Multi-source information fusion method in intelligent question and answer system

Abstract

The invention discloses a multi-source information fusion method in an intelligent question and answer system, which comprises the following steps: acquiring response data and a petroleum information source of a preset system, and preprocessing the response data; the petroleum information source comprises structured data, semi-structured data and unstructured data; matching and screening the petroleum information source by adopting a bidirectional converter model according to the question and answer data to obtain a candidate data set, and performing deterministic optimization according to the information reliability of the candidate data set to obtain a reliability data set; carrying out conflict resolution on the fused data based on a multi-source conflict elimination mechanism to obtain effective data, carrying out dynamic fusion on the effective data according to a multi-source perception strategy to obtain fused data, and constructing a petroleum question and answer multi-source information fusion model according to the fused data; and optimizing the petroleum question and answer multi-source information fusion model according to a feedback optimization mechanism, inputting to-be-fused data into the petroleum question and answer multi-source information fusion model, and outputting a fusion result.

Description

Technical Field

[0001] This invention relates to the field of multi-source information fusion, and more particularly to a method for multi-source information fusion in an intelligent question-answering system. Background Art

[0002] Against the backdrop of rapid development in information technology and continuous improvement in intelligence, human-computer interaction technology has been deeply integrated into various scenarios of modern society, from smartphones and smart homes to autonomous vehicles and intelligent medical systems, with user intent recognition technology playing a core role. However, in the highly specialized and complex data-driven environment of the petroleum industry, the information processing capabilities of existing intelligent question-and-answer systems face significant challenges.

[0003] The information sources in the petroleum industry are highly diverse, encompassing structured, semi-structured, and unstructured data. These data vary significantly in format, content, and reliability, making efficient integration of multi-source information difficult. Furthermore, question-and-answer requests in the petroleum sector often involve complex professional knowledge, demanding extremely high accuracy, relevance, and timeliness of information. Existing methods, when faced with the fusion of multi-dimensional and multi-type information, frequently suffer from insufficient information matching accuracy, difficulty in resolving conflicts, and low reliability of the fusion results, failing to accurately capture users' true intentions and provide precise responses.

[0004] Therefore, developing a multi-source information fusion method for an intelligent question-and-answer system that can effectively integrate multi-source information in the petroleum field, resolve information conflicts, and improve fusion accuracy is of great significance for improving the quality and efficiency of intelligent question-and-answer in the petroleum industry and meeting the industry's demand for precise information services. Summary of the Invention

[0005] The purpose of this invention is to provide a method for multi-source information fusion in an intelligent question-answering system.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: The system collects response data from a preset system and petroleum information sources, and preprocesses the response data. The petroleum information sources include structured data, semi-structured data, and unstructured data. The structured data includes oilfield production data, geological exploration data, and equipment monitoring data. The semi-structured data includes technical reports, industry standard documents, and equipment technical manuals. The unstructured data includes academic papers, patent documents, industry news and conference analyses, and case records. Based on the question-and-answer data, a bidirectional transformer model is used to match and filter the petroleum information sources to obtain candidate datasets. Based on the information reliability of the candidate datasets, deterministic optimization is performed to obtain a reliability dataset. The fused data is resolved based on a multi-source conflict elimination mechanism to obtain effective data. The effective data is then dynamically fused according to a multi-source perception strategy to obtain fused data. A multi-source information fusion model for petroleum question answering is then constructed based on the fused data. The petroleum question-and-answer multi-source information fusion model is optimized according to the feedback optimization mechanism. The data to be fused is input into the petroleum question-and-answer multi-source information fusion model, and the fusion result is output.

[0007] Furthermore, the method for matching and filtering the petroleum information source using a bidirectional transformer model based on the question-and-answer data includes: A bidirectional transformer model is used to extract knowledge from question-and-answer data and petroleum information sources, respectively, to obtain question-and-answer knowledge data and standard knowledge data. Entity matching is performed on the question-and-answer knowledge data and standard knowledge data, and standard knowledge data with an entity matching degree greater than 0.761 is used as primary candidate content. A graph attention network is used to perform relation matching on the question-and-answer knowledge data and standard knowledge data, and primary candidate content with a relation matching degree greater than 0.519 is used as secondary candidate content. Attribute matching is performed on question-and-answer knowledge data and standard knowledge data. Evidence theory is used to handle fuzzy attributes, and primary candidate content with an attribute matching degree greater than 0.604 is used as tertiary candidate. A neighborhood-adaptive petroleum semantic matching model is used to perform semantic matching on question-and-answer knowledge data and standard knowledge data. A multi-granularity attention mechanism is used to jointly calculate word-level, phrase-level, and sentence-level similarity. Tertiary candidates with a similarity degree greater than 0.785 are output as the screening results, and the petroleum information sources of the screening results are output as candidate datasets. Knowledge extraction includes entity recognition, relation extraction, attribute extraction, and semantic extraction.

[0008] Furthermore, the method for obtaining a reliability dataset through deterministic optimization based on the information reliability of the candidate dataset includes: Obtain the authority, time difference, cross-validation rate, and objective correctness score of each element, and calculate the information reliability of the candidate dataset elements: ; The information reliability of the j-th element is: Authority A The time difference is The cross-validation rate is C The objective and correct score is D The first weighting coefficient is The second weighting coefficient is The third weighting coefficient is The fourth weighting coefficient is The attenuation rate is ; Elements with an information reliability score below 0.519 are considered low-reliability data and discarded; elements with an information reliability score greater than 0.519 and less than 0.803 are considered medium-reliability data; and elements with an information reliability score greater than 0.803 are considered high-reliability data. Check if the reliability data contains pressure parameters. If it does, perform pressure range verification. If the pressure exceeds the range, truncate and reduce the weight. If it does not contain pressure parameters, check if it contains physical property parameters. If it does, perform physical property logic verification. If there is a contradiction, mark it as invalid. A reliability dataset will be constructed by verifying physical properties, truncating and reducing the weight of the results, and using high-reliability data.

[0009] Furthermore, a method for resolving conflicts in the fused data based on a multi-source conflict resolution mechanism to obtain valid data includes: Data conflict monitoring is performed on the fused data, key values ​​are extracted from the fused data, and the coefficient of variation is calculated when key values ​​from different sources are inconsistent. , The mean of the z-th key value is The standard deviation of the z-th key value is The coefficient of variation is When the coefficient of variation is greater than 0.15 and less than or equal to 0.15, it is judged as a moderate numerical conflict; when the coefficient of variation is greater than 0.3, it is judged as a severe numerical conflict. Spatiotemporal conflict detection is performed on the fused data, and well location coordinates are extracted from the fused data. When well location coordinates from different sources are inconsistent, a conflict is identified, and the standard deviation of the well location coordinates is calculated. , , The latitude of the i-th well location is The longitude of the i-th well location is The mean of the latitudes is The average longitude is The Earth's radius is R, and the semi-sine distance of the i-th well location is... The standard deviation of the coordinates of the i-th well location is The number of well locations is N The average value of the half-versus distance of the well location is ; When the standard deviation of the well location coordinates is greater than 50 meters but less than 500 meters, the well location is rechecked; when the standard deviation of the well location coordinates is greater than 500 meters, a spatial conflict is triggered; the fused data with numerical and spatial conflicts are removed to obtain the first data. The first data from the sensor data source that has a real-time time of less than two minutes and complies with the API 14B safety standard is selected and output as valid data.

[0010] Furthermore, the method for dynamically fusing the effective data according to a multi-source sensing strategy to obtain fused data includes: The perception dimensions of the multi-source perception strategy include numerical data, spatiotemporal data, and semantic data. The effective data is divided into numerical data, spatiotemporal data, and semantic data through the multi-source perception strategy. The target oilfield is divided into grids, and the data is classified according to the grids. The information reliability, nearest fault distance, and grid permeability of the effective data are obtained. The timeliness compensation, composite attenuation factor, and dynamic weight of the effective data are calculated. , , , The information reliability of the c-th valid data is: The dynamic weight of the c-th valid data is The timeliness compensation for the c-th valid data is The composite attenuation factor of the c-th valid data is The nearest fault distance is The attenuation coefficient is The time difference is Grid penetration rate G The fault distance attenuation coefficient is Penetration rate weighting is Fault weight is The normalization function is ; Based on dynamic weights, the first stage rapidly fuses the effective data to obtain primary fused data. The second stage initiates multiphysics coupling analysis and spatiotemporal consistency verification on the primary fused data. If the spatiotemporal consistency verification result is greater than 0.947, the effective data is fused according to the perception dimension; otherwise, rapid fusion is performed again. The fused result is then output as fused data. The system employs a confidence-weighted average fusion for continuous numerical data, a maximum confidence-first fusion for discrete state data, a knowledge graph semantic aggregation fusion for text reports, and Kalman filtering and confidence-constrained fusion for spatiotemporal trajectories.

[0011] Furthermore, the method for constructing a multi-source information fusion model for petroleum question-and-answer based on the fused data includes: A reasonable evaluation value is obtained by conducting a rationality assessment on the fused data, and the objective function of the petroleum question-and-answer multi-source information fusion model is obtained by objectively weighting the reasonable evaluation value and the loss function. The multi-source information fusion model for petroleum question answering includes autoencoders, Bayesian networks, and recurrent neural networks; An autoencoder compresses input data into low-dimensional features through an encoder, and a decoder reconstructs the original data. The reconstruction error is then used to perform feature dimensionality reduction and anomaly screening to obtain question-and-answer feature data. Bayesian networks express the conditional dependencies between variables based on probabilistic graphical models. They update posterior probabilities using prior probabilities and observed data, and use question-and-answer feature data and petroleum information sources as control groups. Uncertainty inference and causal association analysis are performed on the control groups to obtain screened data. Recurrent neural networks utilize the temporal propagation mechanism of hidden states and combine it with the objective function to capture and optimize long-term dependencies in the data, and then fuse the selected data based on these long-term dependencies.

[0012] Furthermore, the optimization method for the petroleum question-and-answer multi-source information fusion model based on the feedback optimization mechanism includes: By acquiring data freshness, initial confidence, false alarm rate of outdated knowledge, semantic consistency, objective score, structured data weight, unstructured data weight, and timeliness decay factor, the state space of the petroleum industry is reconstructed, and the objective function is given: , , , The weight matrices are Q and B, the transpose of the matrix is ​​T, and the objective function is... The information state matrix is ​​H, a column vector. The first column represents data freshness, the second data reliability, the third semantic consistency, and the fourth objective score. The control matrix is ​​E, a column vector. The first column represents the weight of structured data, the second the weight of unstructured data, and the third the timeliness decay factor. Data reliability is... The initial credibility is The time-varying coefficient at time t is The false alarm rate for outdated knowledge at time t is ; Using real-number encoding, the key parameters of multi-source information fusion are transformed into heritable and optimizeable gene fragments, with each gene corresponding to a parameter and its value range; The candidate solutions of the petroleum question-answering multi-source information fusion model are used as particles, and the population is initialized using a random method and circular mapping, as expressed in the following expression: , The initial population of the (r+1)th dimension is: The chaos value of the r-th dimension The remainder function is The constants less than 1 are w and b, respectively; To obtain the particle's answer accuracy error, response time, and objective satisfaction, given the fitness function, the expression is: , The fitness function of the a-th particle is: The accuracy error of the answer for the a-th particle is The response time of the a-th particle is The objective satisfaction of the a-th particle is The error weight is Time weighting is The satisfaction weight is The objective function weights are ; Calculate the fitness function value of each particle, select the individual with the smallest fitness function value as the best individual, and implement a retention strategy. This involves improving crossover and mutation through adaptive crossover rates, and iteratively updating the population. The expression is: , Where the mutation probability is The fitness function value of the a-th particle is The average fitness function value is Cauchy distribution is The position parameters are The scale parameter is The basic variability rate is , The process continues iterating until the fitness function value reaches its minimum. The particle with the minimum fitness function value is then selected as the optimal combination of parameters.

[0013] The beneficial effects of the present invention are: This invention provides a multi-source information fusion method for intelligent question-answering systems. Compared with existing technologies, this invention has the following technical advantages: This invention improves the accuracy and precision of multi-source information fusion in the petroleum field through steps such as preprocessing, matching and filtering, conflict resolution, dynamic fusion and feedback optimization. It saves resources, improves work efficiency, realizes intelligent real-time information processing, adapts to different fusion needs, and enhances the reliability of results. It is of great significance to the multi-source information fusion of petroleum intelligent question-and-answer systems. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a multi-source information fusion method in an intelligent question-answering system according to the present invention. Detailed Implementation

[0015] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0016] The present invention discloses a multi-source information fusion method in an intelligent question-answering system, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: The system collects response data from a preset system and petroleum information sources, and preprocesses the response data. The petroleum information sources include structured data, semi-structured data, and unstructured data. The structured data includes oilfield production data, geological exploration data, and equipment monitoring data. The semi-structured data includes technical reports, industry standard documents, and equipment technical manuals. The unstructured data includes academic papers, patent documents, industry news and conference analyses, and case records. In the actual assessment, a question-and-answer system of a certain oilfield was used as the research object; Example 1: A multi-source fusion scenario primarily using structured data, with the core being structured oilfield production data (daily oil production, wellhead pressure), supplemented by a small amount of semi-structured equipment technical manuals. Production data and equipment maintenance manuals from the past three months were collected. The production data was deduplicated, missing values ​​were filled, and the manuals were formatted. A bidirectional transformer model was used to extract the entity "wellhead pressure," which was then matched with the entity "equipment pressure resistance parameters" in the manuals to obtain candidate datasets. The information reliability of the production data (authority A=0.9, cross-validation rate C=0.95) was calculated, yielding Cf=0.87 (high reliability), which was directly included in the reliability dataset. Based on a multi-source perception strategy, the production data was determined to be numerical. A reliability-weighted average fusion was used, combined with equipment parameter verification from the manuals, to output the fusion result of "optimal wellhead pressure range." Example 2: Unstructured data conflict resolution scenario; Five papers on "shale oil fracturing technology" contained conflicting fracturing fluid usage parameters; the fracturing fluid usage of each paper was extracted as 100m³, 120m³, 150m³, 180m³, and 200m³, with a calculated mean of 150m³, a standard deviation of 38.7, and a coefficient of variation Cv=25.8%, indicating a moderate numerical conflict; combined with a reliability dataset, the 150m³ value from a high-reliability paper (Cf=0.86, published in a core journal) was selected as the benchmark, and other values ​​were adjusted according to reliability weights, resulting in a final fused usage of 148.5m³; after multiphysics coupling analysis, the spatiotemporal consistency verification value of the fused result was 0.96, confirming its validity; Example 3: Spatiotemporal Conflict Handling Scenario; A spatial conflict exists between the newly measured well location coordinates and the well location descriptions in historical technical reports for a certain block; the newly measured coordinates (φ=30.123°, λ=118.456°) and the coordinates described in the historical reports are extracted, and the standard deviation of the well location semi-versus distance is calculated to be 700 meters, which is judged as a serious spatial conflict; low-reliability historical data (Cf=0.48<0.519) are removed, and the well location coordinates are corrected to (φ=30.125°, λ=118.458°) by reviewing with high-reliability geological exploration reports (Cf=0.91); the timeliness decay factor of the model is adjusted through a feedback mechanism, which improves the semantic consistency score of subsequent well location data to 0.92 and reduces similar conflicts; Example 4: Dynamic Fusion of Mixed Data Types; This scenario involves collecting mixed data from geological exploration, industry standard documents, and shale oil extraction patents. Geological data (porosity, permeability) is categorized as numerical, standard documents (development specifications) as semantic, and patent documents (time series data) as spatiotemporal. For a particular geological data point, due to its high freshness (∆t=30 days, timeliness compensation of 0.86), grid permeability of 0.8, and composite attenuation factor of 0.78, the final dynamic weight is 0.82. For numerical data, a reliability-weighted average is used, resulting in a fused porosity mean of 18.5%. For semantic data, industry standards are aggregated using a knowledge graph to extract key constraints. For spatiotemporal data, Kalman filtering is used to fuse extraction progress data from patents to obtain development cycle predictions. Example 5: Medium-reliability data verification scenario; A reservoir's physical properties (including permeability and oil saturation, structured) have an information reliability Cf = 0.72 (medium reliability). During logical verification of the physical properties, a contradiction was found between permeability (50mD) and oil saturation (80%) (the conventional rule is that high permeability corresponds to medium to low saturation), and this was marked as invalid. High-reliability alternative data (Cf = 0.85) with a permeability of 30mD and an oil saturation of 65% was selected and included in the fused dataset after passing the physical property logical verification. The cross-validation rate weight was adjusted to 0.3 via a feedback mechanism to improve the accuracy of reliability calculation for similar data and reduce the inclusion of invalid data. Based on the question-and-answer data, a bidirectional transformer model is used to match and filter the petroleum information sources to obtain candidate datasets. Based on the information reliability of the candidate datasets, deterministic optimization is performed to obtain a reliability dataset. The fused data is resolved based on a multi-source conflict elimination mechanism to obtain effective data. The effective data is then dynamically fused according to a multi-source perception strategy to obtain fused data. A multi-source information fusion model for petroleum question answering is then constructed based on the fused data. The petroleum question-and-answer multi-source information fusion model is optimized according to the feedback optimization mechanism. The data to be fused is input into the petroleum question-and-answer multi-source information fusion model, and the fusion result is output.

[0017] In this embodiment, the method for matching and filtering the petroleum information source using a bidirectional transformer model based on the question-and-answer data includes: A bidirectional transformer model is used to extract knowledge from question-and-answer data and petroleum information sources, respectively, to obtain question-and-answer knowledge data and standard knowledge data. Entity matching is performed on the question-and-answer knowledge data and standard knowledge data, and standard knowledge data with an entity matching degree greater than 0.761 is used as primary candidate content. A graph attention network is used to perform relation matching on the question-and-answer knowledge data and standard knowledge data, and primary candidate content with a relation matching degree greater than 0.519 is used as secondary candidate content. Attribute matching is performed on question-and-answer knowledge data and standard knowledge data. Evidence theory is used to handle fuzzy attributes, and primary candidate content with an attribute matching degree greater than 0.604 is used as tertiary candidate. A neighborhood-adaptive petroleum semantic matching model is used to perform semantic matching on question-and-answer knowledge data and standard knowledge data. A multi-granularity attention mechanism is used to jointly calculate word-level, phrase-level, and sentence-level similarity. Tertiary candidates with a similarity degree greater than 0.785 are output as the screening results, and the petroleum information sources of the screening results are output as candidate datasets. Knowledge extraction includes entity recognition, relation extraction, attribute extraction, and semantic extraction.

[0018] In this embodiment, the method for obtaining a reliability dataset through deterministic optimization based on the information reliability of the candidate dataset includes: Obtain the authority, time difference, cross-validation rate, and objective correctness score of each element, and calculate the information reliability of the candidate dataset elements: ; The information reliability of the j-th element is: Authority A The time difference is The cross-validation rate is C The objective and correct score is D The first weighting coefficient is The second weighting coefficient is The third weighting coefficient is The fourth weighting coefficient is The attenuation rate is ; Elements with an information reliability score below 0.519 are considered low-reliability data and discarded; elements with an information reliability score greater than 0.519 and less than 0.803 are considered medium-reliability data; and elements with an information reliability score greater than 0.803 are considered high-reliability data. Check if the reliability data contains pressure parameters. If it does, perform pressure range verification. If the pressure exceeds the range, truncate and reduce the weight. If it does not contain pressure parameters, check if it contains physical property parameters. If it does, perform physical property logic verification. If there is a contradiction, mark it as invalid. A reliability dataset will be constructed by verifying physical properties, truncating and reducing the weight of the results, and using high-reliability data.

[0019] In this embodiment, the method for obtaining valid data by resolving conflicts in the fused data based on a multi-source conflict resolution mechanism includes: Data conflict monitoring is performed on the fused data, key values ​​are extracted from the fused data, and the coefficient of variation is calculated when key values ​​from different sources are inconsistent. , The mean of the z-th key value is The standard deviation of the z-th key value is The coefficient of variation is When the coefficient of variation is greater than 0.15 and less than or equal to 0.15, it is judged as a moderate numerical conflict; when the coefficient of variation is greater than 0.3, it is judged as a severe numerical conflict. Spatiotemporal conflict detection is performed on the fused data, and well location coordinates are extracted from the fused data. When well location coordinates from different sources are inconsistent, a conflict is identified, and the standard deviation of the well location coordinates is calculated. , , The latitude of the i-th well location is The longitude of the i-th well location is The mean of the latitudes is The average longitude is The Earth's radius is R, and the semi-sine distance of the i-th well location is... The standard deviation of the coordinates of the i-th well location is The number of well locations is N The average value of the half-versus distance of the well location is ; When the standard deviation of the well location coordinates is greater than 50 meters but less than 500 meters, the well location is rechecked; when the standard deviation of the well location coordinates is greater than 500 meters, a spatial conflict is triggered; the fused data with numerical and spatial conflicts are removed to obtain the first data. The first data from the sensor data source that has a real-time time of less than two minutes and complies with the API 14B safety standard is selected and output as valid data.

[0020] In this embodiment, the method for dynamically fusing the effective data according to a multi-source sensing strategy to obtain fused data includes: The perception dimensions of the multi-source perception strategy include numerical data, spatiotemporal data, and semantic data. The effective data is divided into numerical data, spatiotemporal data, and semantic data through the multi-source perception strategy. The target oilfield is divided into grids, and the data is classified according to the grids. The information reliability, nearest fault distance, and grid permeability of the effective data are obtained. The timeliness compensation, composite attenuation factor, and dynamic weight of the effective data are calculated. , , , The information reliability of the c-th valid data is: The dynamic weight of the c-th valid data is The timeliness compensation for the c-th valid data is The composite attenuation factor of the c-th valid data is The nearest fault distance is The attenuation coefficient is The time difference is Grid penetration rate G The fault distance attenuation coefficient is Penetration rate weighting is Fault weight is The normalization function is ; Based on dynamic weights, the first stage rapidly fuses the effective data to obtain primary fused data. The second stage initiates multiphysics coupling analysis and spatiotemporal consistency verification on the primary fused data. If the spatiotemporal consistency verification result is greater than 0.947, the effective data is fused according to the perception dimension; otherwise, rapid fusion is performed again. The fused result is then output as fused data. The system employs a confidence-weighted average fusion for continuous numerical data, a maximum confidence-first fusion for discrete state data, a knowledge graph semantic aggregation fusion for text reports, and Kalman filtering and confidence-constrained fusion for spatiotemporal trajectories.

[0021] In this embodiment, the method for constructing a multi-source information fusion model for petroleum question-and-answer based on the fused data includes: A reasonable evaluation value is obtained by conducting a rationality assessment on the fused data, and the objective function of the petroleum question-and-answer multi-source information fusion model is obtained by objectively weighting the reasonable evaluation value and the loss function. The multi-source information fusion model for petroleum question answering includes autoencoders, Bayesian networks, and recurrent neural networks; An autoencoder compresses input data into low-dimensional features through an encoder, and a decoder reconstructs the original data. The reconstruction error is then used to perform feature dimensionality reduction and anomaly screening to obtain question-and-answer feature data. Bayesian networks express the conditional dependencies between variables based on probabilistic graphical models. They update posterior probabilities using prior probabilities and observed data, and use question-and-answer feature data and petroleum information sources as control groups. Uncertainty inference and causal association analysis are performed on the control groups to obtain screened data. Recurrent neural networks utilize the temporal propagation mechanism of hidden states and combine it with the objective function to capture and optimize long-term dependencies in the data, and then fuse the selected data based on these long-term dependencies.

[0022] In this embodiment, the method for optimizing the petroleum question-and-answer multi-source information fusion model based on a feedback optimization mechanism includes: By acquiring data freshness, initial confidence, false alarm rate of outdated knowledge, semantic consistency, objective score, structured data weight, unstructured data weight, and timeliness decay factor, the state space of the petroleum industry is reconstructed, and the objective function is given: , , , The weight matrices are Q and B, the transpose of the matrix is ​​T, and the objective function is... The information state matrix is ​​H, a column vector. The first column represents data freshness, the second data reliability, the third semantic consistency, and the fourth objective score. The control matrix is ​​E, a column vector. The first column represents the weight of structured data, the second the weight of unstructured data, and the third the timeliness decay factor. Data reliability is... The initial credibility is The time-varying coefficient at time t is The false alarm rate for outdated knowledge at time t is ; Using real-number encoding, the key parameters of multi-source information fusion are transformed into heritable and optimizeable gene fragments, with each gene corresponding to a parameter and its value range; The candidate solutions of the petroleum question-answering multi-source information fusion model are used as particles, and the population is initialized using a random method and circular mapping, as expressed in the following expression: , The initial population of the (r+1)th dimension is: The chaos value of the r-th dimension The remainder function is The constants less than 1 are w and b, respectively; To obtain the particle's answer accuracy error, response time, and objective satisfaction, given the fitness function, the expression is: , The fitness function of the a-th particle is: The accuracy error of the answer for the a-th particle is The response time of the a-th particle is The objective satisfaction of the a-th particle is The error weight is Time weighting is The satisfaction weight is The objective function weights are ; Calculate the fitness function value of each particle, select the individual with the smallest fitness function value as the best individual, and implement a retention strategy. This involves improving crossover and mutation through adaptive crossover rates, and iteratively updating the population. The expression is: , Where the mutation probability is The fitness function value of the a-th particle is The average fitness function value is Cauchy distribution is The position parameters are The scale parameter is The basic variability rate is , The process continues iterating until the fitness function value reaches its minimum. The particle with the minimum fitness function value is then selected as the optimal combination of parameters.

[0023] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-source information fusion in an intelligent question-answering system, characterized in that, The following steps are involved: The system collects response data from a preset system and petroleum information sources, and preprocesses the response data. The petroleum information sources include structured data, semi-structured data, and unstructured data. The structured data includes oilfield production data, geological exploration data, and equipment monitoring data. The semi-structured data includes technical reports, industry standard documents, and equipment technical manuals. The unstructured data includes academic papers, patent documents, industry news and conference analyses, and case records. Based on the question-and-answer data, a bidirectional transformer model is used to match and filter the petroleum information sources to obtain candidate datasets. Based on the information reliability of the candidate datasets, deterministic optimization is performed to obtain a reliability dataset. The fused data is resolved based on a multi-source conflict elimination mechanism to obtain effective data. The effective data is then dynamically fused according to a multi-source perception strategy to obtain fused data. A multi-source information fusion model for petroleum question answering is then constructed based on the fused data. The petroleum question-and-answer multi-source information fusion model is optimized according to the feedback optimization mechanism. The data to be fused is input into the petroleum question-and-answer multi-source information fusion model, and the fusion result is output.

2. The multi-source information fusion method in an intelligent question-answering system according to claim 1, characterized in that, A method for matching and filtering petroleum information sources using a bidirectional transformer model based on the question-and-answer data includes: A bidirectional transformer model is used to extract knowledge from question-and-answer data and petroleum information sources, respectively, to obtain question-and-answer knowledge data and standard knowledge data. Entity matching is performed on the question-and-answer knowledge data and standard knowledge data, and standard knowledge data with an entity matching degree greater than 0.761 is used as primary candidate content. A graph attention network is used to perform relation matching on the question-and-answer knowledge data and standard knowledge data, and primary candidate content with a relation matching degree greater than 0.519 is used as secondary candidate content. Attribute matching is performed on question-and-answer knowledge data and standard knowledge data. Evidence theory is used to handle fuzzy attributes, and primary candidate content with an attribute matching degree greater than 0.604 is used as tertiary candidate. A neighborhood-adaptive petroleum semantic matching model is used to perform semantic matching on question-and-answer knowledge data and standard knowledge data. A multi-granularity attention mechanism is used to jointly calculate word-level, phrase-level, and sentence-level similarity. Tertiary candidates with a similarity degree greater than 0.785 are output as the screening results, and the petroleum information sources of the screening results are output as candidate datasets. Knowledge extraction includes entity recognition, relation extraction, attribute extraction, and semantic extraction.

3. The multi-source information fusion method in an intelligent question-answering system according to claim 1, characterized in that, A method for obtaining a reliability dataset through deterministic optimization based on the information reliability of the candidate dataset includes: Obtain the authority, time difference, cross-validation rate, and objective correctness score of each element, and calculate the information reliability of the candidate dataset elements: ； The information reliability of the j-th element is: Authority A The time difference is The cross-validation rate is C The objective and correct score is D The first weighting coefficient is The second weighting coefficient is The third weighting coefficient is The fourth weighting coefficient is The attenuation rate is ; Elements with an information reliability score below 0.519 are considered low-reliability data and discarded; elements with an information reliability score greater than 0.519 and less than 0.803 are considered medium-reliability data; and elements with an information reliability score greater than 0.803 are considered high-reliability data. Check if the reliability data contains pressure parameters. If it does, perform pressure range verification. If the pressure exceeds the range, truncate and reduce the weight. If it does not contain pressure parameters, check if it contains physical property parameters. If it does, perform physical property logic verification. If there is a contradiction, mark it as invalid. A reliability dataset will be constructed by verifying physical properties, truncating and reducing the weight of the results, and using high-reliability data.

4. The multi-source information fusion method in an intelligent question-answering system according to claim 1, characterized in that, A method for obtaining effective data by resolving conflicts in the fused data based on a multi-source conflict resolution mechanism includes: Data conflict monitoring is performed on the fused data, key values ​​are extracted from the fused data, and the coefficient of variation is calculated when key values ​​from different sources are inconsistent. ， The mean of the z-th key value is The standard deviation of the z-th key value is The coefficient of variation is When the coefficient of variation is greater than 0.15 and less than or equal to 0.15, it is judged as a moderate numerical conflict; when the coefficient of variation is greater than 0.3, it is judged as a severe numerical conflict. Spatiotemporal conflict detection is performed on the fused data, and well location coordinates are extracted from the fused data. When well location coordinates from different sources are inconsistent, a conflict is identified, and the standard deviation of the well location coordinates is calculated. ， ， The latitude of the i-th well location is The longitude of the i-th well location is The mean of the latitudes is The average longitude is The Earth's radius is R, and the semi-sine distance of the i-th well location is... The standard deviation of the coordinates of the i-th well location is The number of well locations is N The average value of the half-versus distance of the well location is ; When the standard deviation of the well location coordinates is greater than 50 meters but less than 500 meters, the well location is rechecked; when the standard deviation of the well location coordinates is greater than 500 meters, a spatial conflict is triggered; the fused data with numerical and spatial conflicts are removed to obtain the first data. The first data from the sensor data source that has a real-time time of less than two minutes and complies with the API 14B safety standard is selected and output as valid data.

5. The multi-source information fusion method in an intelligent question-answering system according to claim 1, characterized in that, A method for dynamically fusing the effective data according to a multi-source sensing strategy to obtain fused data includes: The perception dimensions of the multi-source perception strategy include numerical data, spatiotemporal data, and semantic data. The effective data is divided into numerical data, spatiotemporal data, and semantic data through the multi-source perception strategy. The target oilfield is divided into grids, and the data is classified according to the grids. The information reliability, nearest fault distance, and grid permeability of the effective data are obtained. The timeliness compensation, composite attenuation factor, and dynamic weight of the effective data are calculated. ， ， ， The information reliability of the c-th valid data is: The dynamic weight of the c-th valid data is The timeliness compensation for the c-th valid data is The composite attenuation factor of the c-th valid data is The nearest fault distance is The attenuation coefficient is The time difference is Grid penetration rate G The fault distance attenuation coefficient is Penetration rate weighting is Fault weight is The normalization function is ; Based on dynamic weights, the first stage rapidly fuses the effective data to obtain primary fused data. The second stage initiates multiphysics coupling analysis and spatiotemporal consistency verification on the primary fused data. If the spatiotemporal consistency verification result is greater than 0.947, the effective data is fused according to the perception dimension; otherwise, rapid fusion is performed again. The fused result is then output as fused data. The system employs a confidence-weighted average fusion for continuous numerical data, a maximum confidence-first fusion for discrete state data, a knowledge graph semantic aggregation fusion for text reports, and Kalman filtering and confidence-constrained fusion for spatiotemporal trajectories.

6. The multi-source information fusion method in an intelligent question-answering system according to claim 1, characterized in that, A method for constructing a multi-source information fusion model for petroleum question-and-answer based on the fused data includes: A reasonable evaluation value is obtained by conducting a rationality assessment on the fused data, and the objective function of the petroleum question-and-answer multi-source information fusion model is obtained by objectively weighting the reasonable evaluation value and the loss function. The multi-source information fusion model for petroleum question answering includes autoencoders, Bayesian networks, and recurrent neural networks; An autoencoder compresses input data into low-dimensional features through an encoder, and a decoder reconstructs the original data. The reconstruction error is then used to perform feature dimensionality reduction and anomaly screening to obtain question-and-answer feature data. Bayesian networks express the conditional dependencies between variables based on probabilistic graphical models. They update posterior probabilities using prior probabilities and observed data, and use question-and-answer feature data and petroleum information sources as control groups. Uncertainty inference and causal association analysis are performed on the control groups to obtain screened data. Recurrent neural networks utilize the temporal propagation mechanism of hidden states and combine it with the objective function to capture and optimize long-term dependencies in the data, and then fuse the selected data based on these long-term dependencies.

7. The multi-source information fusion method in an intelligent question-answering system according to claim 1, characterized in that, The optimization method for the petroleum question-and-answer multi-source information fusion model based on the feedback optimization mechanism includes: By acquiring data freshness, initial confidence, false alarm rate of outdated knowledge, semantic consistency, objective score, structured data weight, unstructured data weight, and timeliness decay factor, the state space of the petroleum industry is reconstructed, and the objective function is given: ， ， ， The weight matrices are Q and B, the transpose of the matrix is ​​T, and the objective function is... The information state matrix is ​​H, a column vector. The first column represents data freshness, the second data reliability, the third semantic consistency, and the fourth objective score. The control matrix is ​​E, a column vector. The first column represents the weight of structured data, the second the weight of unstructured data, and the third the timeliness decay factor. Data reliability is... The initial credibility is The time-varying coefficient at time t is The false alarm rate for outdated knowledge at time t is ; Using real-number encoding, the key parameters of multi-source information fusion are transformed into heritable and optimizeable gene fragments, with each gene corresponding to a parameter and its value range; The candidate solutions of the petroleum question-answering multi-source information fusion model are used as particles, and the population is initialized using a random method and circular mapping, as expressed in the following expression: ， The initial population of the (r+1)th dimension is: The chaos value of the r-th dimension The remainder function is The constants less than 1 are w and b, respectively; To obtain the particle's answer accuracy error, response time, and objective satisfaction, given the fitness function, the expression is: ， The fitness function of the a-th particle is: The accuracy error of the answer for the a-th particle is The response time of the a-th particle is The objective satisfaction of the a-th particle is The error weight is Time weighting is The satisfaction weight is The objective function weights are ; Calculate the fitness function value of each particle, select the individual with the smallest fitness function value as the best individual, and implement a retention strategy. This involves improving crossover and mutation through adaptive crossover rates, and iteratively updating the population. The expression is: ， Where the mutation probability is The fitness function value of the a-th particle is The average fitness function value is Cauchy distribution is The position parameters are The scale parameter is The basic variability rate is , The process continues iterating until the fitness function value reaches its minimum. The particle with the minimum fitness function value is then selected as the optimal combination of parameters.

Images