Feature enhancement method, defect positioning method and system based on multi-agent debate

Abstract

The invention belongs to the technical field of artificial intelligence, and provides a feature enhancement method and system and a defect positioning method and system based on multi-agent debate. The multi-agent debate-based feature enhancement method comprises the following steps: receiving each viewpoint of a current round of debate generated by input of each agent to the same task; determining an effective viewpoint set by using each viewpoint of the round of debate; calculating convergence index values of the viewpoints based on the effective viewpoint set; when it is determined that the set debate round number is not exceeded and the convergence condition is met, outputting a stabilization conclusion and taking the stabilization conclusion as high-order feature enhancement information; and when it is determined that the set debate round number is not reached and the convergence condition is not met, receiving the viewpoint obtained by updating the viewpoint of each agent based on the read viewpoints of other agents, taking the updated viewpoint as the viewpoint of the current round of debate, and skipping to the step of determining the effective viewpoint set. A multi-agent debate mechanism is introduced, the conclusion robustness is improved through a multi-path opinion game, and the reliability of an output result can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a feature enhancement method, a defect positioning method and a system based on multi-agent debate. BACKGROUND

[0002] In complex networks and large-scale systems, fault diagnosis and decision-making often rely on rule bases or reinforcement learning agents. Such methods require training models for specific tasks, for example, AlphaGo, which requires specialized model training for the single task of playing Go to achieve performance beyond human level. However, if switching to another task, such as chess, the model must be retrained, and cross-task generalization cannot be achieved. This approach has obvious limitations in terms of training cost, time consumption, and application flexibility, and has the problems of insufficient generality and high migration cost.

[0003] With the advent of large language models (LLM), the implementation of agents has been improved. LLMs have natural language interaction, extensive knowledge, reasoning ability, and external extensibility, enabling LLM-driven agents to quickly migrate between different tasks.

[0004] However, the output of a single LLM may have uncertainty and illusion, which makes the output of an agent using a single LLM also have uncertainty and illusion, and has the problem of low reliability. SUMMARY

[0005] To solve the above problems, the application provides a feature enhancement method, a defect positioning method and a system based on multi-agent debate, which introduces a multi-agent debate mechanism to improve the robustness of conclusions through multi-path viewpoint game, and can improve the reliability of the output results.

[0006] In a first aspect, a feature enhancement method based on multi-agent debate is provided, comprising: receiving each viewpoint of the current round of debate generated by each agent for the same task, wherein each viewpoint includes: a task conclusion, a reliability, and evidence, and the large models driving each agent are different from each other; determining an effective viewpoint set using each viewpoint of the current round of debate, wherein an effective viewpoint is a representative and mutually distinguishable viewpoint; calculating a convergence index value of the viewpoint based on the effective viewpoint set, wherein the convergence index is: a viewpoint consistency rate and an average reliability; or: a viewpoint consistency rate, a semantic distribution difference, and an average reliability; the viewpoint consistency rate is the number of agents consistent with the mode conclusion in the current round / the number of agents participating in the speech; the semantic distribution difference is the difference degree of the semantic distribution of the viewpoint in the current round and the semantic distribution of the viewpoint in the last round; When it is determined that the set number of debate rounds is not exceeded and the convergence condition is met, a stabilized conclusion is output and used as high-order feature enhancement information, wherein the convergence condition is that any two of the convergence indicators meet the condition that the indicator value is greater than or equal to the corresponding indicator threshold, and the stabilized conclusion includes a task final aggregation conclusion, an aggregation reliability, and support evidence summary; When it is determined that the set number of debate rounds is not reached and the convergence condition is not met, the updated viewpoints of each agent based on reading the viewpoints of other agents are received, the updated viewpoints are used as the viewpoints of the current round of debate, and the step of determining the effective viewpoint set using the viewpoints of the current round of debate is jumped to.

[0007] Further, when it is determined that the set number of debate rounds is reached and the convergence condition is not met, the core viewpoint set is determined using the viewpoints of the current round of debate; The convergence indicator value of the viewpoint is calculated again based on the core viewpoint set; When the recalculated convergence indicator value meets the convergence condition, a stabilized conclusion is output and used as high-order feature enhancement information; when the recalculated convergence indicator value does not meet the convergence condition, an "unstable" label is output and used as high-order feature enhancement information, wherein the "unstable" label includes a primary candidate conclusion and a secondary candidate conclusion, a to-be-reviewed label, and a reviewable suggestion.

[0008] Further, determining the effective viewpoint set using the viewpoints of the current round of debate includes: The comprehensive similarity between the viewpoints of any two agents in the current round of debate is calculated to obtain a comprehensive similarity dataset, wherein the comprehensive similarity is a weighted sum of the task conclusion consistency, the reliability proximity, and the evidence chain similarity; Based on the comprehensive similarity dataset, the viewpoints are de-duplicated and de-noised to obtain the effective viewpoint set.

[0009] Further, based on the comprehensive similarity dataset, the viewpoints are de-duplicated and de-noised to obtain the effective viewpoint set, including: A viewpoint merging graph is constructed: taking the viewpoints as vertices, and establishing an edge between two viewpoints when the comprehensive similarity between the two viewpoints is greater than or equal to a merging threshold; Each connected component is taken as a cluster, and a representative viewpoint is selected therefrom to achieve de-duplication; For each viewpoint obtained after de-duplication, a representative comprehensive similarity score of each viewpoint is determined; When the representative comprehensive similarity score is less than a rejection threshold, the corresponding viewpoint is rejected to achieve de-noising, wherein the rejection threshold is greater than or equal to 0 and less than the merging threshold, and the merging threshold is less than or equal to 1.

[0010] Further, before constructing the viewpoint merging graph, further comprising: based on the comprehensive similarity dataset and the set probability level, determining the merging threshold and the elimination threshold by using the quantile function; Selecting a representative viewpoint from each connected component as a cluster, comprising: Selecting a representative viewpoint from each connected component as a cluster, determining the highest weighted sum score of the centrality and reliability of the cluster, and selecting the viewpoint as the representative viewpoint.

[0011] Further, determining the core viewpoint set by using the viewpoints of the current round of debate, comprising: Calculating the comprehensive similarity between the viewpoints of the current round of debate of any two agents to obtain a comprehensive similarity dataset; Adding each comprehensive similarity between the viewpoints of the current round of debate of each agent and the viewpoints of other agents to obtain the total similarity of the viewpoints of the current round of debate of the agent; Constructing a total similarity dataset by using the total similarity of each agent; Taking the viewpoints corresponding to the top K total similarities in the total similarity dataset as the core viewpoint set.

[0012] Further, the task is a defect positioning task.

[0013] In a second aspect, a defect positioning method based on multi-agent debate is provided, comprising: Each parsing agent in the input layer agent parses the original data of a type matching the parsing type of the agent based on a prompt guide and in combination with a task-related long context data attribute completion strategy to obtain structured input including an original layer, a preprocessing layer, and a label layer, wherein the input types parsed by each parsing agent are different, the prompt is used for semantic constraints on multi-modal input, and the semantic constraints include redundancy, accuracy, and closed loop; The high-order feature enhancement agent processes the structured input by using the feature enhancement method to obtain high-order feature enhancement information, wherein the task is a defect positioning task; The multi-modal feature encoding agent introduces a context completion strategy to perform feature encoding on the high-order feature enhancement information; The decision layer agent selects a discriminator based on a converged label in the feature encoding; and generates a final decision result by using the selected discriminator, wherein when the converged label is not converged, an XGBoost discriminator is selected; and when the converged label is converged, an SVM discriminator is selected.

[0014] In a third aspect, a high-order feature enhancement agent based on multi-agent debate is provided, comprising: receive each viewpoint of the current round of debate generated by inputs of each agent to the same task, wherein each viewpoint comprises a task conclusion, a reliability and evidence, and wherein the large models driving each agent are different from each other; determine a set of effective viewpoints using each viewpoint of the current round of debate, wherein an effective viewpoint is a viewpoint that is representative and distinguishable from each other; calculate a convergence indicator value of the viewpoints based on the set of effective viewpoints, wherein the convergence indicator is: a viewpoint consistency rate and an average reliability; or: a viewpoint consistency rate, a semantic distribution difference and an average reliability; wherein the viewpoint consistency rate is the number of agents in the current round that are consistent with the mode conclusion / the number of agents participating in the speech; and the semantic distribution difference is the difference degree of the semantic distribution of the viewpoints of the current round and the semantic distribution of the viewpoints of the last round; output a stabilized conclusion as high-level feature enhancement information when it is determined that the set number of debate rounds is not exceeded and the convergence condition is met, and output the viewpoint of each agent obtained by updating its own viewpoint based on reading the viewpoints of other agents as the viewpoint of the current round of debate, trigger the determination module when it is determined that the set number of debate rounds is not reached and the convergence condition is not met, wherein the convergence condition is that any two indicators satisfy the indicator value greater than or equal to the corresponding indicator threshold, and the stabilized conclusion comprises a final aggregated conclusion of the task, an aggregated reliability and a support evidence summary.

[0015] In a fourth aspect, a defect positioning system based on multi-agent debate is provided, comprising: an input layer agent, a high-level feature enhancement agent as described above, a multi-modal feature encoding agent and a decision layer agent; wherein the input layer agent is composed of each analysis agent, and each analysis agent analyzes different types of inputs; Each analysis agent is configured to analyze original data of a type matching the analysis type thereof based on a Prompt guide and in combination with a long context data attribute completion strategy related to a task, to obtain structured inputs comprising an original layer, a preprocessing layer and a label layer, wherein the Prompt is configured to perform semantic constraints on multi-modal inputs, and the semantic constraints comprise redundancy, accuracy and closed-loop; The high-level feature enhancement agent is configured to process the structured inputs according to a task based on a multi-agent debate mechanism to obtain high-level feature enhancement information; wherein the task is a defect positioning task; The multi-modal feature encoding agent is configured to introduce a context completion strategy to encode the high-level feature enhancement information; The decision layer agent is configured to select a discriminator based on a convergence label in the feature encoding; when the convergence label is non-convergence, the XGBoost discriminator is selected; and when the convergence label is convergence, the SVM discriminator is selected; and the selected discriminator is used to generate a final decision result.

[0016] In a fifth aspect, an electronic device is provided, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus. The memory is configured to store a computer program. The processor is configured to execute the program stored on the memory to implement the method described above.

[0017] In a sixth aspect, a computer storage medium is provided, and the computer storage medium stores a computer program, and the computer program is executed by a processor to implement the method described above.

[0018] Compared with the prior art, the present application has the following advantages: The present application is based on a LLM-driven multi-agent collaborative debate mechanism, combined with multi-modal feature input, feature encoder layer and adaptive decision layer, to realize efficient and robust reasoning for defect assisted positioning, with the following beneficial effects: (1) Positioning accuracy is improved: through the multi-agent debate mechanism, different reasoning paths and evidence chains are fused, effectively reducing the single model bias and improving the accuracy of defect positioning.

[0019] (2) Intelligent termination and efficiency guarantee: an intelligent termination mechanism based on the consistency rate of viewpoints, semantic distribution difference and average confidence is introduced, so that the debate process can converge within a limited number of rounds, effectively improving the computing efficiency.

[0020] (3) Result stability and interpretability: in the non-convergence scenario, stable core viewpoints are selected through outlier detection and similarity aggregation to avoid conclusion gaps and provide transparent evidence chains, enhancing the reviewability and interpretability of the decision.

[0021] (4) Multi-modal processing capability: the input layer uniformly adopts a three-layer structured representation (RawInput layer, Preprocessed layer, Labels layer), supporting unified parsing and alignment of multi-modal information such as logs, performance indicators and structured data.

[0022] (5) Dynamic adaptability: the feature encoder layer can adaptively encode both convergent and non-convergent scenarios, allowing the decision layer to flexibly select different classifiers such as SVM and XGBoost, improving the adaptability to complex scenarios.

[0023] Other features and advantages of the present application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0025] Figure 1 The layered intelligent agent architecture diagram provided for the embodiments of the present application; Figure 2 The input layer agent architecture diagram provided for the embodiments of the present application; Figure 3-1 The prompt schematic provided for the embodiments of the present application: Figure 3-2 The original information provided for the embodiments of the present application; Figure 3-3 The output of the original information of the embodiments of the present application after being processed by the input layer agent ; Figure 3-4 The output of the original information of the embodiments of the present application after being processed by the input layer agent ; Figure 3-5 The output of the original information of the embodiments of the present application after being processed by the input layer agent ; Figure 4 The flowchart of the high-order feature enhancement method based on the multi-agent debate mechanism provided for the embodiments of the present application; Figure 5 The stability judgment flowchart provided for the embodiments of the present application; Figure 6 The similarity matrix provided for the embodiments of the present application ; Figure 7-1 Multi-agent debate simulation example 1-round 1 provided for the embodiments of the present application; Figure 7-2 Multi-agent debate simulation example 1-round 2 provided for the embodiments of the present application; Figure 7-3 Multi-agent debate simulation example 1-output provided for the embodiments of the present application; Figure 8-1Multi-Agent argumentation simulation example 2- round 1 provided for the embodiments of the present application; Figure 8-2 Multi-Agent argumentation simulation example 2- round 2 provided for the embodiments of the present application; Figure 8-3 Multi-Agent argumentation simulation example 2- round 3 provided for the embodiments of the present application; Figure 8-4 Multi-Agent argumentation simulation example 2- output provided for the embodiments of the present application. DETAILED DESCRIPTION

[0026] The present application proposes a defect aided positioning method based on multi-agent Agent collaborative argumentation mechanism. Unlike traditional reinforcement learning driven Agent, the Agent of the present application is embedded with LLM, which has the following advantages: natural language interaction capability: can directly understand and reason on unstructured data such as system logs and work order feedback; cross-modal semantic alignment capability: can map heterogeneous information such as logs and numerical indicators to a unified semantic space, supporting cross-modal reasoning; deep reasoning and interpretability: simulate expert discussion through multi-Agent argumentation process, compete and integrate different reasoning paths, and gradually enhance feature semantics; flexible extensibility: can call external knowledge and tools, and has the ability of continuous adaptation and evolution when facing complex and variable defect scenarios. By introducing LLM driven Agent, the present application not only overcomes the limitations of traditional Agent relying on specific task training, but also realizes efficient and robust defect aided positioning under multi-modal input and complex reasoning conditions, significantly improving decision accuracy and system robustness.

[0027] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0028] As shown in Figure 1 The overall architecture diagram of the defect aided positioning system based on multi-agent collaborative argumentation mechanism in the embodiments of the present application is shown, which is a layered agent architecture. The system includes four key levels: input layer Agent, high-order feature enhancement Agent, multi-modal feature enhancement encoding Agent and decision layer Agent.

[0029] The input layer Agent, as a semantic structured expert, is composed of various parsing Agents, wherein each parsing Agent parses different types of input. The specific parsing Agents are determined according to the type of the original data. For example, if the original data is of text type, log type, and table (a type of structured type), the input layer Agent can include a text parsing Agent, a structured data parsing Agent, and a log parsing Agent.

[0030] Each parsing Agent is used to parse the original data of a type matching the parsing type of the Agent based on Prompt guidance and in combination with a long context data attribute completion strategy, to obtain structured input including an original layer, a preprocessing layer, and a label layer. The high-order feature enhancement Agent is used to process the structured input based on a multi-agent debate mechanism to obtain high-order feature enhancement information. The multi-modal feature encoding Agent is used to introduce a context completion strategy to encode the high-order feature enhancement information. The decision layer Agent is used to select a discriminator based on the converged label in the feature encoding. When the converged label is not converged, the XGBoost discriminator is selected. When the converged label is converged, the SVM discriminator is selected. The selected discriminator is used to generate a final decision result.

[0031] The following describes each layer: Layer 1: Input layer Agent (semantic structured expert) 1.1 Overview: The input layer Agent can intelligently process multi-modal data. It realizes decoupling and unified structured representation of cross-modal input based on Prompt guidance, and combines a long context data attribute completion strategy related to the task to complete the context or environmental information related to the task in the original input.

[0032] This intelligent processing ensures diversity of input sources while maintaining consistency across modalities and corpora. This allows data from different modalities and sources to be unified in structure and semantics, so that subsequent Agents do not need to be concerned about the type of input source and can be processed based on a unified interface. The specific mechanisms used include: 1) Cross-modality consistency constraint and context attribute completion mechanism based on Prompt: Referring to the research progress of prompt learning, based on the pattern constraint theory in natural language processing (NLP), in combination with the defect auxiliary positioning task, an integrated prompt design mechanism for optical communication equipment testing is introduced. The mechanism introduces the link constraint of the input processing output in the prompt, so as to ensure that the output structures of different analysis agents are consistent.

[0033] At the same time, in combination with the long context data attribute completion strategy related to the task, the missing context information or environmental attributes in the original data (original input) are automatically completed in the analysis stage, so that the input layer agent can still generate complete and comparable standardized results when facing incomplete or ambiguous inputs, and has stronger universality and expansibility.

[0034] 2) Three-layer structure design of cross-modal unification: According to the layered architecture theory in computer science, different types of inputs (text, structured data, logs, etc.) are converted into a unified three-layer structured representation: Raw Input layer (original layer), Preprocessed layer (preprocessing layer), and Labels layer (label layer).

[0035] Raw Input layer: maintain the original input content to ensure traceability; Preprocessed layer: the input layer agent denoises, standardizes, analyzes or extracts features from the data to form an intermediate expression; Labels layer: extracts artificial judgment information, and clearly defines semantic labels and candidate conclusions.

[0036] 3) Decoupled input layer agent design: Through the decoupled mode of input layer agent exclusive analysis + unified structured output, the subsequent processing does not need to care about the data source and modal characteristics; only the unified format interface needs to be called for reasoning and fusion, thereby greatly simplifying the system coupling degree and improving the system robustness and maintainability.

[0037] 1.2 Architecture composition: It is composed of a plurality of full-time agents. Each type of data is adaptively entered into the corresponding agent analysis, and does not interfere with each other, while outputting a unified format. The main components of the input layer agent architecture are as shown in Figure 2 .

[0038] 1) Text analysis agent: responsible for processing text inputs such as natural language description, alarm information, and user feedback.

[0039] 2) Structured Data Parsing Agent: Responsible for processing structured numerical inputs such as tables, sensor data, test records, etc.

[0040] 3) Log Parsing Agent: Responsible for processing system logs, device logs, operation records, etc. time series text.

[0041] 4) Extensible Agent Interface: Reserved interface for future addition of parsing agents for voice, traffic, video, etc. input types.

[0042] 1.3 Prompt-based cross-modal consistency constraints and context attribute completion mechanism: 1) Use Prompt as a constraint language to regulate data flow, similar to defining a cross-modal protocol standard (similar to API DSL such specifications).

[0043] 2) Integrated design mechanism based on Prompt, which realizes low ambiguity and high structuring by uniformly constraining the semantics of multi-modal input; the semantic constraints include redundancy, accuracy, and closed loop, wherein: Redundancy: used to ensure that key information is not missed through context (such as test task background information, test plan, and related technical development materials, etc.) completion; Accuracy: used to constrain the consistency between input semantics and original intent, avoiding semantic drift; Closed loop: as much as possible to form a closed record of input-processing-output, accumulating "phenomenon-essence" related data.

[0044] The integrated design mechanism based on Prompt further customizes and adapts to test task scenarios, ensuring that inputs remain consistent and interpretable under multi-modal and multi-source conditions, thereby improving the reliability and controllability of the defect assistance positioning system.

[0045] Specifically, Figure 3-1 Prompt is given: when used, only replace , the rest of the parsing logic is consistent. 1.4 Specific embodiments: Input: such as Figure 3-2 Example raw information.

[0047] Output: Raw Input: maintains the original integrity, with as much complete context information as possible; such as Figure 3-3 Example raw information after input layer Agent processing .

[0048] Preprocessed: Clean, normalize, and complete the original data in context, including environment, time, tools, and reference thresholds, while extracting key meta-information; such as Figure 3-4 Example of original information output after input layer Agent processing .

[0049] Labels preliminary classification / entity recognition / relation extraction); such as Figure 3-5 Example of original information output after input layer Agent processing .

[0050] Layer 2: High-level feature enhancement Agent 2.1 Overview: The embodiment of the present application proposes a high-level feature enhancement method based on multi-agent debate mechanism. The execution subject of this method is a high-level feature enhancement Agent, which organizes a multi-agent cluster composed of multiple heterogeneous large models to carry out multi-round debate and collaborative reasoning. The input can be a structured input of multi-modal unified representation provided by the input layer Agent. After each round of debate, the high-level feature enhancement Agent performs "comprehensive similarity-based de-redundancy and de-noising": merging overly similar / repeated opinions, removing significantly outlying opinions, and retaining a representative and mutually distinct set of opinions. The set of opinions is then used for subsequent group discussion and feature aggregation, achieving a dynamic balance between individual competition and group optimization, and obtaining more robust high-level semantic feature representations. This scheme can be used in various tasks, such as defect assisted positioning, fault assisted positioning, paper quality review, and innovation point review. The embodiment of the present application takes the defect assisted positioning task as an example for illustration.

[0051] 2.2 Architecture composition: composed of multiple agents. Each agent is deployed with a different large model, with the ability of autonomous language generation, reference, and context memory.

[0052] This layer contains multiple large model Agents equipped with different LLMs, such as: Chinese semantic understanding is deep, and the knowledge base covers all; It has strong multi-modal capabilities (text + picture + code); It has stable dialogue and good reasoning ability; Its mathematical and logical performance is outstanding; It has strong Chinese optimization capabilities.

[0053] These agents receive structured inputs, conduct independent analysis based on their pre-trained background knowledge, and enter the debate process.

[0054] 2.3 Debate mechanism 1) Each agent independently reads the structured input, combines the embedded knowledge in the model to generate preliminary judgments and reasoning; 2) Each agent reads the conclusions and reasoning chains of other agents, updates its own conclusions and gives reasons (refutation, support, supplement, variant view or silence); 3) The debate process can be iterated for several rounds, and the consensus semantics is updated and the context supplement features are added each round; 4) Allow "silence". In each round of debate, some agents may choose "not to speak" because of lack of domain knowledge, lack of current information or that the view is highly repetitive with the existing discussion. This can reduce noise input, avoid forced output of low-quality conclusions when there is no new information, and improve convergence speed.

[0055] The flow of the high-order feature enhancement method based on the above multi-agent debate mechanism is shown in Figure 4 , and specifically includes the following steps: S1: receiving each viewpoint of the current round of debate generated by each agent on the structured input of the same task, wherein each viewpoint includes: task conclusion, reliability and evidence, and the large models driving each agent are different from each other; S2: determining an effective viewpoint set using each viewpoint of the current round of debate, wherein an effective viewpoint is a representative and mutually distinguishable viewpoint; In this step S2, the effective viewpoint set generated by each round of debate reflects the distribution evolution of the viewpoints.

[0056] S3: calculating a convergence index value of the viewpoints based on the effective viewpoint set.

[0057] Wherein, the convergence index is: the viewpoint consistency rate A and the average reliability; or: the viewpoint consistency rate, the semantic distribution difference and the average reliability; The viewpoint consistency rate is the number of agents consistent with the mode conclusion in the current round / the number of agents participating in the speech; the semantic distribution difference is the difference degree of the semantic distribution of the viewpoints in the current round and the semantic distribution of the viewpoints in the last round; the reliability is the reliability of the conclusion considered by the agent, which can be confidence or score. Figure 4 In the example, the average confidence is used.

[0058] S4: judging whether two items in the convergence condition are met, or whether the preset number of debate rounds is reached; Wherein, the convergence condition is that any two items in the index meet the index value greater than or equal to the corresponding index threshold, and the stable conclusion includes: the final aggregated conclusion of the task, the aggregated reliability and the support evidence summary; Figure 4 In the example, the preset number of debate rounds is 3.

[0059] When it is determined that the set number of debate rounds is not exceeded and the convergence condition is met, the debate is terminated, and a stabilized conclusion is output and used as high-order feature enhancement information. When it is determined that the set number of debate rounds is reached and the convergence condition is not met, the debate is terminated, and a stability judgment process is entered Figure 5 as shown in the stability judgment process.

[0060] When it is determined that the set number of debate rounds is not reached and the convergence condition is not met, the next round is continued, the updated opinions of the other agents based on the read opinions of the other agents are received, the updated opinions are used as the opinions of the current round of debate, and the process jumps to step S2.

[0061] Figure 5 To detect that the debate does not converge, a stability judgment process is performed, including the following steps: Outlier detection is triggered, a similarity matrix S is calculated, and a core opinion set C is selected; the convergence index value of the opinion is recalculated in the core opinion set C, and the stability is judged; when it is judged to be stable, a stable conclusion is output; when it is judged to be unstable, an "unstable" label is output.

[0062] The three steps of triggering outlier detection, calculating a similarity matrix, and selecting a core opinion set C are to determine the core opinion set using the opinions of the current round of debate.

[0063] Specifically, when the recalculated convergence index value meets the convergence condition, a stabilized conclusion is output and used as high-order feature enhancement information; when the recalculated convergence index value does not meet the convergence condition, an "unstable" label is output and used as high-order feature enhancement information, wherein the "unstable" label includes: primary candidate conclusions and secondary candidate conclusions, pending review labels, and reviewable suggestions.

[0064] The specific technical details involved in Figure 4 and Figure 5 are described in detail below.

[0065] 1) Initial opinion generation Multiple agents independently give a ternary output for the same structured input , wherein: represents a conclusion / label, which is the judgment of the agent on the target problem or the fault category; represents a confidence score, which is the reliability of the conclusion of the agent; Evidence / reasoning chain, can be input feature, contrast sample, or logical reasoning path. Here encode evidence chain as vector .

[0066] Initial opinion, i.e. opinion generated by first round of debate.

[0067] 2) Similarity measure and double threshold rule For any two agent outputs , , define their comprehensive similarity: where , are adjustable weight parameters, default 0.5 / 0.4 / 0.1; emphasize category consistency.

[0068] Category consistency: ; optionally, based on hierarchical similarity of fault ontology, e.g. "joint wear" and "temperature rise trigger" similarity 0.6; Evidence chain similarity: , take cosine similarity to measure; Confidence proximity: ; Set merging threshold and rejection threshold , and satisfy .

[0069] 3) Merge overly similar opinions (deduplication) Construct opinion merging graph , with opinion as vertex V, if , then connect edge. For each connected component / cluster , select representative opinion , preferred rule is "maximum centrality or highest confidence or weighted median within cluster". To achieve homogeneous opinion merging: .

[0070] 4) Remove significantly outlier opinions (denoising) Calculate representative score of each opinion, e.g. average similarity: ; If or , then consider as outlier, add to rejection set and reject.

[0071] Remaining valid opinion set is: The set removes redundancy and outliers while ensuring diversity (not over-converging).

[0072] 5) Group Optimal Coordination and Progressive Enhancement In the set Continue to discuss and aggregate in subsequent debate rounds (such as fusing the probability distribution of , weighting the evidence of , and calibrating ), update high-level features, and feed back to the next round of input, achieving multi-round progressive semantic enhancement.

[0073] 6) Adaptive Threshold and Representative Selection In different tasks or scenarios, data distribution may differ significantly. If a fixed threshold is used, it may not be able to balance the needs of "suppressing redundancy" and "preserving differences" at the same time. Therefore, the embodiment of the present application proposes an adaptive threshold setting method based on quantile.

[0074] Threshold adaptation: (e.g. ), dynamically adjusted with the scene, ensuring that both redundancy and differences can be suppressed in different tasks.

[0075] Representative selection strategy: select in the cluster , and compromise between "centrality" and "confidence" ( ).

[0076] In detail, the specific calculation process of adaptive threshold calculation is as follows (1) to (4): (1) Adaptive threshold calculation Meaning: take the value of the quantile from the similarity set

[0077] : Similarity set, ; : Quantile, value in ; for example: .

[0078] Merge threshold : When the similarity of two opinions is greater than

[0079] Outlier threshold : When the similarity of an opinion to all other opinions is less than . ​​

[0080] Example: Suppose there are 5 opinions, and the set of pairwise similarity scores in ascending order is: , length n = 10 a. Calculation formula (example with linear interpolation): quantile position index is calculated as: ; here, , i.e., the target position is the 8.1th element (index starts from 0).

[0081] b. Interpolation calculation: integer part (round down): , in the set , index = 8 corresponds to the value: 0.9; decimal part: ; the next number: index = 9 corresponds to the value: 0.95; linear interpolation: ; Similarly, .

[0082] Conclusion: .

[0083] (2) Graph construction and merging determination Now there are 5 opinions, 10 pairwise similarities. Denote the nodes: A, B, C, D, E.

[0084] For convenience of expression, construct an upper triangular similarity matrix consistent with the similarity set : as shown in the example similarity set Figure 6 .

[0085] Graph construction rules: if , connect between (an "merged cluster" is entered); if , node is recorded as an outlier (eliminate) Only satisfy the merging threshold .

[0086] The highest edge is but , not merged.

[0087] Outlier check: : , all , no outlier nodes.

[0088] Merging result:​ cluster } merge into 1 representative opinion .

[0089] the rest are single-node clusters: .

[0090] final set of valid opinions: , from 5 opinions to 4 representative opinions, removing redundancy while preserving diversity.

[0091] (3) Representative selection (how to select "representative opinions" within the cluster) For a cluster , a "centrality + confidence" compromise can be used to select a representative (example formula): .

[0092] "centrality" : average similarity to other members in the cluster (for a 2-person cluster, it is the similarity between the two, 0.95); "confidence" : the confidence of this opinion (from the output of each Agent).

[0093] : weighting coefficient (e.g. 0.7) If is known (e.g. from the previous round ERNIE = 0.78, Hunyuan = 0.62), A is usually chosen as the representative ; if the confidence is close, you can directly take the medoid (both are equivalent, you can choose one).

[0094] (4) Subsequent processes and effects When entering the next round of debate, the 4 "redundancy-removed" representative opinions participate, reducing repetition while preserving the differences between C, D, and E.

[0095] If new opinions are generated in the next round, the "adaptive threshold + merge / remove" step can be repeated to dynamically balance "individual competition vs. group optimization".

[0096] 5) Intelligent termination mechanism Introduce distribution difference (Jensen-Shannon divergence) and opinion consistency rate as judgment criteria; when multiple rounds of debate converge in the feature enhancement space or the marginal gain is insufficient, automatically terminate.

[0097] The key indicators for judging convergence are designed as follows: opinion consistency rate A Definition: Number of Agents consistent with the majority conclusion in this round / Number of Agents participating in the speech denoted as:

[0098] Threshold setting: i.e. can be considered to be converging. This is an empirical setting, "super majority" decision threshold.

[0099] Semantic distribution difference D (JS divergence) Definition: Difference between the semantic distribution of this round and the last round, measuring the change in the distribution of conclusions between rounds.

[0100] denoted as: The smaller it is, the closer the two rounds of opinion distribution are.

[0101] Threshold setting: i.e. indicates semantic stability. In natural language processing, JS divergence is considered to be high similarity.

[0102] ③ Average confidence (Optionally) Definition: Mean of all Agent's confidence in their own opinion; denoted as: where, is the Agent's confidence in their own opinion.

[0103] Threshold setting: i.e. consider the conclusion to be reliable.

[0104] Reference machine learning ensemble model: prediction probability Generally considered as strong signal.

[0105] Optionally, in some low-risk scenarios, only A and D two indicators can be used; When two of the above key indicators for judging convergence are met at the same time, it is considered that the "debate" has stabilized.

[0106] If it does not converge, the set can be executed again to filter the similarity, identify the "core opinion set", eliminate outliers, and aggregate the relatively stable alternative conclusions and confidence.

[0107] Core opinion set selection: Calculate the similarity matrix (diagonal line is 0); Calculate the total similarity: ; Take The largest K is the core set of viewpoints C, where, N is the number of Agents participating in the discussion, maintaining the principle of strict majority; adaptive when N changes, avoiding overfitting.

[0108] Eliminate outliers and re-aggregate: Outlier set Eliminate; Recalculate the viewpoint consistency rate, distribution stability, and aggregation confidence within the core set of viewpoints C; If the threshold is met, output the relatively stable and interpretable conclusion and the residual hypothesis set.

[0109] 2.4 Output (format example) Final conclusion: Defect location result: Physical layer anomaly, manifested as loose fiber joint / severe link attenuation; temperature rise may be a trigger factor.

[0110] Aggregation confidence: Aggregation confidence: 0.86 (obtained by weighted normalization of each Agent output); Residual hypothesis: TX drift or relay parameter mutation of the opposite end light source, confidence about 0.14.

[0111] Key evidence chain: Observation item: LOS alarm, received power -27.1 dBm; Historical association: There is a record of loose joint on the same port; Abnormal exclusion: No hardware alarm on this end.

[0112] Convergence and termination criteria: Distribution difference: ; Viewpoint consistency rate: ; Average confidence: ; Number of rounds <3 (not reaching the upper limit). Finally, it is concluded that the comprehensive convergence condition is met, and the system terminates the debate in advance.

[0113] Reviewable information: Silent reason: Some Agents did not speak because of lack of timing / opposite end data; Suggested additional collection items: opposite end Tx power, OTDR results, power / bit error rate timing data.

[0114] 2.5 Example 1 (convergent) Multi-agent debate process (methodology simulation, N=5 as an example) (convergent) Round 1: Each speaks: such as Figure 7-1 Multi-Agent Debate Simulation Example 1 - Round 1.

[0115] Convergence degree calculation: Round 1: Build distribution P(1) (aligned as two categories: joint / opposite end) 1. Total number of speeches N1=3 (ERNIE, Hunyuan, DeepSeek); 2. Merged weights (weighted by confidence): S1 = 1.40 + 0.55 = 1.95; P(1) =

[0116] Indicator calculation: Number of speeches N1 = 3. Mode support number (Did not reach 0.75) Average confidence (Did not reach 0.7) No previous round comparable (t = 1), D is not calculated.

[0117] Conclusion: , needs to continue.

[0118] Round 2: Read other Agent opinions, as shown in Figure: 7-2 Multi-Agent Debate Simulation Example 1 - Round 2; Convergence degree calculation: Round 2: Build distribution P(2) (after alignment, two categories: ) 1. Total number of speeches N2 = 4 (ERNIE, Qianwen, Hunyuan, DeepSeek) 2. Merged weights (weighted by confidence): ; S2 = 2.18 + 0.35 = 2.53; P(2) = ; Indicator calculation: Number of speeches N2 = 4. Mode support number (Reached 0.75); Average confidence (Did not reach 0.7); Semantic distribution difference . After approximation, take , meet semantic stability); Conclusion: Round 2 meets A and D threshold, system determines convergence and terminates debate early.

[0119] Final output: as shown in Figure 7-3 Multi-Agent Debate Simulation Example 1 - Output.

[0120] 2.6 Example 2 (not convergent) Multi-agent debate process (methodology simulation, N = 5 as an example) (not convergent) Round 1: Each speech as Figure 8-1Multi-Agent Debate Simulation Example 2 - Round 1 Convergence calculation: Round 1: Construct distribution P(1) (weighted and normalized by the confidence level of the speaking agent) 1. Total number of comments N1=4 (ERNIE, Hunyuan, DeepSeek, GLM); 2. Weighted sum: ; P(1) = ; The opposite end = 0.58 / 2.55 0.227; }; Indicator Calculation: Number of participants (N1) = 4. Number of supporters in the mode = 1 (only 1 Agent per category). (far below 0.75); Average confidence level

[0121] Since there is no comparable round (t=1), D is not calculated.

[0122] Conclusion: Round 1 showed significant divergence and failed to converge, requiring further discussion.

[0123] Round 2: Read the opinions of other agents, such as Figure 8-2 Multi-Agent Debate Simulation Example 2 - Round 2; Convergence calculation: Round 2: Constructing distribution P(2) Weighted sum: ;

[0124] Indicator Calculation: Number of participants in the discussion N² = 5. Number of supporters in the mode = 1 (only 1 Agent per category). (far below 0.75); Average confidence level ; Semantic distribution differences After approximate calculation, take (>0.10 indicates that the significant change has not stabilized); Conclusion: Round 2 still does not meet the requirements. (Semantics are not stable), proceed to the third round.

[0125] Round 3: Update your viewpoint after reading other agents again, such as... Figure 8-3 Multi-Agent Debate Simulation Example 2 - Round 3; Convergence calculation: Round 3: Constructing distribution P(3) Weighted sum: S3 = 0.70 + 0.58 + 0.60 + 0.62 + 0.57 = 3.07 (coincidentally the same value as the previous round)

[0126] Indicator Calculation: Number of participants (N3) = 5. Number of supporters in the mode = 1 (only 1 Agent per category). (far below 0.75); Average confidence level ; Semantic distribution differences After approximate calculation, take (Not reaching the stable threshold); Conclusion: Round 3 still does not satisfy A. 0.75, and (Semantics are not stable), the maximum number of rounds T=3 has been reached.

[0127] Since it does not converge, the following stability calculation is required.

[0128] Stability calculation 1. Sorting out To facilitate similarity measurement, key evidence items for each agent's viewpoint in the third round are extracted and binarized into a vector (example feature dimension): Feature dimensions (example): ; Corresponding vector (based on the 3rd round of speaking): ERNIE (connector): ;Hunyuan (temperature rise): GLM (Relay): .

[0129] The confidence scores for each agent (round 3) are as follows: ERNIE 0.70; Hunyuan 0.58; DeepSeek 0.60; GLM 0.62; and QianWen 0.57.

[0130] 2. Pairwise similarity calculation Following the calculation method in Section 2.5, in pairs (Only the upper triangle is listed, retaining 4 digits) ; ; Clearly highest (both share evidence on "topology / relay / two-hop") .

[0131] Note: From evidence and confidence perspective, DeepSeek and GLM are clearly closer in opinion than other two-by-two combinations (as they share "two-hop / relay" related evidence).

[0132] 3. Adaptive threshold (determined by similarity distribution quantile).

[0133] Sort all two-by-two similarity sets above and use the linear interpolation method as before: Take 90% quantile as merge threshold The result is about 0.264 (in this example calculation template, it is ); Take 10% quantile as cull threshold The result is about 0.0897, i.e. ; .

[0134] 4. Merge (deduplication) and cull (denoising) decisions DeepSeek - GLM: Form a merged cluster ; other pairs are not merged.

[0135] No opinion meets culling condition (each opinion is similar to at least one other opinion ), so no culling.

[0136] Merge result: merge DeepSeek and GLM into one representative opinion (choose representative within the cluster: GLM confidence is slightly higher, or choose representative according to centrality / weighting criteria within the cluster).

[0137] 5. Recalculate population distribution after culling / merging Merged representative opinion set (example) after merging: , Hunyuan (temperature rise, 0.58), GLM_rep (relay, 0.62), ; Weighted normalized by confidence, get aggregated probability (only major categories are listed): .

[0138] Recalculating criterion: Consensus rate A = number of people supporting the majority / current number of speakers = the maximum number of supporters is 1, ; Average confidence ; Difference between the distribution of the last round (P(3) after mapping calculation (Significantly less than the threshold ), indicating that the distribution direction has changed little.

[0139] Judgment: Although the distribution difference is small (tends to be stable), the consensus rate A and the average confidence Do not reach the set threshold (for example ), so it cannot be declared completely converged. At the same time, the system has reached the maximum number of rounds 3.

[0140] 6. Touching the top process: Core-Set extraction (guaranteeing the "strict majority" principle) When the upper limit of the number of rounds is reached but the full convergence threshold is not met, the similarity Core-Set extraction is used to output a more stable conclusion. The steps are as follows: 1) Calculate the average similarity of each retained view on the current set (as a representative measure): .

[0141] 2) Set K= . The current N = 4 K = 2. Select the top K views with the highest average similarity as the Core-Set C: that is {Hunyuan, ERNIE} (warm-up Joint).

[0142] 3) Re-aggregate on the Core-Set C: The sum of the weights = .

[0143] On the Core-Set, the joint tends to be dominant but not an absolute majority (there is still no single 75% support within the Core-Set).

[0144] 7. Final conclusion (based on the top-touching + Core-Set strategy) Robust consensus (suboptimal output): Under the top-touching condition, the Core-Set strategy is adopted, and the system gives the primary candidate conclusion as "physical layer - fiber joint loosening / link decay tendency" (dominant on the Core-Set), but at the same time, "warm-up trigger" is retained as an important candidate.

[0145] Certainty / Warning: Since there is no absolute majority within the core viewpoint set and the average confidence level does not reach the most stringent threshold, the system should mark the conclusion as "suboptimal / to be verified" and list the residual hypotheses (peer / relay, module aging).

[0146] Verifiable Recommendation (Required): It is recommended to collect additional data on the peer's transmit power, OTDR path / timing, power / BER timing, and more environmental / room temperature history to eliminate residual assumptions and reassess the situation in subsequent rounds or after manual review.

[0147] Final output: such as Figure 8-4 The output of Example 2 of the multi-agent debate simulation is shown.

[0148] Layer 3: Multimodal Text Encoding Layer (Agent) 3.1 Overview: The conclusions, evidence chains, and distribution data generated by multi-agent debates are characterized to facilitate aggregation, filtering, and stability assessment by the subsequent decision-making layer. This multimodal text encoding layer agent can effectively extract and compress key information in both convergent and non-convergent scenarios, ensuring that the overall system can reach a stable conclusion after multiple rounds of game interaction. Its core tasks are to map the semi-structured output into a unified vector representation; generate classification labels (such as "high convergence" and "low convergence") based on convergence indices; and provide path selection references for the subsequent decision-making layer agent.

[0149] 3.2 Multi-level feature fusion mechanism: Multi-level coding framework: Low-order feature layer: directly retains numerical indicators (power, temperature, bit error rate, etc.); Higher-order semantic layer: semantic embedding of alarm information, feedback, and historical defects (text encoders such as BERT). It not only preserves the original features, but also obtains latent patterns through automatic interaction, thereby improving the discriminative power of downstream SVM / XGBoost.

[0150] 3.3 Context completion enhanced coding: A context completion strategy is introduced, which combines the following during encoding: historical defect records (time series characterization); feature distribution of similar ports / devices (neighbor completion); environmental context (external features such as data center temperature and power status); to avoid individual samples being too isolated, making the encoding closer to the actual operation and maintenance context.

[0151] 3.4 Example The output vector contains three types of features: semantic features, numerical features, and structured flags. Semantic features: Final conclusion embedding (e.g., "loose fiber optic connector / link attenuation") Vector) ; Evidence Chain embedding (e.g. "LOS+Rx-27.1 dBm+Historical Same Port Anomaly" ).

[0152] Numerical Features: Aggregate Confidence (0.86) ; Residual Hypothesis Proportion (0.14) ; Convergence Criteria: Opinion Consensus Rate A = 0.75; Semantic Distribution Difference D = 0.08; Average Confidence .

[0153] Structured Flags: Missing Data Markers (e.g. Missing Timing = 1, Missing Counterparty = 1) ; Convergence Labels: If meets Label = "High-Converged" ; Else Label = "Low-Converged".

[0154] Layer 4: Decision Layer Agent (Integrates Classification Decision Experts) Summary: The Decision Layer Agent receives the vector representation from the Feature Encoder Agent and the Convergence Label, selects the appropriate discriminative path (SVM or XGBoost), and generates the final decision result, including: Fault Category (e.g. Physical Layer / Counterparty Tx / Relay Parameter) ; Confidence (Model Output Probability or Decision Function Value) ; Explainability Evidence (e.g. Feature Contribution, Evidence Chain).

[0155] The Decision Layer supports two implementation paths: one is a simple SVM-based discriminator, and the other is a complex XGBoost-based discriminator. The choice of use can be selected according to the application scenario.

[0156] 4.2 Path Selection Strategy: 1. Path A: SVM Discriminator Applicable Scenario: High Convergence (High-Converged), i.e. converged, most Agents have formed stable consensus. Model Features: Linear or Kernel SVM. Output Form: Binary / Multi-Classification Label + Confidence. Advantages: Lightweight, fast, avoids overfitting.

[0157] 2. Path B: XGBoost Discriminator Applicable Scenario: Low Convergence (Low-Converged), i.e. not converged, there are multiple hypotheses or disagreements. Model Features: Gradient Boosting Tree (XGBoost). Output Form: Candidate Category Distribution + Feature Importance / SHAP Explainability. Advantages: Strong robustness, can form stable conclusions under complex inputs. Disadvantages: Requires relatively more training samples.

[0158] Example: Input (Output of the Debate) Final conclusion: Loose fiber optic connector / link attenuation; Aggregate confidence score: 0.86; Convergence metric: It satisfies convergence.

[0159] Feature encoder Agent: Constructed vector F=[embedding(conclusion),embedding(evidence chain),0.86,0.14,0.08,0.75,0.73,missing time sequence=1,missing counterpart=1]; Label: High-Converged.

[0160] Decision-making agent: Check tags High-Converged Following the SVM path; SVM output: Fault category: Physical layer; Confidence level : 0.91; Explainability: Rx-27.1dBm and historical joint loosening are key factors.

[0161] Another scenario (if convergence fails) Label=Low-Converged): Features enter the XGBoost path; Output: Candidate conclusion set: {Loose fiber optic connector (0.42); Tx drift at the other end (0.33); Sudden change in relay parameters (0.25)}; Interpretability: Ambiguity caused by missing OTDR; Tx power at the other end needs to be sampled.

[0162] In the above process, the feature encoder Agent completes the standardization of features and convergence labels, while the decision layer Agent automatically selects the discrimination path (SVM / XGBoost) based on the labels, which is both lightweight and fast, and can robustly handle complex situations.

[0163] Based on the same inventive concept, this disclosure also provides a high-order feature enhancement agent based on multi-agent debate, corresponding to the above-mentioned high-order feature enhancement method, including: The receiving module is used to receive the various viewpoints generated by each agent in the current round of debate based on the input of the same task. Each viewpoint includes: task conclusion, reliability, and evidence. The large model driving each agent is different. The determination module is used to determine the set of valid viewpoints based on the various viewpoints in this round of debate. Valid viewpoints are those that are representative and distinct from each other. The computing module is configured to calculate a convergence index value of the viewpoint based on the effective viewpoint set, wherein the convergence index is: a viewpoint consistency rate and an average reliability; or: a viewpoint consistency rate, a semantic distribution difference, and an average reliability; wherein the viewpoint consistency rate is the number of agents consistent with the mode conclusion in the current round / the number of agents participating in the speech; and the semantic distribution difference is a difference degree of the semantic distribution of the viewpoint in the current round and the semantic distribution of the viewpoint in the last round. The output module is configured to output a stabilized conclusion as high-order feature enhancement information when it is determined that the set number of debate rounds is not exceeded and the convergence condition is met, and output the updated viewpoint of each agent as the viewpoint of the current round of debate, and trigger the determination module when it is determined that the set number of debate rounds is not reached and the convergence condition is not met, wherein the convergence condition is that any two of the indexes satisfy that the index value is greater than or equal to the corresponding index threshold, and the stabilized conclusion includes: a final aggregated conclusion of the task, an aggregated reliability, and a support evidence summary.

[0164] In addition, the disclosure also provides a defect positioning system based on multi-agent debate, comprising: an input layer Agent, the above-mentioned high-order feature enhancement Agent, a multi-modal feature encoding Agent, and a decision layer Agent; wherein the input layer Agent is composed of each analysis Agent, and each analysis Agent analyzes different types of input; Each analysis Agent is configured to analyze original data of a type matching the analysis type thereof based on a Prompt guide and in combination with a long context data attribute completion strategy related to a task, to obtain structured input including an original layer, a preprocessing layer, and a label layer, wherein the Prompt is used to semantically constrain multi-modal input, and the semantic constraint includes: redundancy, accuracy, and closed loop; The high-order feature enhancement Agent is configured to process the above-mentioned structured input according to a task based on a multi-agent debate mechanism, to obtain high-order feature enhancement information; wherein the task is a defect positioning task; The multi-modal feature encoding Agent is configured to introduce a context completion strategy to encode the high-order feature enhancement information; The decision layer Agent is configured to select a discriminator based on a convergence label in the feature encoding; wherein when the convergence label is not convergent, an XGBoost discriminator is selected; and when the convergence label is convergent, an SVM discriminator is selected; and a final decision result is generated by using the selected discriminator.

[0165] Based on the same inventive concept of the above disclosure, the application also provides an electronic device. The electronic device of the embodiments of the application comprises at least one processor and at least one memory electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method as described above.

[0166] It should be noted that the electrical connection between the above-mentioned units does not necessarily mean the connection between the lines, and the indirect connection mode can also be applied to the embodiments of the application as long as the purpose of the application is achieved.

[0167] Based on the same inventive concept, the application further provides a computer storage medium, wherein the computer storage medium stores a computer program, and the computer program is executed by a processor to implement the method described above.

[0168] Although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A feature enhancement method based on multi-agent debate, characterized in that, include: The system receives input from various agents on the same task and generates various viewpoints for this round of debate. Each viewpoint includes: task conclusion, reliability, and evidence. The large models driving each agent are different. The set of valid viewpoints is determined by using the various viewpoints in this round of debate. Valid viewpoints are those that are representative and distinct from each other. The convergence index of opinions is calculated based on the effective opinion set. The convergence index is: opinion consistency rate and average reliability; or: opinion consistency rate, semantic distribution difference and average reliability. The opinion consistency rate is the number of agents that agree with the mode conclusion in this round / the number of agents that participated in the speech. The semantic distribution difference is the degree of difference between the semantic distribution of opinions in this round and the semantic distribution of opinions in the previous round. When it is determined that the number of debate rounds has not exceeded the set limit and the convergence condition is met, the stabilization conclusion is output and used as higher-order feature enhancement information. The convergence condition is that any two items in the convergence index satisfy that the index value is greater than or equal to the corresponding index threshold. The stabilization conclusion includes: the final aggregation conclusion of the task, the aggregation reliability, and the summary of supporting evidence. If the set number of debate rounds has not been reached and the convergence condition is not met, the system receives the views obtained by each agent based on the views of other agents, updates its own views, and uses the updated views as the views for this round of debate. The system then proceeds to the step of determining the set of valid views using the views of this round of debate.

2. The method according to claim 1, characterized in that, When the set number of debate rounds is reached but the convergence condition is not met, the core set of viewpoints is determined using the viewpoints of this round of debate. The convergence index value of the viewpoints is recalculated based on the core viewpoint set; When the convergence index value is recalculated and satisfies the convergence condition, the stabilization conclusion is output and used as higher-order feature enhancement information. When the convergence index value calculated again does not meet the convergence condition, an "unstable" label is output and used as higher-order feature enhancement information. The "unstable" label includes: primary candidate conclusion and secondary candidate conclusion, verification annotation and verification suggestion.

3. The method according to claim 1 or 2, characterized in that, The valid set of viewpoints identified using the various perspectives from this round of debate includes: Calculate the comprehensive similarity between the viewpoints of any two agents in this round of debate to obtain a comprehensive similarity dataset, where the comprehensive similarity is the weighted sum of the consistency of task conclusions, the closeness of reliability, and the similarity of evidence chains; Based on the comprehensive similarity dataset, the opinions are deredundanted and denoised to obtain an effective set of opinions.

4. The method according to claim 3, characterized in that, Based on the comprehensive similarity dataset, the opinions are deredundanted and denoised to obtain a valid set of opinions, including: Constructing a viewpoint merging graph: Using viewpoints as vertices, establish an edge between two viewpoints when the overall similarity between them is greater than or equal to the merging threshold; Each connected component is treated as a cluster, and representative viewpoints are selected from them to achieve redundancy removal. For each viewpoint obtained after deduplication, determine the representative comprehensive similarity score for each viewpoint; When the representative comprehensive similarity score is less than the elimination threshold, the corresponding viewpoint is eliminated to achieve noise reduction. The elimination threshold is greater than or equal to 0 and less than the merging threshold, and the merging threshold is less than or equal to 1.

5. The method according to claim 2, characterized in that, Use the various viewpoints from this round of debate to identify a core set of viewpoints, including: Calculate the comprehensive similarity between the viewpoints of any two agents in this round of debate to obtain a comprehensive similarity dataset; For each agent's viewpoint in this round of debate, the overall similarity between its viewpoint and the viewpoints of other agents is added together to obtain the total similarity of that agent's viewpoint in this round of debate; Construct a total similarity dataset using the total similarity of each agent; The top K views with the highest total similarity scores extracted from the total similarity dataset are taken as the core view set.

6. A defect localization method based on multi-agent debate, characterized in that, include: Each parsing agent in the input layer, guided by a Prompt and combined with a task-related long context data attribute completion strategy, parses the raw data whose type matches its parsing type, resulting in structured input including the raw layer, preprocessing layer, and label layer. The input types parsed by each parsing agent are different. The Prompt is used to perform semantic constraints on the multimodal input, including redundancy, accuracy, and loop closure. The higher-order feature enhancement agent processes the structured input using the method described in any one of claims 1-5 to obtain higher-order feature enhancement information, wherein the task is a defect localization task; The multimodal feature encoding agent introduces a context completion strategy to encode high-order feature enhancement information; The decision-making agent selects a discriminator based on the convergence label in the feature encoding; the selected discriminator is used to generate the final decision result. When the convergence label is non-convergence, the XGBoos discriminator is selected; when the convergence label is convergence, the SVM discriminator is selected.

7. A high-order feature enhancement agent based on multi-agent debate, characterized in that, include: The receiving module is used to receive the various viewpoints generated by each agent in the current round of debate based on the input of the same task. Each viewpoint includes: task conclusion, reliability, and evidence. The large model driving each agent is different. The determination module is used to determine the set of valid viewpoints based on the various viewpoints in this round of debate. Valid viewpoints are those that are representative and distinct from each other. The calculation module is used to calculate the convergence index value of the viewpoints based on the effective viewpoint set. The convergence index is: viewpoint consistency rate and average reliability; or: viewpoint consistency rate, semantic distribution difference and average reliability. The viewpoint consistency rate is the number of agents that agree with the mode conclusion in this round / the number of agents that participated in the speech; the semantic distribution difference is the degree of difference between the semantic distribution of the viewpoints in this round and the semantic distribution of the viewpoints in the previous round. The output module is used to output a stabilization conclusion and use it as higher-order feature enhancement information when the number of debate rounds has not exceeded the set number and the convergence condition is met. When the number of debate rounds has not been reached and the convergence condition is not met, the module receives the views of each agent based on the views of other agents and updates its own views, and uses the updated views as the views of this round of debate, triggering the determination module. The convergence condition is that any two of the indicators are greater than or equal to the corresponding indicator threshold. The stabilization conclusion includes: the final aggregation conclusion of the task, the aggregation reliability, and the summary of supporting evidence.

8. A defect localization system based on multi-agent debate, characterized in that, include: The input layer agent, the high-order feature enhancement agent based on multi-agent debate as described in claim 7, the multimodal feature encoding agent, and the decision layer agent; wherein, the input layer agent is composed of various parsing agents, and each parsing agent parses different types of inputs; Each parsing agent is used to parse the raw data whose type matches its parsing type, based on the Prompt and combined with the task-related long context data attribute completion strategy, to obtain structured input including the raw layer, preprocessing layer and label layer. The Prompt is used to perform semantic constraints on the multimodal input, and the semantic constraints include: redundancy, accuracy and closure. A higher-order feature enhancement agent is used to process the above-mentioned structured input based on a multi-agent debate mechanism to obtain higher-order feature enhancement information according to the task; wherein, the task is a defect localization task. Multimodal feature encoding agent is used to introduce context completion strategies and encode high-order feature enhancement information; The decision layer (Agent) is used to select a discriminator based on the convergence label in the feature encoding; when the convergence label is non-convergence, the XGBoost discriminator is selected; when the convergence label is convergence, the SVM discriminator is selected; and the final decision result is generated using the selected discriminator.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-6.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images