Federated learning based end-side ai agent cooperative fault diagnosis method and device

Abstract

The application provides a federated learning-based end-side AI agent cooperative fault diagnosis method and device, relates to the technical field of federated learning, and the method comprises the following steps: generating a task requirement description according to a diagnosis task cluster received by a server; each end side extracts a knowledge skeleton from a local agent and performs topology reconstruction through knowledge distillation; matching the local diagnosis target by matching with the task requirement description to determine a cognitive data packet; the server receives the cognitive data packet, activates an analysis engine, and performs double-path operation management; driving an extended group to expand a global knowledge base; and matching and fusing the local diagnosis target through the local cognitive topology and the global knowledge base, and then updating learning and task diagnosis management are performed on the agent. Through the application, the problem that end-side diagnosis cognition is difficult to be structured and aligned in the prior art can be solved, federated cooperative fusion driven by tasks is realized, and the technical effect of improving the global diagnosis reliability is achieved.

Description

Technical Field

[0001] This application relates to the field of federated learning technology, and in particular to a method and device for collaborative fault diagnosis of edge AI agents based on federated learning. Background Technology

[0002] With the large-scale deployment of industrial equipment and the continuous development of edge intelligence technology, edge AI agents are gradually being widely used in equipment operation status monitoring and fault diagnosis scenarios. By completing data analysis and preliminary diagnosis on-site, communication load can be effectively reduced and response efficiency improved. At the same time, combined with collaborative mechanisms such as federated learning, the overall improvement of multi-end diagnostic capabilities can be achieved without aggregating raw data. That is, not only is communication pressure reduced and response speed improved, but with the help of collaborative mechanisms such as federated learning, the comprehensive capabilities of multi-end diagnosis can be improved without aggregating raw data, thereby meeting the comprehensive requirements of complex industrial systems for security, real-time performance and intelligence.

[0003] Currently, existing edge-side collaborative fault diagnosis technologies mostly focus on the periodic interaction of model parameters or gradients. They are largely limited to the periodic interaction of model parameters or gradients, or rely on edge-side uploads of diagnostic results with fixed structures. In practical applications, the diagnostic cognition of edge agents is often treated as homogeneous output, lacking the expression and utilization of the edge's internal diagnostic reasoning structure, feature relationships, and their matching degree with specific diagnostic tasks. In other words, this method ignores the expression and utilization of the edge's internal reasoning logic, feature relationships, and their matching degree with specific diagnostic tasks. This results in the server side typically only being able to perform global fusion based on limited statistical information or model update results, making it difficult to characterize the cognitive differences and effective contributions of different edge agents under different diagnostic tasks.

[0004] In summary, existing technologies suffer from the technical problem that the edge-side diagnostic cognitive structure is difficult to express in a structured way and to be specifically aligned with the specific diagnostic tasks. This leads to the server side being unable to effectively identify the task relevance and effectiveness of the edge-side diagnostic cognition during the federated collaboration process, further affecting the accuracy of global knowledge fusion and the reliability of diagnostic results during edge-side collaborative fault diagnosis. Summary of the Invention

[0005] The purpose of this application is to provide a method and device for collaborative fault diagnosis of edge AI agents based on federated learning, in order to solve the technical problem in the prior art that the edge diagnostic cognitive structure is difficult to express in a structured way and to be specifically aligned with the specific diagnostic task, which makes it impossible for the server side to effectively identify the task relevance and effectiveness of the edge diagnostic cognition in the federated collaboration process, and further affects the accuracy of global knowledge fusion and the reliability of the diagnostic results in the edge collaborative fault diagnosis process.

[0006] In view of the above problems, this application provides a method and device for collaborative fault diagnosis of edge AI agents based on federated learning.

[0007] Firstly, this application provides a method for collaborative fault diagnosis of edge AI agents based on federated learning, implemented through an edge AI agent collaborative fault diagnosis device based on federated learning. The method includes: a server receiving a cluster of diagnostic tasks for the first cycle; generating task requirement descriptions based on the diagnostic task clusters and distributing them to each edge; each edge extracting a knowledge skeleton from its local agent and reconstructing its topology through knowledge distillation; matching this skeleton with the task requirement descriptions; combining this skeleton with local diagnostic targets; determining cognitive data packets; and sending these packets to the server using protocol steganography; the server receiving the cognitive data packets; activating the parsing engine; and performing dual-path operation management; wherein the dual-path operation management includes: determining diagnostic primitives based on edge consensus using local cognitive topology; driving an extension group to perform pseudo-expansion and edge agent verification; expanding the global knowledge base; and matching and fusing local diagnostic targets using local cognitive topology and the global knowledge base, and distributing these targets to the corresponding edge agents for update learning and task diagnosis management.

[0008] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: the first edge extracting a first knowledge skeleton from the first local agent through knowledge distillation; defining nodes by feature concepts, defining hyperedges by diagnostic reasoning paths, and defining edge weights by path confidence based on the first knowledge skeleton to determine a first local topology, wherein the feature concepts at least include key support vectors, feature importance ranking, and anomaly detection threshold range; and performing counterfactual evaluation and pruning on each diagnostic reasoning path in the first local topology to determine a first local cognitive topology.

[0009] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: the server periodically publishing task requirement descriptions; each edge performing a matching of its local cognitive topology with the task requirement description to determine a matching cognitive topology, wherein each edge corresponds to one matching cognitive topology, and each matching cognitive topology has a validity proof; the first type of edge with diagnostic needs generates a local diagnostic target based on the cognitive diagnostic type; each edge integrating the matching cognitive topology and the local diagnostic target as a cognitive data packet is sent to the server via protocol steganography.

[0010] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: introducing a steganography pattern based on dynamic dictionary mapping into the standard read / write command field of the application layer protocol, wherein the steganography pattern uses read / write commands as a carrier; encoding the cognitive data packet into a read / write operation request using the steganography pattern according to the shared key and communication context; and sending the read / write operation request to the server.

[0011] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: the server receiving cognitive data packets from each edge, deconstructing the matching cognitive network according to the embedded parsing engine, and determining multiple sets of cognitive paths; performing structural consistency analysis on the multiple sets of cognitive paths through continuous coherence analysis to determine a cognitive path group, wherein the cognitive path group satisfies the structural consistency condition; and storing the cognitive path group as a diagnostic primitive for edge consensus in the global knowledge base.

[0012] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: developing an extension component in the server and establishing interaction between the extension component and the global knowledge base; the extension component retrieves the diagnostic primitive from the global knowledge base and extends it to generate a false diagnostic path, wherein the extension method includes at least combination and variation; and the false diagnostic path is distributed to a preset proportion of edge agents for local verification.

[0013] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: each edge agent performing local verification performs knowledge distillation and local topology transformation on the local verification data and encapsulates it into a local verification package; the local verification data package is sent to the server through protocol steganography, triggering the parsing engine to perform deconstruction analysis, filtering out valid cognitive paths that have been successfully verified and meet structural consistency, and storing them in the global knowledge base.

[0014] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: the parsing engine reads the first local diagnostic target in the cognitive data packet, matches and fuses the multiple sets of cognitive paths with the global knowledge base to determine the first target diagnostic path; encapsulates the first target diagnostic path and sends it to the first edge via protocol steganography; the agent on the first edge learns and updates according to the first target diagnostic path.

[0015] Preferably, the edge AI agent collaborative fault diagnosis method based on federated learning further includes: the server releasing a batch of task requirement descriptions in the next cycle; and performing knowledge optimization and diagnosis between the edge AI agent and the server for the batch of task requirement descriptions, and performing periodic polling processing.

[0016] Secondly, this application also provides a collaborative fault diagnosis device for edge AI agents based on federated learning, used to execute the collaborative fault diagnosis method for edge AI agents based on federated learning as described in the first aspect, including: a diagnostic task cluster receiving module, used by the server to receive a diagnostic task cluster for the first cycle; a cognitive data packet determination module, used to generate a task requirement description according to the diagnostic task cluster, and distribute it to each edge, each edge extracting a knowledge skeleton from the local agent and performing topology reconstruction through knowledge distillation, matching it with the task requirement description, and jointly determining the cognitive data packet with the local diagnostic target, and sending it to the server using protocol steganography; a dual-path operation management execution module, used by the server to receive the cognitive data packet, activate the parsing engine, and execute dual-path operation management; wherein, the dual-path operation management includes: a global knowledge base expansion unit, used to drive the expansion group to perform pseudo-expansion and edge agent verification by determining the diagnostic primitives based on the edge consensus of the local cognitive topology, and expanding the global knowledge base; a task diagnosis management unit, used to match and fuse the local diagnostic target through the local cognitive topology and the global knowledge base, and distribute it to the corresponding edge agent for update learning and task diagnosis management.

[0017] The technical solution provided in this application has at least the following technical effects or advantages: by achieving the technical goal of structured modeling and federated collaborative fusion of end-side diagnostic cognition driven by task requirements, the server can perform parsable, aligned and verifiable processing of end-side diagnostic cognition without exposing the original end-side data and complete model, thereby improving the technical effect of global knowledge fusion consistency and the reliability of collaborative fault diagnosis results.

[0018] The above description is merely an overview of the technical solution of this application. To enable a clearer understanding of the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1This is a flowchart illustrating the edge AI agent collaborative fault diagnosis method based on federated learning proposed in this application.

[0021] Figure 2 This is a schematic diagram of the edge AI agent collaborative fault diagnosis device based on federated learning, as described in this application.

[0022] Explanation of reference numerals in the attached diagram: Diagnostic task cluster receiving module 1, cognitive data packet determination module 2, dual-path operation management and execution module 3, global knowledge base extension unit 31, and task diagnosis management unit 32. Detailed Implementation

[0023] This application provides a method and device for collaborative fault diagnosis of edge AI agents based on federated learning. It addresses the technical problem in existing technologies where the difficulty in structurally representing the edge-side diagnostic cognitive structure and aligning it with specific diagnostic tasks leads to the server's inability to effectively identify the task relevance and effectiveness of edge-side diagnostic cognition during the federated collaboration process. This further affects the accuracy of global knowledge fusion and the reliability of diagnostic results in edge-side collaborative fault diagnosis. The application achieves the technical goal of task-demand-driven structured modeling and federated collaborative fusion of edge-side diagnostic cognition. It enables the server to perform parsable, alignable, and verifiable processing of edge-side diagnostic cognition without exposing the original edge-side data and complete model, thereby improving the consistency of global knowledge fusion and the reliability of collaborative fault diagnosis results.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides a collaborative fault diagnosis method for edge AI agents based on federated learning, which is applied to an edge AI agent collaborative fault diagnosis device based on federated learning, and specifically includes the following steps: The server receives the diagnostic task cluster for the first cycle.

[0026] Specifically, the server receiving the first-cycle diagnostic task cluster refers to the central server acting as a task scheduling and aggregation node receiving a set of multiple diagnostic tasks issued by the upper-level management module within a preset time period during the startup phase of the collaborative fault diagnosis method. Here, the server represents a centralized processing unit with computing, storage, and task orchestration capabilities. The first cycle is used to define the initial or current round of the diagnostic process to achieve phased federated collaborative processing. The diagnostic task is used to characterize the fault identification, analysis, or reasoning objectives required for the device or its operating status. The task cluster represents a collection of multiple diagnostic tasks that are uniformly organized, scheduled, and managed within the same cycle. This collection provides a unified input basis for subsequent task requirement description generation and end-side collaborative processing.

[0027] The diagnostic task cluster generates a task requirement description, which is then distributed to each endpoint. Each endpoint extracts a knowledge skeleton from the local agent and performs topology reconstruction through knowledge distillation. By matching the skeleton with the task requirement description and combining it with the local diagnostic target, a cognitive data packet is determined and sent to the server using protocol steganography.

[0028] Furthermore, this application also includes: the first end-side extracting a first knowledge skeleton from the first local agent through knowledge distillation; based on the first knowledge skeleton, defining nodes with feature concepts, defining hyperedges with diagnostic reasoning paths, defining edge weights with path confidence, and determining a first local topology, wherein the feature concepts at least include key support vectors, feature importance ranking, and anomaly detection threshold range; performing counterfactual evaluation and pruning on each diagnostic reasoning path in the first local topology to determine a first local cognitive topology.

[0029] Furthermore, this application also includes: the server periodically publishing task requirement descriptions; each end-side performing a matching of its local cognitive topology with the task requirement descriptions to determine a matching cognitive topology, wherein each end-side corresponds to one matching cognitive topology, and each matching cognitive topology identifier has a validity proof; a first type of end-side with diagnostic needs generating a local diagnostic target based on the cognitive diagnostic type; each end-side integrating the matching cognitive topology and the local diagnostic target as a cognitive data packet, and sending it to the server via protocol steganography.

[0030] Furthermore, this application also includes: introducing a steganography pattern based on dynamic dictionary mapping in the standard read / write command field of the application layer protocol, wherein the steganography pattern uses read / write commands as a carrier; encoding the cognitive data packet into a read / write operation request using the steganography pattern according to the shared key and communication context; and sending the read / write operation request to the server.

[0031] Specifically, generating task requirement descriptions based on the diagnostic task cluster and distributing them to each endpoint means that after receiving multiple diagnostic tasks aggregated within the same period, the server performs unified parsing and abstraction of the diagnostic objects, diagnostic objectives, constraints, and priorities corresponding to each diagnostic task to form task requirement descriptions that can be understood and processed by the endpoint agents. The task requirement descriptions are used to characterize the fault types, feature dimensions, and reasoning requirements that need to be considered in the current diagnostic cycle, and are distributed to each endpoint to trigger the local cognitive construction and matching processing of the endpoints.

[0032] The first endpoint is used to represent any specific endpoint node participating in this collaborative diagnosis. Knowledge distillation refers to extracting representative decision knowledge and reasoning structures from the first local agent without exposing the original model parameters and original data. The first local agent is used to represent an artificial intelligence model deployed on the endpoint with local diagnosis and learning capabilities. The first knowledge skeleton is used to represent the core diagnostic logic structure and key feature relationships retained after distillation, thereby providing a foundation for subsequent cognitive topology construction.

[0033] Based on the first knowledge skeleton, defining nodes by feature concepts means mapping the abstract feature units in the knowledge skeleton used to describe device status, signal patterns, or fault representations to node elements in the topological structure. The feature concepts at least include key support vectors, feature importance ranking, and anomaly detection threshold ranges. Key support vectors represent the subset of features that play a major supporting role in diagnostic decisions, feature importance ranking represents the relative contribution of each feature in the diagnostic reasoning process, and anomaly detection threshold ranges limit the judgment range of features deviating from normal states. Defining hyperedges by diagnostic reasoning paths means abstracting the reasoning process involving multiple features in the knowledge skeleton into a high-order association structure connecting multiple nodes, and defining edge weights by path confidence to quantify the reliability of each diagnostic reasoning path in historical diagnosis or model reasoning, thereby comprehensively determining the first local topology. This first local topology is used to structurally express the local diagnostic cognitive relationships of the edge agent.

[0034] Furthermore, counterfactual evaluation and pruning of each diagnostic reasoning path in the first local topology refers to evaluating the stability and effectiveness of each diagnostic reasoning path under changing assumptions by constructing feature perturbation or path substitution scenarios while maintaining the original feature distribution constraints. Reasoning paths with low contribution to diagnostic results, high redundancy, or insufficient stability are eliminated, thereby compressing and optimizing the topology structure. Finally, the first local cognitive topology is determined. This first local cognitive topology is used to represent the screened, more robust, and interpretable end-side diagnostic cognitive structure, providing high-quality cognitive input for subsequent matching with task requirements and federated collaboration.

[0035] The periodic release of task requirement descriptions by the server refers to the server uniformly encapsulating the diagnostic task information to be executed in the current stage and broadcasting it to each end node according to a preset time interval or diagnostic round. Here, the server is used to represent a centralized processing unit with global task orchestration and collaborative control capabilities. The periodicity is used to limit the timing of task release to support the multi-round collaborative processing mechanism under federated learning. The task requirement description is used to abstractly represent the target object, diagnostic type, feature focus scope and inference constraints of the diagnostic task, thereby providing clear and alignable diagnostic guidance for the end-side intelligent agents.

[0036] The matching of local cognitive topology and task requirement description performed by each endpoint refers to each endpoint using its own constructed local cognitive topology to perform structural and semantic alignment analysis on the task requirement description issued by the server, in order to filter cognitive structures that are highly relevant to the current diagnostic task. The local cognitive topology is used to represent the diagnostic knowledge structure and reasoning relationship within the endpoint's intelligent body. The matching process is used to determine the degree of fit between the cognitive structure and the task requirements, thereby determining the matching cognitive topology. Each endpoint corresponds to one matching cognitive topology, which is used to ensure that each endpoint only feeds back one well-structured and task-related cognitive result to the server in this cycle. The validity proof of each matching cognitive topology means that proof information used to characterize its reliability, completeness, or consistency is attached to the matching cognitive topology to support the server's reliable parsing and fusion processing of the endpoint's cognitive results.

[0037] Furthermore, for the first type of endpoint with diagnostic needs, generating local diagnostic targets based on cognitive diagnostic types means that, in the endpoint node, based on the task requirement description and its own operating status, the endpoint category that needs to perform specific diagnostic tasks is identified. The first type of endpoint is used to represent endpoint nodes with clear diagnostic needs or abnormal triggering conditions, while cognitive diagnostic types are used to distinguish different fault modes, diagnostic levels or reasoning strategies. The resulting local diagnostic targets are used to clarify the key diagnostic tasks and output formats that the endpoint needs to complete in the current cycle.

[0038] Subsequently, each end-side integrates and matches the cognitive topology and local diagnostic targets as a cognitive data packet. This means that the cognitive topology structure after task matching and filtering is uniformly encapsulated with the corresponding local diagnostic targets to form a cognitive data unit with structural information and task semantics. The cognitive data packet is used to carry the end-side's cognitive results and target orientation for the diagnostic task.

[0039] In the standard read / write command fields of application layer protocols, a steganography pattern based on dynamic dictionary mapping is introduced. This refers to the use of steganography to encrypt data in standard format fields of protocol communication, such as read / write commands, during data transmission. Steganography encodes data into an inconspicuous format, making it difficult to detect or tamper with during transmission. Dynamic dictionary mapping represents the mechanism for mapping and encoding data based on the communication context and real-time conditions, thereby enhancing data security and concealment. Read / write command fields carry regular operational instructions in normal protocol communication. Steganography embeds encrypted data within these fields, enabling secure data transmission without violating the communication protocol.

[0040] Encoding cognitive data packets into read / write operation requests using steganography, based on a shared key and the communication context, refers to encrypting data using a shared key before transmission to ensure confidentiality and integrity during network transmission. The communication context refers to the communication status, protocol version, and other relevant dynamic information of the transmitting parties, which influences the steganography method and security strategy. By converting cognitive data packets into read / write operation requests using steganography, the data can be disguised as standard protocol commands, thereby avoiding monitoring or interference.

[0041] Sending read / write operation requests to the server means that after steganography, the encoded data is sent to the server via standard network protocols. This ensures that the data arrives at its destination without revealing its actual content, and is then decrypted and parsed on the server side. Steganography effectively prevents malicious interception or analysis of data transmission while maintaining protocol compatibility, thus guaranteeing security during transmission.

[0042] The server receives the cognitive data packet, activates the parsing engine, and performs dual-path operation management.

[0043] Specifically, server receiving cognitive data packets refers to the server's data reception operation after receiving cognitive data packets transmitted from the edge agent via protocol steganography. Cognitive data packets are information packets integrated by the edge agent based on diagnostic task requirements and local cognitive topology. They contain key information such as diagnostic results, target settings, and inference paths, and are transmitted using steganography to ensure data security and confidentiality during transmission. By receiving cognitive data packets, the server obtains the diagnostic information submitted by the edge agent, providing foundational data for further processing.

[0044] Activating the parsing engine refers to the server's initiation of its built-in parsing engine to process the cognitive data packet upon receipt. The parsing engine automatically parses the transmitted data and extracts key diagnostic information, inference paths, and task objectives. By activating the parsing engine, the server can quickly identify valid information within the cognitive data packet, providing support for subsequent fault diagnosis and knowledge fusion.

[0045] Dual-path operation management refers to the server's use of a dual-path operation management mechanism to further process and provide feedback on diagnostic information from edge-side agents after parsing the data. This mechanism comprises two main operation paths: First, the server analyzes the data returned from each edge-side agent, extracts diagnostic primitives for edge-side consensus, and generates pseudo-diagnostic paths to continuously enrich the global knowledge base. Second, the server optimizes local diagnostic targets and distributes them to the corresponding edge-side agents through fusion and matching with the edge-side cognitive topology, guiding the agents to further learn and execute tasks, thus achieving closed-loop management.

[0046] The dual-path operation management includes: determining diagnostic primitives based on the local cognitive topology and end-side consensus, driving the extension group to perform pseudo-expansion and end-side agent verification, and expanding the global knowledge base; matching and fusing local diagnostic targets with the local cognitive topology and the global knowledge base, and distributing them to the corresponding end-side agents for update learning and task diagnostic management.

[0047] Furthermore, this application also includes: the server receiving cognitive data packets from each end side, deconstructing the matching cognitive network according to the embedded parsing engine, and determining multiple sets of cognitive paths; performing structural consistency analysis on the multiple sets of cognitive paths through continuous coherence analysis to determine a cognitive path group, wherein the cognitive path group satisfies the structural consistency condition; and storing the cognitive path group as a diagnostic primitive for end-side consensus in the global knowledge base.

[0048] Furthermore, this application also includes: developing an extension component within the server, establishing interaction between the extension component and the global knowledge base; the extension component retrieving the diagnostic primitive from the global knowledge base, extending and generating a false diagnostic path, wherein the extension method includes at least combination and variation; and distributing the false diagnostic path to a preset proportion of edge agents for local verification.

[0049] Furthermore, this application also includes: each end-side intelligent agent performing local verification performs knowledge distillation and local topology transformation on the local verification data, and encapsulates it into a local verification package; the local verification data package is sent to the server through protocol steganography, triggering the parsing engine to perform deconstruction analysis, filtering out valid cognitive paths that have been successfully verified and meet structural consistency, and storing them in the global knowledge base.

[0050] Furthermore, this application also includes: the parsing engine reads the first local diagnostic target in the cognitive data packet, matches and fuses the multiple sets of cognitive paths with the global knowledge base to determine the first target diagnostic path; encapsulates the first target diagnostic path and sends it to the first terminal side via protocol steganography; the intelligent agent on the first terminal side learns and updates according to the first target diagnostic path.

[0051] Furthermore, this application also includes: the server publishing a batch of task requirements descriptions in the next cycle; and, for the batch of task requirements descriptions, performing knowledge optimization and diagnosis between the edge agent and the server, and performing periodic polling processing.

[0052] Specifically, the server receives cognitive data packets from various endpoints and, based on its embedded parsing engine, deconstructs the matching cognitive network to determine multiple sets of cognitive paths. This means that after receiving cognitive data packets transmitted by the endpoint agents, the server uses its built-in parsing engine to deconstruct the information within the data packets. The deconstruction process includes identifying key reasoning paths, knowledge nodes, and relationships within the data packets, and constructing multiple cognitive paths. The matching cognitive network refers to matching data according to task requirements and the local cognitive topology of the endpoint agents to determine reasoning paths related to the task objective. These reasoning paths reflect the different interpretations and reasoning methods of the endpoint agents for the current diagnostic task. By deconstructing the cognitive network, the server can obtain multiple sets of cognitive paths, serving as the basis for subsequent analysis.

[0053] Continuous coherence analysis (HCA) is used to perform structural consistency analysis on multiple cognitive paths to determine a cognitive path group. This involves the server employing HCA to deeply analyze the extracted cognitive paths after deconstruction, determining whether the paths are structurally consistent. Structural consistency analysis verifies whether different edge agents reach a consensus during the diagnostic process, ensuring that the information contained in different paths is structurally matched and reliable. A cognitive path group, determined through structural consistency analysis by filtering cognitive paths that meet consistency criteria, represents the consistent judgment results of multiple edge agents in this diagnostic task.

[0054] Storing cognitive path groups as diagnostic primitives for edge-side consensus in the global knowledge base means that after the server obtains a structurally consistent cognitive path group, it stores it as the shared cognitive result of edge-side agents on the current diagnostic task. Cognitive paths, as diagnostic primitives, represent the consensus and cognitive conclusions of multiple edge-side agents regarding the fault diagnosis problem. Storing them in the global knowledge base means adding these consensus paths as new knowledge units to the knowledge base for future reference and reuse in similar diagnostic tasks, promoting knowledge accumulation and optimization.

[0055] Developing an extension component within the server and establishing its interaction with the global knowledge base refers to designing and implementing an extension component within the server. This extension component is responsible for data exchange and interaction with the global knowledge base. The extension component enhances the server's functionality, allowing it to communicate with the global knowledge base through an interface to obtain diagnostic information, rules, or paths stored therein, thereby expanding and updating the existing knowledge structure. By establishing an interaction mechanism, the extension component can retrieve relevant data from the global knowledge base and process and expand it as needed to support more complex diagnostic tasks.

[0056] The extension component retrieves the diagnostic primitives from the global knowledge base and extends them to generate pseudo-diagnostic paths. The extension method includes at least combination and mutation. This means the extension component obtains diagnostic primitives from the global knowledge base; these primitives are key cognitive units generated based on the collaborative diagnostic results of the edge agent. By further processing the diagnostic primitives, the extension component can generate pseudo-diagnostic paths. Pseudo-diagnostic paths are possible diagnostic paths generated based on existing knowledge or models. They are not necessarily solutions to real faults, but rather candidate paths generated by extending existing paths. The extension method includes at least combination and mutation. Combination refers to combining multiple existing diagnostic paths to generate a new path, while mutation refers to generating new paths by changing certain parameters or rules of existing paths, providing more solutions or reasoning paths for the diagnostic task.

[0057] Distributing pseudo-diagnostic paths to a predetermined proportion of edge agents for local verification means that after generating pseudo-diagnostic paths, the extension component distributes them to a certain proportion of edge agents for verification. The edge agents perform diagnostic tasks based on their local environment and data, verifying the effectiveness of these pseudo-diagnostic paths. Through this verification process, the edge agents can evaluate the applicability and accuracy of these extended paths in practical applications, thus providing feedback for further optimization of diagnostic paths and improvement of the global knowledge base.

[0058] Each edge agent performing local verification performs knowledge distillation and local topology transformation on the local verification data, encapsulating it into a local verification package. This means that after receiving a false diagnostic path, each edge agent verifies the path based on its local data and environment. The verification process includes extracting key diagnostic knowledge through knowledge distillation and performing local topology transformation to ensure that the generated diagnostic path is effectively consistent with the edge agent's existing knowledge structure. Knowledge distillation refers to transforming the edge agent's complex diagnostic model into a more concise knowledge representation with stronger reasoning capabilities, while local topology transformation refers to mapping the extracted knowledge structure onto the edge agent's specific topology structure for further verification. The edge agent can encapsulate the local verification data into a local verification package, carrying the verification results and related diagnostic information.

[0059] Sending local verification data packets to the server via protocol steganography triggers the parsing engine to perform deconstruction analysis, filtering out valid cognitive paths that have successfully verified and meet structural consistency conditions, and storing them in the global knowledge base. This means that after the edge agent completes local verification, it transmits the local verification data packets to the server via protocol steganography. Protocol steganography is used to ensure the security and confidentiality of data during transmission, preventing sensitive data from being stolen or tampered with during communication. Once the server receives the local verification data packets, it uses its built-in parsing engine to deconstruct and analyze the data, extracting valid diagnostic information and filtering paths to determine those that have successfully verified and meet structural consistency conditions. Valid cognitive paths that meet the conditions are stored in the global knowledge base, providing more accurate knowledge support for subsequent diagnostic tasks and further promoting the optimization of the global knowledge base and the evolution of the intelligent system.

[0060] The parsing engine determines the first target diagnostic path by reading the first local diagnostic target within the cognitive data packet and matching and fusing it with multiple sets of cognitive paths and the global knowledge base. Specifically, upon receiving the cognitive data packet, the parsing engine reads the first local diagnostic target contained within the packet, representing the specific fault diagnosis task set by the edge agent during local diagnosis. Based on this first local diagnostic target, the parsing engine matches and fuses it with multiple sets of cognitive paths stored internally on the server and knowledge in the global knowledge base. The multiple sets of cognitive paths represent candidate paths generated by different edge agents based on different inference strategies or local data, while the global knowledge base contains diagnostic information, rules, and experience shared across multiple edge agents. Through matching and fusion, the parsing engine can identify the most suitable diagnostic path and determine it as the first target diagnostic path, deriving the optimal path based on global information and local requirements.

[0061] Encapsulating the first target diagnostic path and sending it to the first endpoint via protocol steganography means that after determining the first target diagnostic path, the parsing engine encapsulates the path into a data format conforming to the communication protocol, and then sends it to the first endpoint via protocol steganography. Protocol steganography is a technique that hides sensitive data within normal protocol communication to ensure data confidentiality and security. Through steganography, data will not cause abnormal traffic or be identified by external monitoring systems during transmission, thereby enhancing the system's data transmission security.

[0062] The first-side agent learns and updates based on the first target diagnostic path. This means that when the first-side agent receives the encapsulated target diagnostic path, it uses this path to learn and update its local model. The agent reasons and analyzes based on the new diagnostic path, gradually adjusting its reasoning mechanism and decision-making process to optimize its diagnostic capabilities and accuracy. Through a feedback mechanism, the edge agent can continuously optimize its diagnostic model in each iteration, improving the accuracy and reliability of fault diagnosis.

[0063] The server releases a new batch of task requirement descriptions in the next cycle. This means that after a diagnostic cycle ends, the server generates and releases new task requirement descriptions based on the overall system requirements and task progress, serving as a work guide for the next cycle. The new batch of task requirement descriptions details the specific diagnostic tasks, target fault types, and feature analysis scope that each edge agent needs to complete within this cycle. This description is based on feedback from the previous cycle's diagnostic results and updates to the global knowledge base, providing new task objectives for the edge agents and ensuring that the diagnostic objectives and tasks of the entire system can be dynamically adapted and optimized across different cycles.

[0064] For the next batch of task requirements, the process involves knowledge optimization and diagnosis between the edge agent and the server, along with periodic polling. This means that upon receiving the next batch of task requirements, the edge agent updates its local knowledge and performs diagnostic tasks based on the new requirements. The edge agent, based on the task objectives, combines locally stored diagnostic knowledge, device status information, and historical data to perform fault diagnosis or data reasoning, and periodically exchanges information with the server. The server and edge agent continuously optimize knowledge, enhancing the overall diagnostic capability and accuracy of the system by sharing local experience, providing feedback on diagnostic results, and updating the reasoning model. Periodic polling ensures that task execution and knowledge updates continue within each cycle to adapt to changing system environments and task requirements.

[0065] Furthermore, the knowledge skeleton represents the set of core diagnostic rules extracted from the edge agent through knowledge distillation and its corresponding feature association structure. The set of core diagnostic rules includes at least feature selection results, feature weight distribution, and corresponding inference path index relationships. The cognitive topology represents the directed weighted hypergraph structure formed after reconstructing the knowledge skeleton according to a graph structure. Nodes represent feature concepts or state variables, hyperedges represent diagnostic inference paths involving multiple features, and edge weights represent path confidence or historical verification pass rates. The diagnostic primitive represents the minimum reusable diagnostic path unit obtained after server-side structural consistency analysis. The minimum reusable diagnostic path unit includes an initial feature set, inference rules, and an output fault type identifier. The pseudo-extension represents a set of candidate diagnostic paths generated through path combination or parameter perturbation without changing the basic semantic constraints of the original diagnostic primitive, used to verify potential knowledge generalization capabilities. Protocol steganography represents mapping cognitive data to standard read / write instructions through dynamic dictionary mapping within the legal fields of the application layer protocol, thereby achieving separation of data content and communication semantics.

[0066] Furthermore, the dual-path operation management is a parallel processing mechanism on the server side, including a first processing path and a second processing path. The first processing path is used to perform structural deconstruction and consistency analysis on cognitive data from multiple endpoints to generate diagnostic primitives and update the global knowledge base. The second processing path is used to read local diagnostic targets from cognitive data packets, perform matching and fusion in the global knowledge base and multiple sets of cognitive paths, generate target diagnostic paths, and feed them back to the corresponding endpoints. The two processing paths are executed in parallel logically and interact at the data level by sharing the global knowledge base storage structure. The diagnostic primitives output by the first processing path serve as the candidate fusion basis for the second processing path, while the verification results generated by the second processing path can also influence the effectiveness evaluation of the diagnostic primitives, thus forming a closed-loop evolution mechanism.

[0067] In summary, the edge AI agent collaborative fault diagnosis method based on federated learning provided in this application has the following technical effects: by achieving the technical goal of structured modeling and federated collaborative fusion of edge diagnostic cognition driven by task requirements, it enables the server to perform parsable, aligned and verifiable processing of edge diagnostic cognition without exposing the original edge data and complete model, thereby improving the consistency of global knowledge fusion and the reliability of collaborative fault diagnosis results.

[0068] Example 2: Based on the same inventive concept as the federated learning-based collaborative fault diagnosis method for edge AI agents in the foregoing examples, this application also provides a federated learning-based collaborative fault diagnosis device for edge AI agents. Please refer to the appendix. Figure 2The system includes: a diagnostic task cluster receiving module 1, used by the server to receive the diagnostic task cluster for the first cycle; a cognitive data packet determination module 2, used to generate a task requirement description based on the diagnostic task cluster, and distribute it to each end side. Each end side extracts the knowledge skeleton from the local agent and performs topology reconstruction through knowledge distillation. By matching it with the task requirement description and combining it with the local diagnostic target, the cognitive data packet is determined and sent to the server using protocol steganography; a dual-path operation management execution module 3, used by the server to receive the cognitive data packet, activate the parsing engine, and perform dual-path operation management; wherein, the dual-path operation management includes: a global knowledge base extension unit 31, used to drive the extension group to perform pseudo-extension and end-side agent verification by determining the diagnostic primitives based on the end-side consensus of the local cognitive topology, thereby expanding the global knowledge base; and a task diagnosis management unit 32, used to match and fuse the local diagnostic target with the local cognitive topology and the global knowledge base, and distribute it to the corresponding end-side agent for update learning and task diagnosis management.

[0069] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used for: the first edge extracting a first knowledge skeleton from the first local agent through knowledge distillation; defining nodes by feature concepts, defining hyperedges by diagnostic reasoning paths, and defining edge weights by path confidence based on the first knowledge skeleton, thereby determining a first local topology, wherein the feature concepts at least include key support vectors, feature importance ranking, and anomaly detection threshold range; and performing counterfactual evaluation and pruning on each diagnostic reasoning path in the first local topology to determine a first local cognitive topology.

[0070] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used for: the server periodically publishing task requirement descriptions; each edge performing a matching of its local cognitive topology with the task requirement description to determine a matching cognitive topology, wherein each edge corresponds to one matching cognitive topology, and each matching cognitive topology has a validity proof; the first type of edge with diagnostic needs generating a local diagnostic target based on the cognitive diagnostic type; each edge integrating the matching cognitive topology and the local diagnostic target as a cognitive data packet, and sending it to the server through protocol steganography.

[0071] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used to: introduce a steganography pattern based on dynamic dictionary mapping into the standard read / write command field of the application layer protocol, wherein the steganography pattern uses read / write commands as the carrier; encode the cognitive data packet into a read / write operation request using the steganography pattern according to the shared key and communication context; and send the read / write operation request to the server.

[0072] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used for: the server receiving cognitive data packets from each edge, deconstructing the matching cognitive network according to the embedded parsing engine, and determining multiple sets of cognitive paths; performing structural consistency analysis on the multiple sets of cognitive paths through continuous coherence analysis to determine a cognitive path group, wherein the cognitive path group satisfies the structural consistency condition; and storing the cognitive path group as a diagnostic primitive for edge consensus in the global knowledge base.

[0073] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used for: developing an extension component within the server and establishing interaction between the extension component and the global knowledge base; the extension component retrieves the diagnostic primitive from the global knowledge base and extends it to generate a false diagnostic path, wherein the extension method includes at least combination and variation; and the false diagnostic path is distributed to a preset proportion of edge agents for local verification.

[0074] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used for: each edge agent performing local verification performs knowledge distillation and local topology transformation on the local verification data, and encapsulates it into a local verification package; the local verification data package is sent to the server through protocol steganography, triggering the parsing engine to perform deconstruction analysis, filtering out valid cognitive paths that have been successfully verified and meet structural consistency, and storing them in the global knowledge base.

[0075] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used for: the parsing engine reading the first local diagnostic target in the cognitive data packet, matching and fusing the multiple sets of cognitive paths with the global knowledge base to determine the first target diagnostic path; encapsulating the first target diagnostic path and sending it to the first edge via protocol steganography; and the agent on the first edge learning and updating according to the first target diagnostic path.

[0076] Furthermore, the edge AI agent collaborative fault diagnosis device based on federated learning is also used for: the server to release the next batch of task requirements descriptions in the next cycle; for the next batch of task requirements descriptions, to perform knowledge optimization and diagnosis between the edge AI agent and the server, and to perform periodic polling processing.

[0077] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The method and specific examples of edge AI agent collaborative fault diagnosis based on federated learning in the first embodiment are also applicable to the edge AI agent collaborative fault diagnosis device based on federated learning in this embodiment. Through the foregoing detailed description of the edge AI agent collaborative fault diagnosis method based on federated learning, those skilled in the art can clearly understand the edge AI agent collaborative fault diagnosis device based on federated learning in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0079] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A collaborative fault diagnosis method for edge AI agents based on federated learning, characterized in that, The method includes: The server receives the diagnostic task cluster for the first cycle; The task requirement description is generated based on the diagnostic task cluster and distributed to each terminal. Each terminal extracts the knowledge skeleton from the local intelligent agent and performs topology reconstruction through knowledge distillation. By matching it with the task requirement description and combining it with the local diagnostic target, the cognitive data packet is determined and sent to the server using protocol steganography. The server receives the cognitive data packet, activates the parsing engine, and performs dual-path operation management. The dual-path operation management includes: By identifying diagnostic primitives based on local cognitive topology and edge consensus, the extension group is driven to perform pseudo-extension and edge agent verification, thereby expanding the global knowledge base. The local diagnostic targets are matched and fused with the local cognitive topology and the global knowledge base, and then distributed to the corresponding intelligent agents on the edge for update learning and task diagnosis management.

2. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 1, characterized in that, Each endpoint extracts the knowledge skeleton from the local agent and performs topology reconstruction through knowledge distillation, including: The first end-side extracts the first knowledge skeleton from the first local intelligent agent through knowledge distillation. Based on the first knowledge skeleton, nodes are defined by feature concepts, hyperedges are defined by diagnostic reasoning paths, and edge weights are defined by path confidence to determine the first local topology. The feature concepts include at least key support vectors, feature importance ranking, and anomaly detection threshold range. Counterfactual evaluation and pruning are performed on each diagnostic reasoning path in the first local topology to determine the first local cognitive topology.

3. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 2, characterized in that, By matching with the task requirement description and combining it with local diagnostic objectives, a cognitive data package is determined, including: Description of server periodic task publishing requirements; Each terminal performs a matching of its local cognitive topology with the task requirement description to determine a matching cognitive topology. Each terminal corresponds to one matching cognitive topology, and each matching cognitive topology identifier has a validity proof. For the first type of edge device with diagnostic needs, local diagnostic targets based on cognitive diagnostic types are generated. Each endpoint integrates and matches the cognitive topology with the local diagnostic target, which is then sent to the server as a cognitive data packet via protocol steganography.

4. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 3, characterized in that, Sending data to the server via protocol steganography, including: In the standard read and write command fields of the application layer protocol, a steganography encoding mode based on dynamic dictionary mapping is introduced, wherein the steganography encoding mode is carried by read and write commands; Based on the shared key and communication context, the cognitive data packet is encoded into a read / write operation request using the steganography mode; The read / write operation request is sent to the server.

5. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 1, characterized in that, Dual-path operation management, including: The server receives cognitive data packets from each end and, based on the embedded parsing engine, deconstructs the matching cognitive network to determine multiple sets of cognitive paths. Through continuous coherence analysis, structural consistency analysis is performed on the multiple cognitive paths to determine cognitive path groups, wherein the cognitive path groups satisfy the structural consistency condition. The cognitive path group is used as a diagnostic primitive for end-side consensus and stored in the global knowledge base.

6. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 5, characterized in that, After being stored in the global knowledge base, it includes: Develop extended components within the server and establish interaction between the extended components and the global knowledge base; The extended component retrieves the diagnostic primitives from the global knowledge base and extends them to generate false diagnostic paths, wherein the extension methods include at least combination and variation; The false diagnostic path is distributed to a preset proportion of edge agents for local verification.

7. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 6, characterized in that, Each edge agent performing local verification performs knowledge distillation and local topology transformation on the local verification data and encapsulates it into a local verification package. The local verification data packet is sent to the server via protocol steganography, triggering the parsing engine to perform destructive analysis, filtering out valid cognitive paths that have been successfully verified and meet structural consistency, and storing them in the global knowledge base.

8. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 7, characterized in that, Dual-path operation management, including: The parsing engine determines the first target diagnostic path by reading the first local diagnostic target in the cognitive data packet, matching and fusing the multiple sets of cognitive paths with the global knowledge base; The first target diagnostic path is encapsulated and sent to the first end via protocol hidden writing; The agent on the first end learns and updates based on the first target diagnostic path.

9. The edge AI agent collaborative fault diagnosis method based on federated learning as described in claim 1, characterized in that, The server will release the task requirements description for the next batch in the next cycle; Based on the requirements of the second batch of tasks, knowledge optimization and diagnosis are performed between the edge agent and the server, and periodic polling is executed.

10. A fault diagnosis device for edge-side AI agents based on federated learning, characterized in that, The steps for implementing the edge AI agent collaborative fault diagnosis method based on federated learning as described in any one of claims 1 to 9 include: The diagnostic task cluster receiving module is used by the server to receive the diagnostic task cluster for the first cycle. The cognitive data packet determination module is used to generate a task requirement description based on the diagnostic task cluster and distribute it to each terminal. Each terminal extracts the knowledge skeleton from the local intelligent agent and performs topology reconstruction through knowledge distillation. By matching it with the task requirement description and combining it with the local diagnostic target, the cognitive data packet is determined and sent to the server using protocol steganography. The dual-path operation management and execution module is used by the server to receive the cognitive data packet, activate the parsing engine, and perform dual-path operation management. The dual-path operation management includes: The global knowledge base extension unit is used to drive the extension group to perform pseudo-extension and end-side agent verification by determining the diagnostic primitives based on the end-side consensus of the local cognitive topology, thereby expanding the global knowledge base. The task diagnosis management unit is used to match and fuse local diagnostic targets with the local cognitive topology and the global knowledge base, and then distribute them to the corresponding intelligent agents on the edge for update learning and task diagnosis management.

Images