An intention recognition method and system based on dynamic ontology evolution and multi-agent

Abstract

The application relates to the technical field of automation operation and maintenance, and discloses an intention recognition method and system based on dynamic ontology evolution and multi-agent, which comprises the following steps: receiving a natural language instruction, performing entity extraction and probabilistic linking by using a dynamic ontology knowledge base, and generating an initial intention based on predicate analysis; automatically completing missing key slots by using a probabilistic graph model, and generating a standardized intention; decomposing the standardized intention into an atomic subtask sequence, dynamically matching an execution agent based on an agent capability-demand matrix, generating a collaborative workflow, controlling the execution agent to call an atomic tool to execute a task, performing causal correlation analysis on multi-source results according to logical relations between ontology instances, generating a structured reasoning chain, and feeding back; and extracting a new treatment script based on execution feedback by using an evolution engine, and updating an ontology knowledge base and an agent confidence degree. The application can realize accurate understanding, automatic execution and adaptive evolution of a knowledge base of a fuzzy operation and maintenance intention.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and automated operation and maintenance technology, and in particular to an intent recognition method and system based on dynamic ontology evolution and multi-agent systems. Background Technology

[0002] With the widespread adoption of microservice architecture and cloud-native technologies, the scale and complexity of IT systems are increasing exponentially, and the amount of data generated by operations and maintenance (logs, monitoring metrics, call chains) has reached massive levels. To address this challenge, intelligent operations and maintenance (AIOps) and chatbot-based operations and maintenance (ChatOps) have emerged, aiming to lower the barriers to operations and maintenance and improve fault response efficiency through natural language interaction.

[0003] Currently, existing technologies typically employ rule-based matching or static intent recognition templates. That is, the system predefines a series of fixed keyword regular expressions or command templates; when the user's input strictly matches the preset rules, the corresponding script is executed. Some more advanced solutions incorporate general natural language processing (NLP) models for intent classification and entity extraction.

[0004] However, the aforementioned existing technologies still have significant defects and shortcomings in practical applications:

[0005] First, there is a lack of domain-deep semantic understanding and fuzzy reasoning capabilities. Existing solutions often rely on explicit keyword matching. For ambiguous instructions commonly found in operation and maintenance scenarios (such as "the database is a bit slow" without specifying the specific instance IP) or unclear expressions (such as using non-standard aliases or nicknames), the system often cannot perform reasoning and completion based on the conversation context or system topology like human experts. This leads to failure in intent recognition or the need for multiple rounds of mechanical questioning, which seriously affects interaction efficiency.

[0006] Second, the task execution process is rigid and lacks dynamic orchestration capabilities. Existing automated scripts are usually statically written "hard-coded" processes that cannot be dynamically adjusted based on the current system state (such as the unavailability of a tool or service) or the actual capabilities of the executing agent. When faced with complex, combined troubleshooting tasks, it is difficult to flexibly decompose the tasks and schedule the most suitable resources for execution.

[0007] Third, the knowledge base is costly to maintain and lacks an adaptive evolution mechanism. The operational environment is highly dynamic, with new service components, alarm types, and business terms constantly emerging. Existing systems typically rely on manual, periodic offline updates to their knowledge bases. The system cannot automatically learn new terminology mappings (such as new aliases) or accumulate new handling experience by analyzing successful operation logs or user feedback during operation. This "static" characteristic causes the system's ability to identify and resolve new faults to gradually decline over time, making it difficult to maintain a long-term level of intelligence. Summary of the Invention

[0008] This invention provides an intention recognition method and system based on dynamic ontology evolution and multi-agent systems. By constructing a dynamic ontology knowledge base and combining it with probabilistic graph completion, dynamic agent orchestration, and evolutionary learning mechanisms, it effectively solves the problems of difficulty in understanding fuzzy instructions and rigid task execution, and achieves accurate parsing of operation and maintenance intentions, interpretable automated execution with causal reasoning capabilities, and adaptive incremental updates of system knowledge and capabilities.

[0009] This invention provides an intent recognition method based on dynamic ontology evolution and multi-agent systems, including:

[0010] S1. Receive natural language instructions input by the user, and use a pre-built dynamic ontology knowledge base to perform entity recognition and linking on the natural language instructions, so as to map the natural language instructions into a structured initial intent representation; wherein, the initial intent representation includes the intent category corresponding to the natural language instructions and the attribute slots associated with the intent category.

[0011] S2. Detect whether there are any missing key slots in the attribute slots. If so, combine the session-level context graph and real-time alarms and topology data, and use a probabilistic graphical model to fill in the missing key slots and generate a clear standardized intent.

[0012] S3. Decompose the standardized intent into a task sequence composed of multiple atomic subtasks, match a corresponding execution agent for each atomic subtask based on the agent capability-requirement matrix, and generate a collaborative workflow containing the calling order and data dependencies of each execution agent; wherein, the agent capability-requirement matrix includes the confidence of each execution agent in completing a specific type of task;

[0013] S4. Control each execution agent in the collaborative workflow to sequentially call its pre-encapsulated atomic tools to execute the atomic subtasks, so as to obtain multi-source execution results, and perform correlation analysis on the multi-source execution results according to the logical relationship between ontology instances in the dynamic ontology knowledge base, generate response feedback containing a structured reasoning chain and return it to the user.

[0014] S5. Obtain feedback data from task execution, use the evolution engine to extract new knowledge patterns from successful execution trajectories to incrementally update the dynamic ontology knowledge base, and dynamically adjust the confidence level based on the multi-source execution results.

[0015] Furthermore, S1 specifically includes:

[0016] S101. Receive a natural language instruction input by the user, call a pre-trained named entity recognition model to extract mention words from the natural language instruction, query the dynamic ontology knowledge base to obtain ontology concepts or ontology instances associated with the mention words as candidate entities, and calculate the semantic similarity between the candidate entities and the mention words to determine the entity linking result; wherein, the calculation formula is:

[0017]

[0018] in, Let m be the candidate entity, m be the mentioned word, and C be the input context. and These are preset weights, This represents the similarity between the vector embeddings of the candidate entity and the mentioned word. This indicates the semantic similarity between the candidate entity and the input context;

[0019] S102. Analyze the predicate part in the natural language instruction, map the predicate part to the predefined intent class in the dynamic ontology knowledge base using the intent classification model, and activate the attribute slots associated with the intent class to obtain a structured initial intent representation; wherein, the attribute slots include target entity slots, indicator type slots and time range slots.

[0020] Furthermore, S1 also includes:

[0021] S103. If the natural language instruction is found to contain a new mention that is not recorded in the dynamic ontology knowledge base, and a target ontology instance with matching features already exists in the dynamic ontology knowledge base, then an alias mapping relationship between the new mention and the target ontology instance is established through association rule learning of the evolutionary engine.

[0022] Furthermore, S2 specifically includes:

[0023] S201. Based on the definition of the intent category in the dynamic ontology knowledge base, check whether there is a slot in the attribute slot that is marked as required and whose current value is empty. If so, mark it as a missing key slot.

[0024] S202. Start the multi-source data acquisition mechanism, retrieve the most recently mentioned ontology instance from the session-level context graph in the current session history; obtain real-time alarm data from the monitored target system, and filter out the ontology instances corresponding to the alarm objects that are currently active; query the topology database to obtain ontology instances that have a dependency relationship with the entity link results determined in step S1, and use all retrieved ontology instances as candidate fill values ​​for filling missing key slots.

[0025] S203. Construct a probabilistic graphical model, using candidate fill values, contextual evidence, and business rule constraints in the dynamic ontology knowledge base as observation variables. Calculate the joint probability of each candidate fill value filling the missing key slot under the current context conditions, and select the candidate fill value with the highest joint probability to fill the missing key slot to generate standardized intent.

[0026] Furthermore, in S203, the joint probability is calculated as follows:

[0027] (1) Define a set of characteristic functions and their corresponding weights Used to quantify the degree of matching of candidate imputation value v in the current context C; feature score The calculation formula is as follows:

[0028]

[0029] Specifically, the characteristic functions include, but are not limited to:

[0030] Time decay characteristics: This is used to increase the weight of the most recently mentioned ontology instance;

[0031] State association features: The value is 1 when there is an active alarm for the candidate instance, and 0 otherwise.

[0032] Ontology constraint features: The value is 1 when the candidate instance type matches the slot definition; otherwise, it is negative infinity or a minimum value.

[0033] (2) Calculate the joint probability of each candidate filling value using the Softmax function. :

[0034]

[0035] in, As a normalization factor, the joint probability is ultimately chosen. The highest candidate fill value is used as the final completion result.

[0036] Furthermore, S3 specifically includes:

[0037] S301. Call the central planner to receive the standardized intent, query the predefined intent-subtask decomposition rules in the dynamic ontology knowledge base, and decompose the standardized intent into a task sequence composed of a series of ordered or parallel atomic subtasks.

[0038] S302. Based on the agent capability-requirement matrix, select the agent with the highest confidence as the executing agent for each atomic subtask, and arrange the selected executing agents and their corresponding atomic subtasks into a collaborative workflow according to the logical dependencies between the atomic subtasks; wherein, the collaborative workflow is a directed acyclic graph, where nodes represent executing agents and atomic subtasks, and edges represent execution order and data flow direction;

[0039] S303. Before generating the final collaborative workflow, analyze the operation level of each atomic subtask in the collaborative workflow. If a high-risk operation is identified, automatically inject a security control node before the relevant node in the collaborative workflow. The security control node includes at least an authorization verification node, a change window check node, or a manual approval process node.

[0040] Furthermore, S4 specifically includes:

[0041] S401. Activate the execution agents corresponding to each node in sequence according to the topological order of the collaborative workflow. The execution agents parse the assigned atomic subtasks, select matching atomic tools from their pre-encapsulated toolset, and execute the invocation. The execution agents pass parameters to the atomic tools through a unified model context protocol and implement a fault tolerance mechanism during execution. If the atomic tool invocation fails or times out, it is retried according to a preset strategy. If the retry fails, the task status is marked as abnormal.

[0042] S402. The execution agent performs validity verification and cleaning on the raw data returned by the atomic tool, and extracts key information points with business semantics to form intermediate execution results; wherein, extracting key information points includes: extracting indicators of sudden increase or decrease from the raw time series data, or extracting specific error keywords or abnormal patterns from the raw log data.

[0043] S403. Invoke the integrated reasoning agent to perform causal correlation analysis on the aggregated intermediate execution results based on the logical relationships between ontology instances defined in the dynamic ontology knowledge base, and generate a structured reasoning chain; wherein, the integrated reasoning agent generates the structured reasoning chain according to the following logic based on the logical relationships defined in the dynamic ontology knowledge base:

[0044] First, the abnormal phenomenon is confirmed based on the intermediate execution results. Then, the abnormal status of the related service is discovered. Subsequently, the root cause log evidence is locked, and finally a complete reasoning chain containing the phenomenon, the correlation, the root cause and the conclusion is generated.

[0045] S404. Integrate the structured reasoning chain, key data snapshots supporting the structured reasoning chain, and action suggestions generated based on the reasoning conclusions into a response feedback.

[0046] Furthermore, S5 specifically includes:

[0047] S501. Utilize the evolutionary engine to continuously monitor and collect the task execution status in step S4, and obtain multi-dimensional feedback data including system execution feedback and user interaction feedback; wherein, the system execution feedback includes execution status, structured inference chain, execution time and resource consumption data, and the user interaction feedback includes explicit ratings or implicit behavioral data of users on response results.

[0048] S502. When the successful execution of a task is detected, the corresponding collaborative workflow is analyzed, the standardized intent is associated with the atomic tool call sequence that was actually successfully executed, and abstracted into a reusable standardized disposal script, which is then stored as a new disposal strategy in the dynamic ontology knowledge base.

[0049] S503. Perform cluster analysis on unsuccessfully linked mentions or newly emerging association patterns. If frequent and stable new patterns are found, generate an ontology update proposal, which is then merged into the model layer of the dynamic ontology knowledge base after verification.

[0050] S504. Based on the final execution result score of this task, dynamically calibrate the confidence levels of the agents participating in the execution in the agent capability-requirement matrix; wherein, the confidence levels are adjusted using a time-decay-based weighted update formula:

[0051]

[0052] in, For the updated confidence level, R represents the confidence level before the update, and R is the final score of the task execution result. This is a preset forgetting factor.

[0053] This invention also provides an intent recognition system based on dynamic ontology evolution and multi-agent interaction. Based on the intent recognition method based on dynamic ontology evolution and multi-agent interaction described above, the system includes:

[0054] The recognition module is used to receive natural language instructions input by the user, and to perform entity recognition and linking on the natural language instructions using a pre-built dynamic ontology knowledge base, so as to map the natural language instructions into a structured initial intent representation; wherein, the initial intent representation includes the intent category corresponding to the natural language instruction and the attribute slots associated with the intent category;

[0055] The completion module is used to detect whether there are missing key slots in the attribute slots. If there are, it combines the session-level context graph and real-time alarms and topology data, and uses a probabilistic graphical model to complete the missing key slots and generate clear standardized intents.

[0056] The matching module is used to decompose the standardized intent into a task sequence composed of multiple atomic subtasks, match a corresponding execution agent for each atomic subtask based on the agent capability-requirement matrix, and generate a collaborative workflow containing the calling order and data dependencies of each execution agent; wherein, the agent capability-requirement matrix includes the confidence of each execution agent in completing a specific type of task;

[0057] The execution module is used to control each execution agent in the collaborative workflow to sequentially call its pre-encapsulated atomic tools to execute the atomic subtasks in order to obtain multi-source execution results. Based on the logical relationship between ontology instances in the dynamic ontology knowledge base, the module performs correlation analysis on the multi-source execution results, generates response feedback containing a structured reasoning chain, and returns it to the user.

[0058] The update module is used to obtain feedback data from task execution, extract new knowledge patterns from successful execution trajectories using an evolutionary engine to incrementally update the dynamic ontology knowledge base, and dynamically adjust the confidence level based on the multi-source execution results.

[0059] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0060] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0061] The beneficial effects of this invention are as follows:

[0062] This invention constructs a dynamic ontology knowledge base comprising a model layer and a data layer. By combining entity linking with a probabilistic model and a multi-source context slot completion mechanism, it can accurately identify mentions in ambiguous user commands and link them to specific ontology instances, effectively solving the problems of semantic ambiguity and missing key information in natural language commands, and significantly improving the accuracy and robustness of intent recognition. This invention introduces an agent capability-demand matrix and a central planner, which can dynamically orchestrate collaborative workflows composed of multiple atomic subtasks according to task requirements and generate structured reasoning chains based on ontology causal logic. This not only automates the execution of complex operation and maintenance tasks but also provides users with interpretable root cause analysis results. Finally, this invention designs an evolutionary engine that can automatically extract standardized disposal scripts based on execution feedback and dynamically calibrate agent confidence, enabling adaptive incremental updates of the knowledge base and system capabilities, ensuring the system can continuously adapt to the ever-changing operation and maintenance environment. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.

[0065] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0066] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0067] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0068] This invention specifically addresses the following core technical problems:

[0069] 1. How to achieve dynamic evolution and continuous integration of operation and maintenance knowledge models: solve the problem that static knowledge bases cannot automatically adapt to business changes, enable the intent recognition system to autonomously absorb and integrate new knowledge as new equipment goes online, new fault modes emerge, and new operation and maintenance strategies are implemented, and reduce manual maintenance costs.

[0070] 2. How to achieve precision and context completion for vague and fragmented user intents: This addresses the problem that the system struggles to accurately understand the true intent of users when they express their needs in colloquial, abbreviated, or metaphorical ways. For example, automatically complete "Help me check the database slow problem" and associate it with "Analyze the slow query logs and associated CPU and IO metrics of MySQL instance Y on host X over the past 2 hours".

[0071] 3. How to ensure the accuracy, credibility and traceability of decisions while leveraging the generalization capabilities of large models: Overcome the "illusion" and black-box problems of pure LLM solutions, anchor the reasoning capabilities of large models to real, structured business ontology and real-time data, and generate a reasoning chain where every step is verifiable and based on evidence.

[0072] 4. How to safely and automatically convert identified high-level intents into a sequence of low-level executable actions: bridge the gap between "knowing what to do" and "actually doing it" by using programmable agents to decompose complex intents into a series of specific, permission-controlled atomic operations, and ensure the reliability and rollback capability of the execution process.

[0073] like Figure 1 As shown, this invention provides an intent recognition method based on dynamic ontology evolution and multi-agent systems, comprising:

[0074] S1. Receive natural language instructions input by the user, and use a pre-built dynamic ontology knowledge base to perform entity recognition and linking on the natural language instructions, so as to map the natural language instructions into a structured initial intent representation; wherein, the initial intent representation includes the intent category corresponding to the natural language instructions and the attribute slots associated with the intent category.

[0075] In one embodiment, a dynamic ontology knowledge base is pre-built and maintained, including ontology concepts and ontology instances. Ontology concepts (TBox) define categories, including core concepts (such as Host, Service, Metric), attributes (such as cpu_utilization, status), and relationships (such as Host.runs Service, Service.depends_on Service). Ontology instances (ABox) define specific objects (such as Host_A, Service_MySQL). More specifically, step S1 includes the following sub-steps:

[0076] S101, Entity Extraction and Probabilistic Model-Based Entity Disambiguation

[0077] First, the system receives a natural language instruction input by the user (e.g., "database master node is slow to respond"). Then, it calls a Named Entity Recognition (NER) model that has been fine-tuned in the field of operations and maintenance to scan the natural language instruction and extract the mention words (Mention, denoted as m, for example, "database master node").

[0078] Subsequently, using the mentioned term as the search condition, a pre-built dynamic ontology knowledge base is queried to obtain one or more ontology concepts or ontology instances semantically related to the mentioned term as candidate entities (denoted as...). To determine which specific ontology concept a mention refers to, candidate entities are computed. The semantic similarity between the mentioned word m and the reference word m. Specifically, the posterior probability is calculated using the following formula. To make disambiguation decisions:

[0079]

[0080] in, Indicates candidate entities The similarity between the vector embeddings of (ontology concept) and the mention word m (input text), Indicates candidate entities Contextual similarity with the input context C, and These are the corresponding weights. Select the calculated probability. The highest-ranking candidate entity is used as the final entity linking result (e.g., linking "Database Primary Node" to the Database.Primary Node instance in the ontology).

[0081] S102. Intent Classification and Slot Activation Based on Predicate Analysis

[0082] Building upon the completed entity linking, the predicate portion of the natural language instruction (e.g., "slow response") is further analyzed. Using a pre-trained intent classification model, the predicate portion is mapped to a predefined intent class in a dynamic ontology knowledge base. For example, "slow response" is mapped to the intent class "Investigate Performance".

[0083] Once the intent class is determined, the associated attribute slots are automatically activated. These slots are data fields necessary for processing this type of intent in the ontology definition, such as the target entity slot (target_entity), metric_type slot (metric_type), and time range slot (time_range). Thus, the natural language instruction is initially transformed into a structured initial intent representation containing the intent category and the slots to be filled.

[0084] S103, Dynamic Association and Alias ​​Learning Based on Evolutionary Engine

[0085] During the above analysis process, the built-in evolutionary engine continuously interacts with the dynamic ontology knowledge base to handle the dynamic adaptation of new terms. Specifically, if the mention word (m) extracted in step S101 is not recorded in the current dynamic ontology knowledge base, or if its similarity with existing candidate entities is lower than a preset threshold (i.e., a new mention word appears), the evolutionary engine will activate the association learning mechanism.

[0086] If the evolutionary engine, through analysis of context or historical operation logs, discovers that an unrecorded mention (e.g., "database master node") semantically frequently points to a target ontology instance already existing in the dynamic ontology knowledge base that possesses specific characteristics and roles—that is, an existing target ontology instance (e.g., a Server instance with the role of primary and running Database software) that matches the new mention—the evolutionary engine will generate association rules. Based on this, it establishes an alias mapping relationship between the new mention and the target ontology instance, or automatically generates ontology update proposals. This enables adaptive understanding of newly emerging operational terms and incremental updates to the knowledge base, ensuring the accuracy of entity recognition and linking.

[0087] S2. Detect whether there are any missing key slots in the attribute slots. If so, combine the session-level context graph and real-time alarms and topology data, and use a probabilistic graphical model to complete the missing key slots and generate a clear standardized intent.

[0088] In one embodiment, a session-level context graph is maintained in real time to temporarily store ontology instances, events, and relationships mentioned in the current session. Step S2 specifically includes the following sub-steps:

[0089] S201, Critical Missing Slot Detection

[0090] Receive the structured initial intent representation output from step S1, which includes the identified intent categories and activated attribute slots. First, based on the definition of the intent category in the dynamic ontology knowledge base, traverse all attribute slots and check if there are any slots marked as "required" but with an empty current value.

[0091] If such a slot is detected, mark it as a "missing critical slot". For example, "Check Database" is missing the specific database_id and check_type.

[0092] S202. Collection of Multi-Source Contextual Evidence and Retrieval of Candidate Instances

[0093] To fill in the missing key slots, a multi-source data acquisition mechanism is activated to obtain evidence data and retrieve possible candidates for filling (i.e., ontology instances) from the following three dimensions:

[0094] ① Session-level context retrieval: Access the real-time maintained session-level context graph to retrieve the most recently mentioned ontology instance in the current session history (e.g., "Server_001" mentioned by the user in the previous round of conversation).

[0095] ② Real-time status awareness: Connect to the monitored target system, obtain real-time alarm data, and filter out the entity instances corresponding to the alarm objects that are currently active (e.g., the server instance corresponding to the current CPU utilization alarm).

[0096] ③ Topology association retrieval: Query the topology database to obtain ontology instances that have dependencies on the ontology instances identified in step S1 (for example, if S1 identifies "order service", then retrieve the database instances that the service depends on).

[0097] Ultimately, all retrieved ontology instances are used as "candidate fill values" to fill the missing critical slots.

[0098] S203. Slot completion and intent standardization based on probabilistic graphical models.

[0099] Construct a probabilistic graphical model (e.g., a Conditional Random Field (CRF) or a Bayesian network), and use the candidate fill values, contextual evidence, and business rule constraints in the dynamic ontology knowledge base obtained in step S202 as the observed variables of the model.

[0100] The joint probability of each candidate fill value filling the missing critical slot under the current context is calculated using the probabilistic graphical model. The calculation method is as follows:

[0101] (1) Define a set of characteristic functions and their corresponding weights This is used to quantify the degree of matching of the candidate imputation value v in the current context C. Feature score The calculation formula is as follows:

[0102]

[0103] Specifically, the characteristic functions include, but are not limited to:

[0104] ①Time decay characteristics: This is used to increase the weight of the most recently mentioned ontology instance;

[0105] ②State association features: The value is 1 when there is an active alarm for the candidate instance, and 0 otherwise.

[0106] ③ Ontological constraint features: The value is 1 when the candidate instance type matches the slot definition, otherwise it is negative infinity (or a minimum value).

[0107] (2) In order to transform the scores into a probability distribution for decision-making, the joint probability (or posterior probability) of each candidate filler value is calculated using the Softmax function. :

[0108]

[0109] in, This is the normalization factor.

[0110] Choose joint probability The highest candidate fill value is used as the final completion result. Because... For all candidate values ​​to be identical, and with the exponential function monotonically increasing, the highest selection probability is equivalent to the selection feature score. The highest candidate value ultimately transforms the vague initial intent into a standardized intent that is complete in information. For example, combining the context of "there was a CPU alarm just now" with the relationship "the database is running on server A" in the topology, "database" is completed to "MySQL instance on server A", and the intent is refined to "check the CPU utilization of this instance".

[0111] S3. Decompose the standardized intent into a task sequence consisting of multiple atomic subtasks, match a corresponding execution agent for each atomic subtask based on the agent capability-requirement matrix, and generate a collaborative workflow containing the calling order and data dependencies of each execution agent; wherein, the agent capability-requirement matrix includes the confidence of each execution agent in completing a specific type of task.

[0112] In one embodiment, a multi-agent architecture is employed, comprising multiple agents with specific functions, such as a diagnostic agent, a data query agent, a log analysis agent, and an execution agent. These agents operate according to a unified ReAct (reasoning-action) and planning pattern. Step S3 specifically includes the following sub-steps:

[0113] S301. Task decomposition based on ontology rules

[0114] First, the central planner is invoked to receive the standardized intent output from step S2. The central planner queries the model layer in the dynamic ontology knowledge base to obtain predefined "intent-subtask" decomposition rules. Based on these rules, the central planner decomposes the standardized intent into a series of fine-grained atomic subtasks. For example, if the standardized intent is "the root cause of slow access to location service A," it is decomposed into the following ordered or parallel sequence of atomic subtasks:

[0115] : Query the response time metric for service A;

[0116] : Obtain the dependency chain topology of service A;

[0117] Analyze the performance metrics of dependent services B and C;

[0118] : Retrieve relevant error logs.

[0119] S302, Agent Matching and Workflow Orchestration

[0120] After determining the sequence of atomic subtasks, the central planner finds the most suitable agent to execute each subtask. This process relies on a dynamically maintained Agent Capability-Demand Matrix. The elements of this matrix... The confidence level of the i-th agent in completing the j-th atomic subtask is recorded. This confidence level is dynamically updated based on historical success rate and ability description similarity. The central planner traverses the matrix and selects a confidence level for each atomic subtask. The highest level of intelligence acts as the executor.

[0121] Subsequently, the planner determines the logical dependencies between the atomic subtasks (e.g., Execution depends on The returned topology data is used to orchestrate the selected agents and their corresponding atomic subtasks into a collaborative workflow. In this embodiment, the collaborative workflow is specifically represented as a directed acyclic graph (DAG), where nodes represent agents and tasks, and edges represent the execution order and data flow direction.

[0122] S303, Security Policy Injection and Compliance Check

[0123] Before generating the final collaborative workflow, perform a security compliance check. Analyze the operation level of each atomic subtask in the collaborative workflow (e.g., read-only query, configuration change, service restart, etc.).

[0124] If a high-risk operation (such as "restarting the service") is identified, a security control node is automatically injected before the relevant nodes in the collaborative workflow. This node includes, but is not limited to, permission verification nodes, change window check nodes, or manual approval process nodes. The collaborative workflow after security injection becomes the final planned result that can be executed by the system, thereby ensuring the security and rollback capability of the operation and maintenance.

[0125] S4. Control each execution agent in the collaborative workflow to sequentially call its pre-encapsulated atomic tools to execute the atomic subtasks to obtain multi-source execution results. Based on the logical relationships between ontology instances in the dynamic ontology knowledge base, perform correlation analysis on the multi-source execution results, generate response feedback containing a structured reasoning chain, and return it to the user.

[0126] In one embodiment, each agent is backed by several atomic tools that expose their functionality through a unified protocol (such as MCP). These tools range from Prometheus queries and log search engine calls to executing Ansible playbooks. Step S4 specifically includes the following sub-steps:

[0127] S401, Automated Execution and Fault Tolerance of Collaborative Workflows

[0128] Following the topological order of the collaborative workflow (represented as a directed acyclic graph, DAG) generated in step S3, the execution agents of each node are activated sequentially. When a node's dependent prerequisite tasks are completed, the execution agent of that node is triggered.

[0129] The agent parses its assigned atomic subtasks and selects the most suitable atomic tool from its pre-packaged toolset. The agent then passes specific parameters to the atomic tool and initiates the call via a unified tool invocation protocol (e.g., Model Context Protocol, MCP). For example, for the atomic subtask "query the response time of service A," the agent invokes the Prometheus query tool; for the atomic subtask "retrieve error logs," the agent invokes the log retrieval engine tool.

[0130] During execution, a fault-tolerance mechanism is implemented. If an atomic tool call fails or times out, the agent will retry according to a preset strategy; if the retry still fails, the branch task is marked as failed, and the abnormal status is reported to the upper-level planner.

[0131] S402. Verification and Feature Extraction of Multi-Source Execution Results

[0132] When the atomic tool returns raw data (such as time-series data curves or massive log lists), the executing agent first verifies and cleans the data. Subsequently, the agent performs preliminary refinement on the cleaned data, extracting key information points with business semantics to form "intermediate execution results".

[0133] For example, instead of returning the entire log segment directly, the agent extracts the keyword "Error_X" which was found within the time window T; instead of returning the entire monitoring chart directly, it extracts the feature description of "Metric_Y suddenly increases at time T".

[0134] S403, Generation of Reasoning Chains Based on Ontology-Causal Logic

[0135] Once all (or key) atomic subtasks in the collaborative workflow are completed, a comprehensive reasoning agent is invoked to perform correlation analysis on the converged multi-source execution results. This process strictly relies on the logical relationships between ontology instances defined in the dynamic ontology knowledge base (i.e., causal rules defined in the model layer, such as "service dependency relationships" and "resource contention relationships").

[0136] The integrated reasoning agent assembles the scattered intermediate execution results into a logically rigorous structured reasoning chain. For example, based on the Service_A depends_on Service_B relationship defined in the ontology, the following reasoning chain is generated:

[0137] Phenomenon confirmed: The response time of service A slows down at time T (originating from the result of task T1).

[0138] Correlation discovery: Service B, which service A depends on, reaches peak CPU utilization at the same time T (derived from the results of task T3).

[0139] Root cause identification: A large number of abnormal query requests were found in the logs of service B at time T (originating from the result of task T4).

[0140] Final conclusion: The fundamental reason for the performance degradation of service A is that service B, which it depends on, experienced a sudden surge in query load.

[0141] S404, Synthesis and Output of Response Feedback

[0142] The structured inference chain generated in step S403, key data snapshots supporting the inference chain (such as screenshots of key indicators and log fragments), and action suggestions generated based on the inference conclusions (such as "suggest scaling up service instance B") are integrated into the final response feedback. This response feedback is returned to the user in the form of natural language or a visual report, thus completing a full closed loop from user instructions to interpretable results.

[0143] S5. Obtain feedback data from task execution, use the evolution engine to extract new knowledge patterns from successful execution trajectories to incrementally update the dynamic ontology knowledge base, and dynamically adjust the confidence level based on the multi-source execution results.

[0144] In one embodiment, an evolutionary engine is pre-built to continuously monitor interaction logs, execution results, and user feedback (explicit ratings or implicit behaviors). Step S5 specifically includes the following sub-steps:

[0145] S501, Real-time Acquisition of Multidimensional Feedback Data

[0146] An evolutionary engine is continuously run to monitor and collect data on the task execution in step S4. The collected feedback data includes two types:

[0147] ①System execution feedback: Execution status (success / failure), structured inference chain, and execution time and resource consumption data of each atomic subtask from the response feedback generated in step S404;

[0148] ② User interaction feedback: Explicit ratings (such as "likes / dislikes") or implicit behavioral data of users on the final response results (such as whether users repeatedly ask the same question in a short period of time, or whether they adopt the action suggestions given by the system).

[0149] S502, Playbook Extraction Based on Success Trajectory

[0150] When an intent recognition and execution task is detected as "successful" (i.e., it receives a high user rating or passes system self-verification), the evolution engine will analyze the collaborative workflow and its call chain corresponding to the task.

[0151] The evolutionary engine maps the standardized intent generated in step S2 to a sequence of atomic tool calls that were actually successfully executed in step S4. This successful tool call sequence is then abstracted into a reusable standardized playbook. For example, after the system successfully resolves a complex "network jitter" problem, the evolutionary engine solidifies the "network packet capture tool - link quality analysis tool" sequence involved in that execution into a new playbook and stores it as a new handling strategy in the dynamic ontology knowledge base, associating it with the ontology concept of "network jitter".

[0152] S503, Incremental Updates and Concept Discovery of the Ontology Model Layer

[0153] For mentions or newly emerging association patterns that were not successfully linked in step S1 or S4, the evolutionary engine performs cluster analysis after accumulating a certain amount of historical data.

[0154] If a new pattern is found to emerge frequently and stably (e.g., the undefined term "Pod" appears frequently in the logs, and it always appears with the concept of "Container"), the evolution engine generates an ontology update proposal (e.g., "Add an ontology concept Pod, defining it as a collection of Containers"). After the proposal passes manual review or automatic confidence threshold verification, it will be merged into the model layer of the dynamic ontology knowledge base, thereby achieving adaptive evolution of the knowledge base structure.

[0155] S504, Dynamic calibration of the agent capability-demand matrix

[0156] Based on the final execution result score R of this task (e.g., 1 for success and 0 for failure), the evolutionary engine dynamically updates the agent capability-requirement matrix used in step S3.

[0157] Specifically, for each executing agent i participating in this collaborative workflow and its executed atomic subtask type j, its confidence level is adjusted using a time-decay-based weighted update formula. :

[0158]

[0159] in, This is a forgetting factor (e.g., a value of 0.8). This step allows for dynamic adjustment of each agent's ability score based on the latest practical performance, making future S3 planning steps more precise and adaptable to the dynamic changes in agent capabilities.

[0160] The core idea of ​​the steps described in this invention lies in constructing a dynamic, interpretable, and executable intent recognition closed loop. Based on this, some modules have alternative implementation methods, which may have specific advantages under different constraints:

[0161] 1. Alternative implementation of dynamic ontology evolution mechanism:

[0162] Version updates based on centralized manual review: Unlike the semi-automatic evolution based on online learning and threshold triggering proposed in this invention, another approach is to employ a strict "change proposal - expert review - version release" process. All new knowledge and relationships extracted from the logs are first stored in a staging area, where domain experts periodically review them before packaging them into a new version of the ontology model for unified deployment. This approach is more suitable for scenarios with extremely strict change control requirements, such as finance and nuclear power, ensuring the absolute authority and stability of knowledge updates.

[0163] Distributed knowledge fusion based on federated learning: In multi-regional deployments of large multinational enterprises or cloud service providers, the ontology of each region may differ due to localization practices. A federated learning framework can be adopted, where each regional node performs ontology evolution learning locally, and only the model update (not the original data) is encrypted and uploaded to the central server for secure aggregation to generate a global ontology model before being distributed to each node. This approach better balances the contradictions between knowledge unification and data privacy, as well as local adaptability.

[0164] 2. Alternative implementation of planner in multi-agent collaborative decision-making:

[0165] Reinforcement Learning (RL) Based Planner: The rule-based and DAG-based planner in this invention is deterministic and interpretable. A more complex but potentially more flexible alternative is to use deep reinforcement learning to train the planner. Using structured intentions and system states as the state space, agent calls as the action space, and task completion quality and efficiency as rewards, the planning strategy is trained through extensive simulated interactions. This approach may discover efficient task decomposition and orchestration paths not pre-defined by human experts, and is particularly suitable for ultra-large-scale, highly dynamic scenarios (such as supercomputing center scheduling).

[0166] Agent collaboration based on market auction mechanisms: Another decentralized alternative is to introduce a negotiation mechanism based on a market economy. A central planner only issues top-level tasks, and each agent "bids" based on its current load, capabilities, and task requirements. Task allocation and pricing (measured by virtual resources or priority) are determined through a virtual auction market. This approach achieves better global resource load balancing and adapts to highly heterogeneous and dynamically changing agent environments.

[0167] Compared with existing technologies, this invention, by introducing indispensable components such as dynamic ontology, multi-agent collaboration, and learning closed loop, produces the following direct and significant positive effects:

[0168] 1. A fundamental improvement in the accuracy and adaptability of intent recognition: The existence of the core components of this invention—the dynamic ontology knowledge base and evolution engine—enables the system to continuously absorb new knowledge and adapt to business changes. This directly solves the problem of accuracy decay over time caused by model rigidity in traditional static systems. The system no longer needs to perform expensive manual rule adjustments for every minor change, but instead possesses the ability to "learn from experience," maintaining high-precision intent understanding even when faced with new question formats, new ontology instances, and new fault modes. For example, when a new heterogeneous accelerator card is introduced into the data center, relevant monitoring terms and fault characteristics will be automatically or semi-automatically incorporated into the ontology, and subsequent related questions can be immediately and accurately understood.

[0169] 2. Provides highly interpretable and reliable decision support: Through ontology-based semantic anchoring and structured reasoning chain generation, this invention completely changes the drawbacks of LLM black-box decision-making. Every conclusion and suggestion output by the system clearly indicates the real-time data source (such as specific indicator curves), business rules in the ontology (such as service dependencies), and logical reasoning steps on which it is based. This transparency not only greatly enhances the trust of operations and maintenance personnel in the system's suggestions, facilitating rapid verification and decision-making, but also provides complete digital credentials for fault review, audit tracing, and compliance checks, meeting the stringent security and compliance requirements of enterprise-level operations and maintenance.

[0170] 3. Achieving a leap from "passive question-and-answer" to "proactive closed-loop operation": The newly added multi-agent collaborative decision-making and planning layer and atomic tool execution layer constitute a complete "cognition-action" system. The system is no longer merely an information desk answering "what happened" or "what might be the cause," but an actor capable of "diagnosing" and "repairing." For example, after recognizing the intent of "uneven load," the system can automatically orchestrate data queries, analyze agents to complete root cause localization, and finally, the executing agent, after obtaining approval, adjusts the load balancing strategy. This directly reduces the mean time to repair (MTTR) from hours to minutes, achieving a qualitative leap in operational efficiency.

[0171] This invention establishes a virtuous cycle of knowledge assetization and automatic accumulation in operations and maintenance: the creation of a closed loop of execution feedback and dynamic ontology evolution transforms every successful fault handling into a reusable knowledge asset. Traditional expert experience is easily lost with personnel turnover, while this system automatically precipitates effective handling patterns into standardized scripts and enriches them into the ontology model. This allows enterprise operations and maintenance capabilities to no longer rely on individual experts, but rather to be solidified within the system, continuously growing and optimizing over time, forming a virtuous cycle of becoming increasingly intelligent with use, significantly reducing the cost of knowledge transfer and the training threshold for new employees.

[0172] like Figure 2 As shown, the present invention also provides an intent recognition system based on dynamic ontology evolution and multi-agent systems. Based on the intent recognition method based on dynamic ontology evolution and multi-agent systems described above, the system includes:

[0173] The recognition module 1 is used to receive natural language instructions input by the user, and to perform entity recognition and linking on the natural language instructions using a pre-built dynamic ontology knowledge base, so as to map the natural language instructions into a structured initial intent representation; wherein, the initial intent representation includes the intent category corresponding to the natural language instructions and the attribute slots associated with the intent category;

[0174] The completion module 2 is used to detect whether there are missing key slots in the attribute slots. If there are, it combines the session-level context graph and real-time alarms and topology data, and uses a probabilistic graphical model to complete the missing key slots and generate clear standardized intents.

[0175] Matching module 3 is used to decompose the standardized intent into a task sequence composed of multiple atomic subtasks, match a corresponding execution agent for each atomic subtask based on the agent capability-requirement matrix, and generate a collaborative workflow containing the calling order and data dependencies of each execution agent; wherein, the agent capability-requirement matrix includes the confidence of each execution agent in completing a specific type of task;

[0176] The execution module 4 is used to control each execution agent in the collaborative workflow to sequentially call its pre-encapsulated atomic tools to execute the atomic subtasks in order to obtain multi-source execution results. Based on the logical relationship between ontology instances in the dynamic ontology knowledge base, the multi-source execution results are analyzed for correlation, and a response feedback containing a structured reasoning chain is generated and returned to the user.

[0177] Update module 5 is used to obtain feedback data of task execution, use the evolution engine to extract new knowledge patterns from successful execution trajectories to incrementally update the dynamic ontology knowledge base, and dynamically adjust the confidence level according to the multi-source execution results.

[0178] Each of the above modules is used to execute the respective steps in the above intention recognition method based on dynamic ontology evolution and multi-agent. The specific implementation method is as described in the above method embodiment, and will not be repeated here.

[0179] The method provided in this invention is applicable to the field of artificial intelligence and automated operations and maintenance (AIOps), particularly in scenarios involving fault diagnosis and self-healing in cloud-native environments, performance monitoring of large-scale microservice architectures, and intelligent management of enterprise-level IT assets. In practical applications, this invention can be deployed on the central control server of an operations and maintenance management platform, connecting to monitoring and logging systems such as Prometheus and ELK via API interfaces, and connecting to business servers via SSH or K8s APIs. Specifically, some applications are as follows:

[0180] 1. Cloud Computing and Intelligent Data Center Operations and Maintenance: This is the most direct application area. Cloud service providers or large enterprise IT departments can use this system to build a unified intelligent operations and maintenance brain. Application method: Operations engineers or developers submit requests to the system via natural language, such as "Assess the resource requirements of Project A cluster next month and provide expansion suggestions" or "Immediately diagnose the cause of the e-commerce front-end service latency spike and attempt automatic mitigation." The system automatically completes data retrieval, correlation analysis, capacity prediction, root cause identification, and can perform operations such as elastic scaling and service restart after approval.

[0181] 2. Telecommunications Network Operation and Maintenance: Used for fault management, performance optimization, and service activation of backbone networks, metropolitan area networks, and 5G core networks. Application: Network operation and maintenance personnel input "Analyze the cause of packet loss on the Nanjing-Shanghai backbone link last night" or "Quickly activate a cross-border dedicated line for VIP customer XYZ Company." The system can integrate network management data, performance probes, and configuration database information to complete cross-domain fault delimitation and location, and drive automated network configuration tools to complete service orchestration and activation.

[0182] 3. Industrial Internet and Smart Manufacturing Operations and Maintenance: Applied to predictive maintenance of equipment, production process optimization, and energy efficiency management in smart factories and production lines. Application method: Equipment engineers or production supervisors ask questions such as "Predict the next possible failure time of spindle #3 in the CNC machine tool unit" and "Identify the root cause of the abnormally high energy consumption in the welding section this month." The system integrates IoT sensor data, MES production data, and equipment maintenance logs to predict the remaining life of key components, and correlates and analyzes process parameters with energy consumption indicators to provide optimization suggestions.

[0183] 4. Financial Technology System Operation and Security Management: Used for ensuring the stability of critical business systems such as banking and securities trading, and responding to cybersecurity threats. Application: Security analysts input commands such as "Investigate whether abnormal login attempts originating from IP segment A constitute a threat" and "Verify whether the SLOs for all critical transactions in the payment core system meet the standards after this change." The system can correlate and analyze logs, traffic, authentication, and business metrics data to quickly assess security incidents and automatically verify the health of the business after the change, ensuring compliance with financial-grade regulatory requirements.

[0184] 5. Integrated Management of Smart City Infrastructure: Potentially applicable to the collaborative monitoring and scheduling of urban transportation, public facilities, and the energy grid. Application: Command center personnel issue instructions such as "coordinate the handling of traffic congestion around Renmin Road subway station caused by large passenger flow" and "analyze the causes of the peak public electricity load in Binjiang New Area last weekend." The system can break down data barriers between multiple departments such as transportation, police, municipal administration, and power, and perform event correlation and impact analysis based on a unified ontology, assisting in the formulation of cross-departmental joint emergency plans and resource scheduling schemes.

[0185] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all data required for the process based on a dynamic ontology evolution and multi-agent intent recognition method. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the dynamic ontology evolution and multi-agent intent recognition method.

[0186] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0187] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described intention recognition methods based on dynamic ontology evolution and multiple agents.

[0188] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0189] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0190] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intent recognition based on dynamic ontology evolution and multi-agent systems, characterized in that, include: S1. Receive natural language instructions input by the user, and use a pre-built dynamic ontology knowledge base to perform entity recognition and linking on the natural language instructions, so as to map the natural language instructions into a structured initial intent representation; wherein, the initial intent representation includes the intent category corresponding to the natural language instructions and the attribute slots associated with the intent category. S2. Detect whether there are any missing key slots in the attribute slots. If so, combine the session-level context graph and real-time alarms and topology data, and use a probabilistic graphical model to fill in the missing key slots and generate a clear standardized intent. S3. Decompose the standardized intent into a task sequence composed of multiple atomic subtasks, match a corresponding execution agent for each atomic subtask based on the agent capability-requirement matrix, and generate a collaborative workflow containing the calling order and data dependencies of each execution agent; wherein, the agent capability-requirement matrix includes the confidence of each execution agent in completing a specific type of task; S4. Control each execution agent in the collaborative workflow to sequentially call its pre-encapsulated atomic tools to execute the atomic subtasks, so as to obtain multi-source execution results, and perform correlation analysis on the multi-source execution results according to the logical relationship between ontology instances in the dynamic ontology knowledge base, generate response feedback containing a structured reasoning chain and return it to the user. S5. Obtain feedback data from task execution, use the evolution engine to extract new knowledge patterns from successful execution trajectories to incrementally update the dynamic ontology knowledge base, and dynamically adjust the confidence level based on the multi-source execution results.

2. The intention recognition method based on dynamic ontology evolution and multi-agent as described in claim 1, characterized in that, S1 specifically includes: S101. Receive a natural language instruction input by the user, call a pre-trained named entity recognition model to extract mention words from the natural language instruction, query the dynamic ontology knowledge base to obtain ontology concepts or ontology instances associated with the mention words as candidate entities, and calculate the semantic similarity between the candidate entities and the mention words to determine the entity linking result; wherein, the calculation formula is: in, Let m be the candidate entity, m be the mentioned word, and C be the input context. and These are preset weights, This represents the similarity between the vector embeddings of the candidate entity and the mentioned word. This indicates the semantic similarity between the candidate entity and the input context; S102. Analyze the predicate part in the natural language instruction, map the predicate part to the predefined intent class in the dynamic ontology knowledge base using the intent classification model, and activate the attribute slots associated with the intent class to obtain a structured initial intent representation; wherein, the attribute slots include target entity slots, indicator type slots and time range slots.

3. The intention recognition method based on dynamic ontology evolution and multi-agent as described in claim 2, characterized in that, S1 further includes: S103. If the natural language instruction is found to contain a new mention that is not recorded in the dynamic ontology knowledge base, and a target ontology instance with matching features already exists in the dynamic ontology knowledge base, then an alias mapping relationship between the new mention and the target ontology instance is established through association rule learning of the evolutionary engine.

4. The intention recognition method based on dynamic ontology evolution and multi-agent as described in claim 1, characterized in that, S2 specifically includes: S201. Based on the definition of the intent category in the dynamic ontology knowledge base, check whether there is a slot in the attribute slot that is marked as required and whose current value is empty. If so, mark it as a missing key slot. S202. Start the multi-source data acquisition mechanism, retrieve the most recently mentioned ontology instance from the session-level context graph in the current session history; obtain real-time alarm data from the monitored target system, and filter out the ontology instances corresponding to the alarm objects that are currently active; query the topology database to obtain ontology instances that have a dependency relationship with the entity link results determined in step S1, and use all retrieved ontology instances as candidate fill values ​​for filling missing key slots. S203. Construct a probabilistic graphical model, using candidate fill values, contextual evidence, and business rule constraints in the dynamic ontology knowledge base as observation variables. Calculate the joint probability of each candidate fill value filling the missing key slot under the current context conditions, and select the candidate fill value with the highest joint probability to fill the missing key slot to generate standardized intent.

5. The intention recognition method based on dynamic ontology evolution and multi-agent as described in claim 4, characterized in that, In S203, the joint probability is calculated as follows: (1) Define a set of characteristic functions and their corresponding weights , used to quantify the degree of matching of candidate padding value v in the current context C; Feature Score The calculation formula is as follows: Specifically, the feature functions include, but are not limited to: Time decay characteristics: This is used to increase the weight of the most recently mentioned ontology instance; State association features: The value is 1 when there is an active alarm for the candidate instance, and 0 otherwise. Ontology constraint features: The value is 1 when the candidate instance type matches the slot definition; otherwise, it is negative infinity or a minimum value. (2) Calculate the joint probability of each candidate filling value using the Softmax function. : in, As a normalization factor, the joint probability is ultimately chosen. The highest candidate fill value is used as the final completion result.

6. The intention recognition method based on dynamic ontology evolution and multi-agent as described in claim 1, characterized in that, S3 specifically includes: S301. Call the central planner to receive the standardized intent, query the predefined intent-subtask decomposition rules in the dynamic ontology knowledge base, and decompose the standardized intent into a task sequence composed of a series of ordered or parallel atomic subtasks. S302. Based on the agent capability-requirement matrix, select the agent with the highest confidence as the executing agent for each atomic subtask, and arrange the selected executing agents and their corresponding atomic subtasks into a collaborative workflow according to the logical dependencies between the atomic subtasks; wherein, the collaborative workflow is a directed acyclic graph, where nodes represent executing agents and atomic subtasks, and edges represent execution order and data flow direction; S303. Before generating the final collaborative workflow, analyze the operation level of each atomic subtask in the collaborative workflow. If a high-risk operation is identified, automatically inject a security control node before the relevant node in the collaborative workflow. The security control node includes at least an authorization verification node, a change window check node, or a manual approval process node.

7. The intention recognition method based on dynamic ontology evolution and multi-agent as described in claim 1, characterized in that, S4 specifically includes: S401. Activate the execution agents corresponding to each node in sequence according to the topological order of the collaborative workflow. The execution agents parse the assigned atomic subtasks, select matching atomic tools from their pre-encapsulated toolset, and execute the invocation. The execution agents pass parameters to the atomic tools through a unified model context protocol and implement a fault tolerance mechanism during execution. If the atomic tool invocation fails or times out, it is retried according to a preset strategy. If the retry fails, the task status is marked as abnormal. S402. The execution agent performs validity verification and cleaning on the raw data returned by the atomic tool, and extracts key information points with business semantics to form intermediate execution results; wherein, extracting key information points includes: extracting indicators of sudden increase or decrease from the raw time series data, or extracting specific error keywords or abnormal patterns from the raw log data. S403. Invoke the integrated reasoning agent to perform causal correlation analysis on the aggregated intermediate execution results based on the logical relationships between ontology instances defined in the dynamic ontology knowledge base, and generate a structured reasoning chain; wherein, the integrated reasoning agent generates the structured reasoning chain according to the following logic based on the logical relationships defined in the dynamic ontology knowledge base: First, the abnormal phenomenon is confirmed based on the intermediate execution results. Then, the abnormal status of the related service is discovered. Subsequently, the root cause log evidence is locked, and finally a complete reasoning chain containing the phenomenon, the correlation, the root cause and the conclusion is generated. S404. Integrate the structured reasoning chain, key data snapshots supporting the structured reasoning chain, and action suggestions generated based on the reasoning conclusions into a response feedback.

8. The intention recognition method based on dynamic ontology evolution and multi-agent as described in claim 1, characterized in that, S5 specifically includes: S501. Utilize the evolutionary engine to continuously monitor and collect the task execution status in step S4, and obtain multi-dimensional feedback data including system execution feedback and user interaction feedback; wherein, the system execution feedback includes execution status, structured inference chain, execution time and resource consumption data, and the user interaction feedback includes explicit ratings or implicit behavioral data of users on response results. S502. When the successful execution of a task is detected, the corresponding collaborative workflow is analyzed, the standardized intent is associated with the atomic tool call sequence that was actually successfully executed, and abstracted into a reusable standardized disposal script, which is then stored as a new disposal strategy in the dynamic ontology knowledge base. S503. Perform cluster analysis on unsuccessfully linked mentions or newly emerging association patterns. If frequent and stable new patterns are found, generate an ontology update proposal, which is then merged into the model layer of the dynamic ontology knowledge base after verification. S504. Based on the final execution result score of this task, dynamically calibrate the confidence levels of the agents participating in the execution in the agent capability-requirement matrix; wherein, the confidence levels are adjusted using a time-decay-based weighted update formula: in, For the updated confidence level, R represents the confidence level before the update, and R is the final score of the task execution result. This is a preset forgetting factor.

9. An intent recognition system based on dynamic ontology evolution and multi-agent systems, based on the intent recognition method based on dynamic ontology evolution and multi-agent systems according to any one of claims 1-8, characterized in that, The system includes: The recognition module is used to receive natural language instructions input by the user, and to perform entity recognition and linking on the natural language instructions using a pre-built dynamic ontology knowledge base, so as to map the natural language instructions into a structured initial intent representation; wherein, the initial intent representation includes the intent category corresponding to the natural language instruction and the attribute slots associated with the intent category; The completion module is used to detect whether there are missing key slots in the attribute slots. If there are, it combines the session-level context graph and real-time alarms and topology data, and uses a probabilistic graphical model to complete the missing key slots and generate clear standardized intents. The matching module is used to decompose the standardized intent into a task sequence composed of multiple atomic subtasks, match a corresponding execution agent for each atomic subtask based on the agent capability-requirement matrix, and generate a collaborative workflow containing the calling order and data dependencies of each execution agent; wherein, the agent capability-requirement matrix includes the confidence of each execution agent in completing a specific type of task; The execution module is used to control each execution agent in the collaborative workflow to sequentially call its pre-encapsulated atomic tools to execute the atomic subtasks in order to obtain multi-source execution results. Based on the logical relationship between ontology instances in the dynamic ontology knowledge base, the module performs correlation analysis on the multi-source execution results, generates response feedback containing a structured reasoning chain, and returns it to the user. The update module is used to obtain feedback data from task execution, extract new knowledge patterns from successful execution trajectories using an evolutionary engine to incrementally update the dynamic ontology knowledge base, and dynamically adjust the confidence level based on the multi-source execution results.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images