A Smart Review Method for Construction Schemes Based on Adaptive Multi-Agent Collaboration

Abstract

This invention discloses an intelligent review method for construction schemes based on adaptive multi-agent collaboration, belonging to the field of construction scheme review technology. It solves the technical problems of inability to deeply identify key technical points, lack of collaboration among different professional perspectives, absence of a supervisory layer and reflection mechanism, lack of quality monitoring during the review process, and the potential direct output of errors and logical contradictions. Through dynamic agent generation technology, the system can adaptively form agent review teams based on different scheme characteristics. Employing a multi-agent collaboration and shared blackboard mechanism, it simulates a multi-expert review scenario. By introducing a consensus decision-making algorithm based on confidence fusion, the system can autonomously negotiate and adjudicate conflicts of opinion between different agents or regulatory clauses, outputting a unified conclusion. Through the meta-agent layer, the system achieves real-time monitoring and reflection on its own review process, automatically verifying reasoning logic, identifying low-confidence conclusions, and triggering secondary reviews.

Description

Technical Field

[0001] This invention belongs to the field of construction plan review, specifically a construction plan intelligent review method based on adaptive multi-agent collaboration. Background Technology

[0002] The intelligent construction plan review method based on adaptive multi-agent collaboration is an intelligent construction plan review technology that integrates adaptive multi-agent technology, an engineering knowledge base, and dynamic review logic. Its core is achieved through "dynamically generating professional intelligent agents + multi-agent collaborative cooperation + full-process quality control + dynamic closed-loop optimization." Multi-agent collaboration refers to multiple agents working together to complete complex tasks through communication and coordination mechanisms. Each agent possesses autonomy, responsiveness, and initiative, can adjust its behavior according to environmental changes, and avoids conflicts and improves the overall system performance through collaboration.

[0003] Traditional intelligent review schemes for construction plans mostly adopt fixed pipelines, which cannot dynamically adjust the review strategy according to document characteristics. They can only perform surface text review and cannot deeply identify key technical points. They use a single model or unidirectional flow processing, which cannot coordinate different professional perspectives. Without a supervision layer and reflection mechanism, there is no quality monitoring in the review process, and errors and logical contradictions may be directly output. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes an intelligent review method for construction schemes based on adaptive multi-agent collaboration, which is used to solve the technical problems of being unable to deeply identify key technical points, being unable to collaborate from different professional perspectives, lacking a supervision layer and reflection mechanism, having no quality monitoring in the review process, and having errors and logical contradictions that may be directly output.

[0005] To address the above problems, the first aspect of this invention provides an intelligent review method for construction plans based on adaptive multi-agent collaboration, comprising the following steps:

[0006] S1: Extract features from the input construction plan document to obtain a document feature set;

[0007] S2: Based on the document feature set, dynamically generate an agent pool containing at least two different professional types of review agents, wherein the review agents are constructed based on a large language model;

[0008] S3: Each auditing agent in the intelligent agent pool performs parallel auditing and asynchronous collaboration based on a shared blackboard, and writes its audit findings into the shared blackboard.

[0009] S4: Based on the audit findings stored in the shared blackboard, detect audit conflicts between different audit agents; if an audit conflict is detected, initiate a consensus decision process, and conduct weighted voting based on the weights of each audit agent involved in the conflict and the confidence level of their audit findings to form a consensus conclusion; otherwise, directly jump to S5.

[0010] S5: Supervise and control the audit process of the agent pool through meta-agents, and generate a final audit report.

[0011] Optionally, in one example of the above aspects, step S1 includes the following steps:

[0012] Text extraction is performed on construction plan documents in PDF / Word format. AdvancedTextExtractor is used to extract key parameter information from the synchronously uploaded CAD drawings, including parameter tables, text information, pole spacing and wall support positions. At the same time, chapter structure and table of contents information are extracted.

[0013] Extract project type, professional field, key terms, complexity features, and potential features using keyword matching algorithms;

[0014] The extracted features are converted into vector representations and combined into a set to obtain the document feature set.

[0015] Optionally, in one example of the above aspects, step S2 includes the following steps:

[0016] Using the AgentSpawning algorithm, based on the document feature set and the agent activation threshold mapping table, the matching degree and priority between features and each agent type are calculated to generate an agent pool containing at least two of the following: structural audit agents, security audit agents, economic audit agents, regulatory compliance agents, and geometric audit agents. The formula for calculating the matching degree between features and each agent type is as follows: ;

[0017] in, Let be the matching degree between feature fi and agent Aj, ki be the keywords contained in feature fi, Kj be the set of professional keywords of agent Aj, similarity() be the semantic similarity function, λ be the parameter association correction coefficient, and Prelevance(fi,Aj) be the association degree between the parameters in feature fi and the professional domain of agent Aj. Prelevance(fi,Aj) is obtained by calculating the cosine similarity between the parameter keywords in feature fi and the professional domain keywords of agent Aj.

[0018] The formula for calculating the priority of features and agent types is as follows: Where α, β, γ and δ are weight coefficients, weight(fi) is the feature weight of feature fi, Mats(fi,Aj) is the matching degree between feature fi and agent Aj, urgency(fi) is the urgency of feature fi of the current scheme to be reviewed, and Pimpact(fi,Aj) is the correlation influence weight of feature fi parameters corresponding to agent Aj.

[0019] Alternatively, in one example of the above aspects, the feature weights weight(fi) of feature fi are set by the following method:

[0020] Document features are divided into core features and secondary features. A judgment matrix is ​​constructed using the AHP (Analytic Hierarchy Process). Several engineering review experts are invited to score the importance of the features. The expert scores are normalized to the interval between 0 and 1 to obtain the final feature weight (fi).

[0021] The urgency (fi) is set by the following method: based on the ratio of the remaining time to the deadline of the current pending project characteristic fi to the maximum remaining time to the deadline of all pending projects characteristics;

[0022] The association influence weight Pimpact(fi,Aj) of the feature fi parameter corresponding to agent Aj is set by the following method: Based on the standardized parameter database of agent Aj, the association relationship between parameters is determined by experts, the parameter of feature fi is taken as a data node, the data nodes corresponding to the parameters with association relationship are set with association edges, and the degree centrality of each data node is calculated as the association influence weight Pimpact(fi,Aj) of the feature fi parameter corresponding to the data node.

[0023] Optionally, in one example of the above aspects, step S4, detecting review conflicts between different review agents, includes:

[0024] The audit model iterates through all audit findings records using a double loop, compares pairs of records to determine whether they are targeting the same audit object and whether the viewpoints are contradictory, identifies at least one type of conflict among conclusion conflict, numerical conflict, normative reference conflict, and priority conflict, and creates a conflict record.

[0025] Step S4, the consensus decision-making process, includes:

[0026] Collect opinions from participating agents and calculate the weights of agent Aj in the feature fi corresponding to the review conflict. Where H(Aj) is the historical accuracy of the agent, Mats(fi,Aj) is the matching degree between feature fi and agent Aj type, and a1 and a2 are weight coefficients.

[0027] Weighted voting is performed based on agent weights and confidence levels, and opinion scores are calculated using a formula. Where C(Aj) is the confidence level of agent Aj; I(Aj,supportsopinioni) is an indicator function, which takes the value 1 when agent Aj supports opinion opinioni, and 0 otherwise;

[0028] The consensus conclusion is determined based on the comparison between the score difference and the consensus threshold. The confidence level of the final consensus is calculated using the formula... calculate; Σ(W(Aj)) represents the weighted confidence sum of agent Aj, and Σ(W(Aj)) represents the sum of the weights of all agents.

[0029] The review strategy is dynamically adjusted based on real-time monitoring data, and the review results are fed back to step S2.

[0030] Optionally, in one example of the above aspects, a linkage mechanism is established between key parameters and between monitoring data and the audit model. This involves quantifying the correlation between parameters using a new algorithmic formula, dynamically adjusting the audit strategy based on real-time monitoring data, and feeding the audit results back to step S2, including the following steps:

[0031] Dynamic parameters during the construction process are obtained through a real-time monitoring data interface, and the dynamic deviation coefficient is calculated. Where D(kn) is the dynamic deviation coefficient of the kn-th parameter, Preal(kn) is the real-time monitoring value of the kn-th parameter, Pmodel(kn) is the audit model preset value of the kn-th parameter, and Pstd(kn) is the parameter specification standard value.

[0032] Calculate the audit granularity adjustment coefficient based on the parameter dynamic deviation coefficient and the correlation effect results. Where Gadj is the granularity adjustment coefficient for review, and Skn is the importance weight of the kn-th parameter;

[0033] If the dynamic deviation coefficient D(kn) of the kn-th parameter is greater than the threshold, or the audit granularity adjustment coefficient Gadj is greater than the threshold, the agent priority update command is triggered, and the agent priority is updated by using the feature and the priority calculation formula of each agent type in step S2.

[0034] Optionally, in one example of the above aspects, step S5 includes:

[0035] Traverse the sequence of reasoning steps, check the logical coherence between adjacent steps, mark the logical break if there is a logical jump, and trigger a request to supplement details, requesting the supplementation of intermediate steps.

[0036] By calling upon standardized parameter databases, specification databases, and publicly available engineering data resources, a factual consistency score for the generated content is calculated. Among them, Tscore is the factual consistency score. To generate the th-th value in the content, This is the baseline value corresponding to the th-th value in the generated content from the authoritative data source. q represents the importance weight of the th data point, and q represents the total number of data points to be verified. When Tscore ≥ 0.6, the data is deemed to meet the requirement of factual consistency; otherwise, it is deemed not to meet the requirement of factual consistency, and a fact verification request is triggered.

[0037] The authenticity of the source is verified by detecting the explicit data source associated with each conclusion generated by the agent through meta-agent detection.

[0038] Hallucination Identification and Correction: When a hallucination is detected that does not meet the requirements of factual consistency, has no clear data source, or has only partially clear data sources, it is determined to be hallucination information, and the hallucination risk level is calculated. Where Hrisk is the hallucination risk level, θ is the weighting coefficient, and Ssource is the source validity score. The source validity score Ssource is set as follows: 1 when there is no explicit data source, 0.5 when some have explicit data sources, and 0 when all have explicit data sources.

[0039] A secondary review is triggered when the Hrisk exceeds a threshold or when multiple agents have similar scores and cannot reach a consensus.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] This invention utilizes dynamic intelligent agent generation technology, enabling the system to adaptively assemble intelligent agent review teams based on the characteristics of different solutions. This achieves a leap from simple normative text review to in-depth technical rationality examination, automatically identifying key risk points such as the determination of critical engineering projects and the completeness of calculation reports. It innovatively employs a multi-agent collaborative and shared blackboard mechanism, simulating a multi-expert review scenario. This enables cross-disciplinary and cross-chapter information sharing and consistency verification, effectively identifying potential contradictions between design drawings and calculation reports, as well as requirements from different disciplines, ensuring the overall rationality of the solution.

[0042] This invention introduces a consensus decision-making algorithm based on confidence fusion. The system can autonomously negotiate and adjudicate conflicts of opinion between different agents or regulatory provisions, outputting authoritative and unified conclusions with clear reasoning basis, greatly reducing reliance on manual adjudication. Through the meta-agent layer, the system achieves real-time monitoring and reflection on its own review process, automatically verifying reasoning logic, identifying low-confidence conclusions, and triggering secondary reviews, forming a self-correcting quality loop, thereby ensuring the high accuracy and reliability of the final output report.

[0043] This invention dynamically allocates review resources and adjusts the review granularity based on the complexity of the construction plan, avoiding the problems of over-reviewing simple plans and under-reviewing complex plans. It achieves intelligent scheduling of review resources and significantly improves parallel processing efficiency. The entire review process, including the reasoning chains, interaction messages, and conflict negotiation paths of each agent, is fully recorded and structured, making the formation of the final conclusion completely traceable and auditable, greatly enhancing the credibility and acceptability of the intelligent review results. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the core stages of the present invention;

[0046] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The first aspect of this invention provides an intelligent review method for construction plans based on adaptive multi-agent collaboration. Please refer to [link to relevant documentation]. Figure 2 As shown, it includes the following steps:

[0049] S1: Extract features from the input construction plan document to obtain a document feature set;

[0050] S2: Based on the document feature set, dynamically generate an agent pool containing at least two different professional types of review agents, wherein the review agents are constructed based on a large language model;

[0051] S3: Each auditing agent in the agent pool performs parallel auditing and asynchronous collaboration based on a shared blackboard, and writes its audit findings to the shared blackboard.

[0052] S4: Based on the audit findings stored in the shared blackboard, detect audit conflicts between different audit agents; if an audit conflict is detected, initiate a consensus decision process, and conduct weighted voting based on the weights of each audit agent involved in the conflict and the confidence level of their audit findings to form a consensus conclusion; otherwise, directly jump to S5.

[0053] S5: Supervise and control the audit process of the agent pool through meta-agents, and generate a final audit report.

[0054] In this embodiment, as shown in the appendix Figure 1 As shown, it includes the following four core stages:

[0055] Document feature analysis and dynamic agent generation stage: The input construction plan document is parsed and key features are extracted, and an intelligent agent pool is dynamically generated based on the extracted document features;

[0056] Multi-agent parallel review and blackboard collaboration stage: Each agent in the agent pool executes professional review tasks in parallel, and achieves asynchronous collaboration and information sharing by sharing the blackboard;

[0057] Conflict detection and consensus decision-making phase: Conflicts found in the review by each agent are detected, and a consensus decision-making algorithm that combines weighted voting and confidence fusion is used to reach a consensus conclusion;

[0058] Meta-agent reflection and quality control stage: Through meta-agents, the review process and results are verified by reasoning chain verification, contradiction detection and low confidence conclusion processing to achieve quality control and self-correction.

[0059] In this embodiment, the audit results are output in the form of an audit report. The report includes an overall evaluation, the number of issues at each level, the audit confidence level, details of key findings, and information on the reasoning process. The details of key findings include the issue number, issue description, location, basis, and improvement suggestions.

[0060] In one embodiment of the present invention, step S1 includes the following steps:

[0061] Text extraction is performed on construction plan documents in PDF / Word format. AdvancedTextExtractor is used to extract key parameter information from the synchronously uploaded CAD drawings, including parameter tables, text information, pole spacing and wall support positions. At the same time, chapter structure and table of contents information are extracted.

[0062] Extract project type, professional field, key terms, complexity features, and potential features using keyword matching algorithms;

[0063] The formula for analyzing complexity features is as follows:

[0064] ;

[0065] Where Con is the data complexity coefficient, Ncha is the number of chapters, Ncal is the number of calculation formulas, Ndra is the number of drawings, Npar is the number of key parameters, Pcor is the parameter association density, Pcor = number of associated parameter pairs / total number of parameters, and w1, w2, w3, w4 and w5 are weight coefficients.

[0066] Based on the extracted key parameter information, chapter structure, and unlabeled data in the table of contents, a seed feature set is constructed. Candidate features are set for the seed feature set, and latent features are extracted. Where Flatent is the extracted latent feature, Fseed is the seed feature set, Fh is the h-th candidate feature of the seed feature set, sim() is the feature similarity function, and H is the total number of candidate features;

[0067] The extracted features are converted into vector representations and combined into a set to obtain the document feature set.

[0068] In this embodiment, the professional fields of the structural audit agent are structural mechanics and component design. Its audit responsibilities include verifying structural calculations, checking reinforcement design, and assessing structural safety. Its knowledge sources include GB50010 and GB50009 structural codes. The professional fields of the safety audit agent are load verification, construction safety, high-altitude operations, and temporary facilities. Its audit responsibilities include checking safety measures, verifying protection schemes, and assessing risk levels. Its knowledge sources include JGJ80 and JGJ130 safety codes.

[0069] In one embodiment of the present invention, step S2 includes the following steps:

[0070] Using the AgentSpawning algorithm, based on the document feature set and the agent activation threshold mapping table, the matching degree and priority between features and each agent type are calculated to generate an agent pool containing at least two of the following: structural audit agents, security audit agents, economic audit agents, regulatory compliance agents, and geometric audit agents. The formula for calculating the matching degree between features and each agent type is as follows: ;

[0071] in, Let _fi_ be the matching degree between feature_fi and agent_Aj, _ki_ be the keywords contained in feature_fi, _Kj_ be the set of professional keywords of agent_Aj, _similarity()_ be the semantic similarity function, and _λ_ be the parameter association correction coefficient. In this embodiment, the value is set to 0.1-0.3. Prelevance(fi,Aj) is the association degree between the parameters in feature_fi and the professional domain of agent_Aj. Prelevance(fi,Aj) is obtained by calculating the cosine similarity between the parameter keywords in feature_fi and the professional domain keywords of agent_Aj.

[0072] The formula for calculating the priority of features and agent types is as follows: Where α, β, γ and δ are weight coefficients, weight(fi) is the feature weight of feature fi, Mats(fi,Aj) is the matching degree between feature fi and agent Aj, urgency(fi) is the urgency of feature fi of the current scheme to be reviewed, and Pimpact(fi,Aj) is the correlation influence weight of feature fi parameters corresponding to agent Aj.

[0073] In one embodiment of the present invention, the feature weight (fi) of feature fi is set by the following method:

[0074] Document features are divided into core features (including: project type, critical and major project identification, key parameters) and secondary features (including: number of chapters, number of drawings). A judgment matrix is ​​constructed using the AHP (Analytic Hierarchy Process). Several project review experts are invited to score the importance of features. The expert scores are normalized to the range of 0 to 1 to obtain the final feature weight weight(fi) of feature fi.

[0075] The urgency (fi) is set by the following method: based on the ratio of the remaining time to the deadline of the current pending project characteristic fi to the maximum remaining time to the deadline of all pending projects characteristics;

[0076] The correlation influence weight Pimpact(fi,Aj) of the feature fi parameter corresponding to agent Aj is set by the following method: Based on the standardized parameter database of agent Aj, the correlation between parameters is determined by experts (such as the correlation between pole spacing and verticality deviation, and the correlation between wall support position and structural stability). The parameter of feature fi is taken as a data node, and the correlation edges are set between the data nodes corresponding to the parameters with correlation. The degree centrality of each data node is calculated and used as the correlation influence weight Pimpact(fi,Aj) of the feature fi parameter corresponding to the data node.

[0077] In one embodiment of the present invention, in step S3, each reviewing agent in the agent pool performs parallel review and asynchronous collaboration on the construction plan based on the shared blackboard, and writes its review findings into the shared blackboard. The shared blackboard data structure includes the following information:

[0078] Problem List: Stores the problems discovered by each agent and reviews the list of problems discovered. Each problem list record includes a unique identifier for discovery, agent ID, timestamp, problem type, severity level, problem location, detailed description of the problem, supporting evidence, confidence score, and current status indicator.

[0079] The queue of issues to be negotiated stores conflicts that require negotiation among multiple agents. Each negotiation issue record includes a unique identifier for the conflict, a timestamp, a list of agent IDs involved, a description of the conflict topic, a set of opinions from all parties, and the negotiation status.

[0080] Consensus Conclusion Record: Stores the decision results that have reached a consensus, records the consensus conclusion, and each record includes a unique identifier for the conclusion, a timestamp, a topic description, a list of agent IDs that agree with the conclusion, the final consensus content, a comprehensive confidence score, and an inference chain path;

[0081] Reasoning process record: Stores detailed reasoning steps for each agent, records the reasoning process, and each record includes agent ID, timestamp and reasoning step sequence;

[0082] Parameter Relationship Library: Based on the relationship between parameters determined by experts, it stores the relationship rules between key parameters and the relationship influence coefficients calculated by formulas;

[0083] Call Relationship Record Table: Records the iterative nested call paths between agents, feedback trigger conditions, and call results.

[0084] In one embodiment of the present invention, in step S3, the intelligent agents exchange information through a standardized communication protocol. The message data structure of the communication protocol includes a message type identifier, a unique message identifier, a sender intelligent agent ID, a receiver intelligent agent ID, a timestamp, a message content body, and a reference message ID. The message type identifier includes seven types: QUERY, INFORM, REQUEST, PROPOSE, AGREE, REJECT, and CHALLENGE.

[0085] The communication protocols include: standardized communication protocols, ITERATE iterative call instructions, and FEEDBACK feedback instructions;

[0086] The ITERATE iterative call instruction is used to trigger an agent to iteratively execute the same audit task, including the number of iterations, iteration conditions, and parameter update requirements.

[0087] The FEEDBACK instruction is used by an agent to provide feedback on the review results to associated agents, including the feedback object, feedback content, and the review dimensions to be adjusted; the message data structure has been updated to include a call level identifier and iteration count field, supporting the tracing of nested call relationships among multiple agents.

[0088] In one embodiment of the present invention, step S4, detecting review conflicts between different review agents, includes:

[0089] The audit model iterates through all audit findings records using a double loop, compares pairs of records to determine whether they are targeting the same audit object and whether the viewpoints are contradictory, identifies at least one type of conflict among conclusion conflict, numerical conflict, normative reference conflict, and priority conflict, and creates a conflict record.

[0090] Step S4, the consensus decision-making process, includes:

[0091] Collect opinions from participating agents and calculate the weights of agent Aj in the feature fi corresponding to the review conflict. Where H(Aj) is the historical accuracy of the agent, Mats(fi,Aj) is the matching degree between feature fi and agent Aj type, and a1 and a2 are weight coefficients.

[0092] Weighted voting is performed based on agent weights and confidence levels, and opinion scores are calculated using a formula. Where C(Aj) is the confidence level of agent Aj; I(Aj,supportsopinioni) is an indicator function, which takes the value 1 when agent Aj supports opinion opinioni, and 0 otherwise;

[0093] The consensus conclusion is determined based on the comparison between the score difference and the consensus threshold. The confidence level of the final consensus is calculated using the formula... calculate; Σ(W(Aj)) represents the weighted confidence sum of agent Aj, and Σ(W(Aj)) represents the sum of the weights of all agents.

[0094] The review strategy is dynamically adjusted based on real-time monitoring data, and the review results are fed back to step S2.

[0095] In this embodiment, the number of successful tasks (such as making correct decisions and accurately outputting results) and the total number of tasks of agent Aj in recent tasks are counted. The confidence level C(Aj) of agent Aj is calculated as the number of successful tasks / the total number of tasks.

[0096] In one embodiment of the present invention, a linkage mechanism is established between key parameters and between monitoring data and the audit model. The influence of parameter correlation is quantified by adding a new algorithm formula, and the audit strategy is dynamically adjusted in combination with real-time monitoring data. The audit results are then fed back to step S2, including the following steps:

[0097] Dynamic parameters during the construction process are obtained through a real-time monitoring data interface, and the dynamic deviation coefficient is calculated. Where D(kn) is the dynamic deviation coefficient of the kn-th parameter, Preal(kn) is the real-time monitoring value of the kn-th parameter, Pmodel(kn) is the audit model preset value of the kn-th parameter, and Pstd(kn) is the parameter specification standard value.

[0098] Calculate the audit granularity adjustment coefficient based on the parameter dynamic deviation coefficient and the correlation effect results. Where Gadj is the granularity adjustment coefficient for review, and S(kn) is the importance weight of the kn-th parameter; in this embodiment, S(kn) can be set by expert scoring or according to whether the kn-th parameter is a key parameter. When the kn-th parameter is a key parameter, S(kn)=1, and when the kn-th parameter is not a key parameter, S(kn)=0.6, kn∈(1,2,…,P), and P is the total number of parameters;

[0099] If the dynamic deviation coefficient D(kn) of the kn-th parameter is greater than the threshold, or the audit granularity adjustment coefficient Gadj is greater than the threshold, the agent priority update command is triggered, and the agent priority is updated by using the feature and the priority calculation formula of each agent type in step S2.

[0100] In this embodiment, the review granularity of the construction plan document includes three levels: directory level, chapter level, and paragraph level. The system can dynamically adjust the review granularity according to the complexity of the document.

[0101] When D(kn) > Dth (Dth is the deviation threshold, ranging from 0.05 to 0.15), an agent priority update command is triggered, and the agent priority is updated using the features in step S2 and the priority calculation formula for each agent type. When the review granularity adjustment coefficient Gadj > 1.2, paragraph-level refined review is initiated, triggering an agent priority update command, and the agent priority is updated using the features in step S2 and the priority calculation formula for each agent type. When 0.8 ≤ Gadj ≤ 1.2, chapter-level review is maintained, and when Gadj < 0.8, directory-level simplified review is adopted.

[0102] In one embodiment of the present invention, step S5 includes:

[0103] Traverse the sequence of reasoning steps, check the logical coherence between adjacent steps, mark the logical break if there is a logical jump, and trigger a request to supplement details, requesting the supplementation of intermediate steps.

[0104] By calling upon standardized parameter databases, specification databases, and publicly available engineering data resources, a factual consistency score for the generated content is calculated. Among them, Tscore is the factual consistency score. To generate the th-th value in the content, This is the baseline value corresponding to the th-th value in the generated content from the authoritative data source. q represents the importance weight of the th data point, and q represents the total number of data points to be verified. When Tscore ≥ 0.6, the data is deemed to meet the requirement of factual consistency; otherwise, it is deemed not to meet the requirement of factual consistency, and a fact verification request is triggered.

[0105] In this embodiment, the standardized parameter database includes: the NIST Materials Database, publicly available standard data from the official websites of steel, cement, and building materials industry associations, and national standard performance parameters for various products published on the official websites of domestic steel, cement, and building materials industry associations; the specification database includes: information on the formulation and revision of officially published national engineering construction standards (GB), industry standards (JGJ, JG, CJJ, etc.), partial full texts, and authoritative interpretations; and publicly available engineering data resources include: publicly available local engineering construction standards (DB series), which are usually stricter or more detailed than national standards, the national BIM standard component library, and official component libraries of major manufacturers.

[0106] The authenticity of the source is verified by detecting the explicit data source associated with each conclusion generated by the intelligent agent (such as document page number, standard clause number, monitoring data timestamp).

[0107] Hallucination Identification and Correction: When the meta-intelligence determines that the information does not meet the consistency of facts, has no clear data source, or has only partially clear data sources, it is determined to be hallucination information, and the hallucination risk level is calculated as: Hrisk=θ*(1-Tscore)+(1-θ)*Ssource; where Hrisk is the hallucination risk level, θ is the weight coefficient, which is 0.6-0.8 in this embodiment; Ssource is the source validity score, which is set as follows: 1 when there is no clear data source, 0.5 when there are partially clear data sources, and 0 when all have clear data sources;

[0108] A second review is triggered when Hrisk is greater than a threshold or when multiple agents' scores are too close to reach a consensus. In this embodiment, a second review is forced when Hrisk is greater than 0.5.

[0109] In one embodiment of the present invention, the execution process of secondary review includes:

[0110] Generate detailed review prompts, clearly defining review instructions, key areas of focus, required evidence, and review direction;

[0111] Filter out candidate agents that were originally reviewed by the AI ​​agent;

[0112] Invoke the alternative agent to perform the new audit and obtain the results;

[0113] If there are significant differences between the old and new audit results, a manual decision is made; if a consensus is reached, the consensus findings are adopted and the confidence level is increased to the average of the old and new results.

[0114] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for intelligent review of construction plans based on adaptive multi-agent collaboration, characterized in that, Includes the following steps: S1: Extract features from the input construction plan document to obtain a document feature set; S2: Based on the document feature set, dynamically generate an agent pool containing at least two different professional types of review agents, wherein the review agents are constructed based on a large language model; S3: Each auditing agent in the agent pool performs parallel auditing and asynchronous collaboration based on a shared blackboard, and writes its audit findings into the shared blackboard. S4: Based on the audit findings stored in the shared blackboard, detect audit conflicts between different audit agents; if an audit conflict is detected, initiate a consensus decision process, and perform weighted voting based on the weights of each audit agent involved in the conflict and the confidence level of their audit findings to form a consensus conclusion; otherwise, proceed directly to S5. S5: Supervise and control the audit process of the agent pool through meta-agents, and generate a final audit report; In step S4, detecting review conflicts between different review agents includes: The audit model iterates through all audit findings records using a double loop, compares pairs of records to determine whether they are targeting the same audit object and whether the viewpoints are contradictory, identifies at least one type of conflict among conclusion conflict, numerical conflict, normative reference conflict, and priority conflict, and creates a conflict record. In S4, the consensus decision-making process includes: Collect opinions from participating agents and calculate the weights of agent Aj in the feature fi corresponding to the review conflict. Where H(Aj) is the historical accuracy of the agent, Mats(fi,Aj) is the matching degree between feature fi and agent Aj type, and a1 and a2 are weight coefficients. Weighted voting is performed based on agent weights and confidence levels, and opinion scores are calculated using a formula. Where C(Aj) is the confidence level of agent Aj; I(Aj,supportsopinioni) is an indicator function, which takes the value 1 when agent Aj supports opinion opinioni, and 0 otherwise; The consensus conclusion is determined based on the comparison between the score difference and the consensus threshold. The confidence level of the final consensus is calculated using the formula... calculate; Σ(W(Aj)) is the weighted confidence sum of agent Aj, and Σ(W(Aj)) is the sum of the weights of all agents. The review strategy is dynamically adjusted based on real-time monitoring data, and the review results are fed back to step S2. Establish a linkage mechanism between key parameters and between monitoring data and the audit model. Quantify the correlation and impact of parameters through new algorithmic formulas, dynamically adjust the audit strategy based on real-time monitoring data, and feed the audit results back to step S2. This includes the following steps: Dynamic parameters during the construction process are obtained through a real-time monitoring data interface, and the dynamic deviation coefficient is calculated. Where D(kn) is the dynamic deviation coefficient of the kn-th parameter, Preal(kn) is the real-time monitoring value of the kn-th parameter, Pmodel(kn) is the audit model preset value of the kn-th parameter, and Pstd(kn) is the parameter specification standard value. Calculate the audit granularity adjustment coefficient based on the parameter dynamic deviation coefficient and the correlation effect results. Where Gadj is the granularity adjustment coefficient, S(kn) is the importance weight of the kn-th parameter, kn∈(1,2,…,P), and P is the total number of parameters; If the dynamic deviation coefficient D(kn) of the knth parameter is greater than the threshold, or the audit granularity adjustment coefficient Gadj is greater than the threshold, the agent priority update command is triggered, and the agent priority is updated by using the feature and the priority calculation formula of each agent type in step S2. The S5 includes: Traverse the sequence of reasoning steps, check the logical coherence between adjacent steps, mark the logical break if there is a logical jump, and trigger a request to supplement details and request the supplementation of intermediate steps. By calling upon standardized parameter databases, specification databases, and publicly available engineering data resources, a factual consistency score for the generated content is calculated. Among them, Tscore is the factual consistency score. To generate the th-th value in the content, This is the baseline value corresponding to the th-th value in the generated content from the authoritative data source. q represents the importance weight of the th data point, and q represents the total number of data points to be verified. When Tscore ≥ 0.6, the data is deemed to meet the requirement of factual consistency; otherwise, it is deemed not to meet the requirement of factual consistency, and a fact verification request is triggered. The authenticity of the source is verified by detecting the explicit data source associated with each conclusion generated by the agent through meta-agent detection. Hallucination Identification and Correction: When a hallucination is detected that does not meet the requirements of factual consistency, has no clear data source, or has only partially clear data sources, it is determined to be hallucination information, and the hallucination risk level is calculated. Where Hrisk is the hallucination risk level, θ is the weighting coefficient, and Ssource is the source validity score. The source validity score Ssource is set as follows: 1 when there is no explicit data source, 0.5 when some have explicit data sources, and 0 when all have explicit data sources. A secondary review is triggered if the Hrisk exceeds a threshold, or if the difference in scores among multiple agents is less than a threshold and a consensus cannot be reached.

2. The intelligent review method for construction plans based on adaptive multi-agent collaboration as described in claim 1, characterized in that, S1 includes the following steps: Text extraction is performed on construction plan documents in PDF / Word format. AdvancedTextExtractor is used to extract key parameter information from the synchronously uploaded CAD drawings, including parameter tables, text information, pole spacing and wall support positions. At the same time, chapter structure and table of contents information are extracted. Extract project type, professional field, key terms, complexity features, and potential features using keyword matching algorithms; The formula for analyzing complexity features is as follows: ； Where Con is the data complexity coefficient, Ncha is the number of chapters, Ncal is the number of calculation formulas, Ndra is the number of drawings, Npar is the number of key parameters, Pcor is the parameter association density, Pcor = number of associated parameter pairs / total number of parameters, and w1, w2, w3, w4 and w5 are weight coefficients. Based on the extracted key parameter information, chapter structure, and unlabeled data in the table of contents, a seed feature set is constructed. Candidate features are set for the seed feature set, and latent features are extracted. Where Flatent is the extracted latent feature, Fseed is the seed feature set, Fh is the h-th candidate feature of the seed feature set, sim() is the feature similarity function, and H is the total number of candidate features; The extracted features are converted into vector representations and combined into a set to obtain the document feature set.

3. The intelligent review method for construction schemes based on adaptive multi-agent collaboration as described in claim 1, characterized in that, S2 includes the following steps: Using the AgentSpawning algorithm, based on the document feature set and the agent activation threshold mapping table, the matching degree and priority between features and each agent type are calculated to generate an agent pool containing at least two of the following: structural audit agents, security audit agents, economic audit agents, regulatory compliance agents, and geometric audit agents. The formula for calculating the matching degree between features and each agent type is as follows: ; in, Let be the matching degree between feature fi and agent Aj, ki be the keywords contained in feature fi, Kj be the set of professional keywords of agent Aj, similarity() be the semantic similarity function, λ be the parameter association correction coefficient, and Prelevance(fi,Aj) be the correlation degree between the parameters in feature fi and the professional domain of agent Aj. Prelevance(fi,Aj) is obtained by calculating the cosine similarity between the parameter keywords in feature fi and the professional domain keywords of agent Aj. The formula for calculating the priority of features and agent types is as follows: Where α, β, γ and δ are weight coefficients, weight(fi) is the feature weight of feature fi, Mats(fi,Aj) is the matching degree between feature fi and agent Aj type, urgency(fi) is the urgency of feature fi of the current scheme to be reviewed, and Pimpact(fi,Aj) is the correlation influence weight of feature fi parameter corresponding to agent Aj.

4. The intelligent review method for construction plans based on adaptive multi-agent collaboration according to claim 3, characterized in that, The feature weights of feature fi are set using the following method: Document features are divided into core features and secondary features. A judgment matrix is ​​constructed using the AHP (Analytic Hierarchy Process). Several engineering review experts are invited to score the importance of the features. The expert scores are normalized to the interval between 0 and 1 to obtain the final feature weight (fi). The urgency (fi) is set by the following method: based on the ratio of the remaining time to the deadline of the current pending project characteristic fi to the maximum remaining time to the deadline of all pending projects characteristics; The association influence weight Pimpact(fi,Aj) of the feature fi parameter corresponding to agent Aj is set by the following method: Based on the standardized parameter database of agent Aj, the association relationship between parameters is determined by experts, the parameter of feature fi is taken as a data node, the data nodes corresponding to the parameters with association relationship are set with association edges, and the degree centrality of each data node is calculated as the association influence weight Pimpact(fi,Aj) of the feature fi parameter corresponding to the data node.

5. The intelligent review method for construction schemes based on adaptive multi-agent collaboration according to claim 1, characterized in that, In step S3, each reviewing agent in the agent pool performs parallel review and asynchronous collaboration on the construction plan based on the shared blackboard, and writes its review findings into the shared blackboard. The shared blackboard data structure includes the following information: Problem List: Stores the problems discovered by each agent and reviews the list of problems discovered. Each problem list record includes a unique identifier for discovery, agent ID, timestamp, problem type, severity level, problem location, detailed description of the problem, supporting evidence, confidence score, and current status indicator. The queue of issues to be negotiated stores conflicts that require negotiation among multiple agents. Each negotiation issue record includes a unique identifier for the conflict, a timestamp, a list of agent IDs involved, a description of the conflict topic, a set of opinions from all parties, and the negotiation status. Consensus Conclusion Record: Stores the decision results that have reached a consensus, records the consensus conclusion, and each record includes a unique identifier for the conclusion, a timestamp, a topic description, a list of agent IDs that agree with the conclusion, the final consensus content, a comprehensive confidence score, and an inference chain path; Reasoning process record: Stores detailed reasoning steps for each agent, records the reasoning process, and each record includes agent ID, timestamp and reasoning step sequence; Parameter Relationship Library: Based on the relationship between parameters determined by experts, it stores the relationship rules between key parameters and the relationship influence coefficients calculated by formulas; Call Relationship Record Table: Records the iterative nested call paths between agents, feedback trigger conditions, and call results.

6. The intelligent review method for construction schemes based on adaptive multi-agent collaboration according to claim 1, characterized in that, In S3, intelligent agents exchange information through a standardized communication protocol. The message data structure of the communication protocol includes a message type identifier, a unique message identifier, a sender intelligent agent ID, a receiver intelligent agent ID, a timestamp, a message content body, and a reference message ID. The message type identifier includes seven types: QUERY, INFORM, REQUEST, PROPOSE, AGREE, REJECT, and CHALLENGE. The communication protocol includes: a standardized communication protocol, an ITERATE iterative call instruction, and a FEEDBACK feedback instruction; The ITERATE iterative call instruction is used to trigger an agent to iteratively execute the same audit task, including the number of iterations, iteration conditions, and parameter update requirements. The FEEDBACK instruction is used by an agent to provide feedback on the review results to associated agents, including the feedback object, feedback content, and the review dimensions to be adjusted; the message data structure has been updated to include a call level identifier and iteration count field, supporting the tracing of nested call relationships among multiple agents.

7. The intelligent review method for construction plans based on adaptive multi-agent collaboration as described in claim 1, characterized in that, The execution process of the secondary review includes: Generate detailed review prompts, clearly defining review instructions, key areas of focus, required evidence, and review direction; Filter out candidate agents that were originally reviewed by the AI ​​agent; Invoke the alternative agent to perform the new audit and obtain the results; If there are significant differences between the old and new audit results, a manual decision is made; if a consensus is reached, the consensus findings are adopted and the confidence level is increased to the average of the old and new results.

Images