A dynamic context map self-adaptive multi-agent collaborative fault diagnosis method

Abstract

The application provides a kind of dynamic context graph adaptive multi-agent collaborative fault diagnosis method in the field of artificial intelligence technology, comprising: step S1, initialization dynamic context graph, and dynamically update in subsequent diagnosis process;Step S2, based on the entity scale involved in fault, relationship level and reasoning path length, the complexity of fault is quantitatively evaluated;Step S3, according to the complexity of fault, the number of intelligent agents is determined, and the corresponding diagnostic intelligent agent is matched from the intelligent agent pool to form a collaborative set, while the corresponding collaboration mode is selected, to start collaborative fault diagnosis;Step S4, in the process of collaborative diagnosis, conflict identification and resolution are carried out based on dynamic context graph, and the dynamic context graph is optimized and updated;Step S5, the decision conclusion with the highest confidence is extracted from the dynamic context graph as the final diagnosis result.The application has the advantages of greatly improving the resource utilization rate, diagnosis accuracy and system robustness of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a dynamic context graph adaptive multi-agent collaborative fault diagnosis method. Background Technology

[0002] As mobile communication technology evolves towards 5G-A and 6G, network architectures are becoming increasingly complex, making network fault diagnosis a core element in ensuring the quality of mobile communication services. Multi-agent collaborative reasoning technology is widely used in this field, its core purpose being to overcome the limitations of a single agent's fault location capabilities through the collaboration of multiple diagnostic agents, thereby improving the efficiency and accuracy of fault diagnosis. However, existing technologies have significant shortcomings in multi-agent collaborative fault diagnosis, mainly in the following three aspects: First, regarding knowledge graph support, existing technologies primarily rely on static knowledge graphs. Traditional methods construct static knowledge networks by pre-extracting faulty entities (such as base stations, core networks, and transmission links) and their causal relationships in the mobile communication domain, providing basic reasoning support for multi-agent diagnosis. However, static knowledge graphs only record "what the faulty entities and relationships are," failing to capture dynamic contextual information during the fault diagnosis process, such as reasoning trajectories, detection sequences, and data confidence levels. This makes it impossible to trace the decision-making chain to pinpoint the root cause when diagnostic results deviate. Furthermore, static graphs cannot update temporary detection data and intermediate conclusions during the diagnostic process in real time, leading to information asymmetry among multiple agents, potentially causing duplicate detections and reasoning conflicts, thus reducing the reliability and interpretability of the diagnosis.

[0003] Secondly, regarding multi-agent collaboration mechanisms, existing technologies employ a "fixed scale + fixed mode" approach. That is, either the same number of agents are configured for all fault diagnosis tasks, or a limited number of collaboration modes (such as serial or parallel) are preset, with mode switching relying on manual configuration. This fixed collaboration mechanism cannot adapt to differences in fault complexity: for simple faults (such as single-user terminal disconnection or parameter configuration errors), too many agents lead to redundant computing resources and increased communication overhead, reducing diagnostic efficiency; for complex faults (such as regional network congestion or multi-network element linkage faults), a fixed number of agents cannot cover all diagnostic dimensions, including the wireless side, core network side, and transmission side, easily resulting in insufficient fault localization, biased conclusions, or even diagnostic failure due to insufficient reasoning ability.

[0004] Furthermore, in terms of conflict resolution mechanisms, existing technologies rely on pre-defined hard rules for conflict handling, without incorporating contextual information from the fault diagnosis process. This rule-based mechanism cannot accurately identify the root cause of conflicts (such as cascading conflicts caused by low-confidence detection data), and can only handle surface conflicts in a simple way, easily leading to "cascading failure"—that is, small conflicts spread through the diagnostic reasoning chain, ultimately causing overall diagnostic errors. At the same time, after conflict resolution, the diagnostic path cannot be optimized in reverse, and similar conflicts recur, indicating that the system lacks self-iteration and optimization capabilities.

[0005] In summary, existing technologies suffer from three major pain points: static knowledge graphs cannot trace the context of fault diagnosis decisions; multi-agent collaboration models cannot adapt to fault complexity; and conflict resolution mechanisms lack accuracy and self-optimization capabilities. These problems collectively lead to low resource utilization, poor diagnostic accuracy, and insufficient system robustness in multi-agent collaborative fault diagnosis, making it difficult to meet the efficient operation and maintenance requirements of complex mobile communication networks.

[0006] Therefore, how to provide a dynamic context graph adaptive multi-agent collaborative fault diagnosis method to improve the resource utilization, diagnostic accuracy and system robustness of fault diagnosis has become an urgent technical problem to be solved. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a dynamic context graph adaptive multi-agent collaborative fault diagnosis method, which improves the resource utilization, diagnostic accuracy and system robustness of fault diagnosis.

[0008] This invention is implemented as follows: A dynamic context graph adaptive multi-agent cooperative fault diagnosis method, comprising the following steps: Step S1: Construct a dynamic context graph for real-time recording of the inference node set N, context edge set E, decision node set D, time series set T, and weight set W during the fault diagnosis process; Step S2: In response to the triggering of the fault diagnosis task, initialize the dynamic context graph, and in the subsequent diagnosis process, record and update the detection results and decision conclusions of each diagnostic agent in real time to the dynamic context graph. Step S3: Based on the initial information of the fault diagnosis task, assess the entity size, relationship hierarchy and reasoning path length involved in the fault to quantitatively assess the fault complexity of the current fault. Step S4: Based on the fault complexity, adaptively determine the number of agents participating in collaborative diagnosis, and match corresponding diagnostic agents from the agent pool to form a collaborative set based on the number of agents. At the same time, select the corresponding collaborative mode based on the fault complexity to start collaborative fault diagnosis. Step S5: During the collaborative diagnosis process, conflict identification is performed based on the dynamic context graph, the root cause of the conflict is located and the conflict is resolved, and the dynamic context graph is optimized and updated at the same time. Step S6: Extract the decision conclusion with the highest confidence from the optimized and updated dynamic context graph as the final diagnosis result, and archive the final dynamic context graph.

[0009] Furthermore, in step S1, the dynamic context graph is represented as follows: ; DCG stands for Dynamic Context Graph; , ; This represents the i-th inference node; This represents the detection result of the i-th inference node; Represents the detection timestamp of the i-th inference node; , representing the confidence level of the detection result of the i-th inference node; , ; Represents the context edge between the i-th reasoning node and the j-th decision node; This represents the diagnostic reasoning relationship type between the i-th reasoning node and the j-th decision node; This represents the generation time of the context edge between the i-th inference node and the j-th decision node; This represents the unique identifier of the diagnostic reasoning path between the i-th reasoning node and the j-th decision node; , ; This represents the p-th decision node; This represents the decision conclusion of the p-th decision node; This represents the decision timestamp of the p-th decision node; The diagnostic agent identifier represents the decision conclusion generated at the p-th decision node; T records the timestamps of all operations in the entire dynamic context graph; W contains... for The weight.

[0010] Furthermore, in step S2, the formula for initializing the dynamic context graph is: ; in, Represents the initial dynamic context graph; This represents the empty set.

[0011] Furthermore, in step S2, the update formula for the dynamic context graph is: ; ; in, Represents the dynamic context graph at time t; Represents the dynamic context graph at time t-1; This represents the i-th inference node; Represents the context edge between the i-th inference node and the p-th decision node; This represents the p-th decision node; express The weights; , represents the weighting coefficient; The historical fault location accuracy of the diagnostic agent a that generates the detection results; , where represents the confidence level of the detection result of the i-th inference node.

[0012] Furthermore, in step S2, the step of recording and updating the detection results and decision conclusions of each diagnostic agent in real time to the dynamic context graph specifically involves: The detection results and decision conclusions of each diagnostic agent are used as inference nodes and decision nodes, and the inference nodes and decision nodes are associated through context edges. At the same time, the temporal information corresponding to the generation records of each inference node, decision node and context edge is recorded, and the context edge is assigned weight information to update the dynamic context graph.

[0013] Furthermore, in step S3, the formula for calculating the fault complexity is: ; in, This represents the fault complexity of fault Q. For a simple fault, This is a medium-level fault. For complex faults, As a low complexity threshold, This is a high complexity threshold; Indicates the size of the entities involved in fault Q; Indicates the hierarchical relationship involved in fault Q; β represents the length of the inference path involved in fault Q; β, γ, and δ all represent weight coefficients, and .

[0014] Furthermore, step S4 specifically includes: Based on the aforementioned fault complexity, the number of agents participating in collaborative diagnosis is adaptively determined: ; in, This indicates the number of agents required to participate in the collaborative diagnosis of fault Q; Represents the fault complexity of fault Q; This represents the baseline complexity, corresponding to the maximum fault complexity that a single diagnostic agent can handle; round() represents the rounding function. Calculate the matching degree between each diagnostic agent in the agent pool and the fault Q: ; in, This represents the degree of matching between the diagnostic agent a and the fault Q; This represents the capability vector of the diagnostic agent a; The vector representing the diagnostic capability requirement for fault Q; This represents the capability vector of the diagnostic agent a in the i-th diagnostic dimension. This represents the diagnostic capability requirement vector for fault Q in the i-th diagnostic dimension. Based on the matching degree, each diagnostic agent is ranked, and agents ranked before the matching degree are selected from the agent pool. A collaborative set of diagnostic agents; Simultaneously, based on the aforementioned fault complexity, an appropriate collaboration mode is selected to initiate collaborative fault diagnosis: ; in, Indicates the cooperation mode of fault Q; As a low complexity threshold, This is the high complexity threshold.

[0015] Furthermore, step S5 specifically includes: During collaborative diagnosis, if contradictory detection results are found between inference nodes on the same inference path based on the dynamic context graph, a conflict is identified. Calculate the extent of the impact of the conflicting inference node on the inference path to quantify the severity of the conflict; Based on the severity of the conflict and the confidence level of the detection results, the root cause of the conflict is located. The conflict root cause and its associated invalid reasoning path are removed from the dynamic context graph to resolve the conflict. The diagnostic agent is then reassigned to the optimized reasoning path. Based on the optimized reasoning path, the dynamic context graph is re-diagnosed and updated.

[0016] Furthermore, in step S6, the formula for calculating the diagnostic result is: ; ; Where Res represents the diagnostic result; Represents the p-th decision node; DCG represents the optimized and updated dynamic context graph; Indicates the overall confidence level of the decision node; Represents the context edge between the i-th inference node and the p-th decision node; Decision nodes representing conflict associations The set of all context edges on the reasoning path; express The weight.

[0017] The advantages of this invention are: 1. A dynamic context graph is constructed to record the inference node set N, context edge set E, decision node set D, temporal sequence set T, and weight set W in real time during the fault diagnosis process. Then, in response to the triggering of the fault diagnosis task, the dynamic context graph is initialized, and during subsequent diagnosis, the detection results and decision conclusions of each diagnostic agent are recorded and updated to the dynamic context graph in real time. Next, based on the initial information of the fault diagnosis task, the scale of entities involved in the fault, the relationship hierarchy, and the length of the inference path are evaluated to quantify the fault complexity. Based on the fault complexity, the number of agents participating in collaborative diagnosis is adaptively determined to match corresponding diagnostic agents from the agent pool to form a collaborative set. Simultaneously, an appropriate collaborative mode is selected based on the fault complexity to initiate collaborative fault diagnosis. During collaborative diagnosis, conflict identification is performed based on the dynamic context graph, the root cause of the conflict is located and resolved, and the dynamic context graph is optimized and updated. From the optimized and updated dynamic context graph, the decision conclusion with the highest confidence is extracted as the final diagnosis result, and the final dynamic context graph is updated. Archiving is performed; that is, by constructing and updating a dynamic context graph that records the complete reasoning chain in real time, resource utilization is improved first. The key lies in the adaptive allocation of agent scale and cooperation mode based on the quantitative assessment of fault complexity, thereby matching simplified resources for simple faults and allocating sufficient computing power for complex faults, avoiding resource waste or insufficient capacity. Furthermore, the improvement in diagnostic accuracy is due to the fact that the graph provides a shared context information platform with temporal sequence and confidence for multiple agents, which not only avoids information silos and repetitive reasoning, but also enables precise location and resolution of conflicts based on panoramic information in collaborative diagnosis, thus drawing more reliable conclusions. Finally, the enhanced system robustness stems from the system's self-learning ability. After each conflict resolution, the dynamic context graph itself is optimized in reverse (such as adjusting weights and marking weak links), enabling the system to learn iteratively from historical diagnoses, effectively suppressing the recurrence and cascading diffusion of similar conflicts, thus exhibiting stronger adaptability and stability in the face of complex and ever-changing environments, ultimately greatly improving the resource utilization, diagnostic accuracy and system robustness of fault diagnosis.

[0018] 2. By constructing a dynamic context graph that dynamically records inference nodes, context edges, decision nodes, time sequences, and weights, the entire diagnostic process is made traceable and multi-agent information is shared in real time, fundamentally avoiding duplicate detections and inference conflicts caused by information asymmetry. By quantitatively assessing the complexity of the fault and adaptively determining the number of agents, selecting the best matching agents, and dynamically switching the collaboration mode, resource investment is precisely matched with the fault difficulty, effectively eliminating resource redundancy or diagnostic blind spots. Furthermore, based on real-time data from the graph, the root causes of conflicts are accurately identified, conflicts are resolved, and inference paths are optimized, forming a closed-loop self-optimization mechanism, thereby systematically improving the resource utilization, diagnostic accuracy, and system robustness of fault diagnosis.

[0019] 3. By constructing a dynamic context graph (DCG), inference nodes, decision nodes, and their relationships are recorded in real time and updated based on confidence and weight. This structured recording method ensures the integrity and traceability of the diagnostic process, thereby reducing the possibility of misjudgment. For example, by extracting the decision conclusion with the highest confidence and combining it with weight calculation, results with high credibility can be selected first, avoiding the bias or error that may exist in a single intelligent agent, and significantly improving the accuracy and overall reliability of the diagnosis.

[0020] 4. By introducing a fault complexity assessment mechanism and adaptively determining the number of agents and cooperation mode (such as serial, hybrid, or parallel cooperation) based on the complexity, this adaptive capability enables the system to dynamically allocate resources according to the complexity of the fault, avoiding overuse of agents on simple faults (saving computing resources) or insufficient resources on complex faults (ensuring coverage). This not only optimizes diagnostic efficiency and reduces unnecessary latency, but also improves the system's response speed, especially in real-time fault scenarios.

[0021] 5. Through conflict identification and resolution mechanisms, the system can detect contradictory results on the reasoning path based on the dynamic context graph, quantify the severity of the conflict, and then locate and remove the root cause of the conflict. This proactive conflict management avoids inconsistencies in the diagnosis process and ensures the optimization and updating of the diagnosis path. For example, reallocating agents to the optimized path can effectively handle the common problem of disagreements in multi-agent collaboration, improve the stability and fault tolerance of the system, and make it more robust in complex or dynamic environments.

[0022] 6. The modular design, such as the general representation of the dynamic context graph and the matching mechanism of the agent pool, makes the method easy to extend to fault diagnosis scenarios in different fields. The quantitative assessment of fault complexity and the selection of collaboration mode do not depend on specific hardware or environment, but are based on adjustable parameters (such as β, γ, δ weight coefficients), which gives the system a high degree of flexibility. Enterprises or users can customize the graph dimensions and agent capabilities according to actual needs, thereby reducing deployment costs and supporting future technology iterations.

[0023] 7. The dynamic context graph records the timestamps, inference paths, and agent origins of all diagnostic operations in detail, providing a complete diagnostic chain. This transparent recording makes diagnostic decisions no longer a "black box," facilitating user retrospective analysis, auditing, or training. For example, the archived final graph can be used for post-analysis or compliance checks, which not only enhances the credibility of the method but also meets the growing demand for interpretability of fault diagnosis in the industrial field.

[0024] 8. By innovatively combining graph theory, multi-agent systems and adaptive algorithms, a unique collaborative diagnostic framework has been formed. Compared with traditional single-agent or fixed-rule methods, this integration utilizes the structural advantages of graphs (such as capturing complex relationships) and the distributed intelligence of multi-agent systems (such as parallel processing), while avoiding the redundancy of technology stacking through adaptive mechanisms.

[0025] 9. The dynamic context graph fully records the inference nodes, decision nodes, timing, confidence level, and link weights of the fault diagnosis. It can not only output the final fault conclusion, but also clearly trace the entire process of "detection data → inference logic → decision generation". It solves the problem that traditional static graphs cannot explain "why the conclusion is reached", greatly reduces the cost of troubleshooting after fault location deviation, improves system debugging efficiency, and realizes full-link traceability and interpretability of the diagnosis process.

[0026] 10. The number of agents and the collaboration mode are dynamically adjusted according to the complexity of the fault through an adaptive collaboration mechanism. For simple faults such as single terminal configuration errors, the number of agents is reduced and serial collaboration is adopted to reduce computing and communication overhead. For complex faults such as regional multi-base station congestion, the scale of agents is expanded and parallel + feedback collaboration is adopted to ensure full coverage of diagnostic dimensions and effectively improve the resource utilization of diagnostic agents.

[0027] 11. The conflict resolution mechanism based on dynamic context graph can accurately locate the root cause of low-confidence conflicts, rather than just dealing with surface conflicts, effectively avoiding the overall diagnostic failure caused by the cascading spread of small conflicts; at the same time, through link optimization and intelligent weight allocation, the diagnostic path can be self-iterated, the recurrence rate of similar conflicts is significantly reduced, and accurate and efficient conflict resolution is achieved, greatly enhancing the robustness of the system.

[0028] 12. After the fault diagnosis is completed, the complete dynamic context graph will be archived and stored. This historical diagnostic data can be used to optimize the calculation rules of the capability vector and weight coefficient of the diagnostic agent. As diagnostic tasks are continuously accumulated, the accuracy of locating complex faults will gradually improve, adapting to the needs of long-term operation and maintenance of mobile communication networks. Attached Figure Description

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Figure 1 This is a flowchart of a dynamic context graph adaptive multi-agent collaborative fault diagnosis method according to the present invention. Detailed Implementation

[0031] The overall approach of the technical solution in this application is as follows: By constructing and updating a dynamic context graph that records the complete inference chain in real time, resource utilization is first improved. The key lies in the adaptive allocation of agent scale and collaboration mode based on the quantitative assessment of fault complexity, thereby matching simplified resources for simple faults and allocating sufficient computing power for complex faults, avoiding resource waste or insufficient capabilities. Furthermore, the improved diagnostic accuracy benefits from the fact that the graph provides a shared context information platform with temporal sequence and confidence for multiple agents, which not only avoids information silos and repetitive reasoning, but also enables precise location and conflict resolution based on panoramic information in collaborative diagnosis, thus leading to more reliable conclusions. Finally, the enhanced robustness of the system stems from the system's self-learning ability. After each conflict resolution, the dynamic context graph itself is optimized in reverse, enabling the system to iteratively learn from historical diagnoses, effectively suppressing the recurrence and cascading diffusion of similar conflicts, thereby exhibiting stronger adaptability and stability in the face of complex and ever-changing environments.

[0032] Please refer to Figure 1 As shown, a preferred embodiment of the dynamic context graph adaptive multi-agent cooperative fault diagnosis method of the present invention includes the following steps: Step S1: Construct a dynamic context graph (DCG) for real-time recording of the inference node set N, context edge set E, decision node set D, temporal set T, and weight set W during the fault diagnosis process. This invention breaks through the limitations of the traditional static knowledge graph triple structure and constructs a five-element extended graph structure that includes inference nodes, context edges, decision nodes, temporal set, and weight set, thereby realizing the recording of full-dimensional information of the diagnosis process. Step S2: In response to the triggering of the fault diagnosis task, initialize the dynamic context graph, and in the subsequent diagnosis process, record and update the detection results and decision conclusions of each diagnostic agent in real time to the dynamic context graph; the triggering conditions for the fault diagnosis task may be user complaints, network performance indicators exceeding limits, etc. Step S3: Based on the initial information of the fault diagnosis task, assess the entity size, relationship hierarchy and reasoning path length involved in the fault to quantitatively assess the fault complexity of the current fault. Step S4: Based on the fault complexity, adaptively determine the number of agents participating in collaborative diagnosis, and match corresponding diagnostic agents from the agent pool to form a collaborative set based on the number of agents. At the same time, select the corresponding collaborative mode based on the fault complexity to start collaborative fault diagnosis. Step S5: During the collaborative diagnosis process, conflict identification is performed based on the dynamic context graph, the root cause of the conflict is located and the conflict is resolved, and the dynamic context graph is optimized and updated at the same time. Step S6: Extract the decision conclusion with the highest confidence from the optimized and updated dynamic context graph as the final diagnosis result, and archive the final dynamic context graph for subsequent optimization of the diagnostic agent's capabilities and tracing of historical faults.

[0033] In step S1, the dynamic context graph is represented as follows: ; DCG stands for Dynamic Context Graph; , ; This represents the i-th inference node; This represents the detection result of the i-th inference node, such as "base station A signal strength -85dBm" or "core network S1 interface latency 40ms". Represents the detection timestamp of the i-th inference node; , representing the confidence level of the detection result of the i-th inference node; , ; Represents the context edge between the i-th reasoning node and the j-th decision node, i.e., connecting the reasoning node and the decision node, and records the diagnostic reasoning relationship and path; This indicates the diagnostic reasoning relationship type between the i-th reasoning node and the j-th decision node, such as "abnormal signal strength → suspected wireless side fault"; This represents the generation time of the context edge between the i-th inference node and the j-th decision node; This represents the unique identifier of the diagnostic reasoning path between the i-th reasoning node and the j-th decision node; , ; This represents the p-th decision node; This represents the decision conclusion of the p-th decision node, such as "suspected wireless side fault" or "base station A radio frequency module fault". This represents the decision timestamp of the p-th decision node; The identifier of the diagnostic agent that generates the decision conclusion of the p-th decision node; T records the timestamps of all operations in the entire dynamic context graph to ensure the traceability of the fault diagnosis trajectory; W contains... for The weight of the inference path (fault diagnosis link) characterizes its reliability.

[0034] In step S2, the formula for initializing the dynamic context graph is: ; in, Represents the initial dynamic context graph; This represents an empty set, i.e., an initialized empty dynamic context graph.

[0035] In step S2, the update formula for the dynamic context graph is: ; ; in, Represents the dynamic context graph at time t; Represents the dynamic context graph at time t-1; This represents the i-th inference node; Represents the context edge between the i-th inference node and the p-th decision node; This represents the p-th decision node; express The weights; , represents the weighting coefficient, which can be adjusted according to the operation and maintenance needs of mobile communication networks; This represents the historical fault location accuracy of the diagnostic agent a that generates the detection results, such as a historical accuracy of 0.95 for the wireless side diagnostic agent. , representing the confidence level of the detection result of the i-th inference node. It is calculated by weighting the historical fault location accuracy of the diagnostic agent and the confidence level of the current detection results.

[0036] In the specific implementation process, the construction of the dynamic context graph is the foundation for achieving traceable and interpretable diagnosis. Its core lies in expanding the "what" knowledge of the traditional static knowledge graph into a dynamic context that records the "how to reason" throughout the entire diagnostic process.

[0037] The concretization of inference nodes: Taking a 5G base station access failure as an example, the detection result of the inference node can be "the reference signal received power (RSRP) of base station A is -110dBm" (numerical type) or "the S1 interface connection status of base station A is disconnected" (state type). Confidence can be assigned based on the signal-to-noise ratio of the detection data, the stability of historical data, or the diagnostic agent's own evaluation of its detection results. For example, conclusions derived through signaling analysis can be assigned a higher confidence level (e.g., 0.9), while conclusions derived from a single performance counter are assigned a lower confidence level (e.g., 0.6).

[0038] The significance of context edges: Context edges serve as bridges connecting data and conclusions. Relationship types include not only causal relationships (e.g., "RSRP too low" -> "leading to wireless access failure"), but also temporal relationships (e.g., "event B occurs after event A"), spatial relationships (e.g., "multiple users in the same cell experience the same problem"), etc. Unique path identifiers ensure that even in complex reasoning, different diagnostic paths can be clearly distinguished and traced.

[0039] The dynamic assignment logic of weights: The weight calculation formula reflects a dual consideration of the reliability of the diagnostic process. Historical accuracy represents the long-term reliability of the agent, while current confidence represents the immediate reliability of the current detection. The weight coefficient α can be adjusted according to the specific scenario. For example, in networks requiring high stability, a higher α (e.g., 0.7) can be set, relying more on experienced diagnostic agents; in scenarios with extremely high data quality, a lower α (e.g., 0.3) can be set, placing more trust in the current data.

[0040] In step S2, the step of recording and updating the detection results and decision conclusions of each diagnostic agent in real time to the dynamic context graph specifically involves: The detection results and decision conclusions of each diagnostic agent are used as inference nodes and decision nodes, and the inference nodes and decision nodes are associated through context edges. At the same time, the temporal information corresponding to the generation records of each inference node, decision node and context edge is recorded, and the context edge is assigned weight information to update the dynamic context graph.

[0041] In step S3, the formula for calculating the fault complexity is: ; in, This represents the fault complexity of fault Q. For simple faults, such as single-user terminal configuration errors or single-terminal disconnections, This is considered a medium-level fault, such as abnormal signal in a single cell of a base station or fluctuations in local transmission links. For complex faults, such as regional multi-base station congestion and multi-network element linkage faults, Low complexity threshold This is a high complexity threshold; This indicates the scale of entities involved in fault Q, such as the number of base stations, users, transmission links, and core network elements. This indicates the relationship levels involved in fault Q, such as "user terminal → base station → transmission network → core network" having 4 levels; denoted by , represents the length of the reasoning path involved in fault Q, i.e., the minimum number of detection steps from the fault phenomenon to locating the root cause; β, γ, and δ all represent weighting coefficients, and It can be calibrated according to mobile communication network scenarios, such as β=0.5 for base station dense fault scenarios and γ=0.5 for multi-network element linkage fault scenarios.

[0042] Entity size: This refers to the number of network entities affected by the fault. For example, a fault affecting only a single User Equipment (UE) has an entity size of 1; while a fault affecting all users under the entire base station (assuming 100) has an entity size of 100. This value needs to be normalized for weighted calculation with other parameters.

[0043] Relationship level: This refers to the number of network architecture layers that fault reasoning needs to traverse. For example, the path from "user terminal" to "base station" is layer 1, and the path to "Core Network User Plane Function (UPF)" is layer 2. A fault that needs to be traced from the terminal to the core network has a higher relationship level.

[0044] Inference path length: The minimum number of diagnostic steps required to identify the root cause. This can be estimated based on historical cases or domain expert knowledge. For example, a simple parameter configuration error may require only 1-2 steps, while a complex fault involving multiple network element interactions may require more than 5 steps.

[0045] Threshold setting: and The settings are not fixed and can be initialized by performing cluster analysis on historical fault data (such as the K-means algorithm) or based on operation and maintenance experience, and dynamically optimized based on the feedback of diagnostic results during system operation.

[0046] Step S4 specifically involves: Based on the aforementioned fault complexity, the number of agents participating in collaborative diagnosis is adaptively determined: ; in, This indicates the number of agents required for collaborative diagnosis of fault Q. The total number of intelligent agents in the pool shall not exceed the total number of agents in the pool. Represents the fault complexity of fault Q; The base complexity represents the maximum fault complexity that a single diagnostic agent can handle; round() represents the rounding function; for example, simple faults are assigned to 1 terminal agent, medium faults to 2 agents, and complex faults to 4 agents.

[0047] Calculate the matching degree between each diagnostic agent in the agent pool (including radio-side diagnostic agents, core network diagnostic agents, transport network diagnostic agents, service-side diagnostic agents, etc.) and fault Q: ; in, This represents the degree of matching between the diagnostic agent a and the fault Q; This represents the capability vector of the diagnostic agent a; The vector representing the diagnostic capability requirement for fault Q; This represents the capability vector of the diagnostic agent a in the i-th diagnostic dimension. This represents the diagnostic capability requirement vector for fault Q in the i-th diagnostic dimension. The core of agent matching degree is the construction of capability vector and demand vector. It can be defined as the capability score of the diagnostic agent in different diagnostic dimensions (such as wireless side, transmission side, core network side, and service application side). This is determined by the initial symptoms of the fault. For example, if the initial fault report is "video stuttering," then its demand vector values ​​on the "transmission side" (latency, jitter) and the "service side" (video server) will be high, while the value on the "wireless side" may be low. By using dot product calculations, the diagnostic agent best suited to handle such problems can be accurately selected.

[0048] Based on the matching degree, each diagnostic agent is ranked, and agents ranked before the matching degree are selected from the agent pool. A collaborative set of diagnostic agents; Simultaneously, based on the aforementioned fault complexity, an appropriate collaboration mode is selected to initiate collaborative fault diagnosis: ; in, Indicates the cooperation mode of fault Q; As a low complexity threshold, This is the high complexity threshold.

[0049] Serial collaboration: Diagnostic agents diagnose sequentially according to matching degree, such as first being detected by terminal-side agents and then re-detected by base station-side agents, reducing communication overhead.

[0050] Hybrid collaboration: Parallel reasoning for core diagnostic dimensions and serial reasoning for auxiliary dimensions, such as parallel detection of wireless and transmission side agents and serial verification of core network agents.

[0051] Parallel + Feedback Collaboration: Full-dimensional parallel diagnosis, real-time feedback of detection results to DCG, dynamic adjustment of diagnostic direction, such as parallel detection of wireless, core network, transmission, and service-side intelligent agents, and updating diagnostic focus based on real-time graph.

[0052] Step S5 specifically involves: During collaborative diagnosis, if contradictory detection results are found between inference nodes on the same inference path based on the dynamic context graph, a conflict is identified; the conflict determination formula is: ; in, Indicating inference nodes The test results, such as "wireless side fault"; express The unique identifier of the diagnostic reasoning path to which it belongs; The impact range of the conflicting inference node on the inference path is calculated to quantify the severity of the conflict; the formula for calculating the severity of the conflict is: ; in, Fault decision nodes representing conflict associations The set of all edges on the diagnostic reasoning path; Indicates the edge weight; a larger IR indicates a wider impact of conflicts on fault diagnosis; Based on the severity of the conflict and the confidence level of the detection results, the root cause of the conflict is located. The conflict root cause and its associated invalid reasoning path are removed from the dynamic context graph to resolve the conflict. The diagnostic agent is then reassigned to the optimized reasoning path. Based on the optimized reasoning path, the dynamic context graph is re-diagnosed and updated.

[0053] The specific process of conflict resolution and path optimization is as follows: (1) Locating the root cause of the conflict: Select the conflict with the highest IR severity and confidence level. The node is the source of the conflict. For example, “Core network S1 interface signaling lost (confidence level 0.58)”.

[0054] (2) Remove conflict sources and invalid links: .

[0055] (3) Reallocate agent resources: reallocate agents originally assigned to conflict paths to optimized diagnostic paths, such as adjusting core network agents to re-examine S1 interfaces and updating the cooperation set.

[0056] (4) Re-diagnose and update DCG: Re-perform fault diagnosis based on the optimized graph and enter new nodes and edges.

[0057] After the fault diagnosis path is adjusted or the conflict is resolved, update the confidence of the corresponding node / edge or delete invalid links, using the following formula: ; Set a confidence threshold, such as 0.6; delete invalid nodes with low confidence, such as "transmission link delay 50ms (confidence 0.55)", and add optimized new detection nodes, such as "transmission link delay 20ms (confidence 0.93)".

[0058] In step S6, the formula for calculating the diagnostic result is: ; ; Where Res represents the diagnostic result; Represents the p-th decision node; DCG represents the optimized and updated dynamic context graph; Indicates the overall confidence level of the decision node; Represents the context edge between the i-th inference node and the p-th decision node; Decision nodes representing conflict associations The set of all context edges on the reasoning path; express The weight.

[0059] The overall confidence score of a decision node is calculated as the average weight across all reasoning paths leading to that decision. This means that even if a conclusion is proposed by a highly accurate agent, its overall confidence score will be lowered if the supporting detection data (inference nodes) generally have low confidence or are of unreliable origin. Conversely, a conclusion supported by multiple independent and highly reliable pieces of evidence will have a high overall confidence score, even if the weight of each individual piece of evidence is low. This mechanism greatly enhances the scientific rigor and reliability of the final conclusion.

[0060] In summary, the advantages of this invention are: 1. A dynamic context graph is constructed to record the inference node set N, context edge set E, decision node set D, temporal sequence set T, and weight set W in real time during the fault diagnosis process. Then, in response to the triggering of the fault diagnosis task, the dynamic context graph is initialized, and during subsequent diagnosis, the detection results and decision conclusions of each diagnostic agent are recorded and updated to the dynamic context graph in real time. Next, based on the initial information of the fault diagnosis task, the scale of entities involved in the fault, the relationship hierarchy, and the length of the inference path are evaluated to quantify the fault complexity. Based on the fault complexity, the number of agents participating in collaborative diagnosis is adaptively determined to match corresponding diagnostic agents from the agent pool to form a collaborative set. Simultaneously, an appropriate collaborative mode is selected based on the fault complexity to initiate collaborative fault diagnosis. During collaborative diagnosis, conflict identification is performed based on the dynamic context graph, the root cause of the conflict is located and resolved, and the dynamic context graph is optimized and updated. From the optimized and updated dynamic context graph, the decision conclusion with the highest confidence is extracted as the final diagnosis result, and the final dynamic context graph is updated. Archiving is performed; that is, by constructing and updating a dynamic context graph that records the complete reasoning chain in real time, resource utilization is improved first. The key lies in the adaptive allocation of agent scale and cooperation mode based on the quantitative assessment of fault complexity, thereby matching simplified resources for simple faults and allocating sufficient computing power for complex faults, avoiding resource waste or insufficient capacity. Furthermore, the improvement in diagnostic accuracy is due to the fact that the graph provides a shared context information platform with temporal sequence and confidence for multiple agents, which not only avoids information silos and repetitive reasoning, but also enables precise location and resolution of conflicts based on panoramic information in collaborative diagnosis, thus drawing more reliable conclusions. Finally, the enhanced system robustness stems from the system's self-learning ability. After each conflict resolution, the dynamic context graph itself is optimized in reverse (such as adjusting weights and marking weak links), enabling the system to learn iteratively from historical diagnoses, effectively suppressing the recurrence and cascading diffusion of similar conflicts, thus exhibiting stronger adaptability and stability in the face of complex and ever-changing environments, ultimately greatly improving the resource utilization, diagnostic accuracy and system robustness of fault diagnosis.

[0061] 2. By constructing a dynamic context graph that dynamically records inference nodes, context edges, decision nodes, time sequences, and weights, the entire diagnostic process is made traceable and multi-agent information is shared in real time, fundamentally avoiding duplicate detections and inference conflicts caused by information asymmetry. By quantitatively assessing the complexity of the fault and adaptively determining the number of agents, selecting the best matching agents, and dynamically switching the collaboration mode, resource investment is precisely matched with the fault difficulty, effectively eliminating resource redundancy or diagnostic blind spots. Furthermore, based on real-time data from the graph, the root causes of conflicts are accurately identified, conflicts are resolved, and inference paths are optimized, forming a closed-loop self-optimization mechanism, thereby systematically improving the resource utilization, diagnostic accuracy, and system robustness of fault diagnosis.

[0062] 3. By constructing a dynamic context graph (DCG), inference nodes, decision nodes, and their relationships are recorded in real time and updated based on confidence and weight. This structured recording method ensures the integrity and traceability of the diagnostic process, thereby reducing the possibility of misjudgment. For example, by extracting the decision conclusion with the highest confidence and combining it with weight calculation, results with high credibility can be selected first, avoiding the bias or error that may exist in a single intelligent agent, and significantly improving the accuracy and overall reliability of the diagnosis.

[0063] 4. By introducing a fault complexity assessment mechanism and adaptively determining the number of agents and cooperation mode (such as serial, hybrid, or parallel cooperation) based on the complexity, this adaptive capability enables the system to dynamically allocate resources according to the complexity of the fault, avoiding overuse of agents on simple faults (saving computing resources) or insufficient resources on complex faults (ensuring coverage). This not only optimizes diagnostic efficiency and reduces unnecessary latency, but also improves the system's response speed, especially in real-time fault scenarios.

[0064] 5. Through conflict identification and resolution mechanisms, the system can detect contradictory results on the reasoning path based on the dynamic context graph, quantify the severity of the conflict, and then locate and remove the root cause of the conflict. This proactive conflict management avoids inconsistencies in the diagnosis process and ensures the optimization and updating of the diagnosis path. For example, reallocating agents to the optimized path can effectively handle the common problem of disagreements in multi-agent collaboration, improve the stability and fault tolerance of the system, and make it more robust in complex or dynamic environments.

[0065] 6. The modular design, such as the general representation of the dynamic context graph and the matching mechanism of the agent pool, makes the method easy to extend to fault diagnosis scenarios in different fields. The quantitative assessment of fault complexity and the selection of collaboration mode do not depend on specific hardware or environment, but are based on adjustable parameters (such as β, γ, δ weight coefficients), which gives the system a high degree of flexibility. Enterprises or users can customize the graph dimensions and agent capabilities according to actual needs, thereby reducing deployment costs and supporting future technology iterations.

[0066] 7. The dynamic context graph records the timestamps, inference paths, and agent origins of all diagnostic operations in detail, providing a complete diagnostic chain. This transparent recording makes diagnostic decisions no longer a "black box," facilitating user retrospective analysis, auditing, or training. For example, the archived final graph can be used for post-analysis or compliance checks, which not only enhances the credibility of the method but also meets the growing demand for interpretability of fault diagnosis in the industrial field.

[0067] 8. By innovatively combining graph theory, multi-agent systems and adaptive algorithms, a unique collaborative diagnostic framework has been formed. Compared with traditional single-agent or fixed-rule methods, this integration utilizes the structural advantages of graphs (such as capturing complex relationships) and the distributed intelligence of multi-agent systems (such as parallel processing), while avoiding the redundancy of technology stacking through adaptive mechanisms.

[0068] 9. The dynamic context graph fully records the inference nodes, decision nodes, timing, confidence level, and link weights of the fault diagnosis. It can not only output the final fault conclusion, but also clearly trace the entire process of "detection data → inference logic → decision generation". It solves the problem that traditional static graphs cannot explain "why the conclusion is reached", greatly reduces the cost of troubleshooting after fault location deviation, improves system debugging efficiency, and realizes full-link traceability and interpretability of the diagnosis process.

[0069] 10. The number of agents and the collaboration mode are dynamically adjusted according to the complexity of the fault through an adaptive collaboration mechanism. For simple faults such as single terminal configuration errors, the number of agents is reduced and serial collaboration is adopted to reduce computing and communication overhead. For complex faults such as regional multi-base station congestion, the scale of agents is expanded and parallel + feedback collaboration is adopted to ensure full coverage of diagnostic dimensions and effectively improve the resource utilization of diagnostic agents.

[0070] 11. The conflict resolution mechanism based on dynamic context graph can accurately locate the root cause of low-confidence conflicts, rather than just dealing with surface conflicts, effectively avoiding the overall diagnostic failure caused by the cascading spread of small conflicts; at the same time, through link optimization and intelligent weight allocation, the diagnostic path can be self-iterated, the recurrence rate of similar conflicts is significantly reduced, and accurate and efficient conflict resolution is achieved, greatly enhancing the robustness of the system.

[0071] 12. After the fault diagnosis is completed, the complete dynamic context graph will be archived and stored. This historical diagnostic data can be used to optimize the calculation rules of the capability vector and weight coefficient of the diagnostic agent. As diagnostic tasks are continuously accumulated, the accuracy of locating complex faults will gradually improve, adapting to the needs of long-term operation and maintenance of mobile communication networks.

[0072] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A dynamic context graph adaptive multi-agent cooperative fault diagnosis method, characterized in that: Includes the following steps: Step S1: Construct a dynamic context graph for real-time recording of the inference node set N, context edge set E, decision node set D, time series set T, and weight set W during the fault diagnosis process; Step S2: In response to the triggering of the fault diagnosis task, initialize the dynamic context graph, and in the subsequent diagnosis process, record and update the detection results and decision conclusions of each diagnostic agent in real time to the dynamic context graph. Step S3: Based on the initial information of the fault diagnosis task, assess the entity size, relationship hierarchy and reasoning path length involved in the fault to quantify the fault complexity of the current fault. Step S4: Based on the fault complexity, adaptively determine the number of agents participating in collaborative diagnosis, and match corresponding diagnostic agents from the agent pool to form a collaborative set based on the number of agents. At the same time, select the corresponding collaborative mode based on the fault complexity to start collaborative fault diagnosis. Step S5: During the collaborative diagnosis process, conflict identification is performed based on the dynamic context graph, the root cause of the conflict is located and the conflict is resolved, and the dynamic context graph is optimized and updated at the same time. Step S6: Extract the decision conclusion with the highest confidence from the optimized and updated dynamic context graph as the final diagnosis result, and archive the final dynamic context graph.

2. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: In step S1, the dynamic context graph is represented as follows: ; DCG stands for Dynamic Context Graph; , ; This represents the i-th inference node; This represents the detection result of the i-th inference node; Represents the detection timestamp of the i-th inference node; , representing the confidence level of the detection result of the i-th inference node; , ; Represents the context edge between the i-th reasoning node and the j-th decision node; This represents the diagnostic reasoning relationship type between the i-th reasoning node and the j-th decision node; This represents the generation time of the context edge between the i-th inference node and the j-th decision node; This represents the unique identifier of the diagnostic reasoning path between the i-th reasoning node and the j-th decision node; , ; This represents the p-th decision node; This represents the decision conclusion of the p-th decision node; This represents the decision timestamp of the p-th decision node; The diagnostic agent identifier represents the decision conclusion generated at the p-th decision node; T records the timestamps of all operations in the entire dynamic context graph; W contains... for The weight.

3. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: In step S2, the formula for initializing the dynamic context graph is: ; in, Represents the initial dynamic context graph; This represents the empty set.

4. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: In step S2, the update formula for the dynamic context graph is: ; ; in, Represents the dynamic context graph at time t; Represents the dynamic context graph at time t-1; This represents the i-th inference node; Represents the context edge between the i-th inference node and the p-th decision node; This represents the p-th decision node; express The weights; , represents the weighting coefficient; The historical fault location accuracy of the diagnostic agent a that generates the detection results; , where represents the confidence level of the detection result of the i-th inference node.

5. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: In step S2, the step of recording and updating the detection results and decision conclusions of each diagnostic agent in real time to the dynamic context graph specifically involves: The detection results and decision conclusions of each diagnostic agent are used as inference nodes and decision nodes, and the inference nodes and decision nodes are associated through context edges. At the same time, the temporal information corresponding to the generation records of each inference node, decision node and context edge is recorded, and the context edge is assigned weight information to update the dynamic context graph.

6. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: In step S3, the formula for calculating the fault complexity is: ; in, This represents the fault complexity of fault Q. For a simple fault, This is a medium-level fault. For complex faults, As a low complexity threshold, This is a high complexity threshold; Indicates the size of the entities involved in fault Q; Indicates the hierarchical relationship involved in fault Q; β represents the length of the inference path involved in fault Q; β, γ, and δ all represent weight coefficients, and .

7. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: Step S4 specifically involves: Based on the aforementioned fault complexity, the number of agents participating in collaborative diagnosis is adaptively determined: ; in, This indicates the number of agents required to participate in the collaborative diagnosis of fault Q; Represents the fault complexity of fault Q; This represents the baseline complexity, corresponding to the maximum fault complexity that a single diagnostic agent can handle; round() represents the rounding function. Calculate the matching degree between each diagnostic agent in the agent pool and the fault Q: ; in, This represents the degree of matching between the diagnostic agent a and the fault Q; This represents the capability vector of the diagnostic agent a; The vector representing the diagnostic capability requirement for fault Q; This represents the capability vector of the diagnostic agent a in the i-th diagnostic dimension. This represents the diagnostic capability requirement vector for fault Q in the i-th diagnostic dimension. Based on the matching degree, each diagnostic agent is ranked, and agents ranked before the matching degree are selected from the agent pool. A collaborative set of diagnostic agents; Simultaneously, based on the aforementioned fault complexity, an appropriate collaboration mode is selected to initiate collaborative fault diagnosis: ; in, Indicates the cooperation mode of fault Q; As a low complexity threshold, This is the high complexity threshold.

8. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: Step S5 specifically involves: During collaborative diagnosis, if contradictory detection results are found between inference nodes on the same inference path based on the dynamic context graph, a conflict is identified. Calculate the extent of the impact of the conflicting inference node on the inference path to quantify the severity of the conflict; Based on the severity of the conflict and the confidence level of the detection results, the root cause of the conflict is located. The conflict root cause and its associated invalid reasoning path are removed from the dynamic context graph to resolve the conflict. The diagnostic agent is then reassigned to the optimized reasoning path. Based on the optimized reasoning path, the dynamic context graph is re-diagnosed and updated.

9. The dynamic context graph adaptive multi-agent cooperative fault diagnosis method as described in claim 1, characterized in that: In step S6, the formula for calculating the diagnostic result is: ; ; Where Res represents the diagnostic result; Represents the p-th decision node; DCG represents the optimized and updated dynamic context graph; Indicates the overall confidence level of the decision node; Represents the context edge between the i-th inference node and the p-th decision node; Decision nodes representing conflict associations The set of all context edges on the reasoning path; express The weight.

Images