Operation and maintenance authorization self-healing method and system based on intelligent agent-human interactive communication

Abstract

This invention discloses a self-healing method and system for operation and maintenance authorization based on interactive communication between an intelligent agent and a human. The method includes: after receiving a monitoring alarm, the intelligent agent first searches its experience memory system for matching historical self-healing experience; if found, it enters a first processing path, sending a first authorization request containing a summary of the historical solution to a human user through at least one interactive communication channel, and after receiving the human user's authorization, invoking a predefined operation and maintenance tool to perform the self-healing operation; if not found, it enters a second processing path, requesting authorization from the human user through the interactive communication channel to invoke a large model for dynamic reasoning, and invoking the corresponding tool to perform the self-healing operation based on the generated dynamic repair solution; after execution, it feeds back the result and deposits the entire process as new experience into the experience memory system. This invention achieves a safe, low-cost, and evolvable operation and maintenance self-healing closed loop through the organic combination of "experience memory-driven dual-path routing" and "forced human-computer interaction authorization," and does not rely on any specific communication platform, thus having wide applicability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-driven IT operations and maintenance (AIOps) technology, and more specifically, to a method, system, storage medium, and intelligent agent device that combines experience memory retrieval, dual-path human-machine collaborative authorization, large-model dynamic reasoning, and experience closed-loop self-learning in operations and maintenance alarm processing. Background Technology

[0002] With the widespread adoption of distributed systems, microservices, and cloud-native architectures, operational complexity is increasing exponentially. Traditional rule-based alerting and manual intervention models suffer from the following drawbacks: slow response, reliance on expert experience, and long mean time to recovery (MTTR); inability to accumulate and reuse knowledge, leading to repetitive analysis and approval of similar faults; lack of secure human-machine collaboration mechanisms, with existing automation tools either entirely manual or posing security risks due to fully automated execution; instant messaging tools serving only as one-way notification channels, failing to create a two-way closed loop of "approval equals execution"; and the inability of fixed rules or predefined scripts to handle novel, unseen faults, still requiring in-depth manual analysis. Summary of the Invention

[0003] I. Technical problems to be solved This invention aims to solve the following technical problems: (1) How to enable the intelligent agent to prioritize the retrieval of existing operation and maintenance documents and historical solutions after receiving an alarm, so as to avoid unnecessary large model calls and reduce costs and delays; (2) How to implement the dual-path diversion mechanism: known problems follow the predefined tool execution path, and unknown problems follow the large model inference path after authorization, thereby achieving resource optimization; (3) How to achieve a secure and auditable human-machine authorization loop through any interactive communication channel (including but not limited to instant messaging, email, SMS, web pop-ups, etc.) without relying on a specific platform; (4) How to write the complete context of each self-healing (whether from a predefined scheme or a large model generation scheme) into a hierarchical memory system to form experience accumulation, so that subsequent similar alarms can be directly called or automatically executed without authorization, and achieve continuous evolution. Technical solution

[0004] This invention provides a self-healing method for operation and maintenance authorization based on interactive communication between intelligent agents and humans, comprising the following core steps: (i) Alarm Reception and Experience Retrieval (S1-S2): After receiving an alarm event pushed by the monitoring system, the intelligent agent first searches the experience memory system for matching historical self-healing experiences. The experience memory system stores operation and maintenance documents, historical operation records, and patterned experiences extracted from successful self-healing cases.

[0005] (ii) Path splitting (S3): Path splitting is performed based on the search results.

[0006] First processing path (known issue): When a matching historical self-healing experience is retrieved, the agent does not execute it directly. Instead, it sends a first authorization request message to the human user through at least one interactive communication channel. This message includes a summary of the historical plan and interactive authorization controls (such as an approve / deny button). After receiving the first authorization instruction triggered by the human user through the operation of the authorization controls, the agent invokes a predefined maintenance tool associated with that experience to perform the self-healing operation.

[0007] Second processing path (unknown issue): When no matching historical self-healing experience is found, the agent sends a second authorization request message to the human user through the interactive communication channel, requesting authorization to use the large model for inference analysis. After receiving the second authorization instruction from the human user, the agent invokes the large model service, generates a dynamic repair plan based on the alarm context, system status, etc., and then parses and invokes the corresponding operation and maintenance tools to execute it.

[0008] (iii) Result Feedback (S4): The agent obtains the execution result of the self-healing operation and provides feedback to the human user through the interactive communication channel.

[0009] (iv) Experience accumulation (S5): The agent writes the current alarm event, the self-healing scheme adopted (including predefined schemes or schemes generated by large models), authorization records and execution results as new experience into the experience memory system, which is used to update historical self-healing experience for subsequent alarm retrieval.

[0010] (v) Automatic execution mode (preferred): When there is an event in the experience memory system that is more similar to the current alarm event than the preset threshold and the corresponding historical self-healing experience is marked as "verified", the agent can skip the authorization step and directly execute the self-healing operation. After the execution is completed, only a notification message is sent, thereby realizing a smooth evolution from "human-machine collaboration" to "automatic execution".

[0011] Compared with the prior art, the present invention has the following beneficial effects: (1) High security: All automated operations (whether based on historical experience or large model generation schemes) must obtain explicit authorization from human users through interactive communication channels, support identity verification and timeliness control, eliminate the risk of misoperation that may be caused by fully automated operation and maintenance, and meet enterprise-level security compliance requirements.

[0012] (2) Cost and efficiency optimization: Prioritize the retrieval of lightweight experience memory and only call the large model when there is no historical experience, so as to avoid calling the expensive large model for every alarm, significantly reduce the inference cost and shorten the average response time.

[0013] (3) Continuous self-evolution capability: Successful self-healing experience is structured and written into the hierarchical memory system. The more the system is used, the more intelligent it becomes. It can be gradually upgraded from "human-machine collaborative authorization mode" to "authorization-free automatic execution mode", realizing the continuous improvement of the level of operation and maintenance automation.

[0014] (4) Platform independence: By abstracting the interactive communication channel interface (adapter pattern), it can simultaneously support multiple carriers such as instant messaging robots, enterprise-level collaborative software, SMS gateways, email, and web page pop-ups, adapting to the existing infrastructure of different enterprises, and not being bound to any specific vendor platform.

[0015] (5) Complete audit and traceability: The entire process records alarms, retrieval results, authorization instructions, execution process and results to meet compliance and accident traceability requirements. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of an embodiment of the present invention.

[0017] Figure 2 This is a flowchart of an embodiment of the self-healing operation and maintenance method of the present invention.

[0018] Figure 3 This is a flowchart of the first processing path (known problem) of an embodiment of the present invention.

[0019] Figure 4 This is a flowchart of the second processing path (unknown problem) sub-process of an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of experience stratification and sedimentation according to an embodiment of the present invention.

[0021] Figure 6 This invention provides a closed-loop feedback diagram illustrating the process from monitoring → SysOp → operator authorization → SysOp self-healing. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It should be noted that the interactive communication channels described in this invention include, but are not limited to, instant messaging robots, enterprise-level collaborative software robots, SMS gateways, email, webpage pop-ups, and any other communication method capable of enabling bidirectional message passing and interactive control triggering between intelligent agents and human users. For ease of description, the following embodiments use a specific channel as an example, but those skilled in the art should understand that this invention is not limited to that specific channel.

[0023] Example 1: Evolution of Authorized Self-Healing and Automated Execution Based on Known Experience This embodiment focuses on demonstrating the workflow of the first processing path and its evolution towards an automated execution mode.

[0024] Scenario: A company's microservice "order-service" runs on a container orchestration platform (such as Kubernetes). The monitoring system (e.g., Prometheus + Alertmanager) detects that the service instance crashes and restarts multiple times within 10 minutes, triggering an alert: "order-service instance crash rate > 10%".

[0025] System Deployment: The SysOp Agent has been deployed, and the experience memory system is running. The communication module is configured with adapters for various interactive communication channels, including but not limited to instant messaging robots, email, and SMS gateways. This example uses the instant messaging robot channel for demonstration.

[0026] Step 1: Alarm Reception and Experience Retrieval The monitoring system pushes an alert to the intelligent agent. The alert content includes: Service Name=order-service, Instance ID=pod-abc123, Crash Rate=12%.

[0027] The agent extracts the alarm feature fingerprint (service name + "crash" keyword) and searches it in the L2 layer (abstract knowledge memory) and L3 layer (plaintext knowledge base) of the experience memory system. A matching record was successfully retrieved: three months ago, a similar alarm occurred for the same service, and the solution at that time was "restart the service instance," which was executed successfully. This solution has been marked as "verified," and the similarity calculation result is 0.96 (higher than the preset threshold of 0.85).

[0028] Step 2: Path splitting Because a matching historical self-healing experience was retrieved, the agent enters the first processing path.

[0029] Step 3: Send authorization request The intelligent agent invokes the communication module to send an interactive card message to the client of the on-duty maintenance engineer via a pre-configured instant messaging robot channel. The content of the card message is: Alert Name: Frequent crashes of the order-service Pod Severity: High Historical solution: Restart the order-service instance (Estimated impact: Service interruption of approximately 15 seconds) Action: [Approve] [Reject] The "Approve" button in the card is configured to send a request containing the user's identifier, timestamp, and authorization decision to the agent's preset callback interface when the user clicks it.

[0030] Step 4: Receive Authorization and Execute The on-duty engineer clicks "Approve" on the mobile or computer client. The instant messaging robot server sends authorization data to the agent's callback interface.

[0031] After receiving the request, the agent performs authorization verification: ① Verifies that the user identifier has the operation permission for "order-service" (passes); ② Verifies that the authorization time is within the preset valid window after the alarm was generated (passes); ③ Verifies that the operation scope matches the alarm (passes). After successful verification, the agent calls the predefined operation and maintenance tool to execute the command: `restart_service --instance=order-service`. The command executes successfully, and the crash rate drops to 0% after the service instance restarts.

[0032] Step 5: Results Feedback The agent sent a feedback card again through the same instant messaging robot channel: "Self-healing operation has been executed, order-service has been restarted, current crash rate is 0%". Step 6: Accumulating Experience The agent writes the alarm event, the solution adopted (restarting the instance), authorization records, and successful execution results into the experience memory system, increments the success count of the solution, and updates the most recent success timestamp.

[0033] Evolving to automatic execution mode: When the alarm reappears, the agent retrieves the same solution with a similarity of 0.96, and both the "verified" flag and success count for this solution have met the criteria. The agent determines that the preset automatic execution conditions are met (e.g., similarity > 0.95 and success count ≥ 5), so it skips the authorization step and directly executes the restart operation. Upon completion, only one notification message is sent: "Automatic restart of order-service has been successfully performed. For rollback, please contact us through the appropriate channels." Engineers can complete the self-healing process without needing to click for approval.

[0034] The difference from existing technologies: Existing technologies (such as CN121168679B) only disclose static hierarchical storage of memory, but do not disclose the complete dynamic decision-making closed loop of "retrieval results → path diversion → authorized execution → experience reinforcement → automatic execution evolution". In this embodiment, the memory layer is not passively stored, but actively participates in decision-making and supports a smooth evolution from "human-machine collaboration" to "automatic execution", which is a significant advancement of this invention.

[0035] Example 2: Unknown Fault Handling and Experience Generation Based on Large Model Inference This embodiment focuses on demonstrating the workflow of the second processing path and the generation and accumulation of new experiences.

[0036] Scenario: A newly launched service, "recommend-service," is experiencing an alert: "Database connection pool utilization 99%." This issue has never occurred before, and there is no relevant solution in the operations and maintenance documentation. This example uses email as the interactive communication channel to demonstrate the compatibility of this invention with multiple channels.

[0037] Step 1: Alarm Reception and Experience Retrieval After receiving the alarm, the agent searches its experience memory system. Since there are no matching records (similarity is below the preset threshold of 0.6), the search results are empty.

[0038] Step 2: Path splitting Since no matching experience was found, the agent enters the second processing path.

[0039] Step 3: Send an authorization request (request to use the large model) The agent invokes the communication module to send an authorization request email to the senior operations engineer via email. The email content is as follows: Subject: [Authorization Request] Unknown Alert: recommend-service database connection pool usage is 99% text: The system did not find any historical solutions. Requesting authorization to invoke the large model for root cause analysis and dynamic solution generation.

[0040] Please click the following link to authorize (the link is valid for 15 minutes): xxxxx.xxxx.com / authorize?token=xxxx The link leads to the agent's authorization webpage, which contains "Approve" and "Deny" buttons.

[0041] Step 4: Receive Authorization The engineer clicks the authorization link in the email, opens the authorization page in their browser, and clicks "Confirm Authorization." The webpage then calls back the agent's authorization interface, carrying the authorization credentials.

[0042] Step 5: Use the large model to generate a dynamic repair solution After the agent verifies the authorization, it will organize the following information into prompt words and send them to the large model API: Alarm context: "recommend-service database connection pool utilization is 99%" System status data: Current maximum number of connections = 50, active number of connections = 49, "connection leak" warning appears in the service log. Relevant information retrieved from the experience memory system: The service's code recently underwent a commit involving database query logic. The large model returns analysis results and solutions: Root cause analysis: The maximum number of connections in the database connection pool is set too low, and there is a code defect where connections are not released.

[0043] Temporary fix: Temporarily increase the maximum number of connections from 50 to 200 by executing the command: update_db_pool_config --max=200; and restart the recommend-service.

[0044] Long-term recommendation: Notify the development team to investigate connection leaks.

[0045] Step 6: Perform the repair The intelligent agent parses the solution returned by the large model and then sequentially calls the operation and maintenance tools to execute it: Execute the database configuration update command (using a database management tool). Restart the recommended-service instance by calling the container orchestration platform API. After execution, the connection pool utilization rate dropped to 45%.

[0046] Step 7: Results Feedback The agent sends the execution results to the engineer via email: "The database connection pool has been expanded to 200, the service has been restarted, and the current utilization rate is 45%." Step 8: Accumulating Experience The agent will use the characteristics of this alarm (recommend-service + "database connection pool exhausted"), the repair plan generated by the large model (temporarily expanding the connection pool + restarting the service), and the execution result (success) as new self-healing experience, and write them into the experience memory system in a structured way: Writing to L2 layer (abstract knowledge memory): Generating a mapping between alarm feature fingerprints and repair operation sequences. Write to L3 layer (plaintext knowledge base): Store the analysis report of the large model as an operation and maintenance document. If a similar alarm occurs again thereafter, the agent will directly retrieve the experience from the first processing path, enabling rapid authorization and self-healing.

[0047] The difference from existing technologies: Existing solutions centered on large models (such as CN121052587B) treat the large model as a central scheduler for all tasks, calling the large model for every alarm, which is costly and has high latency. In contrast, this invention downgrades the large model to a "last resort solver for unknown problems," only activating it after experience retrieval fails. After successful self-healing, the solution is preserved as experience for direct reuse later, avoiding repeated calls to the large model. This is a significant technological advancement.

[0048] Example 3: Specific Operations of Layered Experience Memory This embodiment combines Figure 5 It details the retrieval and retention mechanisms of the experience memory system and emphasizes its differences from general hierarchical memory.

[0049] The storage structure of L2 layer (abstract knowledge memory): The L2 layer stores "patterned experiences" extracted from historical operation and maintenance events. Each experience record contains three core fields: alert_fingerprint: Alert feature fingerprint (such as a vector representation of service name + fault type keywords) remediation_actions: Successful repair operation sequence (e.g., ["restart_service", "scale_up"]) outcome: Execution result (success / failure) and confidence score Search and matching process: When a new alert arrives, the agent extracts its alert_fingerprint and calculates its similarity (using cosine similarity) with the fingerprints of all experience records in the L2 layer. If the similarity exceeds a preset threshold (e.g., 0.85), it is considered a successful match, and the corresponding remediation_actions are returned.

[0050] The process of accumulating experience: In Example 2, the natural language solution returned by the large model is "Temporarily increase the maximum number of connections to 200, and restart the recommend-service". The agent's parsing module performs the following transformations: Extract key operations from natural language: "update max_connections=200", "restartservice recommend-service" Convert these operations into a standardized sequence of operations: ["db_set_param:max_connections=200","restart_service:recommend-service"] Extract alarm feature fingerprint (service name + "database_connection_pool") Write (fingerprint, operation sequence, result = success) as a new experience into the L2 layer. Meanwhile, the complete analysis report returned by the large model (including root cause analysis and long-term recommendations) is written into the L3 layer (plaintext knowledge base) in structured or unstructured text form for subsequent manual review or RAG retrieval.

[0051] The difference from existing technologies: Existing hierarchical memory technologies (such as CN121168679B) only disclose a general "event-pattern-knowledge" hierarchy, but do not design a dedicated experience triplet structure of "alarm feature fingerprint → repair action sequence → result feedback" for operation and maintenance scenarios. The L2 layer of this invention stores executable operation and maintenance experience, rather than abstract narrative text, which allows the retrieval results to be directly used to call predefined operation and maintenance tools, forming a true automation closed loop.

[0052] Example 4: Multi-channel concurrent notifications and authorization (optional) This embodiment illustrates the invention's ability to simultaneously support multiple interactive communication channels.

[0053] In real-world operations and maintenance scenarios, different maintenance personnel may prefer different communication methods, and some alarms may need to notify multiple people simultaneously. The agent's communication interaction module, through an adapter pattern, can be configured with multiple channels simultaneously (such as instant messaging robots, email, and SMS gateways). When an authorization request needs to be sent, the agent can select one or more channels to send the authorization request message based on preset policies (such as alarm level, duty schedule, and user preferences). For example: For high-priority alarms, card messages are sent via instant messaging robots and authorization link SMS messages are sent via SMS gateways to ensure that on-duty personnel can respond immediately.

[0054] For low-to-medium priority alarms, messages are sent only through enterprise-level collaborative software robots.

[0055] Users can approve or deny authorization through the authorization control of any channel. After receiving the authorization instruction from any channel, the intelligent agent updates the message status on other channels to avoid duplicate processing.

[0056] This multi-channel concurrency capability is a concrete manifestation of the beneficial effect of the "platform independence" of this invention, further improving the availability and flexibility of the system.

[0057] Core Creative Theory The core inventiveness of this invention lies in the overall technical solution defined in claim 1, specifically embodied in the following three aspects: First, a dual-path routing architecture driven by experience memory.

[0058] Existing technologies either employ a centralized scheduling model that calls the entire large model (high cost and latency) or a static memory storage structure (lacking decision-making linkage). This invention, for the first time in the field of self-healing operations, proposes a dual-path routing architecture that uses historical experience retrieval results as decision nodes: if matching experience exists, it follows a lightweight, predefined tool execution path; otherwise, it follows a heavyweight, large model inference path. This architecture is not a simple "if-else" judgment, but a resource optimization strategy designed based on a deep understanding of the "Pareto Principle" (80% of faults are recurring, 20% are new) in operations scenarios, demonstrating significant technological advancement.

[0059] Second, the organic combination of mandatory human-computer interaction authorization and experience closure.

[0060] This invention uses "sending messages containing authorization controls via interactive communication channels" as a pre-operation safety valve for all automated operations. Whether it's the historical experience execution of the first path or the large-scale model solution execution of the second path, human user operation of the authorization controls is required to trigger it. This mechanism solves the problems of high risk in fully automated execution and the inability to form a closed loop through one-way notification in existing technologies. Simultaneously, the results of authorized execution are forcibly written back to the experience memory system, forming a complete closed loop of "retrieval → authorization → execution → storage," enabling the system to continuously evolve from "human-machine collaboration" to "automatic execution."

[0061] Third, a platform-independent communication abstraction layer.

[0062] This invention abstracts interactive communication channels into a unified interface using the adapter pattern, enabling intelligent agents to simultaneously support multiple carriers such as instant messaging robots, enterprise-level collaborative software, email, SMS, and web pop-ups, without relying on any specific vendor platform. This is fundamentally different from existing customized solutions that are only for a single instant messaging tool (such as Lark or DingTalk), and has broader applicability and portability.

[0063] Fourth, the overall solution is not publicly documented in existing technologies.

[0064] A search revealed that existing technologies (including CN121168679B, CN121052587B, etc.) do not disclose or teach the complete technical combination of "S2 experience retrieval → S3 path routing → S3A1 / S3B1 sending authorization controls through any interactive communication channel → S3A3 / S3B4 execution → S5 experience accumulation". The steps of this invention are closely interconnected and synergistic: without experience retrieval, path routing is impossible; without mandatory human-computer interaction authorization, security cannot be guaranteed; without experience accumulation, the system cannot evolve; without platform-independent communication abstraction, the solution cannot be widely applied. This overall combined solution is pioneering in the field of operation and maintenance automation and possesses significant ingenuity.

[0065] In summary, this invention has outstanding substantive features and significant progress compared to the prior art, and meets the requirements of Article 22, Paragraph 3 of the Patent Law regarding inventiveness.

Claims

1. A self-healing operation and maintenance authorization method based on interactive communication between intelligent agents and humans, characterized in that, Includes the following steps: Step S1: The agent receives an alarm event from the monitoring system; Step S2: The agent searches the experience memory system to see if there is any historical self-healing experience that matches the alarm event; Step S3: Perform path routing based on the search results: If a matching historical self-healing experience exists, the first processing path is executed; If no matching historical self-healing experience exists, the second processing path is executed; The first processing path includes: Step S3A1: The intelligent agent sends a first authorization request message to the human user through at least one predefined interactive communication channel associated with the human user. The first authorization request message contains an operation summary of the matched historical self-healing experience and a first authorization control. Step S3A2: The intelligent agent receives a first authorization instruction triggered by the human user operating the first authorization control through the interactive communication channel; Step S3A3: In response to the first authorization instruction, the agent invokes one or more predefined operation and maintenance tools associated with the matched historical self-healing experience to perform self-healing operations; The second processing path includes: Step S3B1: The intelligent agent sends a second authorization request message to the human user through the interactive communication channel. The second authorization request message is used to request authorization to call the large model for inference analysis and includes a second authorization control. Step S3B2: The intelligent agent receives a second authorization instruction triggered by the human user operating the second authorization control through the interactive communication channel; Step S3B3: In response to the second authorization instruction, the agent invokes the large model service to generate a dynamic repair scheme based on the context information of the alarm event; Step S3B4: The intelligent agent parses the dynamic repair scheme and calls the corresponding operation and maintenance tools to perform self-healing operations; Step S4: The intelligent agent obtains the execution result of the self-healing operation and provides feedback to the human user through the interactive communication channel; Step S5: The intelligent agent writes the alarm event, the self-healing scheme adopted, the authorization record, and the execution result as new experience into the experience memory system to update the historical self-healing experience.

2. The method according to claim 1, characterized in that, It also includes an automatic execution mode: when the agent retrieves a similarity to the current alarm event in the experience memory system that exceeds a first preset threshold, and the corresponding historical self-healing experience is marked as "verified", the agent skips steps S3A1 to S3A2, directly executes the self-healing operation, and sends a notification message to the human user after execution.

3. The method according to claim 1, characterized in that, The experience memory system adopts a hierarchical memory architecture oriented towards operational experience, including: L0 layer: Short-term session memory layer, used to store the temporary context of the current operation and maintenance interaction; L1 layer: Long-term user memory layer, used to store the preferences of operation and maintenance personnel and historical operation patterns across sessions; L2 layer: Abstract knowledge memory layer, used to store patterned experience extracted from historical operation and maintenance events. The patterned experience includes alarm feature fingerprints and corresponding successful repair operation sequences. L3 layer: Plaintext knowledge base layer, used to store structured plaintext operation and maintenance documents and operation manuals; L4 layer: The knowledge layer of the large model, used to provide general reasoning capabilities; In step S5, the accumulated experience is preferentially written into the L2 and L3 layers, and an index is established for subsequent retrieval and matching.

4. The method according to claim 1, characterized in that, The first authorization request message and the second authorization request message are interactive card messages or structured messages containing interactive controls. The messages include alarm name, severity, expected impact range, and at least one approval control and at least one denial control. The approval control and denial control are configured to send a request containing user identifier, timestamp and authorization decision to a preset callback interface of the agent when triggered.

5. The method according to claim 1, characterized in that, The step S3B3, which calls the large model service to generate a dynamic repair scheme, specifically includes: the agent organizing the message content of the alarm event, the real-time system status data, and relevant information retrieved from the experience memory system into prompt words, inputting them into the large model, and the large model generating natural language repair steps or executable command sequences.

6. The method according to claim 1, characterized in that, The interactive communication channel is selected from one or more of the following: instant messaging robot, enterprise-level collaborative software robot, SMS gateway, email system, mobile application push notification, web-based message center; the intelligent agent and the interactive communication channel achieve platform-independent bidirectional asynchronous communication through the adapter pattern.

7. The method according to claim 1, characterized in that, The predefined operation and maintenance tools called in step S3A3 include at least one of the following: restarting services, scaling up or down instances, clearing cache, rolling back configurations, switching traffic, terminating abnormal agents, and creating new agent instances.

8. The method according to claim 1, characterized in that, It also includes an authorization verification step: when the intelligent agent receives the first authorization instruction or the second authorization instruction, it verifies the validity of the authorization instruction. The verification includes verifying whether the authorized user has the operation permission, whether the authorization time is within the valid window, and whether the authorized operation scope matches the current alarm. If the verification fails, it refuses to execute and notifies the human user.

9. A self-healing operation and maintenance authorization system based on interactive communication between intelligent agents and humans, characterized in that, include: The monitoring access module is used to receive alarm events from the monitoring system; An experience memory system for storing and retrieving historical self-healing experiences; The path decision module is connected to the experience memory system and is used to select a first processing path or a second processing path based on the retrieval results. A communication interaction module is used to make authorization requests and receive instructions from human users through at least one interactive communication channel. A predefined execution module is used to call a predefined operation and maintenance tool to perform a self-healing operation under the first processing path according to the received first authorization instruction; The large model inference module is used to generate a dynamic repair solution by calling the large model service according to the received second authorization instruction under the second processing path; The tool execution module is used to call the operation and maintenance tools to execute the self-healing operation or the dynamic repair scheme; The result feedback module is used to obtain the execution result and feed it back to the human user through the communication interaction module. The experience accumulation module is used to write the alarm events, the self-healing solutions adopted, authorization records, and execution results into the experience memory system.

10. The system according to claim 9, characterized in that, It also includes an automatic execution control module, which, when the similarity of the matching historical self-healing experience returned by the experience memory system exceeds a preset threshold and is marked as "verified", bypasses the communication interaction module and directly triggers the predefined execution module to perform the self-healing operation, and notifies the human user after the execution is completed.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

12. An intelligent agent device, characterized in that, include: At least one processor; A memory coupled to the processor, wherein instructions are stored in the memory; A communication interface for exchanging data with monitoring systems, operation and maintenance tools, large model services, and at least one interactive communication channel; When the instruction is executed by the processor, it causes the intelligent agent device to perform the method of any one of claims 1 to 8.

Images