Multi-agent brain signal autonomous understanding method based on large language model driving

Abstract

The invention discloses a multi-agent brain signal autonomous understanding method based on large language model driving, and the method comprises the steps: constructing a hierarchical cooperation architecture comprising a central supervisor agent and a specialized sub-agent, and carrying out the autonomous understanding of a brain signal on the basis of reducing the technical threshold of brain signal analysis; the problems that a traditional normal form process is rigid and long-time-history complex tasks are difficult to process are solved. According to the method, a central supervisor agent is used for analyzing a natural language intention of a user and dynamically decomposing a task, and a specialized sub-agent is combined with a global sharing state and a context isolation mechanism to execute domain-specific full-link dynamic planning and accurate tool calling; a comprehensive analysis report with cross-domain causal logic is generated by integrating quantitative calculation results and qualitative clinical knowledge through hierarchical resource allocation and retrieval enhancement generation mechanism, then a three-layer difficulty assessment reference system is established to verify framework performance, and finally autonomy, flexibility and clinical interpretability of a brain signal understanding process are achieved.

Description

Technical Field

[0001] This invention belongs to the field of brain-computer interface and artificial intelligence interdisciplinary technology, specifically involving a multi-agent brain signal autonomous understanding method driven by a large language model. Background Technology

[0002] Brain-computer interface (BCI) technology enables direct interaction between the human brain and external devices by decoding brain signals, and has significant application value in fields such as clinical health monitoring, sleep disorder diagnosis, emotion assessment, and cognitive load analysis. Electroencephalography (EEG), as the most mainstream non-invasive brain signal acquisition modality, has become a core data source for key application scenarios such as sleep staging, fatigue monitoring, and emotion recognition due to its high temporal resolution and convenience.

[0003] The paradigm for understanding brain signals has undergone significant evolution. Early research primarily relied on the feature engineering paradigm, which involved manually designing time-frequency domain features and combining them with traditional classifiers for decoding. This method heavily depended on domain expert experience and had limited generalization ability. With the rise of deep learning, the representation decoding paradigm gradually became mainstream. End-to-end models can automatically learn latent representations from raw signals, achieving performance breakthroughs in various brain signal decoding tasks. However, existing paradigms still suffer from two fundamental bottlenecks that severely restrict their large-scale application in real-world scenarios: The existing technology has high technical barriers, which seriously restricts the popularization and clinical translation of brain signal understanding and analysis technology. A complete brain signal understanding and analysis process requires a background in brain signal research, programming skills, and professional knowledge of signal processing. This makes it difficult for non-technical users such as clinicians to operate and apply it effectively. This knowledge barrier makes it difficult to widely promote EEG technology in primary healthcare institutions and public health scenarios, limiting its practical application and social benefits.

[0004] Existing paradigms for understanding brain signals suffer from static design and functional limitations, making it difficult to support complex and time-consuming end-to-end workflows. Current paradigms are mostly built on single tasks and static processes, enabling only isolated signal decoding operations. They cannot achieve continuous, autonomous, and collaborative processing across the entire diagnostic chain, such as data loading, preprocessing, sleep staging, and report generation. They lack the ability to effectively plan and sequentially execute multi-step, highly dependent analysis tasks, thus limiting the applicability and efficiency of brain signal analysis systems in dynamic, long-term application scenarios.

[0005] In recent years, the rapid evolution of large language models has provided a new technological path to overcome the aforementioned bottlenecks. Their powerful intent understanding, task planning, and tool invocation capabilities make it possible to build intelligent agent systems capable of understanding natural language and autonomously executing complex processes. Some work has attempted to introduce large language models into the field of brain-computer interfaces, but current research mostly focuses on cross-modal translation tasks, such as decoding EEG signals into text, images, or videos. Essentially, this treats the model as a passive "decoder" rather than an "intelligent agent" with autonomous decision-making and planning capabilities. Although there have been preliminary explorations of using intelligent agents to achieve basic event detection and report generation, their workflows are usually predefined patterns, lacking flexibility and the ability to decompose complex user intents layer by layer, as well as cross-domain task collaboration mechanisms. Crucially, a systematic evaluation benchmark has not yet been established in this field, making it difficult to quantitatively evaluate and compare the reliability, stability, and execution efficiency of intelligent agent systems at different complexity levels, from atomic operations and sequential reasoning to cross-domain intent understanding.

[0006] Furthermore, at the system architecture level, single-agent models are easily interfered with by irrelevant tool descriptions when faced with an ever-growing tool library, resulting in distraction and tool illusion, causing the execution path to deviate from the correct path; while multi-agent collaboration has shown advantages in fields such as chemistry and biology, in the field of brain signal analysis, an effective professional division of labor and resource isolation mechanism has not yet been established, making it difficult to balance execution efficiency and system robustness.

[0007] In summary, existing brain signal analysis technologies still have significant shortcomings in terms of accessibility, flexibility and reliability: (1) high barriers to entry and limited technology adoption; (2) rigid analysis process that cannot adapt to open-ended long-term tasks; and (3) lack of a hierarchical and systematic evaluation system that makes it difficult to verify the comprehensive performance of intelligent agents in real-world scenarios.

[0008] Therefore, there is an urgent need in this field for a new technical architecture that can dynamically transform natural language intent into an end-to-end executable process and is equipped with a comprehensive evaluation benchmark, in order to promote the development of brain signal analysis towards autonomy and clinical interpretability. Summary of the Invention

[0009] In view of the above, this invention provides a multi-agent autonomous understanding method for brain signals based on a large language model. By designing a hierarchical multi-agent architecture, a central supervisor agent parses the user's natural language intent and dynamically decomposes the task, which is then handed over to specialized sub-agents to perform domain analysis by calling their dedicated toolsets and knowledge bases. Data flow is realized by relying on a globally shared state, and the reliability of the framework is verified by a three-layer difficulty assessment benchmark system (atomic ability level, sequential reasoning level, and collaborative reasoning level), thereby achieving autonomous brain signal analysis and clinical interpretability.

[0010] A multi-agent brain signal autonomous understanding method based on a large language model includes the following steps: (1) Construct a hierarchical multi-agent architecture that includes a central supervisory agent, multiple specialized sub-agents, and a globally shared state; (2) The central regulator intelligent agent receives the user's natural language query, analyzes the complexity of the user's intent, and decomposes the complex query into multiple sub-tasks; (3) The central supervisory agent distributes sub-tasks to the corresponding specialized sub-agents according to the domain attributes. The specialized sub-agents adopt a context isolation strategy when executing sub-tasks. (4) Upon receiving the subtask instruction, the specialized sub-agent combines the current system state and workflow context to perform logical reasoning and planning to generate the tool execution sequence and tool call instructions; (5) The specialized sub-agents, following the generated tool execution sequence, sequentially call the corresponding tools through the system interface to perform actual brain signal analysis operations and manage the input and output flow of data; (6) Specialized sub-agents trigger the retrieval enhancement mechanism, retrieve relevant knowledge from the general knowledge base and domain-specific knowledge base according to the sub-task, combine the tool execution output results, generate specialized domain sub-reports and return them to the central supervisory agent; (7) The central supervisory agent aggregates the domain sub-reports returned by each sub-agent, performs cross-modal logical reasoning and causal relationship analysis, and generates a comprehensive analysis report to be fed back to the user.

[0011] Furthermore, in step (1), the central supervisor agent is responsible for parsing the user's natural language intent and distributing tasks. Multiple specialized sub-agents include a sleep analysis agent and an emotion analysis agent. The sleep analysis agent is used for specialized sleep analysis tasks, including sleep staging, micro-event detection, sleep structure assessment, and clinical sleep report generation. The emotion analysis agent is used for specialized emotion analysis tasks, including emotion state recognition, psychological fatigue monitoring, cognitive load assessment, and emotion stability analysis. The global shared state is used to manage the brain signal data input by the user, intermediate processing results, and tool execution parameters, realizing data flow and state synchronization between agents. The global shared state is implemented using a key-value dictionary structure. When all tools are executed, they read the input parameters from the global shared state and write the output results back, ensuring that state synchronization can be achieved without explicit communication between agents.

[0012] Furthermore, the specialized sub-agent configuration hierarchical resource access mechanism divides tools into general toolsets and domain-specific toolsets, and knowledge bases into general knowledge bases and domain-specific knowledge bases. This allows sub-agents to access only the subset of resources strongly relevant to their tasks, reducing the proportion of irrelevant tool descriptions in the context window, reducing computational overhead, and supporting sub-agents to perform operations including data loading, preprocessing, feature extraction, and domain analysis. Resource isolation effectively constrains the tool invocation space of sub-agents, avoiding attention distraction and tool illusion problems caused by excessive tool space in complex tasks. The modular design supports seamless expansion of new analysis domains by registering new toolsets and knowledge bases without modifying the core framework logic, achieving "plug and play".

[0013] Furthermore, in step (2), the central supervisory agent first determines the semantic complexity of the query by combining historical conversations, performs user intent parsing through thought chain reasoning, and determines whether to call specialized sub-agents based on the semantic complexity of the query: for simple greetings or ability inquiries, a response is generated directly from the globally shared state; for atomic operations, only a single explicit instruction is distributed; for complex tasks involving high-level semantics, they are decomposed into a sequence of sub-tasks with logical dependencies based on causal reasoning logic. In this process, the supervisory agent adopts a strict tool-agnostic strategy, does not directly access any analysis tools or knowledge bases, and only performs task decomposition and routing decisions based on language reasoning.

[0014] Furthermore, the context isolation strategy in step (3) is as follows: specialized sub-agents only receive specific task instructions distributed by the central supervisor agent and relevant execution result summaries generated in the global shared state after the preceding sub-agents perform tasks, while shielding irrelevant global dialogue history to ensure that the sub-agents focus on performing sub-task domain-specific analysis.

[0015] Furthermore, the specific implementation process of step (4) is as follows: S41. Before initiating tool planning, the specialized sub-agent prioritizes executing the direct response decision logic. It evaluates whether the existing information is sufficient to directly respond to the sub-task instruction by retrieving the global shared state and the execution results of the preceding sub-agent. If the existing information is sufficient, the planning is terminated immediately, the text response containing the reasoning process is directly output, and the execution plan is set to empty. If the existing information is insufficient to directly answer, the subsequent task decomposition and planning stage is entered. S42. When it is necessary to call a tool, the specialized sub-agent breaks down the high-level instructions into a series of indivisible atomic sub-targets, and then determines the logical execution order of each sub-target according to the data flow dependency. Finally, a complete tool execution sequence from the start point to the end point is generated in a single response to ensure the continuity and efficiency of the execution process. S43. When constructing specific tool invocation instructions, specialized sub-agents follow the following parameter configuration rules: if parameters are explicitly specified in the user instruction, they are forcibly mapped to the corresponding tool parameters; for non-critical parameters that are not explicitly specified, the default values ​​defined by the tool or an empty dictionary are automatically used, while the parameter key names are forcibly verified to ensure that they strictly match the variable names in the globally shared state, thus ensuring the correctness of the parameter reading path for tool execution.

[0016] Furthermore, in step (5), the tool called strictly performs corresponding encapsulation processing operations on brain signal data according to the algorithm logic and process steps predefined in its functional description; the tool reads input parameters from the global shared state, and after execution, writes the output results back to the global shared state, thereby updating the current context information of the system for subsequent other tools or sub-intelligent agents to read, so as to achieve seamless synchronization and zero-copy flow of data streams.

[0017] Furthermore, the retrieval enhancement mechanism in step (6) first performs coarse-grained retrieval in the general knowledge base and the domain-specific knowledge base through a dual encoder model, then performs fine-grained reordering using a cross encoder model to select highly relevant physiological definitions or clinical guidelines as domain-specific knowledge; finally, the quantitative output results of the tool are combined with the retrieved domain-specific knowledge to generate an interpretable specialized domain sub-report and return it to the central regulatory agent.

[0018] Furthermore, in step (7), the central supervisory agent receives and parses the domain sub-reports returned by each sub-agent, extracts key physiological indicators, quantitative features, and domain-specific conclusions; then performs deep semantic reasoning to identify the inherent logical connections and physiological mechanism links between the analysis results of different domains, thereby constructing a multi-dimensional view of physiological state. Finally, the central supervisory agent integrates the discrete data insights into a coherent natural language narrative, and ultimately generates a comprehensive analysis report with internal logical coherence to be fed back to the user.

[0019] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the above-described method for autonomous understanding of multi-agent brain signals driven by a large language model.

[0020] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for autonomous understanding of multi-agent brain signals driven by a large language model.

[0021] This invention, by constructing a hierarchical collaborative architecture and hierarchical resource allocation mechanism that includes a "central regulator-specialized sub-intelligent agents," successfully transforms abstract natural language intent into a precise long-term, end-to-end brain signal analysis workflow. This not only solves the problem of rigid processes in traditional paradigms but also significantly reduces the operational threshold for brain science research and clinical applications by shielding the complexity of underlying technologies, thus realizing the democratization and popularization of brain signal understanding technology. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the overall process of the multi-agent brain signal autonomous understanding method of the present invention.

[0023] Figure 2 is a schematic diagram of the overall framework of the multi-agent brain signal autonomous understanding method of the present invention. Detailed Implementation

[0024] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] like Figure 1 As shown, this embodiment provides a multi-agent brain signal autonomous understanding method based on a large language model, including the following steps: (1) System Initialization and Architecture Construction: A hierarchical multi-agent system framework based on a large language model is constructed. This framework includes a central supervisor agent, multiple specialized sub-agents, and a globally shared state. The central supervisor agent, as the top-level scheduling core, is responsible for analyzing user needs, scheduling sub-agents, and outputting the final response report. Specialized sub-agents, based on the specific needs of brain signal analysis, are responsible for domain-specific analysis and return sub-reports to the supervisor agent. The globally shared state enables agents to share data without explicit communication. The specific implementation process of this step is as follows: S11. Configure the supervisor agent: Grant it global session history access permissions and configure it in "tool-agnostic" mode, that is, it does not directly call the underlying signal processing tools, but is only responsible for logical reasoning, task decomposition and routing distribution.

[0026] S12. Configure specialized sub-agents: Based on the domain requirements of brain signal analysis, initialize, but not limited to, sleep analysis agents and emotion analysis agents; allocate hierarchical resources to each sub-agent, including general toolsets (such as data loading and preprocessing), domain-specific toolsets (such as sleep staging models and emotion index calculations), general knowledge bases, and domain-specific knowledge bases.

[0027] S13. Global Shared State Construction: Initialize a global dictionary structure It is used to store the original brain signal file path, metadata, and intermediate processing products (such as filtered signals and feature matrices) to ensure data consistency and zero-copy transfer between intelligent agents.

[0028] (2) User Intent Parsing and Task Decomposition: The central supervisory agent receives the user's natural language query (e.g., "Analyze whether the patient's sleep quality is affected by emotions") and related brain signal data files. The supervisory agent first judges the complexity of the intent based on historical conversations, performs user intent parsing through thought chain reasoning, and decomposes the task into a sequence of sub-tasks with logical dependencies (e.g., "First perform sleep stage analysis, then perform emotional stability assessment"). The specific implementation process of this step is as follows: S21. Atomic Operation Recognition: If the user intent is a single, explicit instruction (such as "load data" or "simple visualization request"), the supervisory agent recognizes it as an atomic task.

[0029] S22. Complex Task Planning: If the user intent involves high-level semantics, the supervisory agent, based on causal reasoning logic, decomposes the query instruction into a sequence of sub-tasks with logical dependencies. .

[0030] During this process, the supervisory agent checks whether there is already an intermediate result in the globally shared state that satisfies the query. If so, it generates a response directly; otherwise, it generates a subtask queue.

[0031] (3) Task Distribution and Context Isolation: The supervisor distributes the decomposed subtask sequence { based on the domain attributes of the subtasks. Distribute to and assign to the corresponding sub-agents { During this process, a context isolation strategy is implemented, whereby the system blocks irrelevant global dialogue history and only transmits specific task instructions and necessary context summaries to the sub-agents, preventing the large model's attention from being distracted and allowing the sub-agents to focus on the correct execution of sub-tasks.

[0032] When sub-agents execute tasks, a context isolation strategy is implemented. Specifically, during task distribution, the system masks the global dialogue history and only transmits specific task instructions and summaries of relevant execution results generated in the shared state after the preceding sub-agents have executed their tasks. This aims to eliminate irrelevant contextual interference, prevent the large model's attention from being distracted, and ensure that sub-agents focus on accurate execution within their domain while considering the execution results of preceding sub-agents to plan their execution strategies.

[0033] (4) Sub-agent Reasoning and Dynamic Programming: Upon receiving the task instruction, the specialized sub-agent combines the current system state (such as whether data has been loaded, whether the sampling rate meets the requirements) with the workflow context to perform logical reasoning. It determines whether to invoke the tool through the execution decision protocol, and autonomously generates a complete tool execution sequence containing precise parameter configurations based on the reasoning when necessary. The specific implementation process of this step is as follows: S41. Rapid State Check: Before initiating tool planning, the sub-agent prioritizes executing the direct response decision logic. It evaluates whether the existing information is sufficient to directly respond to the task instruction by retrieving the globally shared state and the execution results of the preceding sub-agent. If the existing information is sufficient (e.g., querying the sampling rate of the loaded data), the sub-agent immediately terminates the planning, directly outputs a text response containing the reasoning process, and sets the execution plan to null. Only when the existing information is insufficient to directly answer the question will it proceed to the subsequent task decomposition and planning stage.

[0034] S42. End-to-End Task Planning: When a tool needs to be invoked, the sub-agent first performs atomization decomposition, breaking down high-level instructions into a series of indivisible atomic sub-targets, then performs dependency sorting, and determines the logical execution order of each sub-target based on the data flow dependency relationship (such as "load first, then preprocess, then analyze"). Finally, the end-to-end task plan is generated in one go: the sub-agent is configured to generate a complete tool invocation sequence from start to finish in a single response to ensure the continuity and efficiency of the execution process.

[0035] S43. Parameter Mapping and Consistency Verification: When constructing specific tool invocation instructions, the sub-agent follows the following parameter configuration rules: If the user instruction explicitly specifies a parameter (such as "filter 50Hz interference"), it is forcibly mapped to the corresponding tool parameter (such as notch filter frequency 50Hz); for non-critical parameters that are not explicitly specified, the default value or empty dictionary defined by the tool is automatically used, and the parameter key name is forcibly verified to ensure that it strictly matches the variable name in the globally shared state, thus ensuring the correctness of the parameter reading path for tool execution.

[0036] (5) Tool Execution and State Synchronization: The sub-agent sequentially calls the corresponding tools (functions) through the system interface according to the generated tool execution sequence to perform the actual signal analysis operations and manage the input and output flow of data; the input parameters of all tools are read from the global shared state, and the output results (such as the feature matrix) are written back to the shared state in real time, realizing zero-copy data flow. The specific implementation process of this step is as follows: S51. Execute encapsulation logic: The called tool (function) strictly follows the predefined algorithm logic and process steps in its function description to perform corresponding encapsulation processing operations on the brain signal data.

[0037] S52. State Closed-Loop Update: After each tool is executed, the output products (including but not limited to numerical results, feature matrices, visualization objects or file paths) are automatically written back to the global shared state, thereby updating the current context information of the system for subsequent tools or agents to read, and achieving seamless synchronization of data flow.

[0038] (6) Retrieval Enhancement and Sub-report Generation: After the tool is executed, the sub-agent triggers the retrieval enhancement mechanism, queries the general knowledge base and the domain-specific knowledge base according to the received sub-task, retrieves the relevant knowledge, and generates a professional domain sub-report based on the tool execution results and returns it to the central regulator agent.

[0039] In this implementation, the sub-agent triggers a retrieval enhancement generation mechanism. First, a coarse-grained retrieval is performed in a general knowledge base and a domain-specific knowledge base using a dual encoder model to obtain Top-N relevant documents. Then, a fine-grained re-ranking is performed using a cross-encoder model to select and return Top-K highly relevant clinical guidelines or physiological definitions. Subsequently, the sub-agent combines the quantitative results of the tool execution with the retrieved domain-specific knowledge to generate an interpretable domain sub-report, which is then returned to the regulator.

[0040] (7) Comprehensive analysis and final response: The central supervisory agent receives the sub-reports returned by all activated sub-agents, constructs a global view through cross-modal logical reasoning, generates a comprehensive analysis report with internal logical coherence, and feeds it back to the user.

[0041] The supervisor agent receives and parses the set of sub-reports returned by all activated sub-agents. The system extracts key physiological indicators, quantitative features, and domain-specific conclusions. Then, it performs deep semantic reasoning to identify the inherent logical connections and physiological mechanisms between the analysis results of different domains (such as anxiety causing difficulty falling asleep), thereby constructing a multi-dimensional view of physiological state. Finally, the regulatory agent integrates the discrete data insights into a coherent natural language narrative, ultimately generating a comprehensive analysis report containing data insights and correlation analysis to be fed back to the user.

[0042] like Figure 2As shown, the overall system framework of the method of this invention presents a rigorous hierarchical collaborative structure: the top layer is the central supervisory agent, which serves as the decision-making hub of the system and is responsible for parsing the natural language input and historical context from the user side, dynamically decomposing complex analytical intentions into logically related sub-task flows; the middle layer is a cluster of specialized sub-agents, where each sub-agent, when executing tasks, accurately calls the toolsets and knowledge bases of the general layer and the domain-specific layer through a "hierarchical resource allocation mechanism," and achieves data reuse and implicit collaboration between agents based on a globally shared state; the bottom layer is the response synthesis module, which collects and aggregates the domain analysis reports generated by each sub-agent through the supervisory agent, and finally outputs a comprehensive system response with clinical interpretability through cross-modal logical reasoning and causal analysis.

[0043] To quantify and verify the performance and reliability of the method in real-world brain signal analysis scenarios, we constructed a hierarchical difficulty assessment benchmark system. This benchmark covers three capability dimensions, from basic operation verification to complex intent understanding: Atomic Capability Level (L1), which focuses on evaluating the accuracy of tool invocation under a single explicit instruction and the agent's ability to execute explicit single-step instructions; Sequential Reasoning Level (L2), which focuses on examining the agent's ability to plan and execute multi-step workflows with strict logical dependencies within a single domain, involving long-term, highly dependent workflows; and Collaborative Reasoning Level (L3), which is used to test the system's comprehensive intelligence level in processing fuzzy high-level user intents, coordinating multi-agent collaboration, and synthesizing cross-domain analysis results. Based on this assessment benchmark, we conducted extensive comparative experiments on various mainstream large language model bases. The experimental results are shown in Table 1. Table 1

[0044] In the table, TCR represents task completion rate, R-ACC represents routing accuracy, TCA represents tool invocation accuracy, and Lat represents average response latency. The base models selected for the experiment cover representative models from different vendors and with different parameter scales. The Qwen series, including Qwen-Flash (lightweight version), Qwen-Plus (balanced version), and Qwen-Max (high-performance version), are all large language models developed by Alibaba.

[0045] GPT-4o is a flagship multimodal large model developed by OpenAI.

[0046] The Qwen3 series includes Qwen3-8b (8 billion parameters version) and Qwen3-30B-A3B (hybrid expert architecture version).

[0047] Llama 3.3 70B is a high-performance, open-source large model developed by Meta.

[0048] DeepSeek-v3.2 is a high-efficiency large model developed by DeepSeek Inc.

[0049] The experimental data in Table 1 leads to the following conclusions: First, as the task difficulty level increases from L1 to L3, the average response latency of all base models shows a significant upward trend, while the task completion rate and various accuracy indicators (R-ACC, TCA) generally show a downward trend. This aligns with the objective law that the difficulty of complex tasks increases. Second, in terms of model performance comparison, high-performance base models (such as Qwen-Max and GPT-4o) exhibit extremely high accuracy in L1 and L2 levels. Qwen-Max achieved TCRs of 0.95 and 0.98 in L1 and L2 levels, respectively, demonstrating that the architecture of this invention can fully unleash the potential of strong base models. Particularly noteworthy is that in the highest difficulty level, L3, although the performance of most models degrades, Qwen-Max still maintains a TCR of 0.78, while DeepSeek-v3.2, despite having a higher latency (4.06 seconds) in this level, still achieved a TCR of 0.79, demonstrating its processing capability in complex inference scenarios.

[0050] In summary, the experimental data show that the intelligent agent architecture of this invention can run stably on base models of different scales and types, and the stronger the capabilities of the base model, the better the overall system's task completion performance. At the same time, the intelligent agent architecture of this invention exhibits significant performance advantages and excellent robustness in tasks of various difficulty levels, effectively breaking through the bottleneck of existing technologies in complex and long-process analysis.

[0051] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A method for autonomous understanding of multi-agent brain signals driven by a large language model, characterized in that, Includes the following steps: (1) Construct a hierarchical multi-agent architecture that includes a central supervisory agent, multiple specialized sub-agents, and a globally shared state; (2) The central regulator intelligent agent receives the user's natural language query, analyzes the complexity of the user's intent, and decomposes the complex query into multiple sub-tasks; (3) The central supervisory agent distributes sub-tasks to the corresponding specialized sub-agents according to the domain attributes. The specialized sub-agents adopt a context isolation strategy when executing sub-tasks. (4) Upon receiving the subtask instruction, the specialized sub-agent combines the current system state and workflow context to perform logical reasoning and planning to generate the tool execution sequence and tool call instructions; (5) The specialized sub-agents, following the generated tool execution sequence, sequentially call the corresponding tools through the system interface to perform actual brain signal analysis operations and manage the input and output flow of data; (6) Specialized sub-agents trigger the retrieval enhancement mechanism, retrieve relevant knowledge from the general knowledge base and domain-specific knowledge base according to the sub-task, combine the tool execution output results, generate specialized domain sub-reports and return them to the central supervisory agent; (7) The central supervisory agent aggregates the domain sub-reports returned by each sub-agent, performs cross-modal logical reasoning and causal relationship analysis, and generates a comprehensive analysis report to be fed back to the user.

2. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that: In step (1), the central supervisor agent is responsible for parsing the user's natural language intent and distributing tasks. Multiple specialized sub-agents include a sleep analysis agent and an emotion analysis agent. The sleep analysis agent is used for specialized sleep analysis tasks, including sleep staging, micro-event detection, sleep structure assessment, and clinical sleep report generation. The emotion analysis agent is used for specialized emotion analysis tasks, including emotion state recognition, psychological fatigue monitoring, cognitive load assessment, and emotion stability analysis. The global shared state is used to manage the brain signal data input by the user, intermediate processing results, and tool execution parameters, so as to realize data flow and state synchronization between agents. The global shared state is implemented using a key-value dictionary structure. When all tools are executed, they read the input parameters from the global shared state and write the output results back, so as to ensure that state synchronization can be achieved without explicit communication between agents.

3. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that: The specialized sub-agent configuration hierarchical resource access mechanism divides tools into general toolsets and domain-specific toolsets, and knowledge bases into general knowledge bases and domain-specific knowledge bases. This allows sub-agents to access only the subset of resources strongly related to their tasks, reducing the proportion of irrelevant tool descriptions in the context window, reducing computational overhead, and supporting sub-agents to perform operations including data loading, preprocessing, feature extraction, and domain analysis.

4. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that: In step (2), the central supervisory agent first determines the semantic complexity of the query by combining historical conversations, performs user intent parsing through thought chain reasoning, and determines whether to call specialized sub-agents based on the semantic complexity of the query: for simple greetings or ability inquiries, a response is generated directly from the global shared state; for atomic operations, only a single explicit instruction is distributed; for complex tasks involving high-level semantics, they are decomposed into a sequence of sub-tasks with logical dependencies based on causal reasoning logic.

5. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that, The context isolation strategy in step (3) is as follows: specialized sub-agents only receive specific task instructions distributed by the central supervisor agent and relevant execution result summaries generated in the global shared state after the preceding sub-agents execute tasks, while shielding irrelevant global dialogue history to ensure that the sub-agents focus on performing domain-specific analysis of sub-tasks.

6. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that, The specific implementation process of step (4) is as follows: S41. Before initiating tool planning, the specialized sub-agent prioritizes executing the direct response decision logic. It evaluates whether the existing information is sufficient to directly respond to the sub-task instruction by retrieving the global shared state and the execution results of the preceding sub-agent. If the existing information is sufficient, the planning is terminated immediately, the text response containing the reasoning process is directly output, and the execution plan is set to empty. If the existing information is insufficient to answer the question directly, proceed to the subsequent task breakdown and planning stage; S42. When it is necessary to call a tool, the specialized sub-agent breaks down the high-level instructions into a series of indivisible atomic sub-targets, and then determines the logical execution order of each sub-target according to the data flow dependency. Finally, a complete tool execution sequence from the start point to the end point is generated in a single response to ensure the continuity and efficiency of the execution process. S43. When constructing specific tool invocation instructions, specialized sub-agents follow the following parameter configuration rules: if parameters are explicitly specified in the user instruction, they are forcibly mapped to the corresponding tool parameters; for non-critical parameters that are not explicitly specified, the default values ​​defined by the tool or an empty dictionary are automatically used, while the parameter key names are forcibly verified to ensure that they strictly match the variable names in the globally shared state, thus ensuring the correctness of the parameter reading path for tool execution.

7. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that: The tool called in step (5) strictly performs corresponding encapsulation processing operations on brain signal data according to the algorithm logic and process steps predefined in its functional description; The tool reads input parameters from the global shared state, and after execution, writes the output back to the global shared state, thereby updating the current context information of the system for subsequent reading by other tools or sub-agents, achieving seamless synchronization and zero-copy flow of data.

8. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that: The retrieval enhancement mechanism in step (6) first performs coarse-grained retrieval in the general knowledge base and the domain-specific knowledge base through a dual encoder model, then performs fine-grained reordering using a cross encoder model to select highly relevant physiological definitions or clinical guidelines as domain-specific knowledge; finally, the quantitative output results of the tool are combined with the retrieved domain-specific knowledge to generate an interpretable specialized domain sub-report and return it to the central regulatory agent.

9. The multi-agent brain signal autonomous understanding method based on a large language model driven by claim 1, characterized in that: In step (7), the central supervisory agent receives and parses the domain sub-reports returned by each sub-agent, extracting key physiological indicators, quantitative features, and domain-specific conclusions. Then, deep semantic reasoning is performed to identify the inherent logical connections and physiological mechanism links between the analysis results of different domains, thereby constructing a multi-dimensional view of physiological state. Finally, the central supervisory agent integrates the discrete data insights into a coherent natural language narrative, and finally generates a comprehensive analysis report with internal logical coherence to be fed back to the user.

Images