A Multi-Agent Traceable Decision System and Method for Environmental Data Mining
By combining a multi-agent conditional cooperation mechanism with an environmental knowledge graph, the reliability and transparency issues of a single-agent decision-making system are solved, enabling efficient, accurate, and self-optimizing decision-making in environmental data mining.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-10-31
- Publication Date
- 2026-05-26
AI Technical Summary
Existing single-agent decision-making systems suffer from reliability bottlenecks, untraceable decision-making processes, rigid resource allocation, insufficient integration of professional knowledge, and a lack of self-optimization capabilities in environmental data mining, resulting in high uncertainty, high costs, and low efficiency in decision outcomes.
A multi-agent conditional collaboration mechanism is introduced, including a task complexity assessment module, a multi-agent conditional collaboration module, an arbitrator triggering and adjudication module, an environmental knowledge graph module, a human expert correction module, and a feedback learning optimization module, to construct a human-machine collaborative intelligent decision-making framework and realize decision tracing and resource optimization.
Significantly improve the reliability and transparency of decision-making, optimize resource allocation, enhance professionalism and accuracy, and form a decision-making system with self-evolving capabilities.
Smart Images
Figure CN121436176B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a multi-agent traceable decision system and method for environmental data mining, belonging to the interdisciplinary field of artificial intelligence and environmental informatics. It is particularly suitable for scenarios requiring highly reliable decision support in the field of environmental science, such as intelligent literature screening, pollution event analysis, and environmental policy assessment. Background Technology
[0002] In the field of environmental informatics, facing the ever-increasing volume of environmental data (such as academic literature, water plant and pipeline operation data, monitoring reports, satellite remote sensing data, etc.), the key to supporting scientific research innovation and policy making lies in how to quickly and accurately extract valuable information from it. Existing technologies mostly employ a single large language model agent for automated decision-making, which presents the following problems:
[0003] (1) The reliability bottleneck problem of single agent decision-making
[0004] The root of the problem lies in the cognitive blind spots and inherent uncertainties inherent in single models. For the numerous complex, ambiguous, or interdisciplinary cases in environmental science (such as determining whether a new technology belongs to the "advanced oxidation process," or assessing the migration and transformation patterns of a novel pollutant under specific environmental conditions), the judgments often lack sufficient credibility. Specifically, when dealing with cutting-edge technical literature, descriptions of complex pollution mechanisms, or environmental policy analyses involving multiple intertwined factors, the error rate of single models increases significantly, making it difficult to meet the requirements of rigorous scientific research and high reliability in policy formulation.
[0005] (2) The "black box" problem and lack of traceability in the AI decision-making process
[0006] The root of the problem lies in the fact that even if the model provides reasoning, its internal decision-making process remains isolated and unverifiable. Users cannot know whether the decision has undergone sufficient weighing or whether there are obvious evidentiary omissions. Specifically, when the decision outcome is questionable, environmental experts or decision-makers find it difficult to conduct effective audits and reviews, leading to reduced trust in AI systems and hindering their in-depth application in key scenarios such as project review and environmental risk assessment.
[0007] (3) Efficiency and cost issues caused by rigid allocation of computing resources
[0008] The root cause of the problem is that existing solutions cannot intelligently differentiate between the difficulty levels of tasks. Using a "one-size-fits-all" approach with high-performance models to ensure accuracy leads to unnecessary computational waste on many simple tasks, resulting in high costs. Conversely, using only a single model to control costs fails to guarantee the processing quality for complex tasks. This manifests as a mismatch between resource consumption and task difficulty, a low overall system cost-effectiveness ratio, and excessive cost investment.
[0009] (4) The problem of insufficient integration of professional knowledge in the environmental field
[0010] The root of the problem lies in the lack of a deep understanding of the specialized knowledge, terminology, and policy context within the field of environmental science in the general language model, which easily leads to professional misjudgments. Specifically, this may manifest as confusion between similar technical concepts (such as "electrocatalytic oxidation" and "photocatalytic oxidation"), or an inability to accurately understand specific clauses and standard limits in policy documents.
[0011] (5) The problem of the system lacking continuous self-optimization capability
[0012] The root of the problem is that once traditional systems are deployed, their performance tends to become fixed, making it impossible for them to learn and improve from actual use, and thus difficult to adapt to the rapid development of environmental science. Specifically, when faced with new pollutants, treatment technologies, or policies and regulations, the system's decision-making ability may decline, requiring frequent human intervention for adjustments.
[0013] The aforementioned problems reveal significant limitations of this single-agent approach: its decision-making results are uncertain, especially when dealing with complex, ambiguous, or cutting-edge environmental science problems, resulting in a high error rate; furthermore, its decision-making process is like a "black box," lacking effective verification mechanisms and transparency, making it difficult for users to trace and verify its reasoning logic when the results are questionable. Moreover, from a resource utilization perspective, the widespread adoption of high-performance models to ensure accuracy leads to high costs, while using a single model to control costs cannot guarantee the quality of processing complex data. Therefore, there is an urgent need in this field for a technical solution that can significantly improve decision-making reliability, process transparency, and enable intelligent resource allocation. Summary of the Invention
[0014] The purpose of this invention is to address the fundamental shortcomings of existing single-agent-based automated decision-making systems in environmental data mining applications. It provides a multi-agent traceable decision-making system and method for environmental data mining, offering an automated decision-making solution with high reliability, full-process traceability, and resource optimization capabilities. The core idea of this invention is to introduce a multi-agent conditional collaboration mechanism called "Analyst-Analyst-Arbitrator" (RRA), and deeply integrate decision tracing and human expert correction functions to construct a human-machine collaborative intelligent decision-making framework. Decision tracing records and outputs all independent opinions, reasoning chains, and final decisions during the agent collaboration process, forming a complete and auditable decision path.
[0015] The objective of this invention is achieved through the following technical solution:
[0016] A multi-agent traceable decision-making system for environmental data mining includes: a task complexity assessment module, a multi-agent conditional collaboration module, an arbitrator triggering and adjudication module, an environmental knowledge graph module, a human expert correction module, a feedback learning optimization module, and a decision tracing and visualization module.
[0017] The signal output terminals of the task complexity assessment module and the environmental knowledge graph module are both connected to the signal input terminal of the multi-agent conditional collaboration module. The signal output terminal of the multi-agent conditional collaboration module is connected to the signal input terminals of the arbitrator triggering and adjudication module and the decision tracing and visualization module, respectively. The signal output terminal of the arbitrator triggering and adjudication module is connected to the signal input terminals of the human expert correction module and the decision tracing and visualization module, respectively. The signal output terminal of the human expert correction module is connected to the signal input terminals of the feedback learning optimization module and the decision tracing and visualization module, respectively. The signal output terminal of the feedback learning optimization module is connected to the signal output and input terminals of the multi-agent conditional collaboration module.
[0018] The task complexity assessment module is used to analyze the task characteristics of environmental data items, identify task types, perform complexity quantification analysis, and output collaboration mode signals.
[0019] The multi-agent conditional cooperation module is used to realize conditional cooperation between the first and second intelligent analysts. This module dynamically adjusts the cooperation strategy of the two intelligent analysts based on the cooperation mode signal received from the task complexity evaluation module.
[0020] The arbitrator triggering and adjudication module is used to conditionally trigger the intervention of the intelligent arbitrator agent when the final conclusions of two intelligent analysts in the multi-agent conditional collaboration module differ.
[0021] The environmental knowledge graph module is used to store professional knowledge in the environmental domain, providing domain knowledge support for intelligent agent decision-making.
[0022] The human expert correction module is used to correct decisions made by arbitrators in the triggering and adjudication module where the confidence level of the adjudication does not meet the standard through human experts;
[0023] The feedback learning optimization module is used to receive correction feedback from the human expert correction module and optimize the dynamic weight adjustment algorithm and agent decision-making logic based on the feedback data.
[0024] The decision tracing and visualization module is used to record the complete decision chain from task configuration and initial review opinions to arbitration reasons, and generate a visual report.
[0025] Preferably, the collaboration mode signal in the task complexity assessment module is used to indicate whether the subsequent agent should adopt an independent processing mode or an interactive processing mode.
[0026] Preferably, the dynamic adjustment strategy for the collaboration of the two intelligent analysts in the multi-agent conditional collaboration module is as follows:
[0027] When the collaboration mode signal indicates independent processing mode, the two intelligent analysts work in parallel and only compare results in the final output stage. When the collaboration mode signal indicates interactive processing mode, the decision system increases the frequency of information exchange between the two intelligent analysts.
[0028] Preferably, when making a ruling, the intelligent arbitrator in the arbitrator triggering and ruling module will weigh and adopt the arguments of the two intelligent analysts according to the review criteria preset by the task configuration module, and then form a final ruling. If the confidence level given by the intelligent arbitrator is still lower than the threshold, then human experts will be invited to make a final ruling.
[0029] Preferably, the environmental knowledge graph module stores professional knowledge in the environmental field, including: pollutant attributes, treatment technologies, and policy and standard information.
[0030] Preferably, the dynamic weight adjustment algorithm in the feedback learning optimization module is as follows: dynamically updating the weight coefficients of the two intelligent analysts based on the arbitration of the intelligent arbitrator and the corrective feedback from the human expert. This results in analysts whose historical decision outcomes are more closely aligned with correct results receiving higher weight.
[0031] The decision-making method for a multi-agent traceable decision system based on environmental data mining includes the following steps:
[0032] Step 1: Task Complexity Assessment and Collaboration Mode Allocation
[0033] The decision system receives task descriptions and environmental data items, analyzes the task characteristics through the task complexity assessment module, and generates a collaboration mode signal, which is divided into independent processing mode and interactive processing mode.
[0034] Step Two: Parallel Processing and Dynamic Collaboration of Dual-Intelligence Analysts
[0035] After the environmental data item is analyzed in step one, it enters the multi-agent conditional cooperation module. If the cooperation mode signal indicates independent processing mode, the first and second intelligent analysts in the multi-agent conditional cooperation module process the data item in parallel and independently without information exchange. If the cooperation mode signal indicates interactive processing mode, the decision system establishes a communication channel for the first and second intelligent analysts, allowing them to exchange intermediate inference results or conduct a limited number of rounds of negotiation.
[0036] Step 3: Consistency Check and Arbitration Trigger
[0037] Compare the final conclusions of the two analysts obtained in step two. If the conclusions are consistent, output the result and proceed to step six; if the conclusions differ, trigger the arbitrator to intervene and make a ruling, and execute step four.
[0038] Step Four: Arbitrator's Award and Weighting Integration
[0039] Through the arbitrator trigger and award module, the arbitrator intervenes to make the final award. If the confidence level of the award meets the standard, the final decision is output and the process jumps to step six; otherwise, step five is executed.
[0040] Step 5: Human Expert Correction and Weight Optimization Feedback
[0041] The human expert correction module corrects decisions that do not meet the confidence level in step four, and the feedback is recorded by the decision system and used for optimization through the feedback learning optimization module.
[0042] Step Six: Decision Origin Tracing and Weight Visualization
[0043] The decision tracing and visualization module generates a tracing report containing complete information on the decision-making process and collaboration model.
[0044] Preferably, the method for generating the collaboration mode signal in step one is as follows: the task complexity assessment module analyzes the task characteristics and makes a judgment based on a preset rule set to generate a collaboration mode signal; the rule set is configured such that: if the task characteristics satisfy at least one of the preset high complexity conditions, an interactive processing collaboration mode signal is generated; otherwise, an independent processing collaboration mode signal is generated.
[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0046] (1) A multi-agent conditional collaboration mechanism is introduced. Two intelligent analysts are set up to process independently or interactively. In the interactive processing mode, the two intelligent analysts are allowed to exchange intermediate inference results or conduct limited rounds of negotiation. Then, a conclusion is output. When the two conclusions differ, a higher-level intelligent arbitrator is triggered to make a final decision. This "double verification + authoritative arbitration" architecture simulates the "consultation" mode of human experts, providing multiple guarantees for the decision-making results, and significantly improving the decision-making accuracy of difficult cases. In addition, if the confidence level given by the intelligent arbitrator is still lower than the threshold, the human expert correction module is used to request a human expert to make a final decision, which can improve the accuracy of the final decision and achieve more precise decision optimization according to the characteristics of different environmental tasks. Furthermore, the predictive dispute detection mechanism enables the system to actively identify and prevent potential problems, greatly improving the system's intelligence level and decision-making efficiency.
[0047] (2) This invention constructs a full-process decision traceability system. The decision system does not simply output an isolated conclusion, but generates a complete "audit report" that clearly records the entire decision-making path from data input, independent judgment and reasoning or interactive processing by two intelligent analysts, consistency checks, to arbitration rulings (such as triggers). This makes the entire decision-making process transparent, verifiable, and auditable, greatly enhancing the credibility and acceptability of the results. Furthermore, the complete weight visualization traceability system provides transparent audit evidence for the decision-making process.
[0048] (3) This invention implements a difficulty-based dynamic resource allocation strategy. The decision-making system analyzes the task characteristics of environmental data items through the task complexity assessment module, performs task type identification and complexity quantification analysis, assesses the complexity, and outputs a collaboration mode signal to determine whether the two intelligent analysts will handle the tasks independently or interactively. Most simple tasks are assigned to the lower-cost basic model for rapid processing. Furthermore, the high-performance intelligent arbitrator model is only invoked when the two basic analysts disagree. This conditional triggering mechanism ensures that expensive computing resources are precisely allocated to the few most pressing "disputed cases," thereby optimizing system overhead while maintaining overall accuracy.
[0049] (4) This invention includes an intelligent agent that deeply integrates an environmental knowledge graph with specialized optimizations. The decision-making system incorporates an environmental knowledge graph module that includes pollutant attributes, treatment technologies, policy standards, etc., providing authoritative domain context for the intelligent agent's reasoning. Simultaneously, the intelligent analyst or arbitrator model can be fine-tuned for specific sub-domains of environmental science, enabling it to possess a stronger understanding of professional issues. The environmental policy compliance verification function provides important policy dimension guarantees for the decision-making results, and the reasoning enhancement technology based on the environmental knowledge graph significantly improves the professionalism, authority, and accuracy of the decision.
[0050] (5) This invention establishes a closed-loop feedback mechanism of human expert correction and incremental learning. When human experts in the human expert correction module correct the arbitration results of the system, the correction cases and reasons are automatically collected by the feedback learning optimization module of the decision-making system and used for incremental learning and fine-tuning of the model. This enables the decision-making system to continuously learn from the feedback of experts, constantly optimize its decision-making logic, and form a self-evolving system that becomes smarter the more it is used. Attached Figure Description
[0051] Figure 1 This is a diagram illustrating the overall architecture of the multi-agent traceable decision-making system for environmental data mining according to the present invention.
[0052] Figure 2 This is a flowchart of the multi-agent traceable decision system for environmental data mining according to the present invention.
[0053] Figure 3 This is a schematic diagram of the decision tracing and weight visualization interface of the present invention.
[0054] Figure 4 This is a schematic diagram of the environmental knowledge graph query and reasoning enhancement mechanism of the present invention. Detailed Implementation
[0055] The present invention will be further described in detail below with reference to the accompanying drawings: This embodiment is implemented under the premise of the technical solution of the present invention, and detailed implementation methods are given, but the protection scope of the present invention is not limited to the following embodiments.
[0056] like Figure 1 and Figure 2 As shown in the figure, this embodiment involves a multi-agent traceable decision system for environmental data mining, which includes a task complexity assessment module, a multi-agent conditional collaboration module, an arbitrator triggering and adjudication module, an environmental knowledge graph module, a human expert correction module, a feedback learning optimization module, and a decision tracing and visualization module.
[0057] Task complexity assessment module: This module analyzes the task characteristics of environmental data items, identifies task types, performs complexity quantification analysis, and outputs a collaboration mode signal. This signal indicates whether subsequent agents should adopt an "independent processing mode" or an "interactive processing mode."
[0058] A multi-agent conditional collaboration module, connected to the task complexity assessment module, enables conditional collaboration among the first intelligent analyst, the second intelligent analyst, and the arbitrator intelligent agent. This module dynamically adjusts the collaboration strategy of the two intelligent analysts based on received collaboration mode signals: when the collaboration mode signal indicates "independent processing mode," the two intelligent analysts work in parallel, comparing results only in the final output stage. When the collaboration mode signal indicates "interactive processing mode," the decision-making system increases the frequency of information exchange between the two intelligent analysts; for example, allowing them to exchange intermediate results and conduct real-time negotiation to jointly address complex problems. This dynamic collaboration mechanism effectively improves system efficiency while ensuring decision quality. Finally, the consistency of the final conclusions of the two intelligent analysts is verified.
[0059] Arbitrator Trigger and Adjudication Module: The intelligent arbitrator agent is conditionally triggered to intervene only when the final conclusions of the two intelligent analysts differ. During the adjudication process, the intelligent arbitrator will weigh and adopt the arguments of the two intelligent analysts according to the review criteria preset by the task configuration module, thus forming a final ruling. If the confidence level given by the intelligent arbitrator is still below a threshold, human experts will be invited to make the final decision. This invention innovatively proposes a predictive dispute detection method. Before the intelligent arbitrator outputs its results, it predicts cases that may lead to disagreements based on confidence level judgment, and improves the accuracy of the final decision by directly inviting human experts to pay attention.
[0060] The environmental knowledge graph module stores professional knowledge in the environmental field, including information such as pollutant attributes, treatment technologies, policies, and standards, providing domain knowledge support for intelligent agent decision-making; for example... Figure 4 As shown, the decision-making system incorporates an environmental knowledge graph module, containing structured knowledge such as pollutant attribute databases, treatment technology databases, policy and regulatory databases, and case knowledge bases. When the agent makes decisions, it can query and reference authoritative information in the knowledge graph in real time, enhancing the accuracy and professionalism of its reasoning. For example, when determining whether a technology is suitable for the removal of a specific pollutant, the decision-making system automatically verifies the chemical characteristics of the pollutant and existing effective treatment methods in the knowledge graph. Furthermore, based on the policy and regulatory database, it can automatically verify the compliance of the technical solution with current policy requirements. During the decision-making process, the system simultaneously checks the consistency between the technical description and relevant policy provisions, providing compliance assurance for environmental management decisions.
[0061] The human expert correction module is used to correct decisions made by arbitrators in the triggering and adjudication module where the confidence level of the adjudication does not meet the standard through human experts;
[0062] The feedback learning optimization module is used to receive correction feedback from the human expert correction module and optimize the dynamic weight adjustment algorithm and agent decision-making logic based on the feedback data.
[0063] The decision tracing and visualization module is used to record the complete decision chain from task configuration and initial review opinions to arbitration reasons, and generate a visual report.
[0064] A multi-agent traceable decision-making method for environmental data mining includes the following steps:
[0065] Includes the following steps:
[0066] Step 1: Task Complexity Assessment and Collaboration Mode Allocation
[0067] The decision system receives task descriptions and environmental data items, analyzes the task characteristics through the task complexity assessment module, and generates a collaboration mode signal, which is divided into independent processing mode and interactive processing mode.
[0068] Step Two: Parallel Processing and Dynamic Collaboration of Dual-Intelligence Analysts
[0069] After the environmental data item is analyzed in step one, it enters the multi-agent conditional cooperation module. If the cooperation mode signal indicates independent processing mode, the first and second intelligent analysts in the multi-agent conditional cooperation module process the data item in parallel and independently without information exchange. If the cooperation mode signal indicates interactive processing mode, the decision system establishes a communication channel for the first and second intelligent analysts, allowing them to exchange intermediate inference results or conduct a limited number of rounds of negotiation.
[0070] Step 3: Consistency Check and Arbitration Trigger
[0071] Compare the final conclusions of the two analysts obtained in step two. If the conclusions are consistent, output the result and proceed to step six; if the conclusions differ, trigger the arbitrator to intervene and make a ruling, and execute step four.
[0072] Step Four: Arbitrator's Award and Weighting Integration
[0073] Through the arbitrator trigger and award module, the arbitrator intervenes to make the final award. If the confidence level of the award meets the standard, the final decision is output and the process jumps to step six; otherwise, step five is executed.
[0074] Step 5: Human Expert Correction and Weight Optimization Feedback
[0075] The human expert correction module corrects decisions that do not meet the confidence level in step four, and the feedback is recorded by the decision system and used for optimization through the feedback learning optimization module.
[0076] Step Six: Decision Origin Tracing and Weight Visualization
[0077] The decision tracing and visualization module generates a tracing report containing complete information on the decision-making process and collaboration model.
[0078] The method for generating the collaboration mode signal in step one is as follows: the task complexity assessment module analyzes the task characteristics and makes a judgment based on a preset rule set to generate a collaboration mode signal; the rule set is configured such that: if the task characteristics satisfy at least one of the preset high complexity conditions, an interactive processing collaboration mode signal is generated; otherwise, an independent processing collaboration mode signal is generated.
[0079] The pre-defined conditions for high complexity are as follows:
[0080] 1. Vague task type: The task description contains open-ended verbs such as "evaluate", "analyze", and "infer", or the objective involves subjective judgments such as "novelty" and "feasibility".
[0081] 2. Numerous judgment conditions: There are ≥ 5 key judgment conditions required to complete the task.
[0082] 3. High level of professional knowledge: The tasks involve cutting-edge technologies, interdisciplinary knowledge, or require in-depth interpretation of policies and regulations and complex causal reasoning.
[0083] In step four, when making a final ruling, the arbitrator will take into account the conclusions of the two intelligent analysts and their corresponding weighting coefficients.
[0084] In step five, during the correction process involving human experts, the decision-making system records the experts' opinions on adjusting different weight allocation schemes. This feedback information will be used in the dynamic weight adjustment algorithm of the feedback learning optimization module, enabling the system to continuously learn and improve the weight allocation strategy, and gradually enhance its adaptability to environmental data characteristics.
[0085] The dynamic weight adjustment algorithm is as follows: based on the arbitration of the intelligent arbitrator and the corrective feedback from the human expert, the weight coefficients of the two intelligent analysts are dynamically updated. Analysts whose historical decision outcomes more closely align with correct results are given higher weight, as shown in the following formula:
[0086]
[0087] These represent the dynamic weights of the first and second intelligent analysts, respectively. These represent the win counters for the first and second intelligent analysts, respectively. This is a preset smoothing factor constant, and its value is greater than zero. Defined as "system total feedback counter".
[0088] The application process of the dynamic weight adjustment algorithm is illustrated with an example, as follows:
[0089] Set the smoothing factor constant The value is 10, when the decision-making system makes its first judgment. and All values are 0. Assuming that the first and second AI analysts disagree on this judgment, and the first analyst's judgment matches the final answer, then... Then, the weighting factor for the next judgment is dynamically adjusted. Assuming there is no disagreement between the first and second intelligent analysts in the second judgment, the weighting factor remains unchanged, and a third judgment is made. If a disagreement arises in the third judgment, and the first intelligent analyst is correct again, then the weighting factor is updated. The weighting factors will change again and be recalculated. This influences the fourth judgment, and the weights are dynamically adjusted accordingly.
[0090] The decision origination report generated by the decision-making system in step six not only includes traditional decision-making process information, but also adds a visual display of weight allocation and adjustment. Users can clearly see how the weights of each agent are dynamically adjusted throughout the decision-making process, and how these adjustments affect the final decision.
[0091] Example 1
[0092] The following example of environmental literature screening illustrates in detail how the system of the present invention operates through a dynamically weighted agent cooperative network. The task description is to determine whether the following research belongs to the specified topic: experimental research on the removal of organic pollutants in water using a single electro-oxidation / electrocatalytic oxidation technology. The input environmental data is the title and abstract of a literature article, specifically a study on a cost-effective water-manganese ore-silicon solar cell hybrid system for dye decolorization.
[0093] The task complexity assessment module receives the task and literature data. Analysis: The task has a clear theme and well-defined keywords ("electro-oxidation," "electrocatalytic oxidation," "organic pollutants," "experimental research"), belonging to a typical simple classification task. The task type is clear, with fewer than five judgment conditions, and requires minimal professional knowledge. Decision: Based on the preset rule set, the task is classified as low complexity. An "independent processing mode" signal is generated.
[0094] Subsequently, the process enters a parallel, independent processing phase by two intelligent analysts. The first intelligent analyst (e.g., the DeepSeek model) determines the document as "relevant" based on its knowledge base. Its core reasoning is that the document clearly describes the process of removing organic pollutants (methyl orange) from water using a sodium manganese oxide electrode, which aligns with the basic characteristics of water treatment technology. Almost simultaneously, the second intelligent analyst (e.g., the Moonshot model) determines it as "irrelevant," focusing its reasoning on the fact that the technology described is a hybrid system combining photoelectric and electrochemical processes, rather than the "single electro-oxidation / electrocatalytic oxidation" technical path required by the preset standard.
[0095] After detecting a fundamental disagreement between the conclusions of the two intelligent analysts, the system formally activated the arbitrator intelligent agent.
[0096] During the arbitrator's ruling phase, the arbitrator agent received two analysis reports and corresponding weighting schemes. Instead of simply issuing a third vote, the arbitrator focused on reviewing the literature based on the weighting schemes and the analysis processes of the two intelligent analysts. After comprehensive analysis, it concluded that although the study involved the degradation of organic pollutants, its core mechanism was photoelectric synergy, which differed significantly in technical principle from the pre-set standard of "single electro-oxidation / electrocatalytic oxidation." Therefore, the arbitrator ruled the final result as "irrelevant" and provided detailed arbitration reasons. However, because its ruling involved, to some extent, the interpretation of ambiguous standards, the system gave a comprehensive confidence score of 0.65 (lower than the pre-set threshold of 0.7).
[0097] Because the confidence level of the arbitration result did not reach the threshold, the system automatically triggered a human expert correction and weight optimization feedback mechanism. After intervention from environmental engineering experts, the system first reviewed the complete decision tracing report, including records of dynamic weight adjustments. The experts ultimately accepted the arbitrator's "irrelevant" conclusion, while providing more professional and insightful corrective reasons: the water-sodium manganese ore-silicon solar cell hybrid technology differs significantly in technical principle from the preset "single electro-oxidation / electrocatalytic oxidation" standard, failing to meet the screening requirements. Secondly, this technology is mainly suitable for decolorizing specific types of wastewater such as azo dyes, and its mineralization efficiency for persistent organic pollutants (such as PFAS) is limited, with relatively high system energy consumption; therefore, it is not suitable as a replacement or preferred solution for the removal of target high-standard pollutants. After expert confirmation, the system outputs the final conclusion. Based on the final conclusion, the weights of the two intelligent analysts are dynamically adjusted; the second intelligent reviewer, who answered correctly, is given a higher weight, while the first intelligent reviewer, who answered incorrectly, has a lower weight. The decision tracing and visualization module generates a decision tracing report containing the final result (e.g., Figure 3 (As shown).
[0098] Subsequently, in the next round of judgment, the system will extract typical cases and the latest weight factors from the knowledge base to achieve more accurate judgments. This complete process vividly demonstrates how this invention, through a dynamically weighted agent collaborative network, organically integrates multi-agent collaboration, human expert correction, and reinforcement learning to achieve continuous optimization of decision-making accuracy and system adaptability.
[0099] The above description is merely a preferred embodiment of the present invention. These specific embodiments are different implementations based on the overall concept of the present invention, and the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A multi-agent traceable decision-making system for environmental data mining, characterized in that, include: The module includes a task complexity assessment module, a multi-agent conditional collaboration module, an arbitrator triggering and adjudication module, an environmental knowledge graph module, a human expert correction module, a feedback learning optimization module, and a decision tracing and visualization module. The signal output terminals of the task complexity assessment module and the environmental knowledge graph module are both connected to the signal input terminal of the multi-agent conditional collaboration module. The signal output terminal of the multi-agent conditional collaboration module is connected to the signal input terminals of the arbitrator triggering and adjudication module and the decision tracing and visualization module, respectively. The signal output terminal of the arbitrator triggering and adjudication module is connected to the signal input terminals of the human expert correction module and the decision tracing and visualization module, respectively. The signal output terminal of the human expert correction module is connected to the signal input terminals of the feedback learning optimization module and the decision tracing and visualization module, respectively. The signal output terminal of the feedback learning optimization module is connected to the signal output and input terminals of the multi-agent conditional collaboration module. The task complexity assessment module is used to analyze the task characteristics of environmental data items, identify task types, perform complexity quantification analysis, and output collaboration mode signals. The multi-agent conditional cooperation module is used to realize conditional cooperation between the first and second intelligent analysts. This module dynamically adjusts the cooperation strategy of the two intelligent analysts based on the cooperation mode signal received from the task complexity evaluation module. The arbitrator triggering and adjudication module is used to conditionally trigger the intervention of the intelligent arbitrator agent when the final conclusions of two intelligent analysts in the multi-agent conditional collaboration module differ. The environmental knowledge graph module is used to store professional knowledge in the environmental domain, providing domain knowledge support for intelligent agent decision-making. The human expert correction module is used to correct decisions made by arbitrators in the triggering and adjudication module where the confidence level of the adjudication does not meet the standard through human experts; The feedback learning optimization module is used to receive correction feedback from the human expert correction module and optimize the dynamic weight adjustment algorithm and agent decision-making logic based on the feedback data. The decision tracing and visualization module is used to record the complete decision chain from task configuration and initial review opinions to arbitration reasons, and generate a visual report.
2. The multi-agent traceable decision-making system for environmental data mining according to claim 1, characterized in that, The collaboration mode signal in the task complexity assessment module is used to indicate whether subsequent agents should adopt an independent processing mode or an interactive processing mode.
3. The multi-agent traceable decision-making system for environmental data mining according to claim 2, characterized in that, The dynamic adjustment strategy for the collaboration of the two intelligent analysts in the multi-agent conditional collaboration module is as follows: When the collaboration mode signal indicates independent processing mode, the two intelligent analysts work in parallel and only compare results in the final output stage. When the collaboration mode signal indicates interactive processing mode, the decision system increases the frequency of information exchange between the two intelligent analysts.
4. The multi-agent traceable decision-making system for environmental data mining according to claim 1, characterized in that, When making a ruling, the intelligent arbitrator in the arbitrator triggering and adjudication module will weigh and adopt the arguments of the two intelligent analysts according to the review criteria preset by the task configuration module, and then form a final ruling. If the confidence level given by the intelligent arbitrator is still lower than the threshold, human experts will be invited to make a final decision.
5. A multi-agent traceable decision-making system for environmental data mining according to claim 1, characterized in that, The environmental knowledge graph module stores professional knowledge in the environmental field, including: pollutant attributes, treatment technologies, and policy and standard information.
6. A multi-agent traceable decision-making system for environmental data mining according to claim 1, characterized in that, The dynamic weight adjustment algorithm described in the feedback learning optimization module is as follows: based on the arbitration of the intelligent arbitrator and the corrective feedback from the human expert, the weight coefficients of the two intelligent analysts are dynamically updated. This results in analysts whose historical decision outcomes are more closely aligned with correct results receiving higher weight.
7. A decision-making method for a multi-agent traceable decision-making system for environmental data mining based on any one of claims 1-6, characterized in that, Includes the following steps: Step 1: Task Complexity Assessment and Collaboration Mode Allocation The decision system receives task descriptions and environmental data items, analyzes the task characteristics through the task complexity assessment module, and generates a collaboration mode signal, which is divided into independent processing mode and interactive processing mode. Step Two: Parallel Processing and Dynamic Collaboration of Dual-Intelligence Analysts After the environmental data item is analyzed in step one, it enters the multi-agent conditional cooperation module. If the cooperation mode signal indicates independent processing mode, the first and second intelligent analysts in the multi-agent conditional cooperation module process the data item in parallel and independently without information exchange. If the cooperation mode signal indicates interactive processing mode, the decision system establishes a communication channel for the first and second intelligent analysts, allowing them to exchange intermediate inference results or conduct a limited number of rounds of negotiation. Step 3: Consistency Check and Arbitration Trigger Compare the final conclusions of the two analysts obtained in step two. If the conclusions are consistent, output the result and proceed to step six; if the conclusions differ, trigger the arbitrator to intervene and make a ruling, and execute step four. Step Four: Arbitrator's Award and Weighting Integration Through the arbitrator trigger and award module, the arbitrator intervenes to make the final award. If the confidence level of the award meets the standard, the final decision is output and the process jumps to step six; otherwise, step five is executed. Step 5: Human Expert Correction and Weight Optimization Feedback The human expert correction module corrects decisions that fail to meet the confidence level in step four, and the feedback is recorded by the decision system and used for optimization through the feedback learning optimization module. Step Six: Decision Origin Tracing and Weight Visualization The decision tracing and visualization module generates a tracing report containing complete information on the decision-making process and collaboration model.
8. The decision-making method according to claim 7, characterized in that, The method for generating the collaboration mode signal in step one is as follows: the task complexity assessment module analyzes the task characteristics and makes a judgment based on a preset rule set to generate a collaboration mode signal; the rule set is configured such that: if the task characteristics satisfy at least one of the preset high complexity conditions, an interactive processing collaboration mode signal is generated; otherwise, an independent processing collaboration mode signal is generated.