An AI agent correlation analysis method based on a merger and acquisition restructuring scheme

Abstract

This invention discloses an AI-powered intelligent agent correlation analysis method based on merger and acquisition (M&A) restructuring schemes. This invention relates to the field of AI intelligent agent technology. Addressing the industry challenges of fragmented data, low manual efficiency, hidden correlations, and subjective biases and omissions in M&A due diligence, this invention solves core problems in existing technologies such as difficulty in multi-source data fusion, insufficient correlation mining, static due diligence data, and long report generation cycles. This invention constructs a unified knowledge graph foundation through multi-source data governance, achieves multi-agent pipeline collaborative operation through five types of specialized intelligent agents and the Kafka communication protocol, outputs probabilistic valuation ranges through Monte Carlo simulation, and realizes incremental report updates and version tracking through Merkle tree difference comparison and visual annotation. Ultimately, it achieves full-process automation, standardization, and intelligence in M&A restructuring analysis.

Description

Technical Field

[0001] This invention belongs to the field of AI intelligent agent technology, specifically relating to an AI intelligent agent correlation analysis method based on merger and acquisition restructuring schemes. Background Technology

[0002] Mergers and acquisitions are an important scenario for industrial integration, but their due diligence process places extremely high technical demands on the ability to deeply process multi-source heterogeneous data.

[0003] Currently, M&A due diligence data covers structured, semi-structured, and unstructured formats, scattered across multiple heterogeneous systems. Significant differences exist in terminology, data definitions, and coding rules across different data sources. Existing technologies often employ rule mapping for data standardization, failing to address issues such as semantic shifts, polysemy, and synonymy, resulting in low entity extraction accuracy and significant relationship identification bias. Current analysis solutions typically use a single, general-purpose model to handle the entire process, lacking the ability to professionally optimize for the specific technical characteristics of different stages such as data acquisition, entity extraction, association mining, risk assessment, and decision support. Traditional solutions construct knowledge graphs with a fully static structure, lacking the ability to manage the lifecycle of entities and relationships. When data sources change, a complete reconstruction of the knowledge graph is required, resulting in time-consuming updates, high computational costs, and an inability to adapt to the dynamic changes in data over the long cycle of M&A transactions, leading to delayed analysis results.

[0004] In summary, existing technologies have significant shortcomings in four core technical dimensions: multi-source data semantic fusion, specialized task division, dynamic knowledge graph construction, and full lifecycle incremental iteration. They cannot meet the needs of efficient, accurate, and traceable intelligent analysis in mergers and acquisitions scenarios. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the problems in related technologies, this invention provides an AI intelligent agent correlation analysis method based on merger and acquisition restructuring schemes, thereby overcoming the aforementioned technical problems existing in existing related technologies.

[0007] (II) Technical Solution

[0008] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0009] S1. Collect multi-source heterogeneous raw data from the entire M&A process, temporarily store the collected historical raw data to temporary storage nodes, and divide the raw data into fragments according to data type to obtain fragmented data.

[0010] The sharded data is distributed to different parallel processing nodes, and unified semantic mapping and multi-dimensional data governance are performed in parallel to output a standardized set of knowledge entities and relationships.

[0011] Add timestamps to each knowledge entity and relationship in the standardized set of knowledge entities and relationships to mark their lifecycle, and build a dynamic M&A knowledge graph that supports incremental updates;

[0012] S2. Based on the dynamic M&A knowledge graph, initialize multiple specialized intelligent agents according to business dimensions, establish asynchronous communication channels between intelligent agents and conduct connectivity tests, construct schedulable queues, and obtain the configured communication protocol and scheduling mechanism.

[0013] The configured communication protocol and scheduling mechanism are synchronized to all agents, and a communication handshake test between agents is performed. After the test is passed, the cluster status is marked as ready, and a multi-agent execution environment that can collaboratively execute analysis tasks is obtained.

[0014] S3. In a multi-agent execution environment, based on a dynamic merger and acquisition knowledge graph, the overall analysis task is broken down into multiple independent sub-tasks, and all sub-tasks are sorted in the order of the process.

[0015] The system schedules specialized agents to analyze sub-tasks sequentially and outputs an integrated decision support dataset containing the analysis results of each agent.

[0016] S4. Load the standard template for the due diligence report, fill the standard template with the integrated results from the integrated decision support dataset, and generate a standardized due diligence report;

[0017] This invention addresses industry issues such as fragmented M&A due diligence data, low manual efficiency, insufficient correlation mining, and static, outdated reports. It constructs a dynamic M&A knowledge graph supporting incremental updates through multi-source heterogeneous data collection, unified semantic mapping, and time-series lifecycle management. It initializes five types of specialized intelligent agents, configures communication protocols and scheduling mechanisms, forming a stable multi-agent collaborative execution environment. It conducts correlation analysis, risk assessment, and Monte Carlo valuation, and outputs a standardized decision support dataset through confidence-weighted fusion. Finally, it automatically generates due diligence reports through field mapping, and includes a multi-path change monitoring and incremental update mechanism to achieve dynamic iterative support throughout the entire M&A lifecycle.

[0018] Preferably, step S1 includes the following steps:

[0019] S11. Collect historical multi-source heterogeneous raw data of the entire M&A chain through the distributed data acquisition interface, temporarily store the obtained raw data to temporary storage nodes, and divide it into fragments according to data type to obtain terms of different sources and different definitions after fragmentation.

[0020] S12. By using a multi-granularity ontology embedding model and a context-aware graph neural network, the fragmented terms from different sources and with different definitions are mapped to a unified M&A domain ontology library to obtain a standardized set of knowledge entities and related relationship data.

[0021] S13. Store the standardized knowledge entity and relationship data set into a graph database, mark the effective time and expiration time of each entity and relationship with timestamps, and construct a dynamic merger and acquisition knowledge graph with a node-edge-attribute three-layer storage structure that supports incremental updates.

[0022] The system acquires multi-source heterogeneous data from the entire M&A process in real time through a distributed interface, temporarily stores it in temporary nodes, and then shards it according to structured, semi-structured, and unstructured types. It adopts a multi-granularity ontology embedding model and a context-aware graph neural network to map terms from different sources to a unified M&A domain ontology library, outputting standardized knowledge entities and relationships. It stores the knowledge in a graph database and marks its lifecycle with timestamps to build a three-layer dynamic M&A knowledge graph that supports efficient incremental updates.

[0023] Preferably, step S2 includes the following steps:

[0024] S21. Based on the dynamic M&A knowledge graph, according to the business dimensions of M&A analysis, set up data collection intelligent agents, entity relationship extraction intelligent agents, correlation analysis intelligent agents, risk assessment intelligent agents, and decision support intelligent agents, and allocate computing resources and data access permissions to each intelligent agent.

[0025] S22. Establish an inter-agent communication protocol based on Kafka message queues, define a unified task message format and data transmission format, and perform connectivity tests between agents one by one.

[0026] The connectivity test includes network reachability detection, channel latency measurement, message delivery success rate verification, and abnormal disconnection and reconnection test.

[0027] S23. Construct a schedulable queue, configure a priority scheduling strategy based on data readiness status, agent resource utilization rate and historical execution performance, set the maximum queue depth and timeout discard rules, and obtain the configured communication protocol and scheduling mechanism.

[0028] S24. Synchronize the configured communication protocol and scheduling mechanism to all agents, perform a full-duplex communication handshake test, mark the cluster status as ready after the test is passed, establish a periodic heartbeat packet detection mechanism, detect the status of agents in real time, and form a multi-agent execution environment that can collaboratively execute analysis tasks.

[0029] Based on a dynamic merger and acquisition knowledge graph, five types of specialized intelligent agents are initialized, containerized deployment is adopted and differentiated computing resources are matched, and data access permissions are allocated based on the RBAC permission model; an asynchronous communication protocol based on Kafka is built, a unified message and data format is defined, and connectivity testing is completed; a priority scheduling queue is constructed, and weights are determined through the AHP hierarchical analysis method; after handshake testing and verification, the cluster is marked as ready, and a heartbeat detection and failover mechanism is provided to form a stable and reliable multi-agent collaborative execution environment.

[0030] Preferably, step S3 includes the following steps:

[0031] S31. In a multi-agent execution environment, based on the business logic coupling degree, the overall analysis task is decomposed into data collection subtasks, entity relationship extraction subtasks, association analysis subtasks, risk assessment subtasks, and decision support subtasks. Each specialized agent is scheduled to analyze the subtasks it is responsible for, and the subtask analysis results output by each agent are obtained.

[0032] S32. For the subtask analysis results output by each agent, perform cross-agent result alignment, confidence weighted fusion and logical consistency verification, and output an integrated decision support dataset containing multi-dimensional analysis conclusions.

[0033] Based on business logic, the overall analysis task is broken down into five sub-tasks, and the corresponding agents are scheduled to execute them according to the data flow order. Among them, correlation analysis and risk assessment are carried out in parallel to improve efficiency. Cross-agent alignment is performed on the output results of each agent, and discrepancies are detected by using a confidence score matrix and a dynamic conflict threshold. The logical consistency of the conclusions is verified by combining predefined rules of the knowledge graph. Different processing methods are adopted for non-conflicting items and conflicting items. Finally, an integrated decision support dataset with confidence level and traceability identifier is output.

[0034] Preferably, step S31 includes the following steps:

[0035] S311. First, a data collection subtask is distributed to the data collection agent to pull and clean incremental data from the data source and obtain the first processing result. The first processing result is then sent back and the dynamic M&A knowledge graph is updated to obtain the dynamic M&A knowledge graph after the first update.

[0036] S312. Based on the first updated dynamic M&A knowledge graph, distribute entity relation extraction sub-tasks to the entity relation extraction agent, perform multimodal entity and relation extraction, and obtain the second processing result; send the second processing result back and trigger the dynamic M&A knowledge graph update to obtain the second updated dynamic M&A knowledge graph.

[0037] S313. Based on the dynamic M&A knowledge graph updated for the second time, corresponding sub-tasks are distributed to the association analysis agent and the risk assessment agent in parallel. The association analysis agent performs hidden association mining and collaborative value analysis, and the risk assessment agent performs multi-dimensional risk scanning to obtain the third processing result.

[0038] S314. Distribute the result integration subtask to the decision support agent, integrate the first processing result, the second processing result, and the third processing result to obtain the final integrated result; construct a dynamic estimation model that supports Monte Carlo simulation, perform integrity verification on the final integrated result, and return the analysis results.

[0039] The scheduling agent executes sub-tasks: the data acquisition agent completes incremental data retrieval and cleaning, triggering the first knowledge graph update; the entity relationship extraction agent completes entity and relationship extraction based on a multimodal large model, triggering a second knowledge graph update; the parallel scheduling association analysis agent conducts equity penetration and collaborative value calculation, and the risk assessment agent performs multi-dimensional risk scanning; the decision support agent integrates all results, constructs a Monte Carlo dynamic valuation model and performs integrity verification, outputting full-dimensional analysis results.

[0040] Preferably, step S32 includes the following steps:

[0041] S321. Establish a confidence score matrix for the analysis results of each agent; the matrix elements in the confidence score matrix represent the confidence values ​​of each agent for each dimension of the analysis conclusion; calculate the mean and standard deviation of the confidence values ​​for each dimension of the conclusion.

[0042] S322. Calculate the dynamic conflict detection threshold based on the mean and standard deviation of the confidence values ​​of each conclusion dimension; for each dimension, calculate the maximum difference of the confidence values ​​of all agents; when the maximum difference of the confidence values ​​exceeds the dynamic conflict detection threshold, or when the logical conclusions are mutually exclusive, mark the conclusion dimension as a conflict item; otherwise, determine the conclusion dimension as a non-conflict item.

[0043] The logical conclusion mutual exclusion is a logical mutual exclusion detection performed on the agent's conclusion itself, which combines predefined logical rules of dynamic merger and acquisition knowledge graph. Specifically, it includes: triple rule verification and semantic consistency verification.

[0044] S323. The non-conflicting terms are fused using the weighted arithmetic mean method to obtain the confidence value of the conclusion dimension after the first fusion.

[0045] For conflicting items, manual intervention is initiated to make a final judgment, and a judgment weight correction vector is generated; based on the judgment weight correction vector, all conflicting items are weighted and fused to obtain the confidence value of the conclusion dimension after the second fusion.

[0046] The confidence values ​​of the first and second fused conclusion dimensions together constitute the confidence value of the fused conclusion dimension.

[0047] The weights in the weighted arithmetic average method are dynamically adjusted based on the agent's historical accuracy.

[0048] S324. Add a comprehensive confidence level, a data source traceability identifier, and a decision suggestion label to each confidence value of the fused conclusion dimension to form an integrated decision support dataset;

[0049] A confidence score matrix is ​​constructed to quantify the confidence of each agent's conclusions. The mean and standard deviation are calculated to generate a dynamic conflict detection threshold. Logically mutually exclusive conclusions are identified by combining triple rules, transitivity rules, and semantic consistency checks to distinguish between conflicting and non-conflicting items. Non-conflicting items are fused based on the agent's historical accuracy, while conflicting items are fused after manual final review and weight adjustment. A confidence level, data traceability identifier, and decision suggestion label are added to the final conclusion to form a standardized and auditable integrated decision support dataset.

[0050] Preferably, step S4 includes the following steps:

[0051] S41. Using the integrated decision support dataset as input, load the standard template of the due diligence report; automatically fill the structured analysis data and agent conclusions in the integrated decision support dataset into the corresponding positions of the standard template according to the field mapping rules, and generate a standardized due diligence report.

[0052] S42. Deploy multiple asynchronous change listeners to monitor data source change events, knowledge graph entity update events, and agent analysis conclusion correction events in real time. When a change event that meets the preset trigger conditions is detected, automatically acquire the changed data, perform incremental updates to the knowledge graph and partial index reconstruction, and obtain the incrementally updated knowledge graph.

[0053] S43. Based on the incrementally updated knowledge graph, identify the relevant agents affected by the changes, trigger the agents to perform incremental analysis, and automatically merge the incremental analysis results into the standardized due diligence report.

[0054] Using an integrated decision support dataset as input, a standard due diligence report template is loaded. Structured data and agent conclusions are automatically populated through field mapping rules to generate various types of visualization charts and output a standardized due diligence report. Multiple asynchronous change listeners are deployed to listen for four types of change events and perform deduplication. When a change in triggering conditions is detected, the knowledge graph is automatically updated, triggering incremental analysis of relevant agents and merging the incremental results into the report, thus achieving dynamic iterative support for the entire M&A lifecycle.

[0055] Preferably, step S42 includes the following steps:

[0056] S421. Divide change monitoring events into four categories: data source structured changes, unstructured document content changes, external API push updates, and manual review corrections. Based on the change frequency, real-time requirements, and data format of the four types of change events, adapt different monitoring and deduplication schemes respectively.

[0057] S422. Perform a difference comparison between the incremental data generated by the event change and the original data to generate a change difference matrix that includes newly added entities, deleted relationships, attribute corrections and association reconstruction.

[0058] S423. Based on the change difference matrix, the knowledge graph is incrementally updated by using the transaction batch write method to patch node attributes, dynamically rebuild edge relationships, append time-series version chains and clean up expired data, and perform local index reconstruction to obtain the incrementally updated knowledge graph.

[0059] Change events are categorized into four types, each with a differentiated monitoring frequency and corresponding deduplication strategy. A high-efficiency difference comparison algorithm based on Merkle trees is adopted, which quickly verifies the differences through root hashing and recursively locates them layer by layer, generating a change difference matrix that includes newly added entities, deleted relationships, attribute corrections, and association reconstructions. A transactional batch write method is used to perform node attribute patching, edge relationship reconstruction, time-series version chain appending, and expired data cleanup, achieving incremental updates of the knowledge graph without service interruption.

[0060] Preferably, step S43 includes the following steps:

[0061] S431. Configure semantic version number generation rules for the due diligence report. The version number consists of a major version number, a minor version number, and a revision identifier based on the changed content. Each time it is automatically updated, the corresponding version number is automatically incremented according to the change type.

[0062] S432. Visualize the differences in incremental changes in the main text of the due diligence report, using new highlighting, strikethrough and modification comments, and attach the scope of the changes and the index of related chapters.

[0063] S433. Establish a historical version snapshot library, supporting timestamp, version number and change event identifier, allowing you to go back to any historical version, and providing multi-version parallel comparison view and difference analysis report export function;

[0064] S434. Set a timeliness assessment mechanism for the results of dynamic iterative decision support. When the last update time of the key data source exceeds the preset timeliness threshold, a timeliness warning will be automatically triggered and a reassessment suggestion will be pushed to the user.

[0065] A dynamic update mechanism for the entire report lifecycle is constructed, adopting a three-segment semantic version numbering rule that automatically increments the corresponding version number based on the change type; visual difference annotations are implemented through HTML / CSS, with explanations of the scope of change impact and related chapter navigation indexes, and one-click switching of display modes is supported; a historical version snapshot library is established, supporting multi-dimensional backtracking, parallel comparison of multiple versions, and export of difference analysis reports; tiered timeliness thresholds are set, and data exceeding the time limit automatically triggers warnings and pushes reassessment suggestions.

[0066] Preferably, an AI intelligent agent association analysis system based on merger and acquisition schemes is provided to implement the AI ​​intelligent agent association analysis method based on merger and acquisition schemes. The system includes a data acquisition and graph construction module, an intelligent agent cluster management module, a task scheduling and collaborative analysis module, and a report generation and dynamic iteration module.

[0067] The data acquisition and graph construction module includes multi-source heterogeneous data acquisition, data sharding, semantic mapping, multi-dimensional data governance, and dynamic M&A knowledge graph construction, which is used to output a standardized set of knowledge entities and related relationships;

[0068] The intelligent agent cluster management module includes specialized intelligent agent initialization, asynchronous communication channel construction, connectivity testing, schedulable queue configuration, and cluster state management, which are used to form a multi-agent execution environment that can collaboratively execute analysis tasks;

[0069] The task scheduling and collaborative analysis module includes task decomposition, subtask sorting, intelligent collaborative analysis, analysis result integration and conflict resolution, and is used to output an integrated decision support dataset.

[0070] The report generation and dynamic iteration module includes standard template loading, automatic report filling, change monitoring, incremental update triggering, agent incremental analysis, and report version management, which are used to output dynamic iterative decision support results for the entire M&A lifecycle.

[0071] (III) Beneficial Effects

[0072] The present invention has the following beneficial effects:

[0073] This invention constructs a standardized system for building a dynamic M&A knowledge graph; it achieves accurate extraction of entities and relationships in the M&A field through a multi-granularity ontology embedding algorithm, and is equipped with a timestamp lifecycle management mechanism, which solves the problems of fragmented, static, and delayed updates of traditional due diligence data, and provides a complete, real-time, and traceable data foundation for subsequent correlation analysis.

[0074] This invention breaks away from the rigid model of traditional due diligence that relies on human experience. It innovatively constructs five specialized intelligent agents: data collection, entity extraction, correlation analysis, risk assessment, and decision support. It uses the Kafka communication protocol to achieve message interaction and a task scheduling mechanism with dynamic weights set by the AHP (Analytic Hierarchy Process) to achieve pipeline-style collaborative operation of multiple intelligent agents. Compared with traditional manual due diligence, the analysis efficiency is greatly improved, the labor cost is significantly reduced, and the subjective bias and omissions of manual analysis are avoided.

[0075] This invention achieves standardized output of full-process quantitative analysis and dynamic updates; it outputs valuation ranges through Monte Carlo simulation, and is equipped with a Merkle tree efficient difference comparison algorithm and an HTML / CSS visualization difference annotation mechanism, supporting incremental report updates and version traceability, generating standardized due diligence reports, and solving the industry problems of subjective valuation, version confusion, and difficulty in traceability in traditional due diligence reports.

[0076] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0077] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0078] Figure 1 This is a flowchart illustrating an AI intelligent agent correlation analysis method based on merger and acquisition schemes according to the present invention.

[0079] Figure 2 This is a flowchart of the multi-agent collaborative analysis of the present invention;

[0080] Figure 3 This is a schematic diagram illustrating the data flow / control flow separation of the present invention;

[0081] Figure 4 This is a flowchart illustrating the incremental update and dynamic iteration process of the knowledge graph in this invention.

[0082] Figure 5 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0083] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0084] Please see Figure 1 This invention discloses an AI agent correlation analysis method based on merger and acquisition restructuring schemes, comprising the following steps:

[0085] S1. Collect multi-source heterogeneous raw data from the entire M&A process, temporarily store the collected historical raw data to temporary storage nodes, and divide the raw data into fragments according to data type to obtain fragmented data.

[0086] The sharded data is distributed to different parallel processing nodes, and unified semantic mapping and multi-dimensional data governance are performed in parallel to output a standardized set of knowledge entities and relationships.

[0087] Add timestamps to each knowledge entity and relationship in the standardized set of knowledge entities and relationships to mark their lifecycle, and build a dynamic M&A knowledge graph that supports incremental updates;

[0088] S1 includes the following steps:

[0089] S11. Collect historical multi-source heterogeneous raw data of the entire M&A chain through the distributed data acquisition interface, temporarily store the obtained raw data to temporary storage nodes, and divide it into fragments according to data type to obtain terms of different sources and different definitions after fragmentation.

[0090] Specifically, in embodiment S11, in a corporate merger and acquisition project, a distributed data acquisition cluster is deployed to collect historical multi-source heterogeneous raw data of the entire merger and acquisition chain through a standardized API interface;

[0091] The collection frequency is configured differently based on the data type. For example, business data is updated daily, financial data is updated quarterly, public opinion data is collected in real time, and legal litigation data is updated hourly. All collected raw data is temporarily stored in a Redis temporary storage node, and the data expiration time is set to 72 hours.

[0092] Data is partitioned according to data type, into structured data partitioning, semi-structured data partitioning, and unstructured data partitioning, resulting in a set of terms from different sources and with different definitions after partitioning;

[0093] S12. By using a multi-granularity ontology embedding model and a context-aware graph neural network, the fragmented terms from different sources and with different definitions are mapped to a unified M&A domain ontology library to obtain a standardized set of knowledge entities and related relationship data.

[0094] Specifically, in embodiment S12, the original text after being segmented and the term string are first fed into a multi-granularity ontology embedding model with a three-layer embedding architecture for semantic mapping. The model outputs character vectors from bottom to top through a character-level embedding layer based on a BERT-base-Chinese pre-trained model, extracts lexical-level n-gram features through a lexical-level embedding layer composed of a convolutional neural network (CNN), and outputs sentence-level semantic vectors through a sentence-level embedding layer composed of a bidirectional long short-term memory network (Bi-LSTM). Then, the three-layer vectors are weighted and fused through an attention mechanism, and the attention weights of the three-layer granularity are dynamically adjusted according to the data type. Structured data is mainly embedded at the lexical level, and unstructured text is mainly embedded at the sentence level. Finally, a unified semantic vector with a dimension of 768 is output.

[0095] The multi-granularity ontology embedding model is constructed using a two-stage training approach. The first stage is based on a comparative learning pre-training model using a corpus of millions of financial M&A documents, including annual reports, announcements, and legal documents, to optimize the domain semantic representation capability. The second stage is based on 50,000 manually annotated pairs of synonyms in the M&A domain, using a cross-entropy loss function, 30 training rounds, and a learning rate of 2e-5 to complete supervised fine-tuning.

[0096] The obtained semantic vectors are then combined with the entity candidate set and dependency syntax graph corresponding to the text sentences, and input into a context-aware graph neural network improved based on graph attention network GAT to extract entity relations. The network contains 3 graph attention layers and 1 output layer. Each attention layer integrates the entity's own attribute vector, the global context vector of the sentence and the document to which it belongs, and dynamically calculates the attention weight of the neighboring nodes according to the relation direction and relation type. The first-order and second-order neighboring nodes of the target entity are weighted and aggregated to achieve feature extraction of entity-relation-context ternary linkage. Finally, standardized triples are output. This context-aware graph neural network is based on a manually labeled entity relation dataset of 30,000 entities covering enterprises, people, events and other entities in the M&A field. It uses a negative sampling strategy to generate negative samples and completes 50 rounds of supervised training with cross-entropy loss function and Adam optimizer.

[0097] The processed data is distributed to parallel processing nodes, with each node responsible for the semantic mapping and data governance of a data shard, thereby improving processing efficiency; and outputting a standardized set of knowledge entities and relationships.

[0098] S13. Store standardized knowledge entities and related relationship data sets into a graph database, mark the effective time and expiration time of each entity and relationship with timestamps, and construct a dynamic M&A knowledge graph with a node-edge-attribute three-layer storage structure that supports incremental updates.

[0099] Specifically, in embodiment S13, the Neo4j graph database is used to store a standardized set of knowledge entities and related relationships, and a timestamp is added to each knowledge entity and related relationship to mark its lifecycle;

[0100] Knowledge entity records include: create_time (creation time); valid_from (effective start time); valid_to (effective end time, default is 9999-12-31 indicating current validity, which is currently valid in the time dimension); is_active (validity, which is whether it is enabled in the status dimension); and related relationship records contain the same timestamp field.

[0101] A three-layer storage structure of nodes, edges, and attributes is constructed: the node layer stores core entities; the edge layer stores relationships; and the attribute layer stores detailed attribute data of each entity and relationship. An incremental update mechanism is supported, so when a change in the data source is detected, only the affected nodes, edges, and attributes are updated, without the need to rebuild the entire knowledge graph.

[0102] S2. Based on the dynamic M&A knowledge graph, initialize multiple specialized intelligent agents according to business dimensions, establish asynchronous communication channels between intelligent agents and conduct connectivity tests, construct schedulable queues, and obtain the configured communication protocol and scheduling mechanism.

[0103] The configured communication protocol and scheduling mechanism are synchronized to all agents, and a communication handshake test between agents is performed. After the test is passed, the cluster status is marked as ready, and a multi-agent execution environment that can collaboratively execute analysis tasks is obtained.

[0104] S2 includes the following steps:

[0105] S21. Based on the dynamic M&A knowledge graph, according to the business dimensions of M&A analysis, set up data collection intelligent agents, entity relationship extraction intelligent agents, correlation analysis intelligent agents, risk assessment intelligent agents, and decision support intelligent agents, and allocate computing resources and data access permissions to each intelligent agent.

[0106] Specifically, in embodiment S21, based on a dynamic M&A knowledge graph, specialized intelligent agents are initialized according to the five business dimensions of M&A analysis: a data acquisition agent, responsible for incremental data retrieval, cleaning, and format conversion; an entity relationship extraction agent, responsible for entity recognition, relationship extraction, and attribute completion of unstructured text; an association analysis agent, responsible for hidden association mining, equity penetration analysis, and collaborative value calculation; a risk assessment agent, responsible for multi-dimensional risk scanning; and a decision support agent, responsible for result integration, valuation modeling, and report generation.

[0107] The technical implementation path for each intelligent agent is as follows:

[0108] The data acquisition agent is developed based on Python and Scrapy frameworks, and has built-in multi-source API adaptation module, incremental synchronization algorithm module, and data cleaning rule engine. It adopts producer-consumer pattern to realize parallel acquisition of multiple data sources, and has built-in incremental timestamp recording mechanism to pull only the changed data since the last synchronization.

[0109] The entity relation extraction agent is built based on the above multi-granularity ontology embedding model and context-aware graph neural network, and has a built-in multimodal parsing submodule; it supports parallel extraction of entity relations from batch texts and outputs standardized triples.

[0110] The association analysis agent is developed based on the graph computing engine of the graph database Neo4j. It has built-in depth-first traversal algorithm, equity ratio accumulation algorithm, and collaborative value calculation model. It supports multi-level equity penetration, association path tracing, hidden association identification, and collaborative value quantification calculation tasks.

[0111] The risk assessment agent is developed based on a hybrid architecture of rule engine and machine learning. It has a built-in indicator system library, scoring rule engine, and risk level classification model. It uses the rule engine to determine hard compliance risks and uses the gradient boosting tree model to achieve comprehensive scoring of non-linear risks, outputting multi-dimensional risk scores and levels.

[0112] The decision support agent is built on a result aggregation engine, a Monte Carlo simulation module, and a report generation engine. It is only responsible for the alignment and fusion of results from multiple agents, valuation simulation, and report template filling. It does not have the underlying algorithm capabilities for raw data collection, entity extraction, and association mining.

[0113] Each agent is deployed using Kubernetes containerization, and each agent is allocated independent computing resources: the data acquisition agent is configured with 4 cores and 8G, the entity relationship extraction agent is configured with 8 cores and 16G and GPU, the association analysis agent is configured with 16 cores and 32G, the risk assessment agent is configured with 8 cores and 16G, and the decision support agent is configured with 8 cores and 16G.

[0114] In this solution, all five types of intelligent agents support horizontal scaling with multiple instances. Based on Kubernetes' HPA horizontal Pod auto-scaling mechanism, the number of instances of the same type of intelligent agent is dynamically adjusted according to the backlog of the task queue. The number of instances of a single type of intelligent agent ranges from 1 to 20.

[0115] The resource utilization rate dimension of the priority scheduling strategy applies to multiple running instances of the same type of intelligent agent. When the scheduling center assigns a task to a certain type of intelligent agent, it collects the real-time CPU utilization rate and memory utilization rate of all instances of that type and prioritizes assigning the task to the instance with the lowest resource utilization rate to achieve load balancing among instances of the same type.

[0116] The scaling mechanism is as follows: when the average CPU utilization of the same type of intelligent agent exceeds 70% for 3 consecutive minutes and the number of tasks waiting in the queue exceeds 50, one new instance is automatically added; when the average CPU utilization is below 30% for 10 consecutive minutes, the number of instances is automatically reduced, with a minimum of one instance retained.

[0117] Based on the RBAC permission model, each agent is assigned a minimum necessary data access permission. The data acquisition agent only has read and write permissions to the raw data source, but no knowledge graph reading permission; the entity relation extraction agent only has read permissions to unstructured raw data and write permissions to the knowledge graph, but no access permissions to the association analysis and risk assessment modules; the association analysis agent and the risk assessment agent only have read permissions to the knowledge graph, but no access permissions to the raw data or the results of other agents; the decision support agent only has access permissions to the output results of the other four types of agents, but no access permissions to the raw data source, and does not have the underlying algorithm modules for entity extraction, association mining, and risk scanning.

[0118] The computing resources allocated to each agent are the optimal solution derived from three rounds of stress testing and cost budgeting: Data acquisition agent, I / O intensive, minimum CPU computation; Entity relation extraction agent, GPU intensive, speed increased 20 times after adding GPU; Association analysis agent, CPU and memory intensive, highest performance requirements, equity penetration requires parallel searching of millions of paths, more CPU cores result in faster speed; Risk assessment agent, medium computing density, mainly rule and statistics; Decision support agent, medium computing density, more cores in Monte Carlo simulation result in faster convergence.

[0119] S22. Establish an inter-agent communication protocol based on Kafka message queues, define a unified task message format and data transmission format, and perform connectivity tests between agents one by one.

[0120] The connectivity test includes network reachability detection, channel latency measurement, message delivery success rate verification, and abnormal disconnection and reconnection test.

[0121] Specifically, in embodiment S22, an inter-agent communication protocol based on Kafka 2.8 is established, and a 3-node Kafka cluster is deployed; a unified task message format is defined, using the Protobuf serialization protocol, and the message includes: task_id, a unique identifier for the task; task_type, the task type; sender_id, the sender agent ID; receiver_id, the receiver agent ID; priority, the task priority; payload, the task data; and timestamp, the timestamp.

[0122] Define a unified data transmission format: structured data uses Apache Parquet columnar storage, unstructured data uses JSON format, and large files are transmitted in chunks, each chunk being 1MB.

[0123] Connectivity tests between agents are performed one by one: Network reachability is detected using ICMP ping, with each agent sending 100 ICMP packets to all other agents and checking the reception results. The packet loss rate should be 0%. If the packet loss rate is greater than 0%, it indicates a network problem and must be investigated and repaired. Channel latency is measured using the official Kafka performance testing tool, sending 10,000 test messages and counting the end-to-end latency at the receiving end. The average end-to-end latency should be ≤50ms and the P99 latency should be ≤200ms. Message delivery success rate is tested, with the sending end sending 10,000 messages with unique IDs and the receiving end counting the IDs of the received messages and calculating the loss rate, duplication rate, and out-of-order rate. The loss rate should be 0%, i.e., delivery success rate = 100%, duplication rate <0.1%, and out-of-order rate <0.1%. Abnormal disconnection and reconnection test is performed, simulating a 30-second network interruption followed by network recovery and observing for 3 minutes to see if the connection is automatically restored. The automatic reconnection success rate should be 100%. Interrupted messages are resent after recovery.

[0124] S23. Construct a schedulable queue, configure a priority scheduling strategy based on data readiness status, agent resource utilization rate and historical execution performance, set the maximum queue depth and timeout discard rules, and obtain the configured communication protocol and scheduling mechanism.

[0125] Specifically, in embodiment S23, a Redis-based schedulable queue is constructed, divided into three priority sub-queues: high, medium, and low. A priority scheduling strategy based on the Analytic Hierarchy Process (AHP) is configured, and the weight setting process is as follows:

[0126] The first step was the selection of experts and the design of a scoring questionnaire. Fifteen experts in the field of mergers and acquisitions were selected to score the questionnaire, and the importance of the three scheduling dimensions was compared pairwise using the 1-9 scale.

[0127] The second step involves constructing a judgment matrix and cleaning the data. Independent scores from 15 experts were collected, and outliers (scores deviating from the group mean by more than two standard deviations) were removed, leaving 14 valid questionnaires. The arithmetic mean of these valid questionnaires was used to construct a 3×3 judgment matrix A: A1=[1, 1.5, 1.5], A2=[2 / 3, 1, 1], A3=[2 / 3, 1, 1]. The rows and columns of judgment matrix A correspond to: row 1 and column 1 represent data readiness status, row 2 and column 2 represent agent resource utilization rate, and row 3 and column 3 represent historical execution efficiency. ij This indicates how much more important i is than j;

[0128] The third step is to calculate the largest eigenvalue and eigenvector. The sum-product method is used to calculate the largest eigenvalue and weight vector of the judgment matrix: divide each column element of the judgment matrix by the sum of all elements in that column to obtain a normalized matrix; sum the normalized matrix row-wise to obtain column vectors; normalize the column vectors and round them to obtain the weight vectors. ; Calculate the largest eigenvalue =3.0092;

[0129] The fourth step is consistency testing. A rigorous consistency test is performed to ensure the logical consistency of the judgments, and a consistency index is calculated. Look up the random consistency index table, 3rd order matrix Calculate the consistency ratio ;

[0130] Consistency checks passed, the matrix was deemed logically consistent, and weights were determined: data readiness status weight 0.4, prioritizing tasks with 100% data integrity; agent resource utilization weight 0.3, prioritizing agents with low CPU utilization for task allocation; historical execution performance weight 0.3, prioritizing agents with high historical task completion rates and short average execution times. A dynamic calibration mechanism was established, recalculating the AHP matrix monthly based on actual operational data, with weights dynamically adjusted within ±0.05 to ensure continuous optimization of the scheduling strategy. Actual performance scores for each dimension were assigned, and combined with the weight percentages obtained above, a weighted sum was calculated to obtain the total task priority score. Task priorities were then determined by ranking tasks according to their total priority scores.

[0131] The weighted summation formula is:

[0132] in The data readiness score is calculated as follows: The This represents the number of fields retrieved. This represents the total number of fields required. The resource utilization score for the intelligent agent is calculated as follows: CUR refers to CPU utilization. The historical performance score is calculated as follows: , , The For historical accuracy, For the correct number of historical tasks, This represents the total number of historical tasks. Compared to historical speeds, For average time, Set a standard time limit for this type of task; add a capping mechanism to prevent the score from becoming a huge negative number due to excessively slow speed.

[0133] Set the maximum queue depth to 1000; when the depth is exceeded, the queue will enter a waiting state. Set timeout and discard rules: high-priority tasks have a timeout of 30 minutes, medium-priority tasks have a timeout of 1 hour, and low-priority tasks have a timeout of 2 hours. Timeout tasks will be automatically discarded and logged, triggering an alarm notification. Configure a task retry mechanism; failed tasks will be automatically retried up to 3 times, with the retry interval using an exponential backoff algorithm.

[0134] S24. Synchronize the configured communication protocol and scheduling mechanism to all agents, perform a full-duplex communication handshake test, mark the cluster status as ready after the test passes, establish a periodic heartbeat packet detection mechanism, detect the status of agents in real time, and form a multi-agent execution environment that can collaboratively execute analysis tasks.

[0135] Specifically, in embodiment S24, the configured communication protocol parameters and scheduling strategy rules are synchronized to all agents through the configuration center. After each agent loads the configuration, it sends a configuration confirmation message to the scheduling center and performs a full-duplex communication handshake test. Each agent sends a test message to all other agents, requiring all agents to receive and reply correctly. A concurrent stress test is performed, and 100 analysis tasks are submitted to the system simultaneously. All tasks are required to be correctly scheduled and executed without message loss, deadlock, or resource contention. After all tests are passed, the scheduling center marks the cluster status as ready, and the multi-agent execution environment is completed, ready to receive and process merger and acquisition analysis tasks.

[0136] A periodic heartbeat detection mechanism is established. Each agent sends a heartbeat packet to the scheduling center every 5 seconds, containing CPU status information. If no heartbeat is received for 3 consecutive times, the agent is marked as abnormal, and a failover is automatically triggered. The scheduling center reads the raw data of the abnormal agent's task and packages it to a new agent for re-analysis. At the same time, a warning is sent to the operation and maintenance personnel to perform maintenance.

[0137] S3. In a multi-agent execution environment, based on a dynamic merger and acquisition knowledge graph, the overall analysis task is broken down into multiple independent sub-tasks, and all sub-tasks are sorted in the order of the process.

[0138] The system schedules specialized agents to analyze sub-tasks sequentially and outputs an integrated decision support dataset containing the analysis results of each agent.

[0139] Please see Figure 2 and Figure 3 S3 includes the following steps:

[0140] S31. In a multi-agent execution environment, based on the business logic coupling degree, the overall analysis task is decomposed into data collection subtasks, entity relationship extraction subtasks, association analysis subtasks, risk assessment subtasks, and decision support subtasks. Each specialized agent is scheduled to analyze the subtasks it is responsible for, and the subtask analysis results output by each agent are obtained.

[0141] The specific steps in S31 for scheduling each specialized intelligent agent to analyze the sub-tasks they are responsible for include:

[0142] S311. First, a data collection subtask is distributed to the data collection agent to pull and clean incremental data from the data source and obtain the first processing result. The first processing result is then sent back and the dynamic M&A knowledge graph is updated to obtain the dynamic M&A knowledge graph after the first update.

[0143] Specifically, in embodiment S311, the scheduling center first distributes data collection sub-tasks to the data collection agent, specifying the target enterprise list and data time range; the data collection agent performs incremental data retrieval from the data source, only retrieving changed data after the last synchronization time point; the retrieved data is cleaned: missing value filling; outlier detection, using the 3σ principle to identify outliers; format standardization, unifying date format and monetary units; deduplication, based on the enterprise's unified social credit code; a first processing result is obtained; the first processing result is sent back to the scheduling center, triggering a dynamic M&A knowledge graph update, using a transaction batch write method to obtain the first updated dynamic M&A knowledge graph;

[0144] S312. Based on the first updated dynamic M&A knowledge graph, distribute entity relation extraction sub-tasks to the entity relation extraction agent, perform multimodal entity and relation extraction, and obtain the second processing result; send the second processing result back and trigger the dynamic M&A knowledge graph update to obtain the second updated dynamic M&A knowledge graph.

[0145] Specifically, in embodiment S312, based on the first updated dynamic M&A knowledge graph, entity relationship extraction sub-tasks are distributed to the entity relationship extraction agent. A multimodal large model is used to perform entity and relationship extraction, with optimizations specifically for four modalities: text modality (direct input to the large model); PDF version modality (OCR recognition and layout analysis); table modality (table structure parsing → triple conversion); and signature modality (visual recognition and OCR recognition). A second processing result is obtained. The second processing result is sent back and triggers the dynamic M&A knowledge graph update, performing entity merging, merging different expressions of the same enterprise into the same node, removing duplicate relationships, and completing attributes, resulting in the second updated dynamic M&A knowledge graph.

[0146] S313. Based on the dynamic M&A knowledge graph updated for the second time, corresponding sub-tasks are distributed to the association analysis agent and the risk assessment agent in parallel. The association analysis agent performs hidden association mining and collaborative value analysis, and the risk assessment agent performs multi-dimensional risk scanning to obtain the third processing result.

[0147] Specifically, in embodiment S313, based on the second updated dynamic M&A knowledge graph, corresponding sub-tasks are distributed in parallel to the association analysis agent and the risk assessment agent, and the two agents perform analysis simultaneously; the association analysis agent performs hidden association mining: through the graph traversal algorithm depth-first DFS and equity ratio accumulation, equity penetration analysis is performed, starting from the target company, supporting penetrating 10 layers upward to find shareholders and penetrating 10 layers downward to find subsidiaries, and identifying actual controllers, concerted parties, and related parties;

[0148] Perform synergy value analysis to obtain four-dimensional data of both parties in the merger: customer resources, supply chain, technology, and channels; calculate the overlap and complementarity to obtain the synergy value of each dimension, and add up the synergy value of each dimension to obtain the total synergy value;

[0149] The specific calculation method for the collaborative value of each dimension is as follows:

[0150]

[0151] The To create collaborative value for customers The number of non-overlapping customers for both parties. Average order value For customer conversion rate, we assume here that 15% of the acquired customer will be converted after the acquisition;

[0152]

[0153] The For supply chain collaborative value, For the number of overlapping suppliers, This represents the average purchase amount. To calculate the cost reduction rate of centralized procurement, we assume here that centralized procurement can reduce costs by 8%;

[0154]

[0155] The For the value of technological collaboration, This refers to the number of cross-citations of the patents of both parties. The number of overlapping core technical personnel Valuation of single cross-patent items, The valuation for each core talent is as follows: each cross-patent is valued at 500,000, and each core talent is valued at 1 million.

[0156]

[0157] The For channel synergy value, The number of non-overlapping channels. This represents the average revenue per channel. For the new channel conversion rate, we assume it to be 12%.

[0158] The risk assessment agent performs multi-dimensional risk scanning: financial risk, legal risk, industry risk, and integration risk. Each dimension is subdivided into 5 indicators. Each indicator is scored, and the scores are weighted and averaged to obtain the dimension score. Then, the dimension scores are weighted and averaged to obtain the risk score, using a 100-point scale. Based on the risk score, the risk level is output as low / medium / high. The third processing result includes a correlation analysis report and a risk assessment report.

[0159] Each dimension is further subdivided into 5 indicators, as shown in the table below:

[0160] Table 1. Dimensional Breakdown Indicators and Scoring Rules

[0161]

[0162] The weights of indicators within each dimension and the weights of dimensions among the four dimensions are all determined using the same AHP (Analytic Hierarchy Process) method described in S23. The importance of indicators / dimensions is compared pairwise using the 1-9 scaling method. After constructing a judgment matrix and passing the consistency test, the final weights are calculated.

[0163] S314. Distribute the result integration subtask to the decision support agent, integrate the first processing result, the second processing result, and the third processing result to obtain the final integrated result; construct a dynamic estimation model that supports Monte Carlo simulation, perform integrity verification on the final integrated result, and return the analysis results.

[0164] Specifically, S314 involves: distributing a result integration subtask to the decision support agent, integrating all processing results from S311 to S313, constructing a dynamic estimation model for Monte Carlo simulation, performing 10,000 Monte Carlo simulations to obtain an estimation range, integrating the estimation range with all processing results to obtain the final integration result, performing an integrity check on the final integration result, and sending back the analysis results.

[0165] The Monte Carlo simulation specifically involves using the target company's base year revenue... Base year profit margin Sustainable growth rate Based on the given data, a set of parameters is randomly selected to calculate the revenue and net profit for the next 5 years. Then calculate the annual free cash flow. The simplified formula is = ×0.8, here we assume 80% of net profit is converted into free cash flow, using the discount rate in the extracted parameters. Discounting yields the annual Present value, per year Adding the present values ​​together yields Total present value; then calculate future value. The simplified formula is = ×(1+ ) / ( - ),in For the free cash flow in year 5, then calculate the future value. Discounting Present value = / ,get Present value, the The discount rate is used; finally, a simulated valuation is calculated. Total present value and The present value is obtained by adding the present values ​​together. The equity value is obtained by adding the difference between cash and interest-bearing liabilities to the enterprise value. The simulated valuation is obtained after calculation. The above process is repeated to obtain the valuation range.

[0166] The interest-bearing liabilities and cash mentioned are subject to the actual situation of the target company;

[0167] The integrity verification is divided into five results: A-level, verification passed, automatic pass; B-level, there are errors that do not affect the result, the system automatically corrects them, records the problem, and no manual intervention is required; C-level, there are errors that may affect the result, the process is suspended, a problem list is generated, and it is submitted to staff for review and correction; D-level, the process is terminated immediately, a detailed problem report is generated, and it is submitted to experts for review and correction; E-level, the project is terminated.

[0168] S32. For the sub-task analysis results output by each agent, perform cross-agent result alignment, confidence-weighted fusion and logical consistency verification, and output an integrated decision support dataset containing multi-dimensional analysis conclusions.

[0169] The specific steps of confidence-weighted fusion and logical consistency verification in S32 include:

[0170] S321. Establish a confidence score matrix for the analysis results of each agent; the matrix elements in the confidence score matrix represent the confidence values ​​of each agent for each dimension of the analysis conclusion; calculate the mean and standard deviation of the confidence values ​​for each dimension of the conclusion.

[0171] Specifically, in embodiment S321, a 5×8 confidence score matrix is ​​established, with rows representing 5 types of agents and columns representing 8 analysis dimensions. Each agent outputs its own confidence value for each analysis conclusion dimension, with a value range of [0, 1]. This value is automatically generated by the agent based on the uncertainty of its own analysis process. The mean and standard deviation of the confidence values ​​for each conclusion dimension are calculated. The mean reflects the overall confidence level, and the standard deviation reflects the degree of disagreement among agents.

[0172] S322. Calculate the dynamic conflict detection threshold based on the mean and standard deviation of the confidence values ​​of each conclusion dimension; for each dimension, calculate the maximum difference of the confidence values ​​of all agents; when the maximum difference of the confidence values ​​exceeds the dynamic conflict detection threshold, or when the logical conclusions are mutually exclusive, mark the conclusion dimension as a conflict item; otherwise, determine the conclusion dimension as a non-conflict item.

[0173] The logical conclusion mutual exclusion is a logical mutual exclusion detection performed on the agent's conclusion itself, which combines predefined logical rules of dynamic merger and acquisition knowledge graph. Specifically, it includes: triple rule verification and semantic consistency verification.

[0174] Specifically, in embodiment S322, the dynamic conflict detection threshold is calculated based on the mean μ and standard deviation σ of the confidence values ​​for each conclusion dimension: For each dimension, calculate the maximum difference in confidence values ​​among all agents. ,when > If the logical conclusions are mutually exclusive, mark the conclusion dimension as a conflicting item; otherwise, determine that the conclusion dimension is a non-conflicting item.

[0175] Combining the predefined logical rules of the dynamic M&A knowledge graph with triple rule verification, the agent's conclusion is converted into <entity, relation, attribute> triples of the knowledge graph, and it is verified whether it violates the predefined logical rules of the knowledge graph, such as mutual exclusion rules, transitivity rules, type constraints, and semantic consistency.

[0176] The dynamic conflict detection threshold is based on the statistical principle of normal distribution, where 2σ corresponds to a 95.45% confidence interval. That is, when the difference in confidence among agents exceeds 2σ, it is a low-probability event, indicating that there is a significant disagreement among agents.

[0177] In M&A due diligence scenarios, this threshold balances analytical efficiency and risk control. If a 1σ threshold is used, corresponding to a 68.27% confidence interval, it will trigger a large number of meaningless conflict items, and the cost of manual intervention is too high. If a 3σ threshold is used, corresponding to a 99.73% confidence interval, it will miss a lot of substantive disagreements, leading to an increased risk of bias in the analytical conclusions.

[0178] It also supports custom adjustments based on the project's risk level: high-risk projects can be downgraded to [a lower risk level]. Low-risk projects can be upgraded to ;

[0179] S323. The non-conflicting terms are fused using the weighted arithmetic mean method to obtain the confidence value of the conclusion dimension after the first fusion.

[0180] For conflicting items, manual intervention is initiated to make a final judgment, and a judgment weight correction vector is generated; based on the judgment weight correction vector, all conflicting items are weighted and fused to obtain the confidence value of the conclusion dimension after the second fusion.

[0181] The confidence values ​​of the first and second fused conclusion dimensions together constitute the confidence value of the fused conclusion dimension.

[0182] The weights in the weighted arithmetic average method are dynamically adjusted based on the agent's historical accuracy.

[0183] Specifically, in embodiment S323, the conclusion dimensions of non-conflicting items are fused using a weighted arithmetic mean method to obtain the confidence value of the first fused conclusion dimension; the weights are dynamically adjusted based on the agent's historical accuracy, and the system records the matching between the analysis results of each agent's historical tasks and the final actual results, calculating its historical accuracy. The value ranges from [0, 1]. Agents with better historical performance have higher weights. The fusion formula is: , m represents the number of agents participating in the analysis, and the matrix element c ij This represents the confidence value of the i-th agent for the j-th conclusion dimension. This represents the confidence level after fusion of this dimension;

[0184] When a conflict is detected, the system triggers a manual intervention process, pushing all information for that dimension to a human review expert; the expert gives a final decision, generates a decision weight correction vector, adjusts the weights of each agent in that dimension, and obtains the confidence value of the second fused conclusion dimension.

[0185] The original data sources include annual reports, litigation announcements, and industry reports; the specific content of the conflict includes differences in confidence levels and logical contradictions.

[0186] S324. For each conclusion dimension confidence value in the fused conclusion dimension confidence value, add a comprehensive confidence level, a data source traceability identifier, and a decision suggestion label to form an integrated decision support dataset.

[0187] Specifically, in embodiment S324, based on the confidence value of the fused conclusion dimension, a comprehensive credibility level is added to each fused conclusion: confidence level ≥ 0.92 is high credibility, 0.75 ≤ confidence level < 0.92 is medium credibility, and confidence level < 0.75 is low credibility; a data source traceability identifier is added to record the source information of the conclusion, supporting full-link traceability and meeting audit requirements; decision suggestion labels are automatically generated according to the credibility of the conclusion and the processing process; all fused conclusions, confidence levels, and identifier information are integrated into a structured integrated decision support dataset with standardized data format, which can be directly used for report generation;

[0188] S4. Load the standard template for the due diligence report, fill the standard template with the integration results from the integrated decision support dataset, and generate a standardized due diligence report.

[0189] Please see Figure 4 S4 includes the following steps:

[0190] S41. Using the integrated decision support dataset as input, load the standard template of the due diligence report; automatically fill the structured analysis data and agent conclusions in the integrated decision support dataset into the corresponding positions of the standard template according to the field mapping rules, and generate a standardized due diligence report.

[0191] Specifically, in embodiment S41, the integrated decision support dataset is used as input, the standard template for due diligence report is loaded, more than 1,200 field mapping rules are defined, and the structured analysis data and agent conclusions in the integrated decision support dataset are automatically filled into the corresponding positions of the standard template. The text filling accuracy is ≥99% and the numerical filling accuracy is 100%. Visual charts are automatically generated to generate a standardized due diligence report.

[0192] S42. Deploy multiple asynchronous change listeners to monitor data source change events, knowledge graph entity update events, and agent analysis conclusion correction events in real time. When a change event that meets the preset trigger conditions is detected, automatically acquire the changed data, perform incremental updates to the knowledge graph and partial index reconstruction, and obtain the incrementally updated knowledge graph.

[0193] The specific steps for incremental updating of the knowledge graph in S42 include:

[0194] S421. Divide change monitoring events into four categories: data source structured changes, unstructured document content changes, external API push updates, and manual review corrections. Based on the change frequency, real-time requirements, and data format of the four types of change events, adapt different monitoring and deduplication schemes respectively.

[0195] Specifically, in embodiment S421, an asynchronous change listener is deployed, and the change listening events are divided into four categories: structured changes to the data source, listened to once per hour; unstructured document content changes, listened to once per day; external API push updates, listened to in real time; and manual review and corrections, listened to in real time. Differentiated deduplication strategies are configured for each category: structured changes are deduplicated based on the primary key, unstructured changes are deduplicated based on the document hash value, API pushes are deduplicated based on the message ID, and manual corrections are deduplicated based on the version number. This ensures that important changes are not missed, while avoiding redundant processing that would waste resources, and obtains incremental data generated by event changes.

[0196] S422. Perform a difference comparison between the incremental data generated by the event change and the original data to generate a change difference matrix that includes newly added entities, deleted relationships, attribute corrections and association reconstruction.

[0197] Specifically, in embodiment S422, a difference comparison is performed between the incremental data generated by the event change and the original data in the knowledge graph. An efficient difference comparison algorithm based on Merkle tree is used to generate a change difference matrix. The change difference matrix records the change type, value before change, value after change, change time, and change source of each change, supporting change auditing and rollback.

[0198] The efficient difference comparison algorithm for the Merkle tree is as follows: Construct a Merkle tree based on the knowledge graph, calculate the root hash, for example, starting from the bottom layer, calculate the hashes A, B, C, and D of the leaf nodes; move up one level, concatenate the hash values ​​of every two leaf nodes and then hash them again to get hash 1 and hash 2; move up another level, concatenate hash 1 and hash 2 and then hash them again to get the root hash; then compare the root hashes of the old and new knowledge graph Merkle trees. If they are equal, it means there is no change, and we skip this step; if they are not equal, it means there is a change, and we continue to search downwards, checking two child nodes. If the hashes of the left child nodes are the same, it means there is no problem on the left, and we continue to check the child nodes of the right child node until we find which specific node has changed.

[0199] S423. Based on the change difference matrix, the knowledge graph is incrementally updated by using the transaction batch write method to patch node attributes, dynamically rebuild edge relationships, append time-series version chains and clean up expired data, and perform local index reconstruction to obtain the incrementally updated knowledge graph.

[0200] Specifically, in embodiment S423, based on the change difference matrix, an ACID transaction batch write method is adopted, with a single transaction processing ≥1000 changes; four types of update operations are performed: node attribute patching, which only updates the changed attributes without rewriting the entire node; dynamic reconstruction of edge relationships, which deletes old relationships and creates new relationships to maintain the integrity of the association; appending time-series version chains, which creates a new version record for each change to preserve the historical trajectory; and cleaning up expired data, which cleans up data that has expired and exceeded the retention period of valid_to, to obtain the incrementally updated knowledge graph; during this process, the knowledge graph can still provide query services without service interruption;

[0201] The specific local reconstruction of the index involves designing three types of indexes to support the knowledge graph. All of these are stored in shards according to entity type: an inverted index for entity attributes for full-text retrieval of entity attributes; a relation type index for fast query of associations; and a vector index for semantic similarity retrieval.

[0202] Based on the change difference matrix, extract all changed entity IDs and relationship IDs, match the corresponding entity types and relationship types, and locate the affected index shards; the index shards corresponding to entity types that have not changed remain unchanged.

[0203] The rebuild operation is performed only on the affected index shards. The index entries are regenerated based on the changed entity / relationship data, replacing the original shard data. During the rebuild process, the old shards continue to provide query services. After the new shards are built, the update is completed through atomic switching, and the query service is not interrupted throughout the process.

[0204] After the reconstruction is completed, the consistency between the new index and the knowledge graph data is sampled and verified. If the consistency is met, the update is completed; if the consistency is not met, the old index shards are rolled back and the above process is repeated.

[0205] S43. Based on the incrementally updated knowledge graph, identify the relevant agents affected by the changes, trigger the agents to perform incremental analysis, and automatically merge the incremental analysis results into the standardized due diligence report.

[0206] The due diligence report in S43 also includes a dynamic updating mechanism, the specific steps of which include:

[0207] S431. Configure semantic version number generation rules for the due diligence report. The version number consists of a major version number, a minor version number, and a revision identifier based on the changed content. Each time it is automatically updated, the corresponding version number is automatically incremented according to the change type.

[0208] S432. Visualize the differences in incremental changes in the main text of the due diligence report, using new highlighting, strikethrough and modification comments, and attach the scope of the changes and the index of related chapters.

[0209] S433. Establish a historical version snapshot library, supporting timestamp, version number and change event identifier, allowing you to go back to any historical version, and providing multi-version parallel comparison view and difference analysis report export function;

[0210] S434. Set a timeliness assessment mechanism for the results of dynamic iterative decision support. When the last update time of the key data source exceeds the preset timeliness threshold, a timeliness warning will be automatically triggered and a reassessment suggestion will be pushed to the user.

[0211] Specifically, in embodiment S43, a semantic version number generation rule is configured for the due diligence report, with the format vX.YZ, where X is the major version number, which increments when there are major structural changes; Y is the minor version number, which increments when there are important content changes; and Z is the revision number, which increments when there are minor corrections. Each time there is an automatic update, the corresponding version number is automatically incremented according to the type of change: Y increments for changes in core financial data, Z increments for text corrections, and X increments for changes in report structure. The version number generation rule corresponds one-to-one with the type of change, and the version history is clear and traceable.

[0212] The incremental changes are visually marked in the main body of the due diligence report using HTML / CSS: newly added content is highlighted with a green background, deleted content is marked with a red strikethrough, and modified content is marked with a yellow background and a floating comment box; the scope of the changes is explained in the appendix; and a quick index for related chapters is provided, and the difference markings can be toggled on / off with one click. Users can choose to view the clean version or the difference-marked version of the report.

[0213] Establish a historical version snapshot library, generating a complete PDF snapshot and structured data snapshot for each version, with snapshots stored in the cloud; support querying and tracing by timestamp, version number, and change event identifier, and one-click switching to any historical version; provide a multi-version parallel comparison view, supporting up to 3 versions side by side for comparison, automatically highlighting differences between versions, and supporting the export of difference analysis reports;

[0214] A timeliness assessment mechanism is set up for the results of dynamic iterative decision support, with preset timeliness thresholds: 90 days for financial data, 30 days for business registration data, 15 days for legal litigation data, and 7 days for public opinion data. When the last update time of a key data source exceeds the corresponding timeliness threshold, a timeliness warning is automatically triggered, a yellow warning icon is displayed on the report homepage, and an email / SMS notification is sent to the user, suggesting that the assessment be re-executed. User-defined timeliness thresholds are supported to meet the timeliness requirements of different projects.

[0215] Please see Figure 5 Furthermore, an AI-powered intelligent agent correlation analysis system based on merger and acquisition (M&A) schemes is provided to implement the aforementioned AI-powered intelligent agent correlation analysis method based on M&A schemes; the system includes:

[0216] The data acquisition and graph construction module includes multi-source heterogeneous data acquisition, data sharding, semantic mapping, multi-dimensional data governance, and dynamic M&A knowledge graph construction, which is used to output a standardized set of knowledge entities and related relationships;

[0217] The intelligent agent cluster management module includes specialized intelligent agent initialization, asynchronous communication channel construction, connectivity testing, schedulable queue configuration, and cluster state management, which are used to form a multi-agent execution environment that can collaboratively execute analysis tasks;

[0218] The task scheduling and collaborative analysis module includes task decomposition, subtask sorting, intelligent collaborative analysis, analysis result integration and conflict resolution, and is used to output an integrated decision support dataset.

[0219] The report generation and dynamic iteration module includes standard template loading, automatic report filling, change monitoring, incremental update triggering, agent incremental analysis, and report version management, which are used to output dynamic iterative decision support results throughout the entire M&A lifecycle.

[0220] This invention breaks away from the rigid model of traditional M&A due diligence, which relies on manual labor, is inefficient, and has inconsistent standards. It achieves standardized integration of multi-source heterogeneous data through a dynamic M&A knowledge graph, enables multi-dimensional collaborative analysis through multiple agents, and realizes real-time updates and version management of reports through a dynamic iteration mechanism, thus significantly improving the efficiency, accuracy, and standardization of M&A analysis.

[0221] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0222] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for AI-powered intelligent agent correlation analysis based on merger and acquisition restructuring schemes, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous raw data from the entire M&A process, temporarily store the collected historical raw data to temporary storage nodes, and divide the raw data into fragments according to data type to obtain fragmented data. The sharded data is distributed to different parallel processing nodes, and unified semantic mapping and multi-dimensional data governance are performed in parallel to output a standardized set of knowledge entities and relationships. Add timestamps to each knowledge entity and relationship in the standardized set of knowledge entities and relationships to mark their lifecycle, and build a dynamic M&A knowledge graph that supports incremental updates; S2. Based on the dynamic M&A knowledge graph, initialize multiple specialized intelligent agents according to business dimensions, establish asynchronous communication channels between intelligent agents and conduct connectivity tests, construct schedulable queues, and obtain the configured communication protocol and scheduling mechanism. The configured communication protocol and scheduling mechanism are synchronized to all agents, and a communication handshake test between agents is performed. After the test is passed, the cluster status is marked as ready, and a multi-agent execution environment that can collaboratively execute analysis tasks is obtained. S3. In a multi-agent execution environment, based on a dynamic merger and acquisition knowledge graph, the overall analysis task is broken down into multiple independent sub-tasks, and all sub-tasks are sorted in the order of the process. The system schedules specialized agents to analyze sub-tasks sequentially and outputs an integrated decision support dataset containing the analysis results of each agent. S4. Load the standard template for the due diligence report, fill the standard template with the integration results from the integrated decision support dataset, and generate a standardized due diligence report.

2. The AI ​​intelligent agent correlation analysis method based on merger and acquisition schemes according to claim 1, characterized in that, S1 includes the following steps: S11. Collect historical multi-source heterogeneous raw data of the entire M&A chain through the distributed data acquisition interface, temporarily store the obtained raw data to temporary storage nodes, and divide it into fragments according to data type to obtain terms of different sources and different definitions after fragmentation. S12. By using a multi-granularity ontology embedding model and a context-aware graph neural network, the fragmented terms from different sources and with different definitions are mapped to a unified M&A domain ontology library to obtain a standardized set of knowledge entities and related relationship data. S13. Store standardized knowledge entities and related relationship data sets into a graph database, mark the effective time and expiration time of each entity and relationship with timestamps, and construct a dynamic M&A knowledge graph with a node-edge-attribute three-layer storage structure that supports incremental updates.

3. The AI ​​intelligent agent correlation analysis method based on merger and acquisition schemes according to claim 2, characterized in that, The S2 step specifically includes: S21. Based on the dynamic M&A knowledge graph, according to the business dimensions of M&A analysis, set up data collection intelligent agents, entity relationship extraction intelligent agents, correlation analysis intelligent agents, risk assessment intelligent agents, and decision support intelligent agents, and allocate computing resources and data access permissions to each intelligent agent. S22. Establish an inter-agent communication protocol based on Kafka message queues, define a unified task message format and data transmission format, and perform connectivity tests between agents one by one. The connectivity test includes network reachability detection, channel latency measurement, message delivery success rate verification, and abnormal disconnection and reconnection test. S23. Construct a schedulable queue, configure a priority scheduling strategy based on data readiness status, agent resource utilization rate and historical execution performance, set the maximum queue depth and timeout discard rules, and obtain the configured communication protocol and scheduling mechanism. S24. Synchronize the configured communication protocol and scheduling mechanism to all agents, perform a full-duplex communication handshake test, mark the cluster status as ready after the test passes, establish a periodic heartbeat packet detection mechanism, detect the status of agents in real time, and form a multi-agent execution environment that can collaboratively execute analysis tasks.

4. The AI ​​intelligent agent correlation analysis method based on merger and acquisition schemes according to claim 3, characterized in that, The S3 step specifically includes: S31. In a multi-agent execution environment, based on the business logic coupling degree, the overall analysis task is decomposed into data collection subtasks, entity relationship extraction subtasks, association analysis subtasks, risk assessment subtasks, and decision support subtasks. Each specialized agent is scheduled to analyze the subtasks it is responsible for, and the subtask analysis results output by each agent are obtained. S32. For the sub-task analysis results output by each agent, perform cross-agent result alignment, confidence-weighted fusion and logical consistency verification, and output an integrated decision support dataset containing multi-dimensional analysis conclusions.

5. The AI ​​intelligent agent correlation analysis method based on merger and acquisition schemes according to claim 4, characterized in that, The specific steps in S31 for scheduling each specialized intelligent agent to analyze the sub-tasks they are responsible for include: S311. First, a data collection subtask is distributed to the data collection agent to pull and clean incremental data from the data source and obtain the first processing result. The first processing result is then sent back and the dynamic M&A knowledge graph is updated to obtain the dynamic M&A knowledge graph after the first update. S312. Based on the first updated dynamic M&A knowledge graph, distribute entity relation extraction sub-tasks to the entity relation extraction agent, perform multimodal entity and relation extraction, and obtain the second processing result; send the second processing result back and trigger the dynamic M&A knowledge graph update to obtain the second updated dynamic M&A knowledge graph. S313. Based on the dynamic M&A knowledge graph updated for the second time, corresponding sub-tasks are distributed to the association analysis agent and the risk assessment agent in parallel. The association analysis agent performs hidden association mining and collaborative value analysis, and the risk assessment agent performs multi-dimensional risk scanning to obtain the third processing result. S314. Distribute the result integration subtask to the decision support agent, integrate the first processing result, the second processing result, and the third processing result to obtain the final integrated result; construct a dynamic estimation model that supports Monte Carlo simulation, perform integrity verification on the final integrated result, and return the analysis results.

6. The AI ​​intelligent agent correlation analysis method based on merger and acquisition schemes according to claim 4, characterized in that, The specific steps of confidence-weighted fusion and logical consistency verification in S32 include: S321. Establish a confidence score matrix for the analysis results of each agent; the matrix elements in the confidence score matrix represent the confidence values ​​of each agent for each dimension of the analysis conclusion; calculate the mean and standard deviation of the confidence values ​​for each dimension of the conclusion. S322. Calculate the dynamic conflict detection threshold based on the mean and standard deviation of the confidence values ​​of each conclusion dimension; for each dimension, calculate the maximum difference of the confidence values ​​of all agents; when the maximum difference of the confidence values ​​exceeds the dynamic conflict detection threshold, or when the logical conclusions are mutually exclusive, mark the conclusion dimension as a conflict item; otherwise, determine the conclusion dimension as a non-conflict item. The logical conclusion mutual exclusion is a logical mutual exclusion detection performed on the agent's conclusion itself, which combines predefined logical rules of dynamic merger and acquisition knowledge graph. Specifically, it includes: triple rule verification and semantic consistency verification. S323. The non-conflicting terms are fused using the weighted arithmetic mean method to obtain the confidence value of the conclusion dimension after the first fusion. For conflicting items, manual intervention is initiated to make a final judgment, and a judgment weight correction vector is generated; based on the judgment weight correction vector, all conflicting items are weighted and fused to obtain the confidence value of the conclusion dimension after the second fusion. The confidence values ​​of the first and second fused conclusion dimensions together constitute the confidence value of the fused conclusion dimension. The weights in the weighted arithmetic average method are dynamically adjusted based on the agent's historical accuracy. S324. For each conclusion dimension confidence value in the fused conclusion dimension confidence value, add a comprehensive confidence level, a data source traceability identifier, and a decision suggestion label to form an integrated decision support dataset.

7. The AI ​​intelligent agent correlation analysis method based on merger and acquisition schemes according to claim 6, characterized in that, The S4 step specifically includes: S41. Using the integrated decision support dataset as input, load the standard template of the due diligence report; automatically fill the structured analysis data and agent conclusions in the integrated decision support dataset into the corresponding positions of the standard template according to the field mapping rules, and generate a standardized due diligence report. S42. Deploy multiple asynchronous change listeners to monitor data source change events, knowledge graph entity update events, and agent analysis conclusion correction events in real time. When a change event that meets the preset trigger conditions is detected, automatically acquire the changed data, perform incremental updates to the knowledge graph and partial index reconstruction, and obtain the incrementally updated knowledge graph. S43. Based on the incrementally updated knowledge graph, identify the relevant agents affected by the changes, trigger the agents to perform incremental analysis, and automatically merge the incremental analysis results into the standardized due diligence report.

8. The AI ​​intelligent agent correlation analysis method based on merger and acquisition restructuring schemes according to claim 7, characterized in that, The specific steps in S42 include: S421. Divide change monitoring events into four categories: data source structured changes, unstructured document content changes, external API push updates, and manual review corrections. Based on the change frequency, real-time requirements, and data format of the four types of change events, adapt different monitoring and deduplication schemes respectively. S422. Perform a difference comparison between the incremental data generated by the event change and the original data to generate a change difference matrix that includes newly added entities, deleted relationships, attribute corrections and association reconstruction. S423. Based on the change difference matrix, the knowledge graph is incrementally updated by using the transaction batch write method to patch node attributes, dynamically rebuild edge relationships, append time-series version chains, and clean up expired data, and the index is partially rebuilt to obtain the incrementally updated knowledge graph.

9. The AI ​​intelligent agent correlation analysis method based on merger and acquisition schemes according to claim 7, characterized in that, S43 also includes a dynamic updating mechanism for the due diligence report, the specific steps of which include: S431. Configure semantic version number generation rules for the due diligence report. The version number consists of a major version number, a minor version number, and a revision identifier based on the changed content. Each time it is automatically updated, the corresponding version number is automatically incremented according to the change type. S432. Visualize the differences in incremental changes in the main text of the due diligence report, using new highlighting, strikethrough and modification comments, and attach the scope of the changes and the index of related chapters. S433. Establish a historical version snapshot library, supporting timestamp, version number and change event identifier, allowing you to go back to any historical version, and providing multi-version parallel comparison view and difference analysis report export function; S434. Set a timeliness assessment mechanism for the results of dynamic iterative decision support. When the last update time of the key data source exceeds the preset timeliness threshold, a timeliness warning will be automatically triggered and a reassessment suggestion will be pushed to the user.

10. An AI-powered intelligent agent correlation analysis system based on merger and acquisition schemes, characterized in that, The system is used to perform an AI agent correlation analysis method based on a merger and acquisition scheme according to any one of claims 1 to 9, the system comprising: The data acquisition and graph construction module includes multi-source heterogeneous data acquisition, data sharding, semantic mapping, multi-dimensional data governance, and dynamic M&A knowledge graph construction, which is used to output a standardized set of knowledge entities and relationships; The intelligent agent cluster management module includes specialized intelligent agent initialization, asynchronous communication channel construction, connectivity testing, schedulable queue configuration, and cluster state management, which are used to form a multi-agent execution environment that can collaboratively execute analysis tasks; The task scheduling and collaborative analysis module includes task decomposition, subtask sorting, intelligent collaborative analysis, analysis result integration and conflict resolution, which is used to output an integrated decision support dataset. The report generation and dynamic iteration module includes standard template loading, automatic report population, change monitoring, incremental update triggering, agent incremental analysis, and report version management, which are used to output dynamic iterative decision support results throughout the entire M&A lifecycle.

Images