An oil reservoir layered injection and production multi-source heterogeneous data fusion method and system

Abstract

The present application relates to the technical field of oil and gas field development digitization, and in particular to a reservoir layered injection-production multi-source heterogeneous data fusion method and system. The method comprises: constructing a reservoir injection-production space-time topology knowledge graph to give multi-source data standardized space-time and confidence labels; establishing a constraint model library containing space-time aggregation rules, physical mechanism constraints and caliber conversion models; extracting multiple candidate values and confidence of key indicators from unstructured data; based on the candidate values, the constraint library and the knowledge graph, constructing and solving a global constraint satisfaction and optimization model to obtain consistent data that is globally optimal and satisfies physical laws; if there is no solution, locating the smallest constraint set that leads to conflict and its specific space-time location; based on the positioning result, performing adaptive error correction and physically driven data interpolation, and iteratively optimizing, and finally outputting the fusion result, complete evidence chain and audit log. The present application significantly improves the physical consistency and engineering application value of the fused data.

Description

Technical Field

[0001] This invention relates to the field of digital technology for oil and gas field development, specifically to a method and system for fusing multi-source heterogeneous data from reservoir stratification injection and production. Background Technology

[0002] With the deepening of smart oilfield construction, the sources of oilfield stratified injection and production data are increasingly exhibiting multi-source and heterogeneous characteristics. These data mainly include: high-frequency time-series data collected in real-time by devices such as downhole flow meters and pressure sensors (on a second or minute scale); structured or semi-structured data such as manually filled daily, weekly, and monthly production reports; unstructured historical documents such as geological plans, measure summaries, and dynamic analysis reports saved in PDF, Word, and other formats; and professional map data such as injection-production curves and water absorption profiles. Significant differences exist among these data sources in terms of time scale, spatial description granularity, metering standards, and data quality, leading to a severe "data silo" phenomenon and posing a significant challenge to refined analysis and decision-making in reservoir engineering.

[0003] Existing technologies face a series of prominent problems when processing such complex, multi-source, heterogeneous injection and production data. First, there is inconsistency in spatiotemporal scales. Real-time data, report data, and historical data have significant time spans and spatial aggregation level differences, lacking effective multi-scale alignment mechanisms, making direct data fusion difficult. Second, spatial hierarchy is inconsistent. Different data sources describe the same physical entity using different dimensions; some are statistically analyzed by well, while others are by layer or sand body, lacking hierarchical mapping and aggregation rules. Third, there are large differences in measurement calibers. Measurement values ​​under standard wellhead surface conditions and those under high-temperature, high-pressure formation conditions downhole exhibit systematic deviations due to fluid phase changes (PVT effect), lacking unified calibration. Furthermore, when the same indicator shows contradictory values ​​in different data sources, existing methods lack effective conflict detection and precise location methods. Finally, for missing values ​​commonly found in production data, traditional interpolation methods often lack physical mechanism guidance, resulting in insufficient reliability.

[0004] Currently, some general methods exist for fusing multi-source heterogeneous data, such as semantic association and feature extraction based on Large Language Models (LLM), or feature fusion using association graphs and attention mechanisms. However, these general methods have significant limitations when dealing with highly specialized domain data such as reservoir injection and production. They focus on statistical correlations between data, but generally lack the embedding of the specific spatial topology of "well-tube-layer-sand body," and fail to formalize core physical laws of reservoir engineering, such as mass conservation and pressure continuity, into mandatory constraints. Their probabilistic generation characteristics or purely statistical association models cannot guarantee that the fusion results meet the strict physical consistency requirements of injection-production balance and volume conservation, and cannot properly handle the large-diameter conversion problems unique to oilfields that span surface and subsurface conditions. Therefore, there is an urgent need for a specialized technical solution that can deeply integrate reservoir engineering mechanisms to achieve intelligent, consistent, and reliable fusion of multi-source data.

[0005] Therefore, the existing technology still needs further development. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method and system for fusing multi-source heterogeneous data from reservoir stratified injection and production, so as to solve the problems existing in the prior art.

[0007] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a method for fusing multi-source heterogeneous data from reservoir stratified injection and production, comprising: S1. Construct a spatiotemporal topological knowledge graph of reservoir injection and production, and assign spatiotemporal and reliability labels to the data; S2. Establish a library of spatiotemporal multi-scale and physical mechanism-constrained models; S3. Extract multiple candidate values ​​and confidence levels of key indicators from unstructured data; S4. Construct and solve a global constraint satisfaction and optimization model based on candidate values, constraint library and knowledge graph to obtain consistent data; S5. If there is no solution, locate the minimum set of constraints that causes the conflict and its spatiotemporal location. S6. Based on the positioning results, perform adaptive error correction and data interpolation, and iteratively optimize to finally output the fusion result and evidence chain.

[0008] Specifically, the construction of the reservoir injection-production spatiotemporal topological knowledge graph includes: The system integrates multi-source data from downhole sensors, manual reports, historical documents, and professional maps; establishes a multi-level spatial topology covering blocks, wells, layers, and sand bodies, and defines the relationships between levels; uniquely identifies the spatial location of data using a hierarchical coding format that includes block identifiers, well identifiers, and layer identifiers; and assigns each data entry a triplet label consisting of a timestamp, the spatial location of the hierarchical coding, and a reliability weight determined based on the data source type and quality, thus forming a standardized data pool.

[0009] Specifically, the spatiotemporal multi-scale and physical mechanism constraint model library includes: The spatiotemporal aggregation constraints define the aggregation rules from high-frequency to low-frequency data, and the aggregation rules from subspace units to parent space units based on the spatial topology. For cumulative physical quantities, summation is used, and for state-type physical quantities, weighted averaging is used. The physical mechanism constraints include hard constraints that must be satisfied and soft constraints that allow for deviations. The hard constraints include at least the injection-production fluid volume balance equation based on the law of mass conservation, and the soft constraints include pressure gradient range constraints and inflow dynamic relationship curve trend constraints. Finally, a caliber conversion model is used to convert fluid volume between surface standard conditions and subsurface formation conditions; this conversion is a function of downhole pressure and temperature.

[0010] Specifically, the extraction of multiple candidate values ​​and confidence levels of key indicators from unstructured data includes: Multimodal models are used to parse text reports, scanned documents, and graphs. For text and tables, key indicators are extracted and multiple candidate values, units, confidence levels, and original evidence fragments are output. For graphs, the mapping parameters from pixel coordinates to physical coordinates are also parsed when extracting values. Each extracted candidate value is represented as a tuple containing the value, unit, extraction confidence level, and evidence location information, which serves as input for subsequent optimization.

[0011] Specifically, the construction and solution of the global constraint satisfaction and optimization model includes: Each data item and its multiple candidate values ​​are constructed as decision variables, and a binary selection variable is introduced to satisfy the uniqueness constraint. An objective function is constructed to maximize the sum of weighted confidence of all selected candidate values ​​and minimize the degree of violation of soft constraints in the physical mechanism constraints. The spatiotemporal aggregation constraint, physical mechanism hard constraint, and caliber conversion relationship are added to the model as constraints that must be satisfied. A mathematical programming solver is used to solve the model to obtain a globally optimal and consistent data selection scheme that satisfies physical laws.

[0012] Specifically, the minimum set of constraints that cause a conflict due to the location includes: When the model has no feasible solution, a minimum conflict set identification algorithm is used to find a minimum subset of constraints such that the problem becomes feasible after removing any one of the constraints. Each abstract constraint in the minimum constraint subset is mapped back to the specific data item it involves. Based on the spatiotemporal label of the data item, the conflict is located to the specific well, layer and time point, and a diagnostic report containing the conflict type, the data source involved and the specific spatiotemporal location is generated.

[0013] Specifically, the execution of adaptive error correction and data interpolation includes: For conflicts caused by inconsistent measurement standards, the standard conversion model is invoked to unify the data to the same operating conditions; for conflicts caused by inconsistent spatiotemporal aggregation, the time alignment and spatial aggregation paths are re-examined, and correction values ​​are calculated based on conservation constraints; for layers or time points with missing data, the missing values ​​are inversely calculated using known total flow data, measured data from other layers, and combined with mass conservation hard constraints and spatial hierarchical constraints; the new data generated by error correction or interpolation is added as new candidate values, and the solution process of the optimization model is re-triggered for iteration until convergence.

[0014] Specifically, the method of using a mathematical programming solver to solve the model employs a hierarchical solution strategy, including: In the first stage, candidate values ​​with confidence levels higher than a preset threshold are selected first, and only the physical mechanism hard constraints are applied to solve the problem to determine the basic framework for data fusion. In the second stage, low-confidence candidate values ​​and soft constraints are introduced on the basic framework to perform local optimization to improve the results.

[0015] Specifically, the output fusion result and evidence chain include: Output consistent standard data values ​​after global optimization and iterative error correction; assign a comprehensive confidence score and a normalized spatiotemporal label to each output data value; record and output a complete evidence chain for each data value, which points to its original data source and the evidence of the selected candidate values; and output a report on the constraint satisfaction of this fusion and a detailed conflict handling log.

[0016] According to a second aspect of the present invention, a reservoir stratified injection-production multi-source heterogeneous data fusion system is provided, comprising: The data access and standardization module is used to access data and assign standardized labels. The constraint modeling and knowledge base module is used to store spatial topology and physical constraints; The multi-candidate generation engine module is used to extract multiple candidate values ​​from unstructured data; The constraint satisfaction and optimization solution engine module is used to build and solve the global optimization model; The collision detection and localization module is used to calculate the minimum collision set and perform spatiotemporal localization. The adaptive error correction and physical inference module is used to perform error correction and data inference. The results output and auditing module is used to output results and logs.

[0017] Beneficial effects: The reservoir stratified injection and production multi-source heterogeneous data fusion method and system provided by the present invention have the following significant advantages compared with the prior art.

[0018] First, the core benefit of this invention lies in significantly improving the physical consistency and engineering reliability of the fused data. By formalizing core physical mechanisms of reservoir engineering, such as the law of conservation of mass, the pressure continuity equation, and the PVT phase transition law, into explicit mathematical constraints and embedding them into a global optimization solution framework, it is ensured that the final fused data is not only statistically aligned but also completely self-consistent in terms of physical laws. This fundamentally eliminates the data contradictions that may arise from traditional methods—data that is statistically reasonable but physically absurd. This allows the output data to be directly used for refined analyses such as reservoir numerical simulations and studies of remaining oil distribution, which have extremely high data quality requirements, providing a solid and reliable data foundation for development decisions.

[0019] Secondly, this invention enables intelligent and precise diagnosis and localization of data conflicts. By introducing a minimum conflict set algorithm from the field of constraint programming, the system can quickly locate the minimum subset of conflicts causing global inconsistencies in massive amounts of data and complex constraint networks, and accurately map the conflicts to specific "well-layer-time" coordinates. This capability liberates data governance personnel from the dilemma of "knowing there is a problem, but not knowing where the problem is," transforming blind manual screening into targeted review, thereby significantly improving the efficiency of problem diagnosis and repair and significantly reducing the human resource costs of data governance.

[0020] Furthermore, this invention significantly enhances the processing capability and reliability for missing and anomalous data. Regarding the issue of missing data, the system does not employ simple statistical interpolation, but rather performs rigorous mathematical inverse reasoning based on hard physical constraints such as mass conservation. The resulting inferred values ​​have clear physical meaning, and their reliability is far superior to traditional methods. For issues such as inconsistent definitions and spatiotemporal aggregation biases, the system can automatically invoke the built-in PVT transformation model and rule base for adaptive error correction. This "physically driven" data repair and interpolation mechanism results in a qualitative leap in the completeness and accuracy of the output data.

[0021] Furthermore, this invention constructs a complete automated closed-loop process from multi-source access, knowledge modeling, conflict detection, and adaptive error correction, achieving a high degree of intelligence in the data fusion governance process and significantly shortening the data processing cycle. Simultaneously, the system provides a complete chain of evidence and audit logs for each fusion result, ensuring end-to-end traceability of the data processing process and interpretability of decisions, fully meeting the stringent security and reliability requirements of oilfield production data management.

[0022] In summary, this invention not only produces high-quality, highly consistent stratified injection and production data, but also serves as a powerful data quality diagnosis and improvement tool, providing key technical support for the in-depth construction of smart oilfields and the improvement of reservoir development benefits. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the reservoir stratified injection and production multi-source heterogeneous data fusion method provided in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of a spatiotemporal multi-scale physical constraint model provided in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the principle of multi-candidate evidence generation and global optimization solution provided in a specific embodiment of the present invention; Figure 4 This is a flowchart of the minimum conflict set calculation and conflict location provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of PVT aperture conversion and physical drive interpolation provided in a specific embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0025] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0026] Please see Figures 1-5 This invention provides a method for fusing multi-source heterogeneous data from reservoir stratified injection and production, comprising: S1. Construct a spatiotemporal topological knowledge graph of reservoir injection and production, and assign spatiotemporal and reliability labels to the data.

[0027] It should be further explained that this method aims to systematically solve the problem of multi-source heterogeneous data fusion in the field of reservoir stratified injection and production. The overall process of this invention starts with data access, proceeds through knowledge graph construction, constraint modeling, candidate value generation, global optimization solution, conflict detection and localization, adaptive error correction and iterative optimization, and finally outputs the fusion result. The multi-source heterogeneous data includes at least: (1) Real-time data acquisition, such as second-level or minute-level time-series data acquired from devices such as downhole intelligent water distributors, pressure gauges, and flow meters through industrial protocols (such as OPCUA and Modbus); (2) Manually entered data, such as structured or semi-structured report data for daily, weekly, and monthly production reports; (3) Historical document data, such as unstructured texts like geological plans, measures summaries, and dynamic analysis reports in PDF or Word format; (4) Graphical data, such as injection-production curves, water absorption profiles, production profiles, and stratified test diagrams. These data sources have different time scales (from seconds to years), different spatial granularities (from wellhead to sand body), different metering diameters (wellhead and bottom of well), and large differences in data quality.

[0028] This method constructs a unified spatiotemporal topological knowledge graph, assigning a unique triplet label to each piece of data accessed: {timestamp, spatial location, source reliability}. The spatial location uses hierarchical encoding in the format [block ID:well ID:cylinder ID:layer ID:sandbody ID], which supports multi-granularity aggregation queries and association analysis from sandbody to block. The source reliability is pre-weighted based on the reliability of the data source and the collection method; for example, manually measured data has a higher weight than automatically calculated data.

[0029] S2. Establish a library of spatiotemporal multi-scale and physical mechanism constraint models.

[0030] It should be further explained that the aforementioned spatiotemporal multi-scale and physical mechanism constraint model library includes three types of core constraints, each of which has a clearly defined mathematical expression.

[0031] S3. Extract multiple candidate values ​​and confidence levels of key indicators from unstructured data.

[0032] It should be further explained that the extraction of multiple candidate values ​​for key indicators from unstructured data is an innovative processing method addressing the uncertainty inherent in OCR (Optical Character Recognition) or LLM (Large Language Modeling) information extraction from documents and images. Instead of directly outputting a single, definitive value, it outputs a set of Top-K candidate values, each containing a quadruple: <numerical value, unit, extraction confidence level, original evidence fragment>. For image data, the mapping parameters from pixel coordinates to physical coordinates are also extracted.

[0033] S4. Based on candidate values, constraint library and knowledge graph, construct and solve the global constraint satisfaction and optimization model to obtain consistent data.

[0034] It should be further explained that the construction and solution of the global constraint satisfaction and optimization model transforms the data fusion problem into a mathematical programming problem. The decision variable is whether to select a certain candidate value, the objective function is to maximize the global data confidence and minimize the violation of physical soft constraints, and the constraint condition is that all spatiotemporal aggregation constraints and physical hard constraints must be satisfied.

[0035] S5. If there is no solution, locate the minimum set of constraints that causes the conflict and its spatiotemporal location.

[0036] It should be further explained that if no feasible solution is found, a conflict detection and localization algorithm is invoked. The localization of the minimum set of constraints leading to the conflict employs algorithms such as Quick X Plain to find a minimum subset U from all constraint sets, such that the problem remains infeasible even when only constraints in U are retained. U is then mapped to specific wells, layers, and time points.

[0037] S6. Based on the positioning results, perform adaptive error correction and data interpolation, and iteratively optimize to finally output the fusion result and evidence chain.

[0038] It should be further explained that the adaptive error correction and data imputation are performed automatically according to the conflict type. For example, PVT transformation is applied to caliber conflicts, and back-calculation is performed based on quality conservation constraints for missing data. New data is then added to the candidate set for iterative solution. The output fusion result and evidence chain ultimately provide results including fused standard data values, confidence scores, normalized spatiotemporal labels, complete evidence chains, constraint satisfaction reports, and conflict handling logs.

[0039] Understandably, this method constructs a complete, closed-loop, automated processing flow from data access to result output. By introducing hard and soft constraints based on reservoir physics mechanisms and formalizing the data fusion problem into a solvable mathematical optimization problem, it ensures that the output data is not only statistically consistent but also physically self-consistent, fundamentally solving the problems of data contradictions and unreasonable physical meanings in traditional methods. The use of multi-candidate generation and global optimization postpones the decision point, improving the fault tolerance rate for extracting information from unstructured data and the accuracy of the final result. Minimum conflict set location and adaptive error correction mechanisms significantly reduce the workload of manual review and enable intelligent diagnosis and repair of data quality issues.

[0040] Specifically, constructing a spatiotemporal topological knowledge graph of reservoir injection and production includes: The system integrates multi-source data from downhole sensors, manual reports, historical documents, and professional maps; establishes a multi-level spatial topology covering blocks, wells, layers, and sand bodies, and defines the relationships between levels; uniquely identifies the spatial location of data using a hierarchical coding format that includes block identifiers, well identifiers, and layer identifiers; and assigns each data entry a triplet label consisting of a timestamp, the spatial location of the hierarchical coding, and a reliability weight determined based on the data source type and quality, thus forming a standardized data pool.

[0041] It should be further explained that this step is the preprocessing and standardization foundation for data fusion, and specifically includes: (1) Multi-source data access: It can directly access real-time data from downhole sensors by having multiple built-in industrial protocol parsers (such as OPCUA and Modbus); Access structured business data such as daily and monthly production reports through database interfaces (ODBC / JDBC); Access unstructured reports in PDF, Word, and other formats via a document parsing engine (integrating OCR and NLP technologies); The drawing digitization engine processes JPG and other formats of annotation curves and cross-sectional views.

[0042] (2) Establishing a spatial topology: Construct a refined spatial topology of "block-well-tube-layer-sand body" in the form of a directed acyclic graph. Clarify the hierarchical relationships such as inclusion and connection between nodes. For example, a well contains multiple tubes (stems), a tube traverses multiple layers (layers), and a layer can contain multiple sand bodies (sands). Define the parent node function parent:Stem∪Layer∪Sand→Well∪Stem∪Layer to describe the hierarchical affiliation.

[0043] (3) Spatial location hierarchical coding: A unified coding format is used to uniquely identify any spatial location, such as "BlockA:Well102:Stem1:Layer3:Sand2". Even if the data source only provides granularity down to the well or layer, it is incorporated into this coding system by supplementing the default value (such as sand body ID is ALL) to ensure that all data can be aligned under the same spatial framework.

[0044] (4) Data Tagging: Assign a triplet label to each record in the standardized data pool. Timestamps are accurate to the physical point in time the data represents. Spatial location is represented by the hierarchical encoding described above. Source Reliability It is a preset weight used to distinguish the inherent reliability of different data sources. Its preferred assignment rule is, for example, for real-time data from downhole high-precision sensors. Manual on-site measurement and data entry Automated instrument data Data based on model extrapolation or empirical estimation This weight will be used for subsequent optimization of the objective function.

[0045] Understandably, through this step, previously scattered, multi-source data with varying formats and spatiotemporal benchmarks are unified into a standardized data pool with clear semantics and hierarchical relationships. Hierarchical spatial coding and quantified source reliability provide a unique and measurable foundation for subsequent cross-scale and cross-source data association, consistency verification, and fusion computation, which is key to solving the problems of "inconsistent spatial hierarchy" and "large quality differences".

[0046] Specifically, the spatiotemporal multiscale and physical mechanism-constrained model library includes: The spatiotemporal aggregation constraints define the aggregation rules from high-frequency to low-frequency data, and the aggregation rules from subspace units to parent space units based on the spatial topology. For cumulative physical quantities, summation is used, and for state-type physical quantities, weighted averaging is used. The physical mechanism constraints include hard constraints that must be satisfied and soft constraints that allow for deviations. The hard constraints include at least the injection-production fluid volume balance equation based on the law of mass conservation, and the soft constraints include pressure gradient range constraints and inflow dynamic relationship curve trend constraints. Finally, a caliber conversion model is used to convert fluid volume between surface standard conditions and subsurface formation conditions; this conversion is a function of downhole pressure and temperature.

[0047] Further explanation is needed; please refer to [link / reference]. Figure 2 The constraint model library is the core of this invention, incorporating domain knowledge, and consists of three parts: (1) Spatiotemporal aggregation constraints a) Time aggregation constraint: Define the set of time scales and conversion functions. For cumulative indicators (such as water injection volume) (oil production), the data is aggregated from daily data to monthly data using summation: ,in This indicates the cumulative monthly water injection volume. This represents the daily water injection volume on day t. For state-related indicators (such as pressure)... (Moisture content), using arithmetic mean or flow-weighted average: or ,in Indicates the average monthly pressure. Indicates the daily average pressure. It refers to the number of days.

[0048] b) Spatial aggregation constraint: Based on the well-layer-sand body hierarchical structure, the value of a parent node is defined as the aggregation of the values ​​of all its child nodes. For cumulative indices: ,in The parent node's index value. These are the values ​​of the child node indicators. For status-type indicators: ,in The weighting factor is usually taken as the thickness, permeability, or flow rate of the sub-node.

[0049] (2) Physical mechanism constraints a) Hard constraints (must be satisfied): Mass conservation constraint: For a single-well closed system, considering the PVT effect, the injected fluid volume equals the sum of the produced fluid volume and the change in stock. The formula is: ,in This represents the water injection volume under standard ground conditions for each level. The volume factor of water under formation conditions. This represents the oil production rate under standard surface conditions for each layer. This represents the volume factor of crude oil under formation conditions. This refers to the water production.

[0050] Pressure continuity constraint: Within the same connected sand body, the pressure difference between layers is limited by seepage resistance. ,in For formation pressure in different layers, The vertical or spatial distance between two segments. This is the pressure gradient coefficient.

[0051] Value range constraints: such as moisture content ,pressure (Rupture pressure), Flow rate .

[0052] b) Soft constraints (violations are allowed, but at a cost): IPR trend constraint: The relationship between output and flow pressure should follow the dynamic relationship of inflow. ,in For production, For oil production index, For formation static pressure, For bottom hole flowing pressure, This is the liquidity index.

[0053] Permeability gradient constraint: The permeability ratio of adjacent layers should be within a reasonable range (e.g., 1-10).

[0054] (3) The caliber conversion model realizes the conversion of the volume between the wellhead (surface standard conditions) and the bottom of the well (subsurface conditions): ,in This indicates the fluid volume under bottom hole conditions. This indicates the fluid volume below the wellhead (at surface standard conditions). Indicates under pressure and temperature The fluid volume coefficient at that point. For water, ,in is the isothermal compressibility coefficient of water. The coefficient of thermal expansion of water is 1. For ground reference pressure and temperature.

[0055] Understandably, this constraint model library formalizes the principles of reservoir engineering into specific mathematical rules, serving as the "brain" driving intelligent data fusion and error correction. Spatiotemporal aggregation constraints directly address the problems of "inconsistent spatiotemporal scales" and "inconsistent spatial hierarchy," providing criteria for cross-scale data alignment. Physical mechanism constraints, especially hard constraints on mass conservation, ensure that the fusion results conform to the most fundamental physical laws, avoiding absurd results. The caliber conversion model specifically addresses the industry pain point of "large differences in measurement calibers," unifying data benchmarks under different operating conditions through PVT correction. The introduction of soft constraints increases the system's flexibility, enabling it to handle the unavoidable noise and uncertainty in real-world data.

[0056] Specifically, multiple candidate values ​​and confidence levels for key indicators are extracted from unstructured data, including: Multimodal models are used to parse text reports, scanned documents, and graphs. For text and tables, key indicators are extracted and multiple candidate values, units, confidence levels, and original evidence fragments are output. For graphs, the mapping parameters from pixel coordinates to physical coordinates are also parsed when extracting values. Each extracted candidate value is represented as a tuple containing the value, unit, extraction confidence level, and evidence location information, which serves as input for subsequent optimization.

[0057] Further explanation is needed; please refer to [link / reference]. Figure 3 This step targets unstructured information sources such as historical documents, reports, and images. Its innovation lies in outputting multiple candidates rather than a single value. The specific process is as follows: (1) Multimodal parsing: Integrates a multimodal large language model, which can process text, tables and images simultaneously. Input can be PDF, Word document or JPG, PNG image.

[0058] (2) Multiple candidate generation: For the phrase "daily water injection volume of 350 cubic meters" mentioned in the text, the model may output multiple candidates, such as: <350, "m³ / d", 0.95, "located on page 5, paragraph 2">, <35, "m³ / d", 0.60, "located on page 5, paragraph 2"> (incorrect number recognition), <350, "bbl / d", 0.30, "located on page 5, paragraph 2"> (unit recognition ambiguity). The optimal value of K in Top-K is 3, achieving a balance between computational complexity and information preservation.

[0059] (3) Image digitization: For graphs such as annotation curves, the first step is to identify the coordinate axes, scale values, and extract the curves. Then, pixel coordinates are established. Linear mapping relationship to physical coordinates (time, flow rate): Simultaneously, the physical coordinates of key points on the curve (such as peaks and inflection points) are extracted as candidates. Confidence levels are then extracted. It is determined by the probability output of the classifier inside the model, and ranges between 0 and 1.

[0060] (4) Output format: All candidates are organized in a uniform format. For example, for data item d (representing "water injection volume of well A, layer 1 in January 2025"), its candidate set is: Each It is a tuple.

[0061] Understandably, this step is crucial for handling the uncertainty of unstructured data. Traditional methods directly use the single extracted value with the highest confidence level; if an extraction error occurs, it directly contaminates subsequent processes. This invention retains Top-K candidates, combining the decision of "which one is correct" with subsequent physical constraint verification and global optimization. Through global optimization, it searches for combinations in a larger solution space that both conform to the extracted evidence and satisfy physical laws, significantly improving the robustness and accuracy of recovering correct data from noisy and unstructured information.

[0062] Specifically, constructing and solving a global constraint satisfaction and optimization model includes: Each data item and its multiple candidate values ​​are constructed as decision variables, and a binary selection variable is introduced to satisfy the uniqueness constraint. An objective function is constructed to maximize the sum of weighted confidence of all selected candidate values ​​and minimize the degree of violation of the soft constraints in the physical mechanism constraints. The spatiotemporal aggregation constraints, physical mechanism hard constraints, and caliber conversion relationships are added to the model as constraints that must be satisfied. A mathematical programming solver is used to solve the model to obtain a globally optimal and consistent data selection scheme that satisfies the physical laws.

[0063] It should be further explained that this step formalizes the fusion problem into a constraint satisfaction and optimization problem, the mathematical model of which (CSOP) is as follows: (1) Decision variables: for each data item , has Candidates Introducing binary variables Indicate whether to select the first option One candidate. Satisfies If data is missing, a special empty candidate is introduced, with its confidence level set to a minimum value (e.g., 0.01).

[0064] (2) Objective function: in, For data items The source reliability base weight, Extract confidence scores for candidate values. These are decision variables. For the first The penalty coefficient for a soft constraint, This is a constraint violation degree function. For hard constraints, its... Set it to a maximum value (e.g.) This constraint must be satisfied during optimization. For soft constraints, Set according to business importance, such as IPR trend constraints. It can be set to 10, with permeability differential constraints. It can be set to 5. Typically defined as , This represents the allowable tolerance.

[0065] (3) Constraints: These include all equations or inequalities that must be satisfied, including: ① Hard constraint of mass conservation: ,in and These represent the sets of injected and generated data items, respectively. This represents the change in reserves, which can be set to zero under steady state.

[0066] ② Spatial hierarchy constraints: ,in This represents the set of data items belonging to all layers of a certain well.

[0067] ③ Time scale constraints: ,in This represents the collection of all date data items for a given month.

[0068] (4) Solving: Use a mixed integer programming solver (such as Gurobi, CPLEX) or an open source solver (such as SCIP, OR-Tools) to solve the problem.

[0069] Understandably, this global optimization model is the computational core of this invention. It unifies the decentralized data selection problem with strict domain knowledge constraints within a single mathematical framework. The objective function is designed to make the system tend to select original data with high confidence, while penalizing results that deviate from physical empirical laws. By solving this model, the system does not simply perform weighted averaging or voting, but rather searches for the globally optimal solution that simultaneously satisfies "credible evidence" and "physical rationality" within the vast solution space comprised of all possible candidate combinations, thereby achieving intelligent data fusion and arbitration.

[0070] Specifically, the minimum set of constraints that leads to a conflict includes: When the model has no feasible solution, a minimum conflict set identification algorithm is used to find a minimum constraint subset such that the problem becomes feasible after removing any constraint in the subset. Each abstract constraint in the minimum constraint subset is mapped back to the specific data item it involves. Based on the spatiotemporal label of the data item, the conflict is located to the specific well, layer and time point, and a diagnostic report containing the conflict type, the data source involved and the specific spatiotemporal location is generated.

[0071] Further explanation is needed; please refer to [link / reference]. Figure 4 When the optimization model has no solution, it indicates that there are irreconcilable contradictions among the existing candidate data. This step is used to accurately locate the root cause of the contradiction, and the specific algorithm is as follows: (1) Calculate the minimum conflict set: Use the improved Quick X Plain algorithm. The input is the set of all constraints. and the current candidate set The algorithm finds a minimum subset through recursive partitioning and testing. , making Not feasible, but For any It is feasible. This means... It is the "minimum" contradictory combination that leads to infeasibility.

[0072] (2) Spatiotemporal positioning mapping: For Each constraint in Find the set of all data items involved in the constraint. For each Extract the spatiotemporal information (hierarchical encoding and timestamps) from the triplet labels. Aggregate this information to obtain the precise spatiotemporal range where the conflict occurred. For example, if If the data items involved are “#A-Layer 2-2025-01-15-Flowmeter Reading” and “#A-Layer 2-2025-01-Monthly Report”, then the conflict location result is: During January 2025, there is an irreconcilable contradiction between the flowmeter data and the monthly report data of Layer 2 of Well A.

[0073] (3) Generate diagnostic report: The report automatically analyzes the conflict type (such as "mass conservation conflict", "spatial aggregation conflict", "caliber inconsistency conflict"), lists all data sources involved (such as "sensor S101", "manual daily report D202501"), and gives the precise spatiotemporal location (such as "well number: Well102; layer: Layer3; time: 2025-06-10").

[0074] Understandably, traditional conflict detection methods typically only detect discrepancies in total counts, but cannot pinpoint the specific data(s) causing the error at any given location. This invention introduces a minimum conflict set algorithm, which can quickly locate one or a few core data points causing a global conflict from hundreds or thousands of data items and constraints, narrowing the conflict scope from a "vague mess" to a "precise target." This transforms manual review from searching for a needle in a haystack into a targeted process, improving efficiency by over 80% and achieving a breakthrough in conflict location accuracy.

[0075] Specifically, adaptive error correction and data interpolation are performed, including: For conflicts caused by inconsistent metering standards, the metering conversion model is automatically invoked to unify the data to the same operating conditions; for conflicts caused by inconsistent spatiotemporal aggregation, the time alignment and spatial aggregation paths are re-examined, and correction values ​​are calculated or abnormal data points are identified based on conservation constraints; for layers or time points with missing data, the missing values ​​are reverse-calculated using known total flow data, measured data from other layers, and combined with mass conservation hard constraints and spatial hierarchical constraints; the new data generated by error correction or interpolation is added as new candidate values, and the solution process of the optimization model is re-triggered for iteration until convergence.

[0076] Further explanation is needed; please refer to [link / reference]. Figure 5 This step is an intelligent process that automatically executes the repair strategy based on the conflict location results: (1) Calibration conflict correction: Identify the original metering conditions (wellhead or bottom) of conflicting data items and automatically apply the PVT volume factor. Perform the conversion. For example, convert the monthly wellhead water injection volume. Convert to bottom hole conditions Then, the data is compared and integrated with the data from the downhole flow meter.

[0077] (2) Spatiotemporal conflict correction: Check time alignment (e.g., whether the daily report date corresponds to the monthly report month) and spatial aggregation path (e.g., whether the layer list is complete). If it is found that the missing data of a certain layer causes the sum to be not equal to the total wellhead data, calculate the possible value of that layer based on the conservation constraints, or mark that the total wellhead data may be abnormal.

[0078] (3) Physically driven missing data imputation: This is one of the advantages of this invention. For example, given the total water injection volume at the wellhead... (Surface conditions), measured bottom water injection volume in the other n-1 layers. and the PVT coefficient of each layer According to the law of conservation of mass: The amount of water injected into the bottom of the well in the missing section can be calculated in reverse. : This inference is based on strict physical laws, and its error is much smaller than that of traditional mean-filling.

[0079] (4) Iterative optimization: The new values ​​generated by the above error correction or interpolation (with the "physical inference" label and the calculated confidence level) are added to the original data items as new candidate values. Then, the global optimization model is returned to solve it again. This process is repeated until the model obtains a feasible solution and the change in the objective function value is less than a preset threshold (such as 0.1%), or the maximum number of iterations is reached (preferably 10 times).

[0080] Understandably, adaptive error correction and interpolation form a complete closed loop of "detection-localization-repair-verification." The system can not only identify problems but also automatically attempt to repair them based on domain knowledge (physical constraints). Compared to purely mathematical interpolation, the physics-driven interpolation method yields results with clear physical meaning and higher reliability, with the average relative error controlled within 10% in testing. The iterative optimization mechanism ensures that the repaired data undergoes global consistency verification again, thereby outputting a highly self-consistent, complete, and reliable fused data view.

[0081] Specifically, the mathematical programming solver is used to solve the model using a hierarchical solution strategy, including: In the first stage, candidate values ​​with confidence levels higher than a first preset threshold are selected first, and only the physical mechanism hard constraints are applied to solve the problem to determine the basic framework of the data fusion result; In the second stage, on the basic framework, candidate values ​​with confidence levels lower than the first preset threshold but higher than a second preset threshold, as well as the physical mechanism soft constraints, are introduced, and local optimization is performed in the neighborhood of the determined basic framework to refine and improve the fusion result.

[0082] It should be further explained that, to address the efficiency issues of solving large-scale problems (data items > 1000), a hierarchical solution strategy is adopted: (1) First stage: quickly determine the framework.

[0083] Only high-confidence candidates are considered, i.e., confidence levels are extracted. Candidates exceeding a first preset threshold. The preferred value for this threshold is 0.8, because data exceeding this value typically comes from reliable manual measurements or calibration equipment, with a low probability of error.

[0084] Only apply hard constraints based on physical mechanisms (such as mass conservation and range of values) and core spatiotemporal aggregation constraints. All soft constraints are temporarily ignored.

[0085] This simplified model is solved rapidly, yielding a "skeleton" solution composed of highly reliable data that satisfies the most fundamental physical laws. This stage offers fast solution speed and high stability.

[0086] (2) Second stage: Local fine optimization.

[0087] The selection of high-confidence data already determined in the first phase is fixed.

[0088] Introduce candidates with medium to low confidence levels, i.e. Candidates are those falling between the second preset threshold and the first preset threshold. The preferred value for the second preset threshold is 0.3; data below this value is considered too unreliable and is generally not included in the optimization to avoid introducing too much noise.

[0089] At the same time, soft constraints of all physical mechanisms (such as IPR trend and pressure gradient) are introduced.

[0090] Within the neighborhood of the "skeleton" solution, a local search and optimization is performed on the remaining undetermined data items, with the goal of further maximizing the overall confidence and satisfying soft constraints. This stage refines the already established reliable framework, avoiding the complexity of a global search.

[0091] Understandably, the hierarchical solution strategy is an effective engineering practice for large-scale optimization problems. The first stage uses high-confidence data and hard constraints to quickly identify a reasonable solution space region, establishing the "backbone" of the fused results. The second stage, based on this, uses richer but potentially noisier data (medium-to-low confidence candidates) and more refined rules (soft constraints) for "refinement." This strategy significantly reduces the computational complexity of the problem, improves the solution speed and success rate, and enables the method of this invention to handle massive amounts of multi-source data in actual oilfield production.

[0092] Specifically, the output fusion results and evidence chain include: Output consistent standard data values ​​after global optimization and iterative error correction; assign a comprehensive confidence score and a normalized spatiotemporal label to each output data value; record and output a complete evidence chain for each data value, the evidence chain pointing to the original data source on which it is based and the selected candidate value evidence fragments; and output a statistical report on the constraint satisfaction of this fusion process and a detailed conflict handling operation log for auditing and traceability.

[0093] It should be further explained that the final output reflects the value of data fusion, emphasizing auditability and traceability, including: (1) Standard data values: All output values ​​have been standardized to the spatiotemporal granularity and measurement caliber (such as well bottom conditions and daily scale).

[0094] (2) Overall confidence score: A comprehensive score between 0 and 1 is calculated for each output value, which combines its source reliability. Extraction confidence of selected candidates And the situation where constraints are satisfied during optimization.

[0095] (3) Standardized spatiotemporal labels: using standard hierarchical coding and time format.

[0096] (4) Complete evidence chain: For each output value, record its complete "origin", for example: "This data value comes from 'Water injection volume of Layer 22025-01 in Well A'. Its value of 85.3 m³ / d was selected from the candidate set through global optimization. The original evidence is the text 'Approximately 85 cubic meters of water injected per day in Layer 2' on page 5 of the 'Dynamic Analysis Report in January 2025'. It has also undergone mass conservation constraint verification and missing data back-inference interpolation." (5) Audit Report: This includes the "Constraint Satisfaction Report" (listing the satisfaction status of all hard / soft constraints) and the "Conflict Handling Operation Log" (recording the conflicts detected, location results, and error correction operations performed in each iteration). After applying this method, the data consistency accuracy can be improved from 68% of the traditional method to over 96%, the physical law violation rate is less than 1%, the workload of manual review is reduced by 80%, and the entire process time is shortened from 48 hours (mainly manual) to 30 minutes.

[0097] Understandably, the output of this step not only provides high-quality fused data directly applicable to detailed reservoir analysis (such as residual oil studies and numerical simulation history fitting), but also offers a complete "data passport" and "processing history." This significantly enhances the credibility and transparency of the data, meeting the stringent requirements of oilfield production data management for reliability and auditability. Engineers can trust this data and, when in doubt, can quickly trace its origins and understand the basis for system decisions.

[0098] The present invention also provides another embodiment, which provides a reservoir stratified injection and production multi-source heterogeneous data fusion system, characterized in that it includes: (1) Data access and standardization module, used to access data and assign standardized labels.

[0099] It should be further explained that the data access and standardization module has a built-in industrial protocol parser, database interface, document parsing engine (OCR+NLP), and map digitization engine, performs ETL operations, and outputs a standardized data pool with triplet labels.

[0100] (2) Constraint modeling and knowledge base module, used to store spatial topology and physical constraints.

[0101] It should be further explained that the constraint modeling and knowledge base module stores and manages the well-tube-layer-sandbody spatial topology tree, physical mechanism constraint library, spatiotemporal aggregation rule library, PVT parameter table, etc. A visual configuration interface is provided for experts to maintain the knowledge.

[0102] (3) Multi-candidate generation engine module, used to extract multiple candidate values ​​from unstructured data.

[0103] It should be further explained that the multi-candidate generation engine module integrates a multimodal large language model, processes unstructured data, and outputs a set of Top-K candidate tuples.

[0104] (4) Constraint satisfaction and optimization solution engine module, used to build and solve the global optimization model.

[0105] It should be further explained that the constraint satisfaction and optimization solution engine module is the computational core of the system. It integrates a mathematical programming solver (such as Gurobi) and is responsible for modeling the fusion problem as a CSOP and solving it, supporting a hierarchical solution strategy.

[0106] (5) Conflict detection and localization module, used to calculate the minimum conflict set and perform spatiotemporal localization.

[0107] It should be further explained that the conflict detection and localization module is triggered when the solver has no solution, runs algorithms such as QuickX Plain to calculate the minimum conflict set, and maps it to a specific spatiotemporal location to generate a diagnostic report.

[0108] (6) Adaptive error correction and physical inference module, used to perform error correction and data inference.

[0109] It should be further explained that the adaptive error correction and physical inference module is used to automatically invoke strategies such as PVT transformation and inferring missing values ​​based on conservation laws to generate new candidate data based on the diagnostic report.

[0110] (7) Results output and auditing module, used to output results and logs.

[0111] It should be further explained that the results output and audit module is used to generate the final fused data view, evidence chain, various reports and tamper-proof audit logs, and provides a visual interface.

[0112] Understandably, this system instantiates the method flow of this invention into a concrete, deployable software system. Each module has clearly defined responsibilities and connects through standard interfaces, forming an integrated pipeline from data input to intelligent output. This system not only automates data fusion, but its conflict detection and location, and physical-driven inference functions can also serve as independent data quality diagnosis and repair tools, supporting intelligent production management in oilfields and directly supporting refined decision-making such as stratified injection optimization and measure effectiveness evaluation, thereby creating economic benefits.

[0113] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the described reservoir stratified injection-production multi-source heterogeneous data fusion method. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0114] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0115] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0116] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0117] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for fusing multi-source heterogeneous data from reservoir stratified injection and production, comprising: S1. Construct a spatiotemporal topological knowledge graph of reservoir injection and production, and assign spatiotemporal and reliability labels to the data; S2. Establish a library of spatiotemporal multi-scale and physical mechanism-constrained models; S3. Extract multiple candidate values ​​and confidence levels of key indicators from unstructured data; S4. Construct and solve a global constraint satisfaction and optimization model based on candidate values, constraint library and knowledge graph to obtain consistent data; S5. If there is no solution, locate the minimum set of constraints that causes the conflict and its spatiotemporal location. S6. Based on the positioning results, perform adaptive error correction and data interpolation, and iteratively optimize to finally output the fusion result and evidence chain.

2. The method according to claim 1, characterized in that, The construction of the spatiotemporal topological knowledge graph of reservoir injection and production includes: The system integrates multi-source data from downhole sensors, manual reports, historical documents, and professional maps; establishes a multi-level spatial topology covering blocks, wells, layers, and sand bodies, and defines the relationships between levels; uniquely identifies the spatial location of data using a hierarchical coding format that includes block identifiers, well identifiers, and layer identifiers; and assigns each data entry a triplet label consisting of a timestamp, the spatial location of the hierarchical coding, and a reliability weight determined based on the data source type and quality, thus forming a standardized data pool.

3. The method according to claim 2, characterized in that, The spatiotemporal multi-scale and physical mechanism-constrained model library includes: Spatiotemporal aggregation constraints define aggregation rules from high-frequency to low-frequency data, as well as rules for aggregation from subspace units to parent space units based on the spatial topology. Summation is used for cumulative physical quantities, and weighted averaging is used for state-type physical quantities. Physical mechanism constraints include hard constraints that must be satisfied and soft constraints that allow deviations. The hard constraints include at least the injection-production fluid volume balance equation based on the law of mass conservation, and the soft constraints include pressure gradient range constraints and inflow dynamic relationship curve trend constraints. And a caliber conversion model is used to convert fluid volume between surface standard conditions and underground formation conditions, which is a function of downhole pressure and temperature.

4. The method according to claim 3, characterized in that, The extraction of multiple candidate values ​​and confidence levels for key indicators from unstructured data includes: Multimodal models are used to parse text reports, scanned documents, and graphs. For text and tables, key indicators are extracted and multiple candidate values, units, confidence levels, and original evidence fragments are output. For graphs, the mapping parameters from pixel coordinates to physical coordinates are also parsed when extracting values. Each extracted candidate value is represented as a tuple containing the value, unit, extraction confidence level, and evidence location information, which serves as input for subsequent optimization.

5. The method according to claim 4, characterized in that, The construction and solution of the global constraint satisfaction and optimization model includes: Each data item and its multiple candidate values ​​are constructed as decision variables, and a binary selection variable is introduced to satisfy the uniqueness constraint. An objective function is constructed to maximize the sum of weighted confidence of all selected candidate values ​​and minimize the degree of violation of soft constraints in the physical mechanism constraints. The spatiotemporal aggregation constraint, physical mechanism hard constraint, and caliber conversion relationship are added to the model as constraints that must be satisfied. A mathematical programming solver is used to solve the model to obtain a globally optimal and consistent data selection scheme that satisfies physical laws.

6. The method according to claim 5, characterized in that, The minimum set of constraints that cause a conflict in the location includes: When the model has no feasible solution, a minimum conflict set identification algorithm is used to find a minimum subset of constraints such that the problem becomes feasible after removing any one of the constraints. Each abstract constraint in the minimum constraint subset is mapped back to the specific data item it involves. Based on the spatiotemporal label of the data item, the conflict is located to the specific well, layer and time point, and a diagnostic report containing the conflict type, the data source involved and the specific spatiotemporal location is generated.

7. The method according to claim 6, characterized in that, The execution of adaptive error correction and data interpolation includes: For conflicts caused by inconsistent measurement standards, the standard conversion model is invoked to unify the data to the same operating conditions; for conflicts caused by inconsistent spatiotemporal aggregation, the time alignment and spatial aggregation paths are re-examined, and correction values ​​are calculated based on conservation constraints; for layers or time points with missing data, the missing values ​​are inversely calculated using known total flow data, measured data from other layers, and combined with mass conservation hard constraints and spatial hierarchical constraints; the new data generated by error correction or interpolation is added as new candidate values, and the solution process of the optimization model is re-triggered for iteration until convergence.

8. The method according to claim 5, characterized in that, The method of solving the model using a mathematical programming solver employs a hierarchical solution strategy, including: In the first stage, candidate values ​​with confidence levels higher than a preset threshold are selected first, and only the physical mechanism hard constraints are applied to solve the problem to determine the basic framework for data fusion. In the second stage, low-confidence candidate values ​​and soft constraints are introduced on the basic framework to perform local optimization to improve the results.

9. The method according to any one of claims 1-8, characterized in that, The output fusion result and evidence chain include: Output consistent standard data values ​​after global optimization and iterative error correction; assign a comprehensive confidence score and a normalized spatiotemporal label to each output data value; record and output a complete evidence chain for each data value, which points to its original data source and the evidence of the selected candidate values; and output a report on the constraint satisfaction of this fusion and a detailed conflict handling log.

10. A reservoir stratified injection-production multi-source heterogeneous data fusion system, characterized in that, include: The data access and standardization module is used to access data and assign standardized labels. The constraint modeling and knowledge base module is used to store spatial topology and physical constraints; The multi-candidate generation engine module is used to extract multiple candidate values ​​from unstructured data; The constraint satisfaction and optimization solution engine module is used to build and solve the global optimization model; The collision detection and localization module is used to calculate the minimum collision set and perform spatiotemporal localization. The adaptive error correction and physical inference module is used to perform error correction and data inference. The results output and auditing module is used to output results and logs.

Images