial node and an edge set; constructing a first algorithm model to generate a section embedding vector representing lithology pore structure and fluid response; calculating embedding similarity based on the section embedding vector, and generating a gated graph structure combined with the adjacent relationship; constructing a second algorithm model, inputting the graph structure and the section embedding vector, and outputting a section oil content estimate value; filtering high oil content sections according to the estimate value and sorting to form a well internal layering result containing a section identifier. The application constrains the propagation range and order of oil content information, takes into account cross-well information utilization and section-level feature representation, and improves the accuracy and availability of old well layer oil content estimation.

Description

Technical Field

[0001] This application relates to the field of oil and gas exploration and development technology, and in particular to a method, apparatus, equipment and storage medium for predicting oil content. Background Technology

[0002] As oil and gas exploration and development enters a mature stage, many old oil and gas wells, after long-term production, generally face problems such as inconsistencies in inter-well data, complex stratigraphy, and degraded logging responses. There is an urgent need to accurately assess the oil content of each stratigraphic segment at the completion stage to provide a reliable basis for subsequent replenishment plans, selection of intervention well locations, and intra-well stratification decisions. Multi-curve logging data, acquired at depth, possesses multi-channel, temporal, and stratigraphic differences. Cross-well information comparison and constraints on oil-bearing information propagation have become crucial for improving the reliability of oil content prediction in old wells.

[0003] Current oil content prediction methods are mainly divided into two categories: one is based on single-well logging interpretation or general learning models, which only extracts and regresses features from segmented curves of a single well. In scenarios with sparse data or degraded logging curves, it is difficult to maintain prediction stability. The other category constructs candidate adjacency relationships based on well location distance and layer overlap thresholds, and utilizes neighborhood information in the graph framework. However, it mostly uses manual or shallow statistical features as segment-level representations, which makes it difficult to simultaneously characterize lithological pore structure and fluid response. Furthermore, it does not combine segment-level characterization for gating, which easily leads to cross-layer misconnections and makes it difficult to directly form a list of high oil-bearing segments that can be used in engineering.

[0004] Therefore, how to improve the accuracy of oil content estimation in old wells and the usability of well stratification results in complex scenarios, and directly output a set of high oil-bearing strata and their identification to meet the needs of engineering decision-making, is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the oil content prediction method, apparatus, equipment, and storage medium provided in this application can improve the accuracy and usability of oil content estimation in old well formations, and can directly provide engineering basis for supplementary production and in-well stratification decisions. The oil content prediction method, apparatus, equipment, and storage medium provided in this application are implemented as follows:

[0006] This application provides an oil content prediction method, including:

[0007] Acquire multiple first objects, time series data, and adjacency relationships. The first objects include well completion segments and corresponding depth intervals. The time series data includes the multi-curve logging sequence of the first objects. The adjacency relationships include the initial nodes and edge sets of the first objects.

[0008] A first algorithm model is constructed, and a segment embedding vector is determined based on the first algorithm model and the time series data. The segment embedding vector is used to characterize the features of lithological pore structure and fluid response.

[0009] The embedding similarity is determined based on the segment embedding vector.

[0010] The graph structure is determined based on the adjacency relationship and the embedding similarity, and the graph structure includes a gated set of nodes and edges;

[0011] A second algorithm model is constructed, and the graph structure and the segment embedding vector are input into the second algorithm model to obtain the estimated oil content of the segment.

[0012] The well stratification results are determined based on the estimated oil content of the defined intervals. The well stratification results include a set of high oil-bearing intervals and corresponding interval identifiers.

[0013] In some embodiments, the first object, time-series data, and adjacency relationships include:

[0014] The first object is obtained based on the well completion segments and the corresponding depth intervals, wherein the depth intervals are used to characterize the starting and ending depths of the well completion segments.

[0015] A multi-curve logging sequence is acquired, and the multi-curve logging sequence and the first object are correlated to obtain time-series data. The multi-curve logging sequence includes gamma curves, density curves, neutron curves, and sonic curves.

[0016] Obtain distance thresholds and stratigraphic thresholds. The distance thresholds are used to determine the candidate range of inter-well distances, and the stratigraphic thresholds are used to determine the candidate range of stratigraphic relationships in well completion segments.

[0017] The initial node is determined based on multiple first objects;

[0018] Traverse multiple first objects, and when the distance between any two first objects meets the distance threshold and the hierarchical relationship of the completion segments of the two first objects meets the hierarchical threshold, determine the edge corresponding to the two first objects;

[0019] The adjacency relationship is determined based on the initial node and all edges.

[0020] In some embodiments, the first algorithm model includes a patch partitioning module, a masking module, a reconstruction module, and an embedding generation module. The construction of the first algorithm model, and the determination of segment embedding vectors based on the first algorithm model and the time-series data, includes:

[0021] Obtain a patch length threshold and a masking ratio threshold. The patch length threshold is used to determine the length of the depth window, and the masking ratio threshold is used to determine the proportion of masked patches in the patch sequence.

[0022] The patch sequence is obtained by performing deep window segmentation on the time series data according to the patch length threshold by the patch segmentation module.

[0023] The patch sequence is masked by the masking module according to the masking ratio threshold to obtain the masked patch, the unmasked patch, and the masking identifier.

[0024] The depth location code and curve channel code are obtained. The depth location code is used to identify the depth order of each patch in the patch sequence, and the curve channel code is used to identify the source of the gamma curve, density curve, neutron curve and acoustic curve in the time series data.

[0025] The reconstruction module reconstructs the patch sequence, the depth location encoding, the curve channel encoding, and the masking identifier to obtain a reconstructed patch.

[0026] The embedding generation module performs a fusion process on the unmasked patch and the reconstructed patch to obtain a fused patch.

[0027] The layer embedding vector is determined based on the masking ratio threshold and the fused patch.

[0028] In some embodiments, determining the graph structure based on the adjacency relationship and the embedding similarity includes:

[0029] Obtain the embedding similarity corresponding to each edge in the adjacency relationship;

[0030] A first threshold is obtained, and the embedding similarity of each edge is filtered against the first threshold to obtain a gated set of edges.

[0031] The initial graph structure is obtained based on the initial nodes in the adjacency relationship and the gated edge set;

[0032] The edge weight corresponding to each edge is obtained based on the embedding similarity.

[0033] Based on the initial graph structure and the edge weights corresponding to each edge, an adjacency list is obtained;

[0034] The effective neighbor set is obtained based on the first threshold and the adjacency list;

[0035] The final graph structure is obtained based on the set of effective neighbors, the edge weights corresponding to each edge, and the initial graph structure.

[0036] In some embodiments, determining the well stratification result based on the estimated oil content of the strata includes:

[0037] The estimated oil content of the layer is associated with the corresponding first object to obtain the associated first object;

[0038] A high oil content determination threshold is obtained, and the estimated oil content of each layer segment of the first object is compared with the high oil content determination threshold to obtain the high oil content determination result.

[0039] The high oil content determination results are filtered and processed to obtain a set of high oil content layers;

[0040] The high oil-bearing layer set is sorted according to the layer identification rules to obtain the identification of each layer in the high oil-bearing layer set;

[0041] The high oil-bearing layer set is sorted according to the in-well layer sorting rules to obtain the sorted high oil-bearing layer set.

[0042] The well stratification results are obtained based on the sorted set of high oil-bearing sections and the identifiers of each section in the set of high oil-bearing sections.

[0043] In some embodiments, the formula for calculating the embedding similarity based on the segment embedding vector is as follows:

[0044]

[0045] in, To embed similarity, For combined weights, This is the segment embedding vector corresponding to the k-th completion segment of the i-th well. Let f(m) be the segment embedding vector corresponding to the m-th completion segment of the j-th well. and The norm of the embedded vector, The depth overlap ratio, This is the transpose of a vector.

[0046] In some embodiments, the first algorithm model is a time-series patch masking model.

[0047] This application provides an oil content prediction device, comprising:

[0048] The acquisition module is used to acquire multiple first objects, time series data, and adjacency relationships. The first objects include well completion segments and corresponding depth intervals. The time series data includes the multi-curve logging sequence of the first objects. The adjacency relationships include the initial nodes and edge sets of the first objects.

[0049] A construction module is used to construct a first algorithm model and determine a segment embedding vector based on the first algorithm model and the time series data. The segment embedding vector is used to characterize the features of lithological pore structure and fluid response.

[0050] A determination module is used to determine the embedding similarity based on the segment embedding vector;

[0051] The determining module is further configured to determine a graph structure based on the adjacency relationship and the embedding similarity, wherein the graph structure includes a gated set of nodes and edges;

[0052] The construction module is also used to construct a second algorithm model, inputting the graph structure and the segment embedding vector into the second algorithm model to obtain the estimated oil content of the segment;

[0053] The determining module is further configured to determine the well stratification result based on the estimated oil content of the layer, wherein the well stratification result includes a set of high oil-bearing layers and corresponding layer identifiers.

[0054] The computer device provided in this application includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the method described in this application.

[0055] The computer-readable storage medium provided in this application embodiment stores a computer program thereon, which, when executed by a processor, implements the method provided in this application embodiment.

[0056] The oil content prediction method, apparatus, equipment, and storage medium provided in this application acquire a first object containing well completion segments and corresponding depth intervals, multi-curve logging time-series data, and adjacency relationships containing initial node and edge sets; construct a first algorithm model to generate segment embedding vectors characterizing lithological pore structure and fluid response; calculate embedding similarity based on segment embedding vectors, and generate a gated graph structure by combining adjacency relationships; construct a second algorithm model, input the graph structure and segment embedding vectors, and output estimated oil content values ​​for the segments; select and sort high-oil-content segments based on the estimated values ​​to form in-well stratification results with segment identifiers. In this way, by constraining the propagation range and order of oil-content information, and taking into account both cross-well information utilization and segment-level feature characterization, the accuracy and usability of oil content estimation in old well formations are improved, solving the technical problems mentioned in the background art. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A schematic diagram illustrating the implementation process of an oil content prediction method provided in this application embodiment;

[0059] Figure 2 This is a schematic diagram illustrating an implementation process for obtaining multiple first objects, time-series data, and adjacency relationships, provided in an embodiment of this application.

[0060] Figure 3 This is a schematic diagram of an oil content prediction device provided in an embodiment of this application. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0062] The following description of some technologies involved in the embodiments of this application is provided to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, some descriptions of well-known functions and structures are omitted in the following description.

[0063] Figure 1 This is a schematic diagram illustrating the implementation flow of an oil content prediction method provided in an embodiment of this application, including steps 101 to 106. Wherein, Figure 1 This is merely one execution order shown in the embodiments of this application and does not represent the only execution order for an oil content prediction method. Where the final result can be achieved, Figure 1 The steps shown can be performed in parallel or in reverse order.

[0064] Step 101: Obtain multiple first objects, time series data, and adjacency relationships.

[0065] In this embodiment of the application, the acquisition of the first object involves binding the completion segments of a single well with their corresponding depth intervals to form the first object. Specifically, the starting depth and ending depth of each completion segment are first determined, and the two together constitute the depth interval of that segment, thereby defining the spatial range of the first object.

[0066] Regarding the acquisition and correlation of time-series data, gamma curves, density curves, neutron curves, and acoustic curves were collected and arranged in depth order to form a multi-curve logging sequence. The multi-curve logging sequence was then correlated with its corresponding first object, ensuring that each first object has a unique corresponding multi-curve logging sequence, i.e., time-series data.

[0067] The construction of adjacency relationships requires setting two key thresholds: a distance threshold and a stratum threshold. The distance threshold limits the range of well distances participating in adjacency relationship construction; only wells with distances not exceeding this threshold are included in the candidate range. The stratum threshold limits the stratum matching relationship of completion segments, ensuring that segments participating in adjacency are comparable in strata. Using the set of all first objects as initial nodes, any two first objects are evaluated. If the well distance between them satisfies the first distance threshold, and the stratum relationship of the completion segments satisfies the second stratum threshold, then an edge is formed between the initial nodes of these two first objects. All initial nodes and the edges meeting the conditions together constitute a candidate adjacency relationship.

[0068] Step 102: Construct the first algorithm model and determine the segment embedding vector based on the first algorithm model and time series data.

[0069] In this embodiment of the application, a first algorithm model is constructed, which is a time-series patch masking model. The first algorithm model consists of a patch partitioning module, a masking module, a reconstruction module, and an embedding generation module. The time-series data is processed by the first algorithm model to generate segment embedding vectors.

[0070] Specifically, a patch length threshold is set, which determines the length of the depth window. The time series data is segmented in a non-overlapping manner according to this length to obtain several patches. These patches are arranged into a patch sequence in depth order.

[0071] A masking ratio threshold is set to control the number of masks in the patch sequence. According to this ratio threshold, some patches in the patch sequence are randomly marked for masking to distinguish between masked and unmasked patches, and corresponding masking identifiers are generated. At the same time, the masked patches are replaced with a uniform masking placeholder vector to form a masked patch sequence.

[0072] To preserve depth order information and curve origin information, depth location encoding and curve channel encoding are defined. Depth location encoding identifies the depth order of each patch in the patch sequence, while curve channel encoding distinguishes the origins of four types of curves: gamma, density, neutron, and acoustic. The masked patch sequence, depth location encoding, curve channel encoding, and masking identifier are input into the reconstruction module. The reconstruction module then reconstructs the masked patches based on information from the unmasked patches, resulting in the reconstructed patch sequence.

[0073] According to the patch aggregation rule, which determines the fusion weights of unmasked patches and reconstructed patches based on depth location encoding and curve channel encoding, the sum of the fusion weights of all patches is 1. Under the constraint of a masking ratio threshold, unmasked patches and reconstructed patches are weighted and fused to generate a segment embedding vector. This vector can simultaneously characterize the features of lithological pore structure and fluid response, and is associated with the corresponding first object to form a set of segment embedding vectors.

[0074] Step 103: Determine the embedding similarity based on the segment embedding vector.

[0075] In this embodiment, based on the generated segment embedding vector, the embedding similarity between different first objects is calculated, and then a graph structure is generated by combining the candidate adjacency relationship.

[0076] Embedsion similarity is calculated using a joint metric, combining the correlation between layer embedding vectors and the depth overlap ratio of the first object. The depth overlap ratio is the ratio of the intersection length of the depth intervals corresponding to two first objects to the sum of the lengths of these two depth intervals, used to quantify the degree of matching in layer relationships. By comprehensively considering both vector correlation and depth overlap ratio, an embedding similarity that fully reflects the comparability between first objects is obtained.

[0077] Step 104: Determine the graph structure based on adjacency relationships and embedding similarity.

[0078] In this embodiment, each edge in the candidate adjacency relationship is traversed, and the embedding similarity between the two first objects corresponding to the edge is compared with a first threshold. If the embedding similarity meets the first threshold, the edge is retained, and the corresponding edge weight is determined according to the magnitude of the embedding similarity; if it does not meet the threshold, the edge is removed, forming a gated edge set.

[0079] The initial nodes in the candidate adjacency relationships are combined with the gated edge set to obtain the initial graph structure. For each node in the initial graph structure, an adjacency list is established, with nodes arranged in descending order of edge weight to limit the propagation order of oil-containing information. Under the constraint of a first threshold, the adjacency list is truncated, retaining only the adjacent nodes whose embedding similarity meets the first threshold, forming the effective neighbor set. The effective neighbor set and its corresponding edge weights are then incorporated into the initial graph structure to obtain the final graph structure, which can constrain the propagation path and range of oil-containing information.

[0080] Step 105: Construct a second algorithm model by inputting the graph structure and the segment embedding vector into the second algorithm model to obtain the estimated oil content of the segment.

[0081] In this embodiment of the application, a second algorithm model, namely a graph transformation network, is constructed. This model consists of a node feature input module, an edge weight gating module, a neighborhood aggregation module, and a regression output module. The graph structure and the layer embedding vector are input into the model, and the estimated oil content of the layer is output.

[0082] Specifically, the segment embedding vector is used as a node feature and bound to the nodes in the graph structure to ensure that each node has a corresponding feature vector.

[0083] A second threshold is set to limit the lower bound of edge weight propagation. For each node's effective neighbor set, the propagation coefficient is calculated based on the relationship between the edge weight and the second threshold. If the edge weight satisfies the second threshold, the propagation coefficient is calculated based on the relative proportion of that edge weight to other edge weights that satisfy the condition; if the edge weight does not satisfy the second threshold, the corresponding propagation coefficient is set to zero, thereby filtering out weakly correlated information propagation.

[0084] Using the calculated propagation coefficient as weights, the node features of each neighbor in the effective neighbor set are weighted and aggregated to obtain a node aggregation vector. This vector reflects the comprehensive influence of neighboring nodes on the current node. It is then connected with the current node's own segment embedding vector to form a regression input vector, which simultaneously contains the node's own features and neighborhood influence features.

[0085] The regression input vector is fed into a preset regression model to perform oil content regression calculations on the segments corresponding to each first object, outputting the estimated oil content value for each segment. The estimated oil content value of each segment is then correlated with the corresponding first object to form an oil content distribution, clearly presenting the oil content situation of each segment.

[0086] Step 106: Determine the well stratification results based on the estimated oil content of the strata. The well stratification results include a set of high oil-bearing strata and corresponding strata identifiers.

[0087] In this embodiment, a high oil content threshold is first set to distinguish between high oil content layers and non-high oil content layers. The estimated oil content of each layer corresponding to the first object is compared with the threshold. If the estimated value is greater than or equal to the high oil content threshold, the layer is determined to be a high oil content layer; otherwise, it is a non-high oil content layer. Based on the determination result, all high oil content layers are selected to form a set of high oil content layers.

[0088] Subsequently, segment identifiers are generated. According to the segment identifier rules, the segment identifier is composed of the completion segment number of the first object and the corresponding depth interval, ensuring that each high oil-bearing segment has a unique and identifiable identifier.

[0089] Finally, a sorting process is performed. Following the in-well stratification sorting rules, the high-oil-cut intervals are first sorted in descending order of estimated oil content. If intervals have the same estimated oil content, they are then sorted in ascending order of the starting depth of the depth range to ensure the stability and rationality of the sorting results. The sorted high-oil-cut intervals are then combined with their corresponding interval identifiers to form the final in-well stratification result.

[0090] This application's embodiments, by embedding similarity-gated adjacency relationships and graph structure constraints on propagation range, ensure that oil-bearing information is only transmitted between comparable and corresponding well-layer segments, reducing errors caused by static adjacency and weakly related information diffusion. This adapts to scenarios with inconsistent data between old wells and complex stratigraphic systems. Based on the segment embedding vector generated by the first algorithm model, it simultaneously characterizes lithological pore structure and fluid response features, overcoming the shortcomings of existing technologies that rely on manual methods and shallow features to simultaneously capture two types of key information. Without additional post-processing, it directly generates in-well stratification results containing a set of high-oil-bearing segments and corresponding identifiers, solving the problem of relying on numerous rules for screening and sorting, which hinders rapid engineering decision-making. This improves the practicality and efficiency of the results, providing direct basis for re-production and layer selection in old wells. Addressing the issues of degraded logging responses and data sparsity in old wells, it enhances feature robustness through masking reconstruction and aggregates effective neighborhood information through graph transformation networks, reducing the impact of data defects on prediction results. Compared to single-well modeling or static graph methods, it significantly improves the accuracy and stability of oil-bearing estimation and in-well stratification.

[0091] In the above Figure 1 Based on this, this application also provides a schematic diagram of the implementation process for obtaining multiple first objects, time-series data, and adjacency relationships, such as... Figure 2 As shown, steps 201 to 206 are included:

[0092] Step 201: Obtain the first object based on the well completion segments and the corresponding depth ranges.

[0093] In this embodiment, the completion segment information of a single well is extracted to clarify the physical extent of each completion segment. That is, the start and end positions of the segment downhole are determined to define the depth boundary of the completion segment.

[0094] By accurately obtaining the starting and ending depths of each completion segment through well logging depth calibration data, the two together constitute the depth range of that completion segment. For example, the third completion segment of a single well, after calibration, has a starting depth of 1200 meters and an ending depth of 1250 meters, so its depth range is 1200-1250 meters.

[0095] Each well completion segment is bound to its corresponding depth range to form a first object. Each first object must contain unique identification information and its corresponding depth range.

[0096] Step 202: Obtain the multi-curve logging sequence, perform correlation processing on the multi-curve logging sequence and the first object, and obtain time series data.

[0097] In this embodiment of the application, four types of key logging curves related to the first object are retrieved, including gamma curve, density curve, neutron curve, and sonic curve. These four types of curves are data continuously collected at depth, and can respectively reflect the lithology of the formation (gamma curve), rock density (density curve), porosity (neutron curve), and sonic propagation velocity (sonic curve), which are the core data characterizing the lithological pore structure and fluid response.

[0098] Since well logging curves are acquired based on depth, each type of curve needs to be arranged in a uniform depth order, either from shallow to deep or from deep to shallow, to ensure that the temporal sequence of the curve data is consistent with the actual downhole depth distribution. For example, the well logging values ​​corresponding to depths of 1200 meters, 1201 meters, ..., 1250 meters in the gamma curves can be arranged in ascending order of depth to form a gamma curve sequence within that depth range.

[0099] Using the depth range of the first object as the matching criterion, the sorted multi-curve logging sequences are bound to the corresponding first object. Specifically, if the depth range of a first object is 1200-1250 meters, all data from the four types of curves falling within the 1200-1250 meter range are extracted to form a dedicated multi-curve logging sequence for that first object, i.e., time-series data. Each first object corresponds to only one set of time-series data, ensuring accurate correspondence between data and layers and avoiding data confusion across layers.

[0100] Step 203: Obtain the distance threshold and the layer threshold.

[0101] In this embodiment, the distance threshold is used to limit the range of wells participating in the construction of adjacency relationships, avoiding the inclusion of wells with excessively large spatial distances or significant differences in geological conditions in the calculation. The threshold should be set with reference to the well network density within the block. For example, if the well spacing in an old well block is generally within 500 meters, the distance threshold is set to 500 meters, allowing only wells with a distance ≤ 500 meters to form candidate adjacencies; if the well network in the block is sparse, the distance threshold can be adjusted to 800 meters to ensure a sufficient number of candidate adjacency objects.

[0102] Stratification thresholds are used to define the stratigraphic matching relationship of well completion segments, ensuring that adjacent segments belong to the same or similar geological strata. The stratigraphic threshold is set based on the regional geological stratification report, with depth overlap ratio as the core quantitative indicator. For example, if the regional geological report specifies that the depth overlap of the same stratum must be ≥50%, then the stratigraphic threshold is set to a depth overlap ratio ≥50%. Only when the depth overlap ratio of two first-object segments meets this condition is the stratigraphic relationship considered to meet the requirements.

[0103] Step 204: Determine the initial node based on multiple first objects.

[0104] In this embodiment of the application, all first objects with bound depth ranges within the same exploration block are aggregated to form a first object set, ensuring coverage of all old well completion segments with predictable oil content within the block.

[0105] Each first object in the first object set is directly mapped to an initial node. Each initial node must carry the core information of its corresponding first object, including the well number, completion segment number, and depth range, so as to quickly retrieve data when calculating inter-well distances and stratigraphic relationships later. For example, the first object corresponding to well A-segment 3 is mapped to the initial node V-A3, and associated with the depth range information of 1200-1250 meters.

[0106] All the mapped initial nodes form the initial node set, which is the set of vertices with adjacency relationships.

[0107] Step 205: Traverse multiple first objects. When the well distance between any two first objects meets the distance threshold and the stratigraphic relationship of the well completion segments of the two first objects meets the stratigraphic threshold, determine the edge corresponding to the two first objects.

[0108] In this embodiment, a pairwise traversal method is used to combine all the first objects corresponding to the initial node set. That is, two different first objects are selected sequentially from the first object set to form first object pairs, ensuring that all possible combinations are covered without repetition or omission.

[0109] For each pair of first objects, retrieve the wellhead coordinates of the corresponding well and calculate the horizontal distance between the two wells (i.e., the distance between wells). Compare the calculated distance between wells with a preset distance threshold: if the distance between wells is less than or equal to the distance threshold, the distance constraint is satisfied; if the distance between wells is greater than the distance threshold, the first object pair is directly excluded and no edge is generated.

[0110] For the first object pair that satisfies the distance constraint, the layer relationship is further determined. The depth interval overlap ratio of the two first objects (i.e., the intersection length of the two depth intervals divided by the sum of the lengths of the two depth intervals) is calculated and compared with a preset layer threshold: if the depth overlap ratio is ≥ the layer threshold, the layer constraint is satisfied; if the depth overlap ratio is < the layer threshold, the object pair is excluded and no edge is generated.

[0111] An edge is generated between the two initial nodes of the first object pair only if both the distance constraint and the layer constraint are satisfied. Each edge must be labeled with the identifiers of its two associated initial nodes, and the basis for the edge's generation must be recorded.

[0112] Step 206: Determine the adjacency relationship based on the initial node and all edges.

[0113] In this embodiment of the application, all edges generated that satisfy the constraints are aggregated to form an edge set, ensuring that each edge is associated with two initial nodes and that there are no invalid edges.

[0114] The initial set of nodes and the set of edges are combined to form adjacency relationships. Adjacency relationships include all associated edges corresponding to each initial node and the two initial nodes corresponding to each edge. Essentially, they describe which initial nodes have associations that meet certain conditions.

[0115] The adjacency relationships after combination are validated. This involves checking whether each initial node has associated edges and whether each edge is associated with two valid initial nodes, ensuring the integrity and correctness of the adjacency relationships.

[0116] This application's embodiments employ a dual-threshold constraint of distance threshold and stratigraphic threshold to eliminate weakly correlated edges with excessively large inter-well distances and significant stratigraphic differences. This addresses the problem of invalid nodes in adjacency relationships causing interference in subsequent graph propagation, providing an effective initial set of associated nodes and edges for graph structure construction. The distance threshold can be adjusted based on the well network density of old well blocks, while the stratigraphic threshold can be set in conjunction with regional geological stratification patterns. This flexibly adapts to the inter-well distribution and stratigraphic complexity of different blocks. Compared to existing technologies with fixed adjacency rules, this approach better fits the actual geological scenarios of old wells, improving the scenario adaptability of adjacency relationships.

[0117] In some embodiments, the first algorithm model includes a patch partitioning module, a masking module, a reconstruction module, and an embedding generation module. The first algorithm model is constructed, and the layer embedding vector is determined based on the first algorithm model and time series data, including: obtaining a patch length threshold and a masking ratio threshold. The patch length threshold is used to determine the length of the depth window, and the masking ratio threshold is used to determine the proportion of masked patches in the patch sequence.

[0118] Specifically, the first algorithm model needs to determine two core thresholds before it can run.

[0119] The patch length threshold determines the length of the depth window, and its value must match the depth sampling interval of the logging data and the actual thickness of the completion section. For example, if the depth sampling interval of the logging data for an old well is 0.1 meters (i.e., one data point is collected every 0.1 meters), and the thickness of the completion section is mostly in the range of 10-50 meters, then setting the patch length threshold to 10 meters ensures that each patch contains enough sampling points (100 data points) to reflect the formation characteristics, while also preventing the lithological differences within the section from being masked by an excessively long window. Simultaneously, this threshold is explicitly used to define the non-overlapping division of the depth window; that is, the ending depth of the previous window is the starting depth of the next window.

[0120] The masking ratio threshold is used to control the number of masks in a patch sequence, with the core purpose of enhancing the model's robustness to missing and noisy data. Setting the masking ratio threshold to 20%-30% is recommended: if the ratio is too low (e.g., <10%), the model will struggle to learn to complete missing data; if the ratio is too high (e.g., >40%), insufficient effective data will lead to unstable model training. For example, if the time-series data of a first object is divided into 10 patches, a 30% masking ratio will randomly mask 3 patches, ensuring the model can still accurately extract features even when some data is missing.

[0121] Furthermore, the time-series data is segmented into deep windows according to the patch length threshold by the patch segmentation module to obtain the patch sequence.

[0122] Specifically, the time series data corresponding to the first object is first depth aligned. This ensures that the depth sampling points of the four types of curves are completely matched, avoiding subsequent segmentation confusion due to depth deviations. For example, if the depth range of a certain first object is 1200-1250 meters, all four types of curves must include the sampling value of every 0.1 meters within this range, forming a time series data matrix with a unified depth dimension.

[0123] Using a set patch length threshold as the depth window length, the time series data is segmented in ascending order of depth. For example, with a patch length threshold of 10 meters, starting from 1200 meters, the first window covers the time series data from 1200 to 1210 meters, the second window covers 1210 to 1220 meters, and so on, until the entire depth range of 1200 to 1250 meters is covered, resulting in 5 patches. Each patch contains all sampled data of four types of curves within the corresponding depth window, forming a three-level structure of patch-curve-data point.

[0124] Arrange all the patches obtained from the segmentation in depth order to form a patch sequence specific to the first object. For example, the above 5 patches are arranged in the order of 1200-1210 m patch → 1210-1220 m patch → … → 1240-1250 m patch to ensure that the order of the patch sequence is consistent with the actual depth order downhole.

[0125] Furthermore, the patch sequence is masked according to the masking ratio threshold by the masking module to obtain the masked patch, the unmasked patch, and the masking identifier.

[0126] Specifically, based on a set masking ratio threshold, a corresponding number of patches are randomly selected from the patch sequence as masked patches, and the remaining patches are unmasked patches. For example, if the patch sequence contains 10 patches and the masking ratio is 30%, 3 patches (such as the 2nd, 5th, and 8th patches) are randomly selected as masked patches, and the remaining 7 are unmasked patches. The selection process uses a uniform random strategy to avoid model learning bias caused by fixed masking positions.

[0127] Assign a masking identifier to each patch in the patch sequence to indicate whether it is masked. For example, use 1 to represent a masked patch and 0 to represent an unmasked patch. The masking identifiers must correspond one-to-one with the patch sequence to form a patch-identifier mapping relationship.

[0128] All original logging data for masked patches are replaced with masked placeholder data (such as all-zero data or industry-standard missing logging placeholder values) to prevent leakage of original data from affecting model reconstruction capabilities. After processing, the patch sequence simultaneously contains the real data of unmasked patches, the placeholder data of masked patches, and the corresponding masking identifier.

[0129] Furthermore, depth location encoding and curve channel encoding are obtained. Depth location encoding is used to identify the depth order of each patch in the patch sequence, and curve channel encoding is used to identify the source of gamma curves, density curves, neutron curves and acoustic curves in the time series data.

[0130] Specifically, depth position encoding is used to identify the depth order of each patch in the patch sequence, and its generation logic is directly related to the depth position of the patch. For example, when the patch sequence is arranged from shallow to deep, the first patch (shallowest depth) is assigned the code 001, the second patch is assigned 002, and so on until the last patch (deepest depth) is assigned 005 (if the sequence contains 5 patches); the encoding can be in the form of integer identifiers or vector identifiers, the core of which is to enable the model to identify the relative depth relationship of the patches through the encoding.

[0131] Curve channel coding is used to distinguish the origins of four types of curves in time-series data: gamma, density, neutron, and acoustic curves, assigning a unique code to each type. For example, gamma curves are assigned code G, density curves D, neutron curves N, and acoustic curves S. The code must be bound to the curve data within a patch. For instance, all gamma curve data within a patch carries code G, and density curve data carries code D, ensuring that the model can recognize the physical meaning of different curves and avoid confusing the characteristics of the four types of curves.

[0132] Furthermore, the patch sequence, depth location encoding, curve channel encoding, and masking identifier are reconstructed using the reconstruction module to obtain the reconstructed patch.

[0133] Specifically, the processed patch sequence (including the real data of unmasked patches and placeholder data of masked patches), depth location encoding, curve channel encoding, and masking identifiers are input into the reconstruction module. The module first locates the positions of all masked patches using the masking identifiers, and only performs reconstruction on these patches, while the original data of unmasked patches remains unchanged.

[0134] The model learns the characteristic distribution patterns of masked patches based on the logging characteristics of unmasked patches (such as the lithological response of gamma curves and the porosity correlation characteristics of density curves), combined with depth sequence information encoded by depth location (such as the depth correlation between adjacent patches) and curve source information encoded by curve channels (such as the correlation between gamma curves and neutron curves). For example, if a masked patch is located between two unmasked patches (depth 1210-1220 meters), the model can infer the gamma curve characteristics of the masked patch by combining the gamma curve trends of the 1200-1210 meter and 1220-1230 meter patches with the geological stratification patterns of that depth range.

[0135] Based on the learned feature patterns, the model generates a reconstruction patch for each masked patch. The error between its data distribution and the actual logging data is controlled within an acceptable engineering range (e.g., error < 5%). The reconstruction patch must have the same structure as the original patch, that is, it must contain the corresponding depth window data of the four types of curves to ensure a consistent format during subsequent fusion processing.

[0136] Furthermore, the unmasked patch and the reconstructed patch are merged by embedding a generation module to obtain the merged patch.

[0137] Specifically, patches in key depth regions (such as depth windows corresponding to known oil-bearing sections) have higher weights. For example, if a patch corresponds to a depth range of a known oil-bearing section, its weight is 20% higher than that of a patch in a non-oil-bearing section.

[0138] Curves that are more sensitive to lithology and fluid response (such as neutron curves being more sensitive to porosity than gamma curves) have higher patch weights. For example, patches with a higher proportion of neutron curves have a 15% higher patch weight than patches with a higher proportion of gamma curves.

[0139] The total fusion weight of all patches is 1, ensuring that the features of the fused data are not excessively amplified or reduced.

[0140] Based on the calculated fusion weights, the original data of the unmasked patch and the reconstructed data of the reconstructed patch are weighted and summed to obtain the fused patch. For example, if an unmasked patch has a weight of 0.8 and the corresponding reconstructed patch (which is mistakenly masked due to local noise) has a weight of 0.2, then the fused patch data = 0.8 × original data + 0.2 × reconstructed data. This preserves the authenticity of the original data while correcting for noise through the reconstructed data.

[0141] Furthermore, the layer embedding vector is determined based on the masking ratio threshold and the fused patch.

[0142] Specifically, all fused patches are integrated in depth order and compressed into a fixed-dimensional vector through the model's embedding layer—the segment embedding vector. This vector must simultaneously represent lithological pore structure (such as density and porosity characteristics reflected in neutron curves) and fluid response (such as oil-bearing response characteristics in sonic curves), and be associated one-to-one with the corresponding first object (completion segment + depth interval), forming a mapping relationship between the first object and the segment embedding vector, providing a feature basis for subsequent embedding similarity calculations. Meanwhile, the masking ratio threshold plays a constraining role in this stage. During the fusion process, it must be ensured that the fusion weight of the masked patch does not exceed the upper limit corresponding to the masking ratio threshold (e.g., 30%), to avoid excessive influence of reconstructed data on the original features.

[0143] This application's embodiments control masking reconstruction through a masking ratio threshold, simulating scenarios of local missing data and noise interference in logging data from old wells. This enables the model to learn and complete missing features, addressing the poor robustness of segment-level features to data defects and improving the reliability of embedding vectors under complex data conditions in old wells. Depth location encoding clearly defines the depth order of patches, preventing the model from losing depth-related features of the layers. Curve channel encoding distinguishes the sources of curves such as gamma and density, preventing the model from confusing curve features with different physical meanings. This compensates for the deficiency of segment-level features not explicitly encoding depth and channel information, making the embedding vectors more accurately reflect the true properties of the formation. Based on depth location encoding and curve channel encoding, the fusion weights are determined, allowing the true features of unmasked patches and the completed features of reconstructed patches to be fused according to their physical meaning weights. This avoids feature distortion caused by simple mean fusion, and the generated segment embedding vectors can more accurately represent the joint features of lithological pore structure and fluid response.

[0144] In some embodiments, determining the graph structure based on adjacency relationships and embedding similarity includes: obtaining the embedding similarity corresponding to each edge in the adjacency relationships.

[0145] Specifically, first identify the associated objects of each edge in the adjacency relationship. Each edge corresponds to two first objects (denoted as object A and object B). From the first object-segment embedding vector mapping relationship, retrieve the segment embedding vectors corresponding to these two objects, and at the same time obtain the depth range information of the two objects (e.g., object A depth 1200-1250 meters, object B depth 1210-1260 meters).

[0146] The degree of correlation between the embedding vectors of the layers is used to determine the similarity. For example, if the embedding vectors of object A and object B are highly consistent in terms of lithological characteristics (such as the sandstone and mudstone properties reflected by the gamma curve) and fluid response (such as the oil-bearing response of the sonic curve), then the feature similarity is high.

[0147] The overlap ratio is used to determine the depth. It is the ratio of the intersection length of the depth interval to the sum of the lengths of the two. For example, the depth intersection of objects A and B is 1210-1250 meters (length 40 meters), and the total length of the two is 50+50=100 meters. The overlap ratio is 40%. The higher the overlap ratio, the better the layer matching.

[0148] The embedding similarity is obtained by combining the results of both. For example, if the feature similarity accounts for 60% and the deep overlap ratio accounts for 40%, and the feature similarity score is 0.9 and the overlap ratio score is 0.8, then the embedding similarity is 0.9×0.6+0.8×0.4=0.86. The score range is 0-1, and the closer it is to 1, the stronger the comparability of the two first objects.

[0149] The calculated embedding similarity is bound to the corresponding edge one by one. For example, the edge connecting object A and object B in the adjacency relationship is bound to a similarity of 0.86, forming a mapping relationship between edge and embedding similarity.

[0150] Furthermore, a first threshold is obtained, and the embedding similarity of each edge is filtered against the first threshold to obtain a gated set of edges.

[0151] Specifically, the first threshold needs to be set based on the geological consistency of the exploration block and the verification results of historical data. If the geological conditions within the block are stable (such as continuous distribution of the same sand body), the threshold can be appropriately reduced (e.g., 0.6) to retain more potential comparable layers; if the block has complex geology (e.g., many faults and large lithological variations), the threshold needs to be increased (e.g., 0.7) to strictly screen strongly correlated edges. For example, in an old well block dominated by stable sand bodies, the first threshold is set to 0.65 to ensure that no effective comparable layers are missed while avoiding interference from weakly correlated edges across layers.

[0152] Traverse all edges in the adjacency relationship, and compare the embedding similarity of each edge with a first threshold. If the similarity is greater than or equal to the first threshold (e.g., 0.65), the two first objects associated with the edge are determined to be highly comparable, and the edge is retained. If the similarity is less than the first threshold (e.g., 0.6), the two are determined to be poorly comparable (e.g., large differences in lithology or mismatch in stratigraphy), and the edge is removed.

[0153] All retained edges are aggregated to form a gated edge set. This set contains only strongly related edges. For example, if there are 10 edges in an initial adjacency relationship, after filtering, 7 edges with a similarity of ≥0.65 are retained. Subsequent graph structure construction is based solely on these 7 edges, thus constraining the effectiveness of oil-related information propagation from the source.

[0154] Furthermore, the initial graph structure is obtained based on the initial nodes in the adjacency relationships and the gated edge set.

[0155] Specifically, the initial nodes in the adjacency relationship (all nodes mapped to the first object) remain unchanged, and only isolated nodes caused by edge removal are removed (if a node has no associated edges and the well corresponding to the first object has no other comparable layers, it can be retained but will not participate in information propagation later), ensuring that the nodes cover all well completion segments with the oil content to be predicted.

[0156] The gated edge set is associated with the initial nodes. The edges connected to each node are the filtered strongly related edges. For example, if node object A originally has 3 connected edges, after filtering, only 2 are retained (connecting objects B and C). Then, the associated edges of object A in the initial graph structure are these 2 edges.

[0157] The initial graph structure can be presented through a node-edge association table, such as object A-edge 1-object B, object A-edge 2-object C, object B-edge 3-object D. This allows for the verification of invalid connections (such as self-loop edges or duplicate edges) to ensure that the initial graph structure is logically clear and free of data errors.

[0158] Furthermore, the edge weight corresponding to each edge is obtained based on the embedding similarity.

[0159] Specifically, edge weights are positively correlated with embedding similarity. The higher the embedding similarity, the larger the edge weight, and the higher the weight percentage of that edge during information propagation. Mapping rules need to be simplified and conform to engineering logic. For example, embedding similarity can be directly used as edge weight (0.86 similarity corresponds to edge weight 0.86), or mapping can be done according to similarity intervals (e.g., 0.65-0.75 corresponds to edge weight 0.7, 0.75-0.85 corresponds to edge weight 0.8, and 0.85-1.0 corresponds to edge weight 0.9), avoiding complex calculations.

[0160] The mapped edge weights are bound to each edge in the gated edge set. For example, the edge connecting object A and object B (similarity 0.86) is bound with an edge weight of 0.86; the edge connecting object B and object D (similarity 0.72) is bound with an edge weight of 0.72, forming a fixed association between edge and edge weight. In subsequent information propagation, the propagation weight is allocated according to the edge weight.

[0161] Furthermore, an adjacency list is obtained based on the initial graph structure and the edge weights corresponding to each edge.

[0162] Specifically, taking each node in the initial graph structure as the core, we sort out all its associated edges and corresponding associated nodes. For example, the associated edges of node object B include object A-object B (edge ​​weight 0.86) and object B-object D (edge ​​weight 0.72), and the associated nodes are object A and object D.

[0163] Sort the associated nodes of each node in descending order of edge weight to form the adjacency list of that node. For example, the adjacency list of node object B is from object A (0.86) to object D (0.72), which means that when information is propagated, information is first obtained from object A with higher edge weight, and then from object D, to avoid errors caused by the priority propagation of information from weakly related nodes.

[0164] Bind each node's adjacency list to it. For example, the adjacency list of node object A is from object B (0.86) to object C (0.78), ensuring that the subsequent graph transformation network can quickly retrieve the propagation objects and order of each node.

[0165] Furthermore, the effective neighbor set is obtained based on the first threshold and the adjacency list.

[0166] Specifically, using a first threshold (e.g., 0.65) as the filtering criterion, the adjacency list of each node is traversed, and associated nodes with edge weights less than the first threshold are removed. For example, the adjacency list of node object C originally consists of object A (0.78) to object E (0.62). Since the edge weight of object E is 0.62 < 0.65, it is removed from the list, and the truncated list only retains object A (0.78).

[0167] The associated nodes in the truncated adjacency list are defined as valid neighbors. For example, the only valid neighbor of node object C is object A, and the only valid neighbors of node object B are objects A and D (both edge weights are ≥0.65). The valid neighbors of each node form a valid neighbor set. If the set is empty, it means that the node has no strongly related neighbors, and subsequent information propagation depends only on its own characteristics.

[0168] Furthermore, the final graph structure is obtained based on the effective neighbor set, the edge weight corresponding to each edge, and the initial graph structure.

[0169] Specifically, the initial node, the gated edge set, the edge weight of each edge, and the effective neighbor set of each node are uniformly integrated. Each node is associated with its own effective neighbor set and corresponding edge weight, and each edge is associated with its corresponding embedding similarity and edge weight. For example, node object A is associated with effective neighbor object B (edge ​​weight 0.86) and object C (edge ​​weight 0.78), and edge object A-object B is associated with similarity 0.86 and edge weight 0.86.

[0170] Verify whether the valid neighbors of each node are all from the gated edge set and whether the edge weights correspond to the embedding similarity. For example, if the edge corresponding to the valid neighbor object B of node object D is not in the gated edge set, the filtering process needs to be checked again to ensure that there are no invalid associations in the graph structure.

[0171] The integrated graph structure can be directly used as input for graph transformation networks. Its core function is to constrain the propagation range of oil-bearing information through effective neighbor sets (propagating only between strongly correlated nodes) and constrain the propagation order through edge weights and adjacency lists (prioritizing information from nodes with high edge weights), thereby avoiding interference from weakly correlated and cross-layer information and providing structural support for accurate estimation of oil content in old well sections.

[0172] This application's embodiments use a first threshold to filter edges based on embedding similarity, retaining only strongly correlated edges. This addresses the problem of static graph structures lacking filtering and containing weakly correlated edges. Simultaneously, edge weights reflect the strength of node associations, and the effective neighbor set constrains the propagation range, ensuring that oil-bearing information is only transmitted between comparable layers, reducing the risk of cross-layer mispropagation and enhancing the graph structure's constraint on information propagation. The adjacency list is arranged from high to low edge weights, allowing the graph transformation network to prioritize information from neighboring nodes with high association strength. This avoids errors caused by disordered graph propagation and the priority transmission of weakly correlated information, ensuring that oil-bearing information transmission conforms to the logic of prioritizing strong associations and improving the effectiveness of neighborhood aggregation. Edge weights are generated by embedding similarity mapping, achieving linkage between feature similarity and information propagation weights. This allows the graph structure's edge weights to reflect the degree of association of layer features, solving the problem of graph structure edge weights being unrelated to features. This provides structure-feature matching support for the regression calculation of the second algorithm model (graph transformation network), improving the accuracy of oil content estimation.

[0173] In some embodiments, determining the well stratification result based on the estimated oil content of the strata includes: associating the estimated oil content of the strata with the corresponding first object to obtain the associated first object.

[0174] Specifically, the unique identifier of the first object is used as the association benchmark. This identifier consists of the well number and the completion segment number, such as well A - segment 2, well B - segment 5, ensuring that each first object is unique within the block and can be quickly located to a specific well and specific segment.

[0175] From the oil content distribution output by the graph transform network, the estimated oil content of each segment corresponding to the first object is extracted, and the unique identifier, the estimated oil content, and the depth range are bound one by one. For example, if the first object, well A-segment 2, corresponds to the depth range of 1200-1250 meters and has an estimated oil content of 62%, then the association result is recorded as well A-segment 2 (1200-1250m): 62%.

[0176] Iterate through all related data to check for cases where there is no corresponding first object with an estimated oil content or where an estimated oil content has no corresponding first object. If such cases exist, backtrack the regression process of the graph transformation network to correct data matching errors and ensure that each first object has a unique corresponding estimated oil content.

[0177] Furthermore, a high oil content determination threshold is obtained, and the estimated oil content of each layer of the first object is compared with the high oil content determination threshold to obtain the high oil content determination result.

[0178] Specifically, thresholds are set by referring to historical oil-bearing data of the target old well block, the oil content range of developed high-oil-bearing zones, and engineering application requirements. For example, if the oil content of developed high-oil-bearing zones in a certain block is generally ≥50%, and the exploitation of zones with oil content <50% is less economical, then the high-oil-bearing threshold is set at 50%. If the geological conditions of the block are special (such as tight oil reservoirs), the threshold can be appropriately reduced (e.g., 35%) to ensure that no exploitable zones are missed.

[0179] Taking the first associated object as a unit, compare the estimated oil content of its segment with the high oil content judgment threshold one by one: if the estimated oil content of the segment is ≥ the high oil content judgment threshold (e.g., 50%), the segment corresponding to the first object is judged as a high oil content segment, and recorded as judgment result 1; if the estimated oil content of the segment is < the high oil content judgment threshold (e.g., 45%), the segment is judged as a non-high oil content segment, and recorded as judgment result 0.

[0180] The determination results of all first objects are bound with unique identifiers to form a first object-determination result mapping table. For example, well A-segment 2:1, well A-segment 4:0, well B-segment 5:1, clearly presenting the oil-bearing properties of each segment.

[0181] Furthermore, the results of the high oil content determination were screened and processed to obtain a set of high oil content layers.

[0182] Specifically, the mapping table of first objects and judgment results is traversed, and only first objects with a judgment result of 1 (high oil-bearing layer segment) are retained, while non-high oil-bearing layer segments with a judgment result of 0 are excluded. For example, if a block is associated with 20 first objects, of which 12 are judged as high oil-bearing layer segments, then only these 12 first objects are retained.

[0183] Check if there are duplicates in the first object after screening. If so, take the first object corresponding to the average value of the oil content estimate. At the same time, complete the core information of each first object, including the well completion segment number, depth range, and oil content estimate, to ensure the data in the set is complete. For example, well A - segment 2 (1200-1250m, 62%) and well B - segment 5 (1300-1350m, 58%).

[0184] The first objects of the completed high oil-bearing segments are summarized to form a set of high oil-bearing segments. This set is the basic data for well stratification, and subsequent identification and sorting are based on this set. For example, the set can be represented as {Well A - Segment 2, Well B - Segment 5, Well C - Segment 3, ...}, and each element carries complete segment information and oil content data.

[0185] Furthermore, the high oil-bearing layer set is sorted according to the layer identification rules to obtain the identification of each layer in the high oil-bearing layer set.

[0186] Specifically, the segment identification rule is defined as well number - completion segment number (depth interval start depth - end depth), where: well number is a unique number within the block of old wells, completion segment number is the sequential number of completion segments within a single well, and depth interval is the start-end depth corresponding to the first object.

[0187] A unique identifier is generated for each object in the set of high-oil-bearing layers, according to certain rules. For example:

[0188] The third completion section of well J123, with a depth range of 1200-1250 meters, is marked as J123-3 (1200-1250m).

[0189] The fifth completion section of well J124, with a depth range of 1320-1380 meters, is marked as J124-5 (1320-1380m).

[0190] The generated segment identifiers are bound one by one to the first object in the set of high oil-bearing segments, forming a correlation between segment identifier, oil content estimate, and depth range. For example, J123-3 (1200-1250m): 62%, which provides identifiable segment identifiers for subsequent sorting and result output.

[0191] Furthermore, the high oil-bearing layer set is sorted according to the well-layer sorting rules to obtain the sorted high oil-bearing layer set.

[0192] Specifically, the sorting rules are divided into two priority levels. The first priority is the estimated oil content of the formation, which is sorted from high to low. The higher the oil content of the formation, the more priority is given to supplementary extraction or other measures in engineering. For example, a formation with an oil content of 65% is given priority over a formation with an oil content of 60%.

[0193] Second priority: starting depth of the layers, sorted from shallowest to deepest. If two layers have the same estimated oil content (e.g., both are 60%), the layer with the shallower starting depth is given priority, because shallower layers have lower operating costs and are easier to construct (e.g., a layer with a starting depth of 1200 meters is given priority over a layer with a starting depth of 1300 meters).

[0194] Sort the set of high-oil-bearing zones for each well on a per-well basis (to avoid confusion in engineering logic caused by sorting across wells).

[0195] The first step is to extract the estimated oil content of the high oil-bearing sections of the well and sort them from high to low. For example, the high oil-bearing sections of well J123 are initially sorted as section 3 (62%) → section 5 (58%) → section 2 (55%).

[0196] The second step is to check if there are any segments with the same oil content in the preliminary sorting. If so (e.g., both segment 4 and segment 6 have 56%), then compare their starting depths and place segment 4, with a starting depth of 1280 meters, before segment 6, with a starting depth of 1350 meters, to form the final sorting result for the well.

[0197] The sorted set is validated to ensure there are no sorting logic errors (such as high oil content segments not being ranked first, or the depth of segments with the same oil content being reversed). For example, after well J124 is sorted as segment 5 (60%, 1320m) → segment 2 (58%, 1250m) → segment 7 (56%, 1400m), it needs to be confirmed that segment 2 (58%) is ranked second because its oil content is lower than that of segment 5 (60%), even though it is shallower.

[0198] Furthermore, the well stratification results are obtained based on the sorted set of high oil-bearing layers and the identifiers of each layer in the set of high oil-bearing layers.

[0199] Specifically, the high-oil-bearing intervals within a single well are sorted and matched one-to-one with their corresponding interval identifiers, oil-bearing estimates, and depth ranges to form a stratification result list. For example, the stratification result list for well J123 is as follows:

[0200] Marking: J123-3 (1200-1250m), estimated oil content: 62%.

[0201] Marking: J123-5 (1300-1350m), estimated oil content: 58%.

[0202] Marking: J123-2 (1250-1300m), estimated oil content: 55%.

[0203] Check whether the stratification results of each well cover all high oil-bearing segments of the well, whether the identification and oil content match, and whether the sorting conforms to the rules. For example, confirm that there are no missing high oil-bearing segments in the stratification results of well J124, and that the oil content corresponding to the identification of J124-5 (1320-1380m) is correct. Finally, form a set of in-well stratification results of all old wells in the block.

[0204] This application's embodiments quantitatively screen high-oil-content zones using a high-oil-content threshold, avoiding errors from subjective judgments based on experience. This ensures the objectivity and consistency of the high-oil-content zone set, solves the problem of ambiguous zone selection criteria in old wells, and improves the reliability of zone screening. The zone identification rules combine well-zone location and depth range information, allowing direct location to the depth range of a specific well and zone, resolving issues of ambiguous identification and the need for additional data lookups, thus improving the operability of the results. The in-well stratification sorting rules prioritize high-oil-content, shallow-depth zones, aligning with the engineering requirements of prioritizing high-value and low-cost shallow-layer operations in old well re-production. This directly supports the formulation of the operation sequence without additional adjustments, solving the problem of post-processing sorting of results and improving engineering application efficiency.

[0205] While this application provides the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in this embodiment is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the methods shown in this embodiment or the accompanying drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0206] like Figure 3 As shown in the illustration, this application also provides an oil content prediction device 300. The device includes:

[0207] The acquisition module 301 is used to acquire multiple first objects, time series data, and adjacency relationships. The first objects include well completion segments and corresponding depth intervals, the time series data includes the multi-curve logging sequence of the first objects, and the adjacency relationships include the initial nodes and edge sets of the first objects.

[0208] The construction module 302 is used to construct the first algorithm model and determine the layer embedding vector based on the first algorithm model and time series data. The layer embedding vector is used to characterize the features of lithological pore structure and fluid response.

[0209] The determination module 303 is used to determine the embedding similarity based on the segment embedding vector.

[0210] The determination module 303 is also used to determine the graph structure based on adjacency relationships and embedding similarity, the graph structure including a gated set of nodes and edges.

[0211] The construction module 302 is also used to construct a second algorithm model, inputting the graph structure and the segment embedding vector into the second algorithm model to obtain the estimated oil content of the segment.

[0212] The determination module 303 is also used to determine the well stratification results based on the estimated oil content of the strata. The well stratification results include a set of high oil-bearing strata and corresponding strata identifiers.

[0213] In some embodiments, the acquisition module 301 is further configured to obtain a first object based on the well completion segment and the corresponding depth interval, wherein the depth interval is used to characterize the starting depth and ending depth of the well completion segment.

[0214] The acquisition module 301 is also used to acquire a multi-curve logging sequence, perform correlation processing on the multi-curve logging sequence and the first object to obtain time-series data. The multi-curve logging sequence includes gamma curve, density curve, neutron curve and sonic curve.

[0215] The acquisition module 301 is also used to acquire distance threshold and layer threshold. The distance threshold is used to determine the candidate range of inter-well distance and the layer threshold is used to determine the candidate range of layer relationship of well completion segment.

[0216] The determination module 303 is also used to determine the initial node based on multiple first objects.

[0217] The determination module 303 is also used to traverse multiple first objects, and when the well distance between any two first objects meets the distance threshold and the stratigraphic relationship of the well completion segments of the two first objects meets the stratigraphic threshold, the edge corresponding to the two first objects is determined.

[0218] The determination module 303 is also used to determine adjacency relationships based on the initial node and all edges.

[0219] In some embodiments, the acquisition module 301 is further configured to acquire a patch length threshold and a masking ratio threshold, wherein the patch length threshold is used to determine the length of the depth window and the masking ratio threshold is used to determine the proportion of masked patches in the patch sequence.

[0220] The construction module 302 is also used to perform deep window segmentation on the time series data according to the patch length threshold through the patch segmentation module to obtain the patch sequence.

[0221] The construction module 302 is also used to perform masking processing on the patch sequence according to the masking ratio threshold through the masking module, so as to obtain the masked patch, the unmasked patch and the masking identifier.

[0222] The acquisition module 301 is also used to acquire depth position encoding and curve channel encoding. The depth position encoding is used to identify the depth order of each patch in the patch sequence, and the curve channel encoding is used to identify the source of the gamma curve, density curve, neutron curve and acoustic curve in the time series data.

[0223] The construction module 302 is also used to reconstruct the patch sequence, depth position encoding, curve channel encoding and masking identifier through the reconstruction module to obtain the reconstructed patch.

[0224] The building module 302 is also used to merge the unmasked patch and the reconstructed patch by embedding the generation module to obtain the merged patch.

[0225] The determination module 303 is also used to determine the segment embedding vector based on the masking ratio threshold and the fused patch.

[0226] In some embodiments, the acquisition module 301 is further configured to acquire the embedding similarity corresponding to each edge in the adjacency relationship.

[0227] The acquisition module 301 is also used to acquire a first threshold, and to filter the embedding similarity of each edge against the first threshold to obtain a gated edge set.

[0228] The acquisition module 301 is also used to obtain the initial graph structure based on the initial nodes in the adjacency relationship and the gated edge set.

[0229] The determination module 303 is also used to obtain the edge weight corresponding to each edge based on the embedding similarity.

[0230] The acquisition module 301 is also used to obtain the adjacency list based on the initial graph structure and the edge weight corresponding to each edge.

[0231] The acquisition module 301 is also used to obtain a set of valid neighbors based on the first threshold and the adjacency list.

[0232] The acquisition module 301 is also used to obtain the final graph structure based on the effective neighbor set, the edge weight corresponding to each edge, and the initial graph structure.

[0233] In some embodiments, the construction module 302 is further configured to associate the estimated oil content of the layer with the corresponding first object to obtain the associated first object.

[0234] The acquisition module 301 is also used to acquire a high oil content determination threshold, compare the estimated oil content of each layer segment of the first object with the high oil content determination threshold, and obtain a high oil content determination result.

[0235] The acquisition module 301 is also used to filter and process the high oil content determination results to obtain a set of high oil content layers.

[0236] The acquisition module 301 is also used to sort the high oil-bearing layer set according to the layer identification rules to obtain the identification of each layer in the high oil-bearing layer set.

[0237] The acquisition module 301 is also used to sort the set of high oil-bearing sections according to the well layer sorting rules to obtain the sorted set of high oil-bearing sections.

[0238] The acquisition module 301 is also used to obtain the well stratification results based on the sorted set of high oil-bearing segments and the identifiers of each segment in the set of high oil-bearing segments.

[0239] Some modules in the apparatus described in this application can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0240] The apparatus or module described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. For ease of description, the above apparatus is described by dividing it into various modules according to their functions. When implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0241] The methods, apparatus, or modules described in this application can be implemented in a computer-readable program code manner. The controller can be implemented in any suitable manner, such as a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of a memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code manner, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included within it for implementing various functions can also be considered as structures within the hardware component. Alternatively, the device used to implement various functions can be viewed as either a software module that implements the method or a structure within a hardware component.

[0242] This application also provides an apparatus, the apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method described in this application.

[0243] This application also provides a non-volatile computer-readable storage medium storing a computer program or instructions thereon, which, when executed, enables the method described in this application embodiment to be implemented.

[0244] Furthermore, in the various embodiments of the present invention, each functional module can be integrated into a processing module, or each module can exist independently, or two or more modules can be integrated into a single module.

[0245] The aforementioned storage media include, but are not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions.

[0246] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, or it can be embodied in the process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0247] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0248] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A method of predicting oil content, characterized by, The method comprises the following steps: obtaining a plurality of first objects, time series data and adjacency relations, wherein the first objects comprise well completion segments and corresponding depth intervals, the time series data comprise a plurality of curve logging sequences of the first objects, and the adjacency relations comprise initial nodes and edge sets of the first objects; constructing a first algorithm model, determining a layer segment embedding vector according to the first algorithm model and the time series data, and using the layer segment embedding vector to represent characteristics of lithology pore structure and fluid response; determining an embedding similarity according to the layer segment embedding vector; determining a graph structure according to the adjacency relations and the embedding similarity, wherein the graph structure comprises a set of nodes and edges that are gated; constructing a second algorithm model, inputting the graph structure and the layer segment embedding vector into the second algorithm model, and obtaining a layer oil content estimation value, wherein the second algorithm model is a graph transformation network, and the second algorithm model comprises a node feature input module, an edge weight gating module, a neighborhood aggregation module and a regression output module; determining an in-well layering result according to the layer oil content estimation value, wherein the in-well layering result comprises a set of high oil-bearing layer segments and corresponding layer segment identifiers; the first algorithm model comprises a patch division module, a masking module, a reconstruction module and an embedding generation module, and the construction of the first algorithm model and the determination of the layer segment embedding vector according to the first algorithm model and the time series data comprise the following steps: obtaining a patch length threshold and a masking proportion threshold, wherein the patch length threshold is used to determine the length of a depth window, and the masking proportion threshold is used to determine the proportion of masked patches in a patch sequence; performing depth window segmentation processing on the time series data according to the patch length threshold by using the patch division module, and obtaining the patch sequence; performing masking processing on the patch sequence according to the masking proportion threshold by using the masking module, and obtaining masked patches, unmasked patches and a masking identifier; obtaining depth position encoding and curve channel encoding, wherein the depth position encoding is used to identify the depth order of each patch in the patch sequence, and the curve channel encoding is used to identify the sources of gamma curves, density curves, neutron curves and sonic curves in the time series data; performing reconstruction processing on the patch sequence, the depth position encoding, the curve channel encoding and the masking identifier by using the reconstruction module, and obtaining reconstructed patches; performing fusion processing on the unmasked patches and the reconstructed patches by using the embedding generation module, and obtaining fused patches; determining a layer segment embedding vector according to the masking proportion threshold and the fused patches; the first algorithm model is a time series patch masking model.

2. The method of claim 1, wherein, The method comprises the following steps: obtaining a plurality of first objects, time series data and adjacency relations, wherein the first objects comprise well completion segments and corresponding depth intervals, the time series data comprise a plurality of curve logging sequences of the first objects, and the adjacency relations comprise initial nodes and edge sets of the first objects; constructing a first algorithm model, determining a layer segment embedding vector according to the first algorithm model and the time series data, and using the layer segment embedding vector to represent characteristics of lithology pore structure and fluid response; determining an embedding similarity according to the layer segment embedding vector; determining a graph structure according to the adjacency relations and the embedding similarity, wherein the graph structure comprises a set of nodes and edges that are gated; constructing a second algorithm model, inputting the graph structure and the layer segment embedding vector into the second algorithm model, and obtaining a layer oil content estimation value, wherein the second algorithm model is a graph transformation network, and the second algorithm model comprises a node feature input module, an edge weight gating module, a neighborhood aggregation module and a regression output module; determining an in-well layering result according to the layer oil content estimation value, wherein the in-well layering result comprises a set of high oil-bearing layer segments and corresponding layer segment identifiers; the first algorithm model comprises a patch division module, a masking module, a reconstruction module and an embedding generation module, and the construction of the first algorithm model and the determination of the layer segment embedding vector according to the first algorithm model and the time series data comprise the following steps: obtaining a patch length threshold and a masking proportion threshold, wherein the patch length threshold is used to determine the length of a depth window, and the masking proportion threshold is used to determine the proportion of masked patches in a patch sequence; performing depth window segmentation processing on the time series data according to the patch length threshold by using the patch division module, and obtaining the patch sequence; performing masking processing on the patch sequence according to the masking proportion threshold by using the masking module, and obtaining masked patches, unmasked patches and a masking identifier; obtaining depth position encoding and curve channel encoding, wherein the depth position encoding is used to identify the depth order of each patch in the patch sequence, and the curve channel encoding is used to identify the sources of gamma curves, density curves, neutron curves and sonic curves in the time series data; performing reconstruction processing on the patch sequence, the depth position encoding, the curve channel encoding and the masking identifier by using the reconstruction module, and obtaining reconstructed patches; performing fusion processing on the unmasked patches and the reconstructed patches by using the embedding generation module, and obtaining fused patches; determining a layer segment embedding vector according to the masking proportion threshold and the fused patches; the first algorithm model is a time series patch masking model. The method comprises the following steps: obtaining a plurality of first objects, time series data and adjacency relations, wherein the first objects comprise well completion segments and corresponding depth intervals, the time series data comprise a plurality of curve logging sequences of the first objects, and the adjacency relations comprise initial nodes and edge sets of the first objects; constructing a first algorithm model, determining a layer segment embedding vector according to the first algorithm model and the time series data, and using the layer segment embedding vector to represent characteristics of lithology pore structure and fluid response; determining an embedding similarity according to the layer segment embedding vector; determining a graph structure according to the adjacency relations and the embedding similarity, wherein the graph structure comprises a set of nodes and edges that are gated; constructing a second algorithm model, inputting the graph structure and the layer segment embedding vector into the second algorithm model, and obtaining a layer oil content estimation value, wherein the second algorithm model is a graph transformation network, and the second algorithm model comprises a node feature input module, an edge weight gating module, a neighborhood aggregation module and a regression output module; determining an in-well layering result according to the layer oil content estimation value, wherein the in-well layering result comprises a set of high oil-bearing layer segments and corresponding layer segment identifiers; the first algorithm model comprises a patch division module, a masking module, a reconstruction module and an embedding generation module, and the construction of the first algorithm model and the determination of the layer segment embedding vector according to the first algorithm model and the time series data comprise the following steps: obtaining a patch length threshold and a masking proportion threshold, wherein the patch length threshold is used to determine the length of a depth window, and the masking proportion threshold is used to determine the proportion of masked patches in a patch sequence; performing depth window segmentation processing on the time series data according to the patch length threshold by using the patch division module, and obtaining the patch sequence; performing masking processing on the patch sequence according to the masking proportion threshold by using the masking module, and obtaining masked patches, unmasked patches and a masking identifier; obtaining depth position encoding and curve channel encoding, wherein the depth position encoding is used to identify the depth order of each patch in the patch sequence, and the curve channel encoding is used to identify the sources of gamma curves, density curves, neutron curves and sonic curves in the time series data; performing reconstruction processing on the patch sequence, the depth position encoding, the curve channel encoding and the masking identifier by using the reconstruction module, and obtaining reconstructed patches; performing fusion processing on the unmasked patches and the reconstructed patches by using the embedding generation module, and obtaining fused patches; determining a layer segment embedding vector according to the masking proportion threshold and the fused patches; the first algorithm model is a time series patch masking model. obtain a distance threshold value and a horizon threshold value, the distance threshold value is used to determine a candidate range of interwell distance, and the horizon threshold value is used to determine a candidate range of horizon relationship of completion section; determine an initial node according to a plurality of first objects; traverse the plurality of first objects, and determine an edge corresponding to any two first objects when an interwell distance corresponding to the two first objects meets the distance threshold value and a horizon relationship of completion section of the two first objects meets the horizon threshold value; determine the adjacency relationship according to the initial node and all edges.

3. The method of claim 1, wherein, the graph structure is determined according to the adjacency relationship and the embedding similarity, including: obtain an embedding similarity corresponding to each edge in the adjacency relationship; obtain a first threshold value, and perform screening processing on the embedding similarity of each edge and the first threshold value to obtain a gated edge set; obtain an initial graph structure according to the initial node in the adjacency relationship and the gated edge set; obtain an edge weight corresponding to each edge according to the embedding similarity; obtain an adjacency list according to the initial graph structure and the edge weight corresponding to each edge; obtain an effective neighbor set according to the first threshold value and the adjacency list; obtain a final graph structure according to the effective neighbor set, the edge weight corresponding to each edge and the initial graph structure.

4. The method of claim 1, wherein, the well internal layering result is determined according to the layer segment oil content estimation value, including: associate the layer segment oil content estimation value with the corresponding first object to obtain an associated first object; obtain a high oil content determination threshold value, compare the layer segment oil content estimation value of each first object with the high oil content determination threshold value, and obtain a high oil content determination result; screen the high oil content determination result to obtain a high oil content layer segment set; sort the high oil content layer segment set according to a layer segment identification rule to obtain an identification of each layer segment in the high oil content layer segment set; sort the high oil content layer segment set according to a well internal layering sorting rule to obtain a sorted high oil content layer segment set; obtain a well internal layering result according to the sorted high oil content layer segment set and the identification of each layer segment in the high oil content layer segment set.

5. The method of claim 1, wherein, the calculation formula of the embedding similarity determined according to the layer segment embedding vector is: wherein, is an embedding similarity, is a combination weight, is a layer embedding vector corresponding to the i-th well and the k-th completion section, is a layer embedding vector corresponding to the j-th well and the m-th completion section, and is a norm of an embedding vector, is a depth overlap ratio, is a vector transpose.

6. An oil content prediction device characterized by comprising: including: an obtaining module, configured to obtain a plurality of first objects, time series data and an adjacency relationship, the first object including a completion section and a corresponding depth interval, the time series data including a multi-curve logging sequence of the first object, and the adjacency relationship including an initial node and an edge set of the first object; a constructing module, configured to construct a first algorithm model, determine a layer segment embedding vector according to the first algorithm model and the time series data, and use the layer segment embedding vector to represent characteristics of lithology pore structure and fluid response; a determining module, configured to determine an embedding similarity according to the layer segment embedding vector; the determining module is further configured to determine a graph structure according to the adjacency relationship and the embedding similarity, and the graph structure includes a gated node and an edge set; The construction module is further configured to construct a second algorithm model, input the graph structure and the layer section embedding vector into the second algorithm model to obtain a layer section oil content estimation value, the second algorithm model being a graph transformation network, and the second algorithm model being composed of a node feature input module, an edge weight gate module, a neighborhood aggregation module and a regression output module; The determination module is further configured to determine a well internal layering result according to the layer section oil content estimation value, the well internal layering result including a high oil-bearing layer section set and corresponding layer section identifiers; The first algorithm model includes a patch division module, a masking module, a reconstruction module and an embedding generation module, and the construction module is further configured to construct a first algorithm model and determine a layer section embedding vector according to the first algorithm model and the time series data, wherein: A patch length threshold and a masking proportion threshold are obtained, the patch length threshold being used to determine the length of a depth window, and the masking proportion threshold being used to determine the proportion of masked patches in a patch sequence; The patch division module is used to perform depth window segmentation processing on the time series data according to the patch length threshold to obtain the patch sequence; The masking module is used to perform masking processing on the patch sequence according to the masking proportion threshold to obtain masked patches, unmasked patches and a masking identifier; A depth position encoding and a curve channel encoding are obtained, the depth position encoding being used to identify the depth order of each patch in the patch sequence, and the curve channel encoding being used to identify the sources of gamma curves, density curves, neutron curves and sonic curves in the time series data; The reconstruction module is used to perform reconstruction processing on the patch sequence, the depth position encoding, the curve channel encoding and the masking identifier to obtain reconstructed patches; The embedding generation module is used to perform fusion processing on the unmasked patches and the reconstructed patches to obtain fused patches; The layer section embedding vector is determined according to the masking proportion threshold and the fused patches; The first algorithm model is a time series patch masking model.

7. A computer device comprising a memory and a processor, the memory storing a computer program operable on the processor, characterized in that, The processor executes the program to implement the steps of the method in any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 5.

Images