Information fusion and reasoning method and system based on multi-agent collaborative networking search

Abstract

The invention discloses an information fusion and reasoning method and system based on multi-agent collaborative networking search, and relates to the technical field of network information, and the method comprises the steps: a deep research strategy scheduling agent deconstructs an unstructured request through a task formalization mechanism, generates a high-dimensional instruction code, and dynamically optimizes an execution path. And the network intelligence index generation agent extracts and screens the trusted URL by using a rule engine and a semantic evaluator based on the code, and generates a structured link instruction stream. The structured analysis intelligent agent implements multi-level semantic distillation and conversion on an original document to obtain key value pair representations, and the key value pair representations are packaged into machine readable data. And generating a comprehensive intelligence abstract with consistent logic through the information fusion analysis agent. And the natural language generation engine converts the information into a professional report, and returns the professional report to the scheduling module through a feedback interface to realize cognitive closed-loop iterative optimization. The technical problem that in the prior art, it is difficult for geologists to obtain accurate, comprehensive and credible professional conclusions is solved.

Description

Technical Field

[0001] This invention relates to the field of network information technology, and specifically to an information fusion and reasoning method and system based on multi-agent collaborative network search. Background Technology

[0002] With the rapid development of Large Language Models (LLMs) and agent technology, information retrieval and analysis systems based on natural language interaction have demonstrated strong generalization capabilities in general domains. However, when facing highly specialized and knowledge-intensive research tasks in vertical fields such as geology, single agents exhibit significant limitations in network search and information fusion, including insufficient autonomous planning capabilities, weak internalization of domain knowledge, and a lack of multi-source information collaborative discrimination mechanisms. These limitations may lead to a chain reaction of resource assessment biases and distorted structural interpretations.

[0003] Currently, large language models or agents for network search typically incorporate basic mechanisms such as query construction, result ranking, and preliminary assessment of source reliability to improve the efficiency and relevance of information acquisition. However, these internal mechanisms often rely on static strategies or isolated models. When faced with complex, dynamic, and heterogeneous multi-source information in a network environment, they often fail to globally balance the complementarity and contradictions of information from multiple sources, making it difficult to guarantee the quality, reliability, and completeness of the information obtained.

[0004] To improve the quality and efficiency of intelligent agents in network searches, existing technologies typically rely on knowledge graph-based retrieval enhancement, pre-defined fixed task processes, or tuning mechanisms based on human feedback. However, these methods often exhibit limitations in domain adaptability, dynamic collaborative decision-making, and autonomous evolution capabilities when facing open, dynamic, and heterogeneous network research environments. Therefore, addressing the shortcomings of existing intelligent agents in autonomous planning for complex tasks, deep embedding of domain knowledge, reliable fusion of multi-source information, and dynamic collaborative decision-making, and continuously optimizing their performance through multi-agent collaboration and closed-loop feedback, has become a key technical challenge in overcoming the bottleneck of intelligentization in specialized research. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an information fusion and reasoning method and system based on multi-agent collaborative network search.

[0006] The objective of this invention is achieved through the following technical solution: In a first aspect, this application discloses an information fusion and reasoning method based on multi-agent cooperative network search, comprising the following steps: S1. The deep research strategy scheduling agent deconstructs and parses the input unstructured semantic requests through the task formalization mechanism embedded in the knowledge ontology, generates high-dimensional instruction codes, and performs dynamic replanning and adaptive optimization of the execution path through feedback signals. The high-dimensional instruction codes include subtask-dependent topology, spatiotemporal constraint boundary conditions, and source priority weight vectors. S2. Based on the high-dimensional instruction encoding, the network intelligence index generating agent extracts, aggregates, and filters the credibility of the unified resource locator associated with potential tasks in the open network ecosystem by integrating a patterned rule engine and a context-aware semantic vector approximation evaluator, thereby generating a structured link instruction stream. S3. Based on the link instruction stream, the original network document payload is obtained. The structured parsing agent performs multi-level semantic distillation and structure transformation on it through a domain adaptive parser cluster to obtain a key value pair structure representation that conforms to international standards. After completing integrity verification and unit standardization, it is encapsulated into a machine-readable associated data format. S4. Based on multi-source heterogeneous data in the associated data format, the information fusion and analysis intelligent agent performs ontology alignment, spatiotemporal benchmark unification processing, prior knowledge constraints, and deep association mining and conflict resolution based on probabilistic graphical models to generate a comprehensive intelligence summary with logical consistency. S5. The comprehensive intelligence summary is transformed into a professional narrative report through a natural language generation engine, and then sent back to the deep research strategy scheduling module through a preset feedback application interface to realize the iterative optimization of the cognitive loop and the dynamic adjustment of strategy execution.

[0007] Based on the first aspect, the task formalization mechanism described in step S1 includes: constructing a multi-hop inference graph, assigning execution confidence thresholds based on Bayesian networks to subtask nodes, and then converting instruction encoding into constraint encoding signals, wherein the constraint encoding signals include sparse encoding format and normalized weight sequence.

[0008] Based on the first aspect, the sparse coding format uses a hash mapping table to store the combination relationship between entity type and attribute key-value pairs; the normalized weight coefficient sequence is generated by calculating the product of the node centrality index and the semantic path decay value in the topological network, and the effective constraint conditions are screened through a dynamic weight threshold filtering mechanism, finally forming a parameterized constraint signal with a hierarchical structure.

[0009] Based on the first aspect, the high-dimensional instruction encoding described in step S1 uses a feedback-driven dynamic replanning algorithm to continuously adaptively optimize and iteratively adjust the cognitive path, mapping the original semantic request into a multi-hop reasoning graph with consistent geological process logic.

[0010] Based on the first aspect, the pattern rule engine mentioned in step S2 includes a library of document link structure patterns formally defined by the extended Backus paradigm, and the semantic vector approximation evaluator includes calculating the cosine similarity between the semantic embedding vector of the user query and the semantic embedding vector of the webpage title and meta-description in a high-dimensional feature space, wherein the semantic embedding vector is generated by a deep semantic encoder finely tuned to a large-scale geological corpus.

[0011] Based on the first aspect, in step S3, the domain adaptive parser cluster includes a dedicated rule engine for multiple typical geological data formats, which is triggered in collaboration with a dynamic template matching mechanism and a document object model path learning algorithm. The key value pair structure representation includes spatial location data under a geographic coordinate reference system, a standardized text description of lithological composition based on a stratigraphic lithology terminology classification system, pore structure parameter ranges, permeability order of magnitude, and digital object identifiers. Based on the first aspect, in step S4, the ontology alignment adopts semantic mapping rules based on description logic, and performs entity parsing and attribute fusion under a shared concept hierarchy. The spatiotemporal benchmark unification process includes converting and normalizing time markers and spatial coordinates in various types of data according to the internationally accepted spatiotemporal reference system. The deep association mining and conflict resolution are achieved by combining a hybrid reasoning framework of Bayesian networks and Markov logic networks. The prior knowledge constraints come from the axiom set and rule base in the domain knowledge graph.

[0012] Secondly, this application discloses an information fusion and reasoning system based on multi-agent cooperative network search, which utilizes the aforementioned information fusion and reasoning method based on multi-agent cooperative network search, including: The in-depth research strategy scheduling module, based on the deconstruction and parsing of multi-granularity semantic units, is used to parse user queries into a structured task sequence, perform semantic enhancement and path planning, and generate execution instructions with constraints. The network intelligence index generation module performs intelligent retrieval and credibility filtering in open networks based on execution commands, and outputs a set of information sources; The structured parsing module performs domain-adaptive parsing and structured extraction on the acquired raw network information, transforming it into standardized machine-readable data. The information fusion and analysis module aligns, merges, and infers multi-source structured data to generate logically consistent professional reports and supports closed-loop feedback optimization.

[0013] The beneficial effects of this invention are: 1) The technical solution provided by this invention can effectively solve the problems of information fragmentation, unreliable information sources, and difficulty in fusion of multi-source heterogeneous data faced by multi-agents in complex query tasks by constructing a multi-agent collaborative closed-loop system of "deep research strategy scheduling - network intelligence index generation - structured parsing - information fusion and analysis", thereby realizing intelligent, structured and reliable processing of unstructured semantic requests.

[0014] 2) Through task formalization and multi-hop reasoning graph construction, the system can decompose macro-research tasks into logically clear and well-constrained sub-task sequences, and optimize task scheduling based on Bayesian networks and dynamic weighting mechanisms, thereby improving the accuracy of parsing and execution. The network intelligence index generation module integrates rule matching and semantic understanding mechanisms to achieve intelligent screening and credibility assessment of high-quality information sources in open networks, ensuring information quality from the source. The structured parsing module relies on domain-adaptive parser clusters and ontology mapping mechanisms to achieve automated and standardized extraction of multi-source heterogeneous data, providing machine-readable structured data for subsequent fusion. The information fusion and analysis module achieves deep correlation and conflict resolution of multi-source evidence through ontology alignment, spatiotemporal unification, and probabilistic graphical model reasoning, generating logically consistent and confidence-assessable comprehensive intelligence.

[0015] 3) This application not only significantly improves the automation level and reliability of geological information retrieval and fusion, but also provides a scalable collaborative architecture and closed-loop optimization paradigm for high-quality information intelligent processing in other vertical fields. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the steps of the information fusion and reasoning method based on multi-agent collaborative network search according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an information fusion and reasoning system based on multi-agent collaborative network search according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This application discloses an information fusion and reasoning method based on multi-agent collaborative networked search, which addresses the problems of low credibility and poor completeness of geological research conclusions caused by existing general-purpose agents in complex geological query tasks due to information fragmentation, unreliable information sources, lack of professional constraints, and insufficient multi-source information collaboration capabilities. Through multi-agent task collaborative planning, domain-knowledge-enhanced semantic understanding and retrieval, structured fusion of multi-source heterogeneous information, and a feedback-driven closed-loop optimization mechanism, the method achieves intelligent processing of the entire geological professional query process. A schematic diagram of the steps of the method is shown below. Figure 1 As shown, the specific steps include: S1. The deep research strategy scheduling agent deconstructs and parses the input unstructured semantic requests through the task formalization mechanism embedded in the knowledge ontology, generates high-dimensional instruction codes, and performs dynamic replanning and adaptive optimization of the execution path through feedback signals. The high-dimensional instruction codes include subtask-dependent topology, spatiotemporal constraint boundary conditions, and source priority weight vectors. S2. Based on the high-dimensional instruction encoding, the network intelligence index generating agent extracts, aggregates, and filters the credibility of the unified resource locator associated with potential tasks in the open network ecosystem by integrating a patterned rule engine and a context-aware semantic vector approximation evaluator, thereby generating a structured link instruction stream. S3. Based on the link instruction stream, the original network document payload is obtained. The structured parsing agent performs multi-level semantic distillation and structure transformation on it through a domain adaptive parser cluster to obtain a key value pair structure representation that conforms to international standards. After completing integrity verification and unit standardization, it is encapsulated into a machine-readable associated data format. S4. Based on multi-source heterogeneous data in the associated data format, the information fusion and analysis intelligent agent performs ontology alignment, spatiotemporal benchmark unification processing, prior knowledge constraints, and deep association mining and conflict resolution based on probabilistic graphical models to generate a comprehensive intelligence summary with logical consistency. S5. The comprehensive intelligence summary is transformed into a professional narrative report through a natural language generation engine, and then sent back to the deep research strategy scheduling module through a preset feedback application interface to realize the iterative optimization of the cognitive loop and the dynamic adjustment of strategy execution.

[0019] Specifically, the task formalization mechanism described in step S1 includes: constructing a multi-hop inference graph, assigning execution confidence thresholds based on Bayesian networks to subtask nodes, then converting instruction encoding into constraint encoding signals, forming hierarchical parameter constraints after dynamic threshold filtering, and triggering downstream intelligence collection sequences through bidirectional interfaces; the constraint encoding signals include sparse encoding formats and normalized weight sequences.

[0020] Specifically, the sparse coding format uses a hash mapping table to store the combination relationship between entity type and attribute key-value pairs; the normalized weight coefficient sequence is generated by calculating the product of the node centrality index and the semantic path decay value in the topological network, and then the effective constraint conditions are screened through a dynamic weight threshold filtering mechanism to finally form a parameterized constraint signal with a hierarchical structure.

[0021] Specifically, the high-dimensional instruction encoding described in step S1 uses a feedback-driven dynamic replanning algorithm to continuously adaptively optimize and iteratively adjust the cognitive path, mapping the original semantic request into a multi-hop reasoning graph with consistent geological process logic.

[0022] Specifically, the pattern rule engine mentioned in step S2 includes a library of document link structure patterns formally defined by the extended Backus paradigm, and the semantic vector approximation evaluator includes calculating the cosine similarity between the semantic embedding vector of the user query and the semantic embedding vector of the webpage title and meta-description in a high-dimensional feature space, wherein the semantic embedding vector is generated by a deep semantic encoder fine-tuned by a large-scale geological corpus.

[0023] Specifically, in step S3, the domain adaptive parser cluster includes a dedicated rule engine for multiple typical geological data formats, which is triggered in collaboration with a dynamic template matching mechanism and a document object model path learning algorithm. The key value pair structure representation includes spatial location data under a geographic coordinate reference system, a standardized text description of lithological composition based on a stratigraphic lithology terminology classification system, pore structure parameter ranges, permeability order of magnitude, and digital object identifiers.

[0024] Specifically, in step S4, the ontology alignment adopts semantic mapping rules based on description logic, and performs entity parsing and attribute fusion under a shared concept hierarchy. The spatiotemporal benchmark unification process includes converting and normalizing time markers and spatial coordinates in various types of data according to internationally accepted spatiotemporal reference systems. The deep association mining and conflict resolution are achieved by combining a hybrid reasoning framework of Bayesian networks and Markov logic networks. The prior knowledge constraints are derived from the axiom set and rule base in the domain knowledge graph.

[0025] This application also discloses an information fusion and reasoning system based on multi-agent cooperative network search, which utilizes the aforementioned information fusion and reasoning method based on multi-agent cooperative network search. Its structural diagram is shown below. Figure 2 As shown, it includes: This study delves into the strategy scheduling module, focusing on task-driven cognitive graph construction and probabilistic scheduling principles. It utilizes semantic role labeling and slot-filling mechanisms embedded with domain knowledge ontology to deconstruct unstructured semantic requests, generating atomic-level semantic units. A two-layer graph model integrating Bayesian networks and conditional random fields (CRFs) is employed for strategy inference. The Bayesian network infers execution confidence thresholds for each subtask node based on prior probabilities and evidence, while the CRF integrates multivariate constraints for global optimization, jointly generating multivariate graphs with geological process logic and probabilistic edge weights. A jump inference graph is generated. To transform this graph into schedulable instructions, it is multiplied by the inverse of the key semantic path decay value to generate normalized weight coefficients. The semantic path decay value is dynamically determined by the product of the path hop count and the confidence level passed along the path. A dynamic weight threshold filtering mechanism, combined with real-time feedback, simultaneously considers this weight coefficient and the execution confidence threshold output by the Bayesian network to select a set of high-priority, high-reliability core subtasks, which are encoded into parameterized constraint signals with clear hierarchies. This enables the intelligent generation and scheduling of research strategies from fuzzy semantic requests to precise executable research strategies. The network intelligence index generation module is responsible for intelligently locating the most relevant high-quality information sources for geological research tasks from massive and complex network information. Specifically, it first works in concert with an intelligent engine embedded with professional rules and a content evaluator with deep semantic understanding capabilities: the rule engine quickly filters out potential URLs with structural norms based on a link feature library of geological academic resources, while the semantic evaluator analyzes the matching degree between web page content and user queries in terms of professional meaning. Subsequently, by fusing the evaluation results of the two methods and further analyzing the link authority of these candidate URLs on the Internet and the consistency of their content themes with the core tasks, a comprehensive quantification of the credibility of the information sources is achieved. Finally, the set of high-value URLs selected is transformed into a structured sequence of collection instructions, thereby accurately driving the subsequent structured parsing module. The structured parsing module focuses on domain-intelligent information extraction and standardized reshaping of raw online documents. Specifically, it first activates its domain-adaptive parser cluster, which includes dedicated parsing engines for various typical geological data types (such as research papers, exploration reports, and data tables). Through a combination of dynamic template matching and intelligent document structure analysis, it accurately locates and extracts key information fragments from the documents. Subsequently, the module invokes a structured mapping protocol based on geological domain knowledge ontology to forcibly convert the extracted unstructured text, numerical values, and metadata into a unified key-value pair structure according to international standards and industry specifications. During this process, core geological attributes such as spatial coordinates, standardized lithological descriptions, physical property parameter ranges, and numerical identifiers are accurately separated and assigned values. Finally, all structured data undergoes rigorous logical integrity verification and unit standardization, and is encapsulated into machine-readable standardized data packets, providing high-quality data input that can be directly utilized by downstream information fusion and analysis modules. The information fusion and analysis module generates logically consistent comprehensive intelligence through intelligent fusion and deep reasoning of multi-source heterogeneous structured data. Specifically, it achieves this by: first, using semantic mapping rules based on descriptive logic to align the ontology of multi-source data, enabling unified parsing and fusion of heterogeneous entities and attributes, and normalizing the time stamps and spatial coordinates in the data according to an internationally recognized spatiotemporal reference system; second, employing a hybrid reasoning framework combining Bayesian networks and Markov logic networks, under the constraints of axioms and rules provided by the domain knowledge graph, performing probabilistic association mining and automatic conflict resolution on the aligned data to form an internally consistent data view; finally, based on this view, a natural language generation engine transforms it into an intelligence summary conforming to professional stylistic norms, while key insights from the analysis process are fed back to the strategy scheduling module via a feedback application interface, thus forming a cognitive closed loop driving continuous strategy optimization.

[0026] For example, when a user inputs an unstructured professional query (e.g., "evaluate the fracturability and controlling factors of the Longmaxi Formation shale gas reservoir in the Sichuan Basin"), the multi-agent collaborative process of this invention is immediately initiated. The deep research strategy scheduling module, acting as the system's "central control," first performs deep semantic deconstruction of the query. This module embeds a geological knowledge ontology (e.g., covering relationships between concepts such as lithology, structure, and hydrocarbon accumulation), and can identify the implicit research intent, key geological entities ("Sichuan Basin," "Longmaxi Formation," "shale gas reservoir"), target attributes ("fracturability"), and potential constraints (e.g., depth range, engineering parameters) in the query. The module decomposes the macro-task into a sequence of sub-tasks with clear logical dependencies and spatiotemporal constraints, such as "retrieve regional geological background and tectonic stress field data," "collect rock mechanical parameters and mineral composition data of the target strata," and "collect fracturing cases and effect evaluations under similar geological conditions." Each subtask is assigned an initial execution confidence based on Bayesian network inference, and combined with source priority weights (such as priority academic databases and authoritative institution reports) to generate parameterized high-dimensional instruction codes to accurately guide downstream retrieval behavior.

[0027] For example, after receiving the above instructions, the network intelligence index generation module initiates a targeted search in open networks and geological professional databases (such as CNKI, Wanfang, USGS, and oilfield internal databases). Its implementation relies on a dual-track strategy of "rule matching and semantic understanding": the rule engine performs rapid initial screening based on URL feature libraries from geological literature and data platforms; simultaneously, a semantic model fine-tuned from geological corpora calculates the deep relevance of the query and webpage content within the professional context. Further, a comprehensive credibility assessment is conducted on candidate links, considering factors such as the authority of the source site, the consistency of the content with the geological task theme, and its importance within the professional link network. Finally, a sorted and labeled set of credible links is output, forming a structured intelligence gathering instruction flow.

[0028] For example, the structured parsing module retrieves original documents (such as HTML-formatted exploration reports, PDF-formatted core analysis papers, and Excel-formatted well logging data tables) based on the instruction stream. Its built-in domain-adaptive parser cluster can identify and process document structures and data representations specific to the geological field. For instance, for well logging curves, the parser can extract numerical sequences of curves such as depth, gamma, and resistivity; for lithological description text, it can identify and standardize terms such as "grayish-black siliceous shale interbedded with thin layers of siltstone"; and for data tables, it can accurately associate parameters such as "porosity," "permeability," and "brittleness index" with their respective stratigraphic layers and sample numbers. All extracted information is mapped to a standardized key-value pair structure.

[0029] For example, the information fusion and analysis module receives structured data streams from different sources. First, ontology alignment is performed, such as unifying the "bottom of the Longmaxi Formation" mentioned in different documents into a standard stratigraphic sequence, and uniformly transforming the spatial locations in various coordinate systems. Subsequently, the module uses a probabilistic graphical model (integrating Bayesian networks and Markov logic networks) to perform correlation analysis and conflict resolution of multi-source evidence under the constraints of knowledge rules in fields such as geomechanics and reservoir evaluation.

[0030] For example, when different sources provide differing data on the brittleness index of rocks in the same region, the system performs weighted fusion and optimal inference based on the reliability of the data source, the measurement method, and its consistency with indirect evidence such as mineral composition and sonic transit time. The natural language generation engine automatically organizes the comprehensive intelligence view, with confidence assessment, formed after the above fusion and inference, into a well-structured and terminologically standardized geological report. The report content can cover geological background overview, key parameter analysis, discussion of main controlling factors, potential evaluation, and uncertainty explanation. Key metadata generated throughout the analysis process, such as the source adoption status of each sub-task, data quality evaluation, inference path, and conflict resolution log, is fed back to the deep research strategy scheduling module through a feedback interface. The scheduling module uses this feedback information to dynamically optimize its task decomposition strategy, confidence assessment model, and source weight allocation, thereby achieving a closed-loop learning capability that continuously improves retrieval accuracy and conclusion reliability during iterative execution. This example fully demonstrates the entire process from geological query input to multi-agent collaborative completion of "strategy planning - targeted retrieval - refined analysis - intelligent fusion - report generation - feedback optimization". This invention effectively addresses the challenges of information fragmentation, multi-source heterogeneity, and professional and reliable integration in the field of geological research by constructing this closed-loop framework.

[0031] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An information fusion and reasoning method based on multi-agent collaborative networked search, characterized in that, Includes the following steps: S1. The deep research strategy scheduling agent deconstructs and parses the input unstructured semantic requests through the task formalization mechanism embedded in the knowledge ontology, generates high-dimensional instruction codes, and performs dynamic replanning and adaptive optimization of the execution path through feedback signals. The high-dimensional instruction codes include subtask-dependent topology, spatiotemporal constraint boundary conditions, and source priority weight vectors. S2. Based on the high-dimensional instruction encoding, the network intelligence index generating agent extracts, aggregates, and filters the credibility of the unified resource locator associated with potential tasks in the open network ecosystem by integrating a patterned rule engine and a context-aware semantic vector approximation evaluator, thereby generating a structured link instruction stream. S3. Based on the link instruction stream, the original network document payload is obtained. The structured parsing agent performs multi-level semantic distillation and structure transformation on it through a domain adaptive parser cluster to obtain a key value pair structure representation that conforms to international standards. After completing integrity verification and unit standardization, it is encapsulated into a machine-readable associated data format. S4. Based on multi-source heterogeneous data in the associated data format, the information fusion and analysis intelligent agent performs ontology alignment, spatiotemporal benchmark unification processing, prior knowledge constraints, and deep association mining and conflict resolution based on probabilistic graphical models to generate a comprehensive intelligence summary with logical consistency. S5. The comprehensive intelligence summary is transformed into a professional narrative report through a natural language generation engine, and then sent back to the deep research strategy scheduling module through a preset feedback application interface to realize the iterative optimization of the cognitive loop and the dynamic adjustment of strategy execution.

2. The information fusion and reasoning method based on multi-agent collaborative networked search according to claim 1, characterized in that, The task formalization mechanism described in step S1 includes: constructing a multi-hop inference graph, assigning execution confidence thresholds based on Bayesian networks to subtask nodes, and then converting instruction encoding into constraint encoding signals, wherein the constraint encoding signals include sparse encoding format and normalized weight sequence.

3. The information fusion and reasoning method based on multi-agent cooperative networked search according to claim 2, characterized in that: The sparse coding format uses a hash mapping table to store the combination relationship between entity type and attribute key-value pairs; the normalized weight coefficient sequence is generated by calculating the product of the node centrality index and the semantic path decay value in the topological network, and then the effective constraint conditions are screened through a dynamic weight threshold filtering mechanism to finally form a parameterized constraint signal with a hierarchical structure.

4. The information fusion and reasoning method based on multi-agent cooperative networked search according to claim 2, characterized in that: The high-dimensional instruction encoding described in step S1 uses a feedback-driven dynamic replanning algorithm to continuously adaptively optimize and iteratively adjust the cognitive path, mapping the original semantic request into a multi-hop reasoning graph with consistent geological process logic.

5. The information fusion and reasoning method based on multi-agent collaborative networked search according to claim 1, characterized in that: The pattern rule engine described in step S2 includes a library of document link structure patterns formally defined by the extended Backus paradigm. The semantic vector approximation evaluator includes calculating the cosine similarity between the semantic embedding vector of the user query and the semantic embedding vector of the webpage title and meta-description in a high-dimensional feature space. The semantic embedding vector is generated by a deep semantic encoder fine-tuned by a large-scale geological corpus.

6. The information fusion and reasoning method based on multi-agent cooperative networked search according to claim 1, characterized in that: In step S3, the domain adaptive parser cluster includes a dedicated rule engine for multiple typical geological data formats, which is triggered in collaboration with the document object model path learning algorithm through a dynamic template matching mechanism. The key value pair structure representation includes spatial location data under a geographic coordinate reference system, lithological composition text description standardized according to the stratigraphic lithology terminology classification system, pore structure parameter range, permeability order of magnitude, and digital object identifier.

7. The information fusion and reasoning method based on multi-agent cooperative networked search according to claim 1, characterized in that: In step S4, the ontology alignment adopts semantic mapping rules based on description logic, and performs entity parsing and attribute fusion under a shared concept hierarchy. The spatiotemporal benchmark unification process includes converting and normalizing time markers and spatial coordinates in various types of data according to internationally accepted spatiotemporal reference systems. The deep association mining and conflict resolution are achieved by combining a hybrid reasoning framework of Bayesian networks and Markov logic networks. The prior knowledge constraints are derived from the axiom set and rule base in the domain knowledge graph.

8. An information fusion and reasoning system based on multi-agent collaborative networked search, employing the information fusion and reasoning method based on multi-agent collaborative networked search as described in any one of claims 1-7, characterized in that, include: The in-depth research strategy scheduling module, based on the deconstruction and parsing of multi-granularity semantic units, is used to parse user queries into a structured task sequence, perform semantic enhancement and path planning, and generate execution instructions with constraints. The network intelligence index generation module performs intelligent retrieval and credibility filtering in open networks based on execution commands, and outputs a set of information sources; The structured parsing module performs domain-adaptive parsing and structured extraction on the acquired raw network information, transforming it into standardized machine-readable data. The information fusion and analysis module aligns, merges, and infers multi-source structured data to generate logically consistent professional reports and supports closed-loop feedback optimization.

Images