DPSIR agent-based offshore ecological health assessment method and system

Abstract

The present application relates to the field of marine ecological health evaluation, and discloses a kind of offshore ecological health evaluation method and system based on DPSIR intelligent agent, method includes: input user evaluation task, after identifying evaluation object, evaluation target and scale constraint, carry out spatial scale selection and time scale selection, and combination;Overall intelligent agent receives evaluation task, dispatches five kinds of intelligent agents of DPSIR to carry out each dimension evaluation;And set up big language model preliminary evaluation and multilayer perception machine data-driven supervision correction mechanism in S state intelligent agent, output final S dimension evaluation result;After five kinds of intelligent agents complete each dimension evaluation, according to the adjacent causal relationship between each dimension, in turn execute causal consistency detection, form comprehensive evaluation result.The present application improves the stability and reviewability of S state evaluation result, accurately identifies ecological stress source, state abnormal reason and subsequent influence path.

Description

Technical Field

[0001] This invention belongs to the field of marine ecological health assessment, and specifically relates to a method and system for assessing nearshore ecological health based on DPSIR intelligent agents. Background Technology

[0002] Marine ecological health assessment is a crucial foundation for nearshore marine ecological supervision, risk early warning, and governance decisions. Its ecological health status is influenced by a combination of factors, including terrestrial inputs, industrial development, port and shipping activities, hydrodynamic exchange, nutrient structure, biological responses, and governance measures, exhibiting characteristics of multi-source coupling and significant spatiotemporal differences. It should be noted that the information disclosed in the above background section is solely for enhancing understanding of the background of this invention and therefore may include information that does not constitute prior art known to those skilled in the art.

[0003] Currently, existing technologies related to intelligent assessment of marine ecological health mainly fall into two categories: one is coastal zone or aquatic ecosystem health assessment methods based on the DPSIR model, and the other is multi-agent collaborative processing methods based on large language models. The former provides an indicator organization and comprehensive evaluation framework for ecological health assessment, while the latter provides an intelligent technological foundation for natural language task parsing, task decomposition, multi-agent collaboration, and result aggregation.

[0004] Existing coastal ecosystem health assessment methods based on the DPSIR model typically divide ecosystem health assessment indicators into five dimensions: drivers, stresses, states, impacts, and responses. They then obtain the ecological health assessment results for the target sea area through indicator standardization, weight calculation, comprehensive index calculation, and health level classification. This new method can utilize the DPSIR framework to systematically organize coastal ecosystem health assessment indicators and obtain ecological health assessment results through standardization, weight calculation, comprehensive evaluation, and trend prediction.

[0005] However, the above method still has the following two shortcomings, which will be discussed in detail below:

[0006] The dimensions of DPSIR are often calculated as parallel indicators, lacking a mechanism for detecting consistency between adjacent dimensions based on the ecological causal chain. Although the DPSIR framework theoretically has a causal progression of "driving force—stress—state—impact—response," existing calculation processes typically do not further determine whether the causal relationship between adjacent dimensions holds. Furthermore, existing DPSIR evaluation methods usually do not detect the strength, type, and coverage consistency between adjacent links such as driving force (D) and stress (P), stress (P) and state (S), state (S) and impact (I), and impact (I) and response (R). They also do not trigger review and correction when adjacent dimension results are mismatched, which may lead to inconsistencies in the causal chain of the evaluation results.

[0007] When large language models are used in S-state dimension evaluation, a numerical supervision and correction mechanism based on multilayer perceptrons is lacking. Existing multi-agent collaborative methods based on large language models are mainly geared towards general task processing. Their agent collaboration mechanisms typically focus on task dependencies, execution order, resource scheduling, and output fusion, without establishing dedicated numerical supervision mechanisms for specific state indicators in marine ecological health assessment. This makes it difficult to simultaneously leverage the knowledge reasoning capabilities of large language models and the numerical constraint capabilities of data-driven models, potentially leading to insufficient stability of S-state evaluation results and further affecting subsequent ecological impact identification and response suggestion generation. Summary of the Invention

[0008] To address or at least alleviate one or more of the above problems, a nearshore ecological health assessment method and system based on DPSIR agents is provided. This method retains the knowledge understanding and interpretation generation capabilities of large language models, introduces the data-driven numerical constraint capabilities of multilayer perceptrons, and improves the stability and verifiability of S-state evaluation results.

[0009] To achieve the above objectives, according to a first aspect of the present invention, a method for assessing nearshore ecological health based on a DPSIR agent is provided, comprising the following steps:

[0010] S1. Multi-temporal and spatial scale task analysis:

[0011] The evaluation task input by the user is obtained, the evaluation object, evaluation target and scale constraint contained in the evaluation task are parsed, and spatial scale selection and time scale selection are performed respectively. The selected spatial scale and time scale are combined into a target evaluation unit for the DPSIR agent to call.

[0012] S2, DPSIR agent collaborative evaluation and data-driven supervised correction of state agents:

[0013] The evaluation is carried out collaboratively by DPSIR agents, which include a coordinating agent, a driving force agent, a stress agent, a state agent, an influencing agent, and a response agent. Each agent performs a multi-dimensional evaluation based on the evaluation objective within the spatial range and time span defined by the target evaluation unit, and obtains the evaluation results for each dimension.

[0014] The state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervision and correction channel. The output results of the two channels are judged for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent.

[0015] S3. Causality Consistency Detection:

[0016] According to the adjacent causal relationships between the driving force (D) dimension and the pressure (P) dimension, the pressure (P) dimension and the state (S) dimension, the state (S) dimension and the influence (I) dimension, and the influence (I) dimension and the response (R) dimension, causal consistency checks are performed sequentially. After the consistency check of each causal edge passes, the coordinating agent summarizes the evaluation results of each dimension to form a comprehensive evaluation result.

[0017] Furthermore, the selection of spatial scale and time scale includes:

[0018] In terms of spatial scale, a multi-level spatial scale organization method from coarse to fine is adopted. The multi-level spatial scale organization method includes the overall nearshore zone, typical marine units, regional unit nearshore zones, and functional shorelines. Among them, the overall nearshore zone is used for ecological health assessment of the entire area, typical marine units are used for assessment of key or representative marine areas, regional unit nearshore zones are used for regional differentiated assessment, and functional shorelines are used for refined assessment of different shoreline functional types such as ports, estuaries, industrial towns, aquaculture, and natural ecology.

[0019] In terms of time scale, a multi-level time scale organization method is adopted, from long to short. The multi-level time scale organization method includes annual scale, quarterly scale and monthly scale. Among them, the annual scale is used for long-term trend and annual comprehensive evaluation, the quarterly scale is used for seasonal process and stage risk identification, and the monthly scale is used for short-term dynamics, abnormal state capture and early warning clue extraction.

[0020] Furthermore, within the spatial range and time span defined by the target evaluation unit, each agent performs a multi-dimensional evaluation based on the evaluation target to obtain evaluation results for each dimension, including:

[0021] The driving force agent receives driving force indicators, which include per capita disposable income, gross domestic product, marine catch, and number of tourists; these indicators are used to identify the preceding driving factors that affect the ecological environment from socio-economic activities, resource development, and human activities.

[0022] The pressure agent receives pressure indicators, including river runoff, land-based total phosphorus emissions, land-based ammonia nitrogen emissions, and pesticide and fertilizer application rates; these indicators are used to identify the environmental pressures directly affecting the marine ecosystem caused by land-based pollution inputs, nutrient pressures, and hydrodynamic changes.

[0023] The state agent receives state indicators, including phosphate, chemical oxygen demand, inorganic nitrogen, dissolved oxygen, nitrogen-to-phosphorus ratio, chlorophyll a, salinity, and water temperature, which are used to evaluate the current water environment and ecological status of the target assessment unit.

[0024] The influence agent receives red tide events to identify ecological risks or abnormal events caused by state changes.

[0025] The response agent receives wastewater treatment volume information to evaluate the pollution control response level of the target assessment unit.

[0026] Furthermore, the state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervised correction channel; the outputs of the two channels are judged for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent, including:

[0027] The state agent receives the state indicators of the target evaluation unit, and simultaneously receives the spatial scale, time scale and data quality information corresponding to the current evaluation task, which are respectively input into the large language model initial evaluation channel and the multilayer perceptron data-driven supervision and correction channel.

[0028] The initial evaluation channel of the large language model, based on state indicators, evaluation tasks, time scales, spatial scales, evaluation criteria, historical evaluation examples, and the current task context, performs state semantic understanding, abnormal pattern recognition, and evaluation rule matching on the state indicators and their evaluation contexts. It generates an initial score for the S-dimensional state and determines the initial evaluation state level according to preset level classification rules, outputting the initial evaluation result of the large language model. The initial evaluation result of the large language model includes the initial score for the S-dimensional state, the initial evaluation state level, and an explanation of abnormal states.

[0029] The evaluation criteria include ecological health evaluation-related standards, indicator threshold rules, and grading rules; historical evaluation examples include historical samples that have completed ecological health evaluations and have scores, grades, and anomaly explanations; the current task context includes user-input evaluation objects, evaluation objectives, spatial scale, time scale, and objective assessment unit information.

[0030] The multilayer perceptron data-driven supervised correction channel constructs a state feature vector from the state indicators, time scale, spatial scale, and data quality information of the target evaluation unit, and performs data preprocessing on the state feature vector; the preprocessed state feature vector is input into a pre-trained multilayer perceptron, which outputs a state S-dimensional reference score and determines a reference state level according to a preset level classification rule; wherein, the state S-dimensional reference score and the reference state level are the output results of the multilayer perceptron.

[0031] After the initial evaluation results of the large language model and the output results of the multilayer perceptron are processed through a consistency discrimination operation, the final state S-dimensional evaluation results are fused and corrected.

[0032] Furthermore, after the initial evaluation results of the large language model and the output results of the multilayer perceptron undergo a consistency discrimination operation, they are corrected and fused to output the final state S-dimensional evaluation results, including:

[0033] Consistency determination operation: Compare the score difference between the initial score of state S dimension and the reference score of state S dimension;

[0034] When the score difference does not exceed the preset threshold and the initial evaluation status level is consistent with the reference status level, the status S dimension evaluation result is directly formed.

[0035] When the score difference exceeds the preset threshold, or the initial evaluation status level is inconsistent with the reference status level, the review and correction process includes: reviewing and locating the reasons for the difference between the initial evaluation channel of the large language model and the multilayer perceptron data-driven supervision correction channel, and taking corresponding correction strategies respectively, and completing the correction by numerical correction and abnormal state interpretation regeneration or correction followed by re-prediction and re-comparison or re-checking and re-correction.

[0036] Correction and Fusion: Re-execute the consistency discrimination operation to output the consistency discrimination result. When the consistency discrimination result is consistent, retain the initial score, initial evaluation state level, and abnormal state explanation of the initial evaluation channel of the large language model to form the final state S dimension evaluation result. When the consistency discrimination result is inconsistent, locate the cause of conflict through rule review, sample comparison, and data backtracking, and correct the evaluation result of the corresponding channel according to the cause of conflict determined after review to obtain the final state S dimension evaluation result.

[0037] Furthermore, the reasons for the differences between the initial evaluation channel of the large language model and the multilayer perceptron data-driven supervised correction channel are identified, and corresponding correction strategies are adopted. Correction is achieved through numerical correction and abnormal state explanation, regeneration, or correction followed by re-prediction and re-comparison, or re-checking and re-correction. These strategies include:

[0038] The initial evaluation results of the large language model and the output results of the multilayer perceptron are sequentially reviewed through evaluation criteria and level thresholds, comparison with historical similar evaluation samples, backtracking of state index data, and checking of data preprocessing results, and the output review results are then output.

[0039] Based on the review results, determine the reasons for the differences between the initial evaluation channel of the large language model and the data-driven supervision and correction channel of the multilayer perceptron.

[0040] When the review results indicate that the large language model has errors in calling evaluation rules, misunderstanding of indicator thresholds, or numerical judgment, the reference score for the state S dimension output by the multilayer perceptron and the reference state level corresponding to the state S dimension reference score are used as the basis for numerical correction. After determining the corresponding state level according to the preset level division rules, the large language model combines the state indicators, evaluation criteria, and historical evaluation examples to regenerate the abnormal state explanation; wherein the abnormal state explanation matches the state S dimension reference score and the reference state level corresponding to the state S dimension reference score.

[0041] When the review results show that the status indicators have missing measurements, outliers, inconsistent dimensions, or insufficient data quality, the status indicators and preprocessed status feature vectors are corrected, the multilayer perceptron prediction is re-executed, the multilayer perceptron output results are output, and the scoring differences and status levels of the large language model initial evaluation channel and the multilayer perceptron data-driven supervision and correction channel are compared again.

[0042] The multilayer perceptron prediction includes inputting the preprocessed state feature vector into a pre-trained multilayer perceptron, outputting a state S-dimensional reference score, and determining the reference state level according to a preset level division rule.

[0043] When the review results show that the current multilayer perceptron output is significantly inconsistent with historical similar evaluation examples or evaluation criteria, the state feature vector of the model input, the data preprocessing process, and the model configuration called are re-examined. After the correction is completed, the state S-dimensional reference score is regenerated, and the reference state level is determined according to the preset level division rules.

[0044] Furthermore, based on the adjacent causal relationships between the driving force (D) dimension and the stress (P) dimension, the stress (P) dimension and the state (S) dimension, the state (S) dimension and the influence (I) dimension, and the influence (I) dimension and the response (R) dimension, causal consistency checks are performed sequentially, including:

[0045] The elements to be verified are extracted from the dimensional evaluation results of the preceding agents and the subsequent agents at both ends of the current adjacent causal edge. The elements to be verified at the preceding agent end include the preceding dimensional score, the preceding causal explanation, and the preceding key element set. The elements to be verified at the subsequent agent end include the subsequent dimensional score, the subsequent causal explanation, and the subsequent type set.

[0046] Invoke the causal mapping rules corresponding to the current adjacent causal edge, match the key elements of the preceding agent with the set of possible subsequent agent types, and then perform strength consistency detection, type consistency detection and coverage consistency detection in parallel.

[0047] Among them, intensity consistency detection is used to determine whether the direction of change pointed to by the preceding dimension matches the score or level of the subsequent dimension; type consistency detection is used to determine whether the type output by the subsequent agent belongs to the set of allowed types obtained by the causal mapping rule of the key elements of the preceding agent; coverage consistency detection is used to determine whether the causal explanation of the subsequent agent covers the main key elements identified by the preceding agent and their corresponding subsequent types.

[0048] When the strength consistency test, type consistency test, and coverage consistency test all pass, the current adjacent causal edge is determined to be consistent, and the evaluation results of subsequent agents are allowed to continue to be passed to the consistency test of the next adjacent causal edge.

[0049] If any of the strength consistency test, type consistency test, or coverage consistency test fails, the coordinating agent locates the adjacent causal edge that is in conflict and the agents at both ends of it. After conflict review and result correction, the causal consistency test of the current adjacent causal edge is re-executed, and the result of the test that passes is passed and then passed on.

[0050] Furthermore, the conflict review and result correction process includes:

[0051] Re-examine the evaluation rules and relevant basis, verify the monitoring data with historical evaluation examples, and regenerate or revise the causal explanations of the relevant intelligent agents.

[0052] To achieve the above objectives, according to a second aspect of the present invention, a nearshore ecological health assessment system based on a DPSIR agent is provided, comprising:

[0053] The multi-spatiotemporal scale task parsing module is used for multi-spatiotemporal scale task parsing: acquiring the evaluation task input by the user, parsing the evaluation object, evaluation target and scale constraints contained in the evaluation task, performing spatial scale selection and temporal scale selection respectively, and combining the selected spatial scale and temporal scale into a target evaluation unit for the DPSIR agent to call.

[0054] The DPSIR agent collaborative evaluation and state agent data-driven supervision and correction module is used for DPSIR agent collaborative evaluation and state agent data-driven supervision and correction: the evaluation is performed collaboratively by DPSIR agents, which include a coordinating agent, a driving force agent, a stress agent, a state agent, an influencing agent, and a response agent. Each agent performs a multi-dimensional evaluation according to the evaluation target within the spatial range and time span defined by the target evaluation unit, and obtains the evaluation results of each dimension.

[0055] The state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervision and correction channel. The output results of the two channels are judged for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent.

[0056] The causal consistency detection module is used for causal consistency detection: according to the adjacent causal relationships between the driving force (D dimension) and the pressure (P dimension), the pressure (P dimension) and the state (S dimension), the state (S dimension) and the influence (I dimension), and the influence (I dimension) and the response (R dimension), causal consistency detection is performed sequentially; after the consistency detection of each causal edge passes, the overall planning agent summarizes the evaluation results of each dimension to form a comprehensive evaluation result.

[0057] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0058] This invention improves the causal coherence of the comprehensive evaluation results through the DPSIR agent collaborative evaluation and causal consistency detection mechanism, and enhances the stability and verifiability of the S-state evaluation results through the initial evaluation of the large language model within the S-state agent and the data-driven supervision and correction mechanism of the multilayer perceptron.

[0059] Traditional DPSIR evaluation methods, while incorporating five dimensions—driving force, stress, state, impact, and response—often manifest as a weighted aggregation of parallel indicators in practical calculations. This can lead to inconsistencies in the overall causal chain, where individual results for each dimension (D, P, S, I, and R) are valid. This invention establishes adjacent causal edges between each dimension: driving force (D) and stress (P), stress and state (S), state and impact (I), and impact and response (R). It also performs strength, type, and coverage consistency checks on the outputs of adjacent agents. When causal inconsistencies are detected, the system triggers conflict review and result correction, ensuring that the evaluation results not only have dimensional scoring but also cross-dimensional causal verification. Consequently, the system can more accurately identify sources of ecological stress, causes of abnormal states, and subsequent impact paths, improving the relevance of ecological risk attribution and governance recommendations.

[0060] The large language model is responsible for state semantic understanding and abnormal state interpretation, while the multilayer perceptron learns the nonlinear mapping relationship between state indicators such as phosphate, chemical oxygen demand, inorganic nitrogen, dissolved oxygen, nitrogen-to-phosphorus ratio, chlorophyll a, salinity, and water temperature and the S-dimensional score based on historical monitoring data, and outputs a reference score and a reference state level. When the score difference between two channels exceeds a preset threshold or the state levels are inconsistent, the system triggers a review of the evaluation rules, historical examples, and state data, and corrects the evaluation results of the corresponding channel based on the review conclusions. Thus, the system retains the knowledge understanding and interpretation generation capabilities of the large language model while introducing the data-driven numerical constraint capabilities of the multilayer perceptron.

[0061] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0062] The accompanying drawings, as part of this invention, are provided to further illustrate the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation thereof. Clearly, the drawings described below are merely some embodiments; those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0063] In the attached diagram:

[0064] Figure 1 This is a flowchart illustrating the nearshore ecological health assessment method based on DPSIR agents in this specific embodiment.

[0065] Figure 2 This is a schematic diagram of the multi-temporal and spatial scale organization of marine ecological health assessment in this specific embodiment;

[0066] Figure 3 This is a flowchart illustrating the initial evaluation of the large language model and the data-driven supervision and correction mechanism of the multilayer perceptron in this specific implementation.

[0067] Figure 4 This is a schematic diagram of the causal consistency detection process in this specific embodiment;

[0068] Figure 5 This is a diagram showing the timing consistency verification results in this specific implementation.

[0069] Figure 6 This is a diagram showing the traceable interpretation and decision support output of the ecological health assessment results in this specific implementation method.

[0070] Figure 7 This is a heatmap showing the attribution semantic consistency in this specific implementation.

[0071] Figure 8 This is a diagram showing the pre-identification results of typical abnormal events A and B in this specific implementation method;

[0072] Figure 9 This is a risk warning result diagram for a typical abnormal event C in this specific implementation method;

[0073] Figure 10 This is a diagram showing the consistency results of the ecological health score in this specific implementation embodiment;

[0074] Figure 11 The diagram shows the results of the consistency verification of the ecological health level classification in this specific embodiment; where (a) is the classification consistency verification: distribution comparison, and (b) is the classification consistency verification: confusion matrix (normalized).

[0075] Figure 12 This is a graph showing the score changes during multiple reasoning processes in different regions in this specific implementation method;

[0076] Figure 13 This is a diagram showing the consistency of health level results in the repeated reasoning process described in this specific embodiment.

[0077] Figure 14 This is a system framework diagram of the nearshore ecological health assessment system based on DPSIR agents in this specific embodiment. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0079] In the following description, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0080] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0081] Please see Figure 1 This invention provides a nearshore ecological health assessment method based on DPSIR intelligent agents, comprising the following steps:

[0082] S1. Multi-spatiotemporal scale task analysis:

[0083] The evaluation task input by the user is obtained, the evaluation object, evaluation target and scale constraint contained in the evaluation task are parsed, and spatial scale selection and time scale selection are performed respectively. The selected spatial scale and time scale are combined into a target evaluation unit for the DPSIR agent to call.

[0084] S2, DPSIR agent collaborative evaluation and data-driven supervised correction of state agents:

[0085] The evaluation is carried out collaboratively by DPSIR agents, which include a coordinating agent, a driving force agent, a stress agent, a state agent, an influencing agent, and a response agent. Each agent performs a multi-dimensional evaluation based on the evaluation objective within the spatial range and time span defined by the target evaluation unit, and obtains the evaluation results for each dimension.

[0086] The state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervision and correction channel. The output results of the two channels are judged for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent.

[0087] S3. Causality Consistency Detection:

[0088] According to the adjacent causal relationships between the driving force (D) dimension and the pressure (P) dimension, the pressure (P) dimension and the state (S) dimension, the state (S) dimension and the influence (I) dimension, and the influence (I) dimension and the response (R) dimension, causal consistency checks are performed sequentially. After the consistency check of each causal edge passes, the coordinating agent summarizes the evaluation results of each dimension to form a comprehensive evaluation result.

[0089] Please see Figure 2 In some embodiments, the selection of spatial scale and time scale includes:

[0090] In terms of spatial scale, a multi-level spatial scale organization method from coarse to fine is adopted, which includes the overall nearshore zone, typical marine units, regional unit nearshore zones, and functional shorelines. Among them, the overall nearshore zone is used for ecological health assessment of the entire area, typical marine units are used for assessment of key or representative marine areas, regional unit nearshore zones are used for regional differentiation assessment, and functional shorelines are used for refined assessment of different shoreline functional types such as ports, estuaries, industrial towns, aquaculture, and natural ecology.

[0091] In terms of time scale, a multi-level time scale organization method is adopted, from long to short. The multi-level time scale organization method includes annual scale, quarterly scale and monthly scale. Among them, the annual scale is used for long-term trend and annual comprehensive evaluation, the quarterly scale is used for seasonal process and stage risk identification, and the monthly scale is used for short-term dynamics, abnormal state capture and early warning clue extraction.

[0092] Specifically, the overall nearshore zone is used to represent a large area of ​​comprehensive ecological health assessment objects, and can be divided into different distance zones according to the distance from the shore; typical marine units are used to extract several representative marine units from the overall nearshore zone to support key area comparison and spatial difference identification; regional unit nearshore zones are used to conduct nearshore zone assessments for several regional units to support regional differentiated management and problem identification; functional shorelines are used to subdivide shorelines according to their functions or ecological attributes, such as port shorelines, estuary shorelines, industrial or urban shorelines, aquaculture shorelines, and natural ecological shorelines, thereby supporting more refined pollution source identification, ecological risk identification, and governance strategy formulation.

[0093] Specifically, the annual scale is mainly used for long-term trend assessment, annual comprehensive evaluation, and structural change identification; the quarterly scale is mainly used for seasonal process analysis, phased risk identification, and response measure evaluation; and the monthly scale is mainly used for short-term dynamic analysis, anomaly detection, and early warning clue extraction. By setting the time scale, users can switch between annual, quarterly, and monthly scales according to different task requirements.

[0094] Specifically, regarding the combination of spatial and temporal scales, spatial and temporal scales are not used in isolation. Instead, the selected spatial and temporal scales are combined into target evaluation units based on the user evaluation task. A spatiotemporal combination matrix represents the typical evaluation focus under various spatiotemporal combinations. In actual implementation, the spatial and temporal scales are flexibly selected according to task requirements, data conditions, and evaluation objectives. Among them, the combination of the overall nearshore zone with an annual scale can be used to identify long-term spatial patterns; the combination of the overall nearshore zone with a monthly scale can be used to capture short-term anomalies within the overall area; the combination of typical marine units with an annual scale can be used to analyze the annual change trends of typical units; and the combination of functional shorelines with a monthly scale can be used for rapid identification and early warning of local risks.

[0095] Specifically, through step S1, the present invention unifies the originally scattered evaluation areas, evaluation times, and evaluation targets into a computable, transferable, and agent-callable target evaluation unit, which serves as the evaluation entry point for subsequent DPSIR five-agent causal consistency detection, providing clear task boundaries and scale constraints for subsequent DPSIR five-agent collaborative evaluation.

[0096] In other embodiments, each agent, within the spatial range and time span defined by the target evaluation unit, performs multi-dimensional evaluations based on the evaluation target to obtain evaluation results for each dimension, including:

[0097] The driving force agent receives driving force indicators, including per capita disposable income, gross domestic product, marine catch, and number of tourists; these indicators are used to identify the preceding driving factors that affect the ecological environment from socio-economic activities, resource development, and human activities.

[0098] The stress agent receives stress indicators, including river runoff, land-based total phosphorus emissions, land-based ammonia nitrogen emissions, and pesticide and fertilizer application rates; these indicators are used to identify the environmental stresses directly affecting marine ecosystems from land-based pollution inputs, nutrient stress, and hydrodynamic changes.

[0099] The state agent receives state indicators, including phosphate, chemical oxygen demand, inorganic nitrogen, dissolved oxygen, nitrogen-to-phosphorus ratio, chlorophyll a, salinity, and water temperature, which are used to evaluate the current water environment and ecological status of the target assessment unit.

[0100] The influence of intelligent agents on receiving red tide events is used to identify ecological risks or abnormal events caused by state changes;

[0101] The responding agent receives wastewater treatment volume information to evaluate the pollution control response level of the target assessment unit.

[0102] Please see Figure 3 In some embodiments, the state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervised correction channel; the outputs of the two channels are judged for consistency, and the final state S-dimensional evaluation result is passed to the influencing agent, including:

[0103] The state agent receives the state indicators of the target evaluation unit, and simultaneously receives the spatial scale, time scale and data quality information corresponding to the current evaluation task, which are respectively input into the large language model initial evaluation channel and the multilayer perceptron data-driven supervision and correction channel.

[0104] The initial evaluation channel of the large language model, based on state indicators, evaluation tasks, time scales, spatial scales, evaluation criteria, historical evaluation examples, and the current task context, performs state semantic understanding, abnormal pattern recognition, and evaluation rule matching on the state indicators and their evaluation contexts. It generates an initial score for the S-dimensional state and determines the initial evaluation state level according to preset level classification rules, outputting the initial evaluation result of the large language model. The initial evaluation result of the large language model includes the initial score for the S-dimensional state, the initial evaluation state level, and an explanation of abnormal states.

[0105] The evaluation criteria include ecological health evaluation-related standards, indicator threshold rules, and grading rules; historical evaluation examples include historical samples that have completed ecological health evaluations and have scores, grades, and anomaly explanations; the current task context includes user-input evaluation objects, evaluation objectives, spatial scale, time scale, and objective assessment unit information.

[0106] The multilayer perceptron data-driven supervised correction channel constructs a state feature vector from the state indicators, time scale, spatial scale, and data quality information of the target evaluation unit, and performs data preprocessing on the state feature vector; the preprocessed state feature vector is input into a pre-trained multilayer perceptron, which outputs a state S-dimensional reference score and determines a reference state level according to a preset level classification rule; wherein, the state S-dimensional reference score and the reference state level are the output results of the multilayer perceptron.

[0107] After the initial evaluation results of the large language model and the output results of the multilayer perceptron are processed through a consistency discrimination operation, the final state S-dimensional evaluation results are fused and corrected.

[0108] Specifically, the data quality information is generated by the preprocessing of multi-source monitoring data from the target assessment unit, including missing indicator markers, outlier markers, dimensional consistency check results, time matching degree, spatial matching degree, and data source reliability markers.

[0109] Specifically, the initial evaluation channel for large language models mainly utilizes the understanding capabilities of large language models of domain knowledge, evaluation criteria, and multi-indicator relationships to generate readable and interpretable initial evaluation results for large oracle models.

[0110] Specifically, state indicators, evaluation tasks, time scales, spatial scales, evaluation criteria, historical evaluation examples, and current task context are used to assist the large language model in completing state semantic understanding, evaluation rule matching, anomaly interpretation, and preliminary level determination.

[0111] Specifically, the state feature vector data preprocessing includes missing test labeling, outlier handling, normalization, and data quality weighting.

[0112] Specifically, the multilayer perceptron, or multilayer perceptron expert small model, includes: an input layer, multiple fully connected hidden layers, and an output layer; the input layer is used to receive the preprocessed state feature vector; the fully connected hidden layers learn the combination relationship between different state indicators and the nonlinear mapping relationship between state indicators and state S-dimensional scores through nonlinear activation functions; the output layer is used to generate state S-dimensional reference scores.

[0113] Specifically, the training samples and update basis of the multilayer perceptron are as follows: The multilayer perceptron undergoes supervised training based on training samples, which include historical monitoring data, existing state evaluation samples, level labels corresponding to standard guidelines, and typical state examples. Input features include state indices, spatial scale, temporal scale, and data quality information. Output labels include state S-dimensional scores and reference state levels. During training, the weights and biases of each fully connected layer are iteratively adjusted through backpropagation, so that the predicted scores output by the multilayer perceptron gradually approximate the state S-dimensional scores and reference state levels corresponding to the samples. In some embodiments, the multilayer perceptron may also output model confidence, index contribution, or sensitivity factors to help determine the reliability of the reference scores and identify the main state indices that have a significant impact on the state S-dimensional evaluation results.

[0114] In some embodiments, the initial evaluation results of the large language model and the output results of the multilayer perceptron are corrected and fused after a consistency discrimination operation to output the final state S-dimensional evaluation results, including:

[0115] Compare the score differences between the initial score of state S dimension and the reference score of state S dimension;

[0116] When the score difference does not exceed the preset threshold and the initial evaluation status level is consistent with the reference status level, the status S dimension evaluation result is directly formed.

[0117] When the score difference exceeds the preset threshold, or the initial evaluation status level is inconsistent with the reference status level, the review and correction process includes: reviewing and locating the reasons for the difference between the initial evaluation channel of the large language model and the multilayer perceptron data-driven supervision correction channel, and taking corresponding correction strategies respectively, and completing the correction by numerical correction and abnormal state interpretation regeneration or correction followed by re-prediction and re-comparison or re-checking and re-correction.

[0118] Correction and Fusion: Re-execute the consistency discrimination operation to output the consistency discrimination result. When the consistency discrimination result is consistent, retain the initial score, initial evaluation state level, and abnormal state explanation of the initial evaluation channel of the large language model to form the final state S dimension evaluation result. When the consistency discrimination result is inconsistent, locate the cause of conflict through rule review, sample comparison, and data backtracking, and correct the evaluation result of the corresponding channel according to the cause of conflict determined after review to obtain the final state S dimension evaluation result.

[0119] Specifically, the difference between the initial score and the reference score of state S dimension is expressed as follows:

[0120] ;

[0121] in, This represents the initial score of the state S dimension output by the large language model. This represents the S-dimensional reference score of the multilayer perceptron output. This represents the score difference between the initial score and the reference score for the S-dimensional state.

[0122] Specifically, when the score difference does not exceed the preset threshold and the initial evaluation state level is consistent with the reference state level, the initial score, initial evaluation state level and abnormal state explanation of the state S dimension output by the large language model are retained, and the state S dimension evaluation result is directly formed without triggering the review and correction process.

[0123] Specifically, re-executing the consistency judgment operation means, after completing the review and correction process, comparing the score difference between the corrected initial score of the S dimension and the current reference score of the S dimension, and determining whether the corrected initial score of the S dimension is consistent with the current reference score of the S dimension.

[0124] Specifically, the consistency judgment result includes three parts: First, the comparison result of the score difference between the revised initial score of state S dimension and the reference score of the current state S dimension; Second, the level judgment result of whether the revised initial state level is consistent with the current reference state level; Third, the comprehensive judgment result of "consistent" or "inconsistent" formed based on the score difference comparison result and the level judgment result.

[0125] When the consistency judgment result is consistent, it means that the difference in the corrected score does not exceed the preset threshold, and the corrected initial evaluation status level is consistent with the current reference status level. The system retains the corrected initial evaluation result of the large language model and forms the final status S-dimensional evaluation result.

[0126] When the consistency judgment result is inconsistent, it means that the difference in the corrected score exceeds the preset threshold, or the corrected initial evaluation state level is inconsistent with the current reference state level. If the review result confirms that the current multilayer perceptron output result conforms to the historical data pattern and evaluation criteria, the current state S-dimensional reference score is used as the final numerical score, and the final state level is determined according to the preset level division rules. At the same time, the large language model regenerates the abnormal state explanation that matches the final numerical score and the final state level.

[0127] Specifically, the correction fusion does not involve averaging the scores of the initial evaluation channel of the large language model and the data-driven supervised correction channel of the multilayer perceptron with fixed weights. Instead, it determines the evaluation results of the corresponding channels based on the consistency discrimination results and the verification results, thereby obtaining the final state S-dimensional evaluation results, which include: the final numerical score, the state level, and the explanation of abnormal states. The final state S-dimensional evaluation results are then passed on to the influencing agents for subsequent ecological risk identification.

[0128] In some embodiments, the reasons for the differences between the initial evaluation channel of the localization large language model and the data-driven supervised correction channel of the multilayer perceptron are reviewed, and corresponding correction strategies are adopted respectively. The correction is completed through numerical correction and abnormal state interpretation regeneration or correction followed by re-prediction and re-comparison or re-checking and re-correction, including:

[0129] The initial evaluation results of the large language model and the output results of the multilayer perceptron are successively checked and verified by evaluation criteria and level thresholds, comparison of historical similar evaluation samples, backtracking of state index data, and data preprocessing results.

[0130] Based on the review results, determine the reasons for the differences between the initial evaluation channel of the large language model and the data-driven supervision and correction channel of the multilayer perceptron.

[0131] When the review results indicate that the large language model has errors in calling evaluation rules, misunderstandings of indicator thresholds, or numerical judgment errors, the reference score for the state S dimension output by the multilayer perceptron and the reference state level corresponding to the state S dimension reference score are used as the basis for numerical correction. After determining the corresponding state level according to the preset level division rules, the large language model combines the state indicators, evaluation criteria, and historical evaluation examples to regenerate the abnormal state explanation; wherein the abnormal state explanation matches the reference score for the state S dimension and the reference state level corresponding to the state S dimension reference score.

[0132] When the review results show that the status indicators have missing measurements, outliers, inconsistent dimensions, or insufficient data quality, the status indicators and preprocessed status feature vectors are corrected, the multilayer perceptron prediction is re-executed, the multilayer perceptron output results are output, and the scoring differences and status levels of the large language model initial evaluation channel and the multilayer perceptron data-driven supervision and correction channel are compared again.

[0133] Among them, the multilayer perceptron prediction includes inputting the preprocessed state feature vector into a pre-trained multilayer perceptron, outputting a state S-dimensional reference score, and determining the reference state level according to a preset level classification rule.

[0134] When the verification results show that the current multilayer perceptron output is significantly inconsistent with historical similar evaluation examples or evaluation criteria, the state feature vector of the model input, the data preprocessing process, and the model configuration called are re-examined. After the correction is completed, the state S-dimensional reference score is regenerated, and the reference state level is determined according to the preset level division rules.

[0135] Specifically, historical similar evaluation examples are similar samples retrieved from the "historical evaluation examples".

[0136] Specifically, the scoring differences include the initial score and the reference score for the S-dimensional status, and the status level includes the initial status level and the reference status level.

[0137] Please see Figure 4 In other embodiments, causal consistency checks are performed sequentially according to the adjacent causal relationships between the driving force (D) dimension and the pressure (P) dimension, the pressure (P) dimension and the state (S) dimension, the state (S) dimension and the influence (I) dimension, and the influence (I) dimension and the response (R) dimension. This includes:

[0138] The elements to be verified are extracted from the dimensional evaluation results of the preceding agents and the subsequent agents at both ends of the current adjacent causal edge. The elements to be verified at the preceding agent end include the preceding dimensional score, the preceding causal explanation, and the preceding key element set. The elements to be verified at the subsequent agent end include the subsequent dimensional score, the subsequent causal explanation, and the subsequent type set.

[0139] Invoke the causal mapping rules corresponding to the current adjacent causal edge, match the key elements of the preceding agent with the set of possible subsequent agent types, and then perform strength consistency detection, type consistency detection and coverage consistency detection in parallel.

[0140] Among them, intensity consistency detection is used to determine whether the direction of change pointed to by the preceding dimension matches the score or level of the subsequent dimension; type consistency detection is used to determine whether the type output by the subsequent agent belongs to the set of allowed types obtained by the causal mapping rule of the key elements of the preceding agent; coverage consistency detection is used to determine whether the causal explanation of the subsequent agent covers the main key elements identified by the preceding agent and their corresponding subsequent types.

[0141] When the strength consistency test, type consistency test, and coverage consistency test all pass, the current adjacent causal edge is determined to be consistent, and the evaluation results of subsequent agents are allowed to continue to be passed to the consistency test of the next adjacent causal edge.

[0142] If any of the strength consistency test, type consistency test, or coverage consistency test fails, the coordinating agent locates the adjacent causal edge that is in conflict and the agents at both ends of it. After conflict review and result correction, the causal consistency test of the current adjacent causal edge is re-executed, and the results of the successful test are passed on to the next edge until the causal consistency test of all adjacent causal edges is completed.

[0143] Specifically, taking the driving force D dimension and the stress P dimension as examples: extract the elements to be verified from the evaluation of each dimension of the driving force agent and the stress agent respectively. The elements to be verified at the driving force agent side are the output elements of the driving force agent, including the driving force D dimension score, causal explanation and driving factor set. The elements to be verified at the stress agent side are the output elements of the stress agent, including the stress P dimension score, stress explanation and stress type set.

[0144] The causal mapping rules of the driving force D dimension and the pressure P dimension are invoked to match the driving factors at the driving force agent end with the pressure types that may be caused at the pressure agent end. Then, intensity consistency detection, type consistency detection and coverage consistency detection are performed in parallel.

[0145] Among them, intensity consistency detection is used to determine whether the direction of change pointed to by the preceding dimension matches the score or stress level of the subsequent dimension; type consistency detection is used to determine whether the stress type output by the stress agent belongs to the stress type set obtained by the driving factors of the driving force agent through the causal mapping rule; and coverage consistency detection is used to determine whether the causal explanation of the stress agent covers the main driving factors and corresponding stress types identified by the driving force agent.

[0146] When the strength consistency test, type consistency test, and coverage consistency test all pass, it is determined that the causal edge between the driving force D dimension and the pressure P dimension is consistent, and the pressure result of the pressure agent is allowed to continue to be transmitted to the causal consistency test between the pressure P dimension and the state S dimension. At the same time, the pressure result of the pressure agent is used as a prerequisite constraint for the state agent to identify state anomalies.

[0147] When any of the consistency checks—strength consistency check, type consistency check, and coverage consistency check—fail, the coordinating agent locates the conflicting causal edge and the agents at both ends. After conflict review and result correction, the causal consistency check of the current causal edge is re-executed, and the result that passes the check is passed on to the next DPSIR chain.

[0148] Specifically, the same processing method is used between the pressure P dimension and the state S dimension, between the state S dimension and the influence I dimension, and between the influence I dimension and the response R dimension, as is used between the driving force D dimension and the pressure P dimension.

[0149] Specifically, the set of driving factors includes dominant driving factors such as the intensity of economic activity and the intensity of tourism activity; the set of pressure types includes land-based pollution input, nutrient pressure, and hydrodynamic changes.

[0150] Specifically, in the example of causal consistency testing between the driving force D dimension and the pressure P dimension, the driving force D dimension scored 84 points, which was causally explained as the high intensity of economic and tourism activities, which may increase the pollution input pressure formed by production and living activities along the river. The set of driving factors included the dominant driving factors such as the intensity of economic and tourism activities. The pressure P dimension scored 58 points, which was explained as the high pollution input and nutrient pressure formed by land-based total phosphorus, land-based ammonia nitrogen emissions and pesticide and fertilizer application. Changes in river runoff affected the dilution and transport of pollutants. The set of pressure types included land-based pollution input, nutrient pressure and hydrodynamic changes.

[0151] Specifically, by invoking the causal mapping rules of the D dimension and the P dimension of stress, the driving factors at the driving agent end are matched with the stress types that may be caused at the stress agent end. For example, an increase in the intensity of economic activities can be mapped to land-based pollution input or nutrient stress, and an increase in the intensity of tourism activities can be mapped to domestic pollution input or land-based pollution input; the intensity of marine fishing activities can be used as driving information on the degree of marine resource development and utilization.

[0152] The system performs intensity consistency, type consistency, and coverage consistency checks in parallel. Intensity consistency checks whether the direction of change indicated by the preceding dimension matches the score or stress level of the subsequent dimension. Type consistency checks whether the stress type output by the stress agent belongs to the set of stress types obtained by the driving factors from the driving force side through causal mapping rules. Coverage consistency checks whether the causal explanation from the stress agent side covers the main driving factors and corresponding stress types identified by the driving force side. For example, when the causal explanation from the driving force side points to increased environmental pressure, and the stress agent side's evaluation result shows a high stress level, they are considered consistent in the intensity direction. Type consistency checks whether the stress type output by the stress agent side belongs to the set of permissible stress types obtained by the driving factors from the driving force side through causal mapping rules. Coverage consistency checks whether the causal explanation from the stress agent side covers the main driving factors and corresponding stress types identified by the driving force side.

[0153] In some embodiments, conflict review and result correction processing includes:

[0154] Re-examine the evaluation rules and relevant basis, verify the monitoring data with historical evaluation examples, and regenerate or revise the causal explanations of the relevant intelligent agents.

[0155] Example 2:

[0156] Based on the same inventive concept, this invention also provides a method and system for assessing nearshore ecological health based on DPSIR intelligent agents, including:

[0157] The multi-spatiotemporal scale task parsing module is used for parsing tasks at multiple spatiotemporal scales.

[0158] The evaluation task input by the user is obtained, the evaluation object, evaluation target and scale constraint contained in the evaluation task are parsed, and spatial scale selection and time scale selection are performed respectively. The selected spatial scale and time scale are combined into a target evaluation unit for the DPSIR agent to call.

[0159] The DPSIR agent collaborative evaluation and state agent data-driven supervised correction module is used for DPSIR agent collaborative evaluation and state agent data-driven supervised correction.

[0160] The evaluation is carried out collaboratively by DPSIR agents, which include a coordinating agent, a driving force agent, a stress agent, a state agent, an influencing agent, and a response agent. Each agent performs a multi-dimensional evaluation based on the evaluation objective within the spatial range and time span defined by the target evaluation unit, and obtains the evaluation results for each dimension.

[0161] The state agent receives state indicators and evaluates them in parallel through the large language model initial evaluation channel and the multilayer perceptron data-driven supervision and correction channel. The output results of the two channels are judged for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent.

[0162] The causality consistency detection module is used for causality consistency detection.

[0163] According to the adjacent causal relationships between the driving force (D) dimension and the pressure (P) dimension, the pressure (P) dimension and the state (S) dimension, the state (S) dimension and the influence (I) dimension, and the influence (I) dimension and the response (R) dimension, causal consistency checks are performed sequentially. After the consistency check of each causal edge passes, the coordinating agent summarizes the evaluation results of each dimension to form a comprehensive evaluation result.

[0164] Example 3:

[0165] The model invocation and forward learning process are based on nearshore ecological monitoring data and historical ecological health assessment samples from 2011 to 2020. These data include driving force indicators, stress indicators, state indicators, impact indicators, and response indicators. State indicators include phosphate, chemical oxygen demand, inorganic nitrogen, dissolved oxygen, nitrogen-to-phosphorus ratio, chlorophyll a, salinity, and water temperature. The system organizes the aforementioned historical data, evaluation criteria, historical evaluation examples, and the current evaluation task into model inputs to support DPSIR multi-agent collaborative evaluation and state S-dimensional supervised correction.

[0166] In this embodiment, instead of fine-tuning the parameters of the large language model, a pre-trained large language model is used for task parsing, state semantic understanding, abnormal pattern recognition, causal explanation generation, and evaluation text generation. Specifically, the large language model uses Qwen3-8B to generate evaluation results, causal explanations, and corrected evaluation text for each agent. The system implementation environment includes LangGraph and PyTorch, where LangGraph is used to organize the multi-agent collaborative process, task transfer, and result feedback, and PyTorch is used to support numerical computation and the invocation of multilayer perceptron-related modules.

[0167] Regarding the model usage strategy, this embodiment employs a forward learning strategy. The system first generates initial evaluation method text based on historical samples and evaluation rules, and performs forward inference on training or historical samples. Subsequently, the generated evaluation method text is used for inference on validation or backtesting samples to determine the consistency between the inference results and historical evaluation conclusions, evaluation criteria, and expert experience. When deviations are detected in evaluation rule calls, abnormal interpretations, or numerical judgments, the feedback results are sent back to the previous round of evaluation method text, and the next round of evaluation methods is regenerated. This process does not involve backpropagation updates of the large language model parameters, nor does it involve optimization of the loss function of the large language model.

[0168] In this embodiment, the forward learning process is executed in 20 rounds. Each round includes four steps: evaluation method generation, historical sample reasoning, consistency check, and feedback correction. Through multiple rounds of forward reasoning and feedback correction, the system gradually forms an evaluation method text suitable for nearshore ecological health assessment tasks, which is then used for subsequent DPSIR agent evaluation, S-state supervision correction, and causal consistency detection.

[0169] It should be noted that the model training in this embodiment does not refer to fine-tuning or backpropagation training of a large language model, but rather to the process of forward updating the evaluation method text based on historical data, evaluation criteria, and feedback results. Therefore, this embodiment does not include traditional loss functions, gradient backpropagation processes, and parameter update processes.

[0170] Example 4:

[0171] To verify the effectiveness of this invention in intelligent assessment of marine ecosystem health, this invention conducts experimental verification focusing on multi-scale spatiotemporal task parsing and adaptive construction of target assessment units, DPSIR agent collaborative assessment and data-driven supervised correction of S-state agents, and causal consistency detection. The verification includes: 1) verifying whether multi-scale spatiotemporal selection partitioning can automatically construct suitable target assessment units based on the assessment task, improving the spatial adaptability and temporal response capability of the assessment task; 2) verifying whether DPSIR agent collaborative assessment and causal consistency detection can achieve a complete reasoning process from multi-dimensional ecological element assessment to causal chain explanation; and 3) verifying whether the "large language model initial assessment—expert small model supervised correction" in the S-state agent can improve the stability, reliability, and verifiability of the state assessment results.

[0172] First, the evaluation effect of multi-spatial-temporal scale organization and target evaluation unit is verified. To verify the multi-spatial-temporal scale organization proposed in this invention, experiments were conducted on ecological health samples from different spatial levels and evaluation periods. First, the corresponding spatial and temporal scales were selected according to the evaluation objects, evaluation objectives, and scale constraints in the user task, and the two were combined into a target evaluation unit. Then, the DPSIR agent was invoked to complete the multi-dimensional evaluation.

[0173] In terms of spatial scale, the experimental samples cover different levels, including the overall nearshore zone, typical marine units, regional nearshore units, and functional shorelines; in terms of temporal scale, the experimental samples cover annual, quarterly, and monthly scales. Through the above spatiotemporal combination, assessments can be conducted on long-term trends, phased ecological risks, and short-term abnormal states.

[0174] Figure 5The graph showing the temporal consistency verification results illustrates the consistency verification results of the regional average health score change, i.e., the trend of change over time. The average absolute error between the system's predictions and actual results for multiple periods is 3.91, indicating that the system can accurately trace and characterize the trend of ecological health score changes over time. The predicted curve and the measured curve are consistent in the overall direction of change, reflecting the stage-specific changes in ecological health status at different time points, demonstrating that the system of this invention can adapt to evaluation tasks at different time scales. Experimental results show that this invention can unify the evaluation object, evaluation time, and evaluation objective into a computable and transferable target evaluation unit, and maintain good evaluation adaptability at different spatial and temporal scales. This mechanism provides clear task boundaries and scale constraints for subsequent DPSIR agent collaborative evaluation.

[0175] Second, the effectiveness of DPSIR agent collaborative evaluation and causal consistency detection was verified. To verify the proposed DPSIR agent collaborative evaluation mechanism, a typical ecological health evaluation task was selected for closed-loop testing. After the user inputs the evaluation task, the system first uses a coordinating agent to parse the task, converting the natural language task into a structured evaluation task, which is then distributed to the driving agent, stress agent, state agent, influencing agent, and response agent. The driving agent mainly analyzes driving indicators such as per capita disposable income, GDP, marine catch, and tourist numbers; the stress agent mainly analyzes stress indicators such as river runoff, land-based total phosphorus emissions, land-based ammonia nitrogen emissions, and pesticide and fertilizer application rates; the state agent mainly analyzes state indicators such as phosphate, chemical oxygen demand, inorganic nitrogen, dissolved oxygen, nitrogen-phosphorus ratio, chlorophyll a, salinity, and water temperature; the influencing agent mainly identifies red tide events and their ecological risks; and the response agent mainly evaluates the pollution control response level based on wastewater treatment volume.

[0176] Figure 6 The system demonstrates the traceable interpretation and decision support output of the ecological health assessment results. It first parses the input task, then invokes the DPSIR intelligent agent collaborative evaluation link to independently analyze different dimensions. Finally, the coordinating intelligent agent aggregates the output results of each agent to form a comprehensive health score, health level determination, key risk factor identification, and governance recommendations.

[0177] Experimental results show that the system outputs a comprehensive score of 71 points, classifying the ecological health level as sub-healthy. Specifically, the scores for the D (driving force) dimension are 61 points, the P (stress) dimension is 58 points, the S (state) dimension is 66 points, the I (influence) dimension is 63 points, and the R (response) dimension is 72 points. These results demonstrate that the system can independently analyze the influencing factors at different levels of the ecosystem through different agents and comprehensively assess the ecological health status.

[0178] The system further conducts interpretive analysis along the DPSIR causal chain to form a traceable assessment chain from "outcome" to "interpretation," identifying the relationships between drivers, pressures, states, impacts, and responses. For example, increased economic and tourism activity intensity may increase land-based pollution input and nutrient pressure; changes in river runoff may further affect the dilution and transport of pollutants; pollution input and nutrient pressure may cause changes in state indicators such as nitrogen and phosphorus structure, chlorophyll a, and dissolved oxygen; state anomalies may further increase the risk of red tides; and wastewater treatment volume is used to characterize the level of response to related pollution pressures. Thus, the system can expand a single state score into an interpretable decision support output chain of "health level assessment—key issue identification—governance recommendation generation—automatic policy archiving."

[0179] In terms of causal consistency detection, the system performs consistency checks on adjacent causal relationships between the driving force (D) dimension and the pressure (P) dimension, between the pressure (P) dimension and the state (S) dimension, between the state (S) dimension and the influence (I) dimension, and between the influence (I) dimension and the response (R) dimension. Specifically, the system judges whether the outputs of adjacent agents conform to preset causal relationships from three aspects: intensity consistency, type consistency, and coverage consistency. When all three types of detections meet the preset conditions, the system determines that the adjacent causal edges are consistent and allows the evaluation results to continue to be passed on. When any detection item fails, the system triggers a review mechanism to re-analyze and correct the relevant evaluation rules, monitoring data, historical examples, and agent outputs.

[0180] To verify the stability of the causal explanations generated by the system, typical evaluation samples were selected to conduct repeated inferences, and risk attribution keywords were extracted from each evaluation report. The explanatory stability between different inference results was compared by semantic consistency calculation.

[0181] Figure 7 The heatmap of attribution semantic consistency for typical evaluation samples is presented. The horizontal and vertical axes represent different inference rounds for the same evaluation sample, respectively. Each cell represents the degree of semantic consistency between two inference results; the closer the value is to 1, the more consistent the risk attribution content generated by the two inferences. Experimental results show that the diagonal positions in the graph are all 1.00, indicating that the same inference result is completely consistent with itself; most values ​​on the off-diagonal positions are above 0.80, indicating that the risk attribution results across different inference rounds maintain a high overall consistency, with only a few cells showing relatively low consistency values. The risk attribution results generated by multiple inferences exhibit high semantic consistency, with an average attribution semantic consistency of 88.32%. This result demonstrates that the system of this invention can not only output health scores and ratings but also reliably identify the causes of ecological risks. By using DPSIR causal chain constraints and adjacent causal edge consistency detection, the uncertainty of the large language model's interpretation results can be reduced, improving the credibility of risk attribution.

[0182] To further verify the system's ability to identify ecological anomaly risks along the "pressure-state-impact" causal chain, three sets of historical typical ecological anomaly events were selected for pre-event perception experiments, and these were designated as Typical Anomaly Event A, Typical Anomaly Event B, and Typical Anomaly Event C, respectively. For each event, the system only inputs ecological environment monitoring data from the adjacent evaluation period before the event occurred, without using post-event anomaly conclusions in the evaluation. After the system completes risk identification, subsequent actual ecological anomaly events are used as verification criteria to determine whether the system can extract risk signals from early environmental changes.

[0183] Figure 8 The results of the pre-identification of typical anomalies A and B are presented. Taking the state dimension results of typical anomaly A as an example, the state agent outputs a state score of 63.9, with an evaluation level of "general" and a risk level of "medium risk." The figure further provides key evidence of the anomaly, including a nitrogen-to-phosphorus ratio of 432.7:1, significantly deviating from the reference ratio of 16:1, indicating an extreme imbalance in the nitrogen-phosphorus structure; a chlorophyll a concentration of 5.45 μg / L, exceeding the baseline value by 82%, the highest value in the three-stage evaluation, indicating an early risk of abnormal algal proliferation; a dissolved inorganic nitrogen concentration of 0.6235 mg / L, exceeding the baseline value by 108%, indicating strong nitrogen nutrient pressure; meanwhile, water quality parameters such as COD, DO, and salinity are still within a relatively safe range, indicating that the anomaly risk is mainly driven by changes in nutrient structure and chlorophyll a. The figure also shows an eutrophication index of 0.59, close to the activation threshold, further supporting the system's judgment of potential abnormal algal proliferation risk. For typical anomaly B, the system identified a significant increase in chlorophyll a, accompanied by changes such as rising water temperature, decreasing salinity, increased spatial variation in inorganic nitrogen, and dissolved oxygen approaching the risk range. This indicates that the assessment unit is transitioning from the early environmental accumulation stage to a rapid algal growth stage. These results correspond to the subsequent ecological anomaly, demonstrating that the system can identify potential risks from changes in multiple indicators prior to the event and form an explanatory chain from changes in environmental pressure to state anomalies, and then to ecological impacts.

[0184] Figure 9The system displays the risk warning results for a typical abnormal event, C. The overall system score is 65.3, indicating a medium-to-high risk level. The stress dimension has reached an unhealthy level (score: 52), and the ecological state is at a critical sub-healthy state (score: 78.5). The warning further indicates that the target assessment unit is classified as a high-risk area for red tides. Key criteria include a chlorophyll a concentration of 9.77 μg / L, approximately 2.4 times the baseline value, with strong spatial heterogeneity; a eutrophication index (E) of approximately 4.76, indicating moderate eutrophication; and favorable conditions of low salinity, high temperature, high nitrogen, and medium phosphorus levels. The system also incorporates historical evidence to indicate that this area is a key monitoring zone for eutrophication and red tide warnings, generating recommendations to immediately activate the red tide emergency monitoring plan, increase the frequency of nearshore patrols, and prepare for emergency response.

[0185] The experimental results above demonstrate that the system of this invention can utilize monitoring data prior to ecological anomalies to comprehensively identify synergistic changes in indicators such as nutrient structure, chlorophyll a, water temperature, salinity, and dissolved oxygen, rather than relying solely on whether a single indicator exceeds a threshold. Through result transmission and causal correlation analysis among stress agents, state agents, and influencing agents, the system can organize prior environmental pressures, ecological anomalies, and potential ecological impacts into a continuous risk evolution chain, thereby achieving pre-event perception and explainable early warning of typical ecological anomalies.

[0186] Third, verification of the state agent supervision and correction effect. To verify the effect of the state agent supervision and correction mechanism proposed in this invention on improving the accuracy and stability of ecological state assessment, experiments were conducted on scoring consistency, health level consistency, and repeated inference stability.

[0187] The S-dimensional state directly reflects the current ecological environment status of the target assessment unit, and its evaluation results further influence the risk identification and governance suggestion generation of the response agent. Therefore, this invention sets up a large language model initial evaluation channel and a multilayer perceptron data-driven expert small model supervision and correction channel within the state agent. The large language model initial evaluation channel is responsible for state semantic understanding, abnormal pattern identification, and preliminary scoring; the expert small model learns the numerical mapping relationship between state indicators and S-dimensional state scores and state levels based on historical monitoring data and existing evaluation samples, and provides independent numerical supervision for the large language model's initial evaluation results.

[0188] To verify the system's predictive accuracy for ecological health scores, a consistency verification experiment was conducted. The mean absolute error (MAE) was used as the evaluation index for the consistency verification, and its calculation formula is as follows:

[0189] ;

[0190] in, This represents the ecological health score of the i-th sample predicted by the system. The value represents the actual ecological health score of the corresponding sample, and N represents the total number of samples. The smaller the mean absolute error, the smaller the average deviation between the system's predicted score and the actual score.

[0191] The comprehensive ecological health score is calculated using the following formula:

[0192] ;

[0193] in, This represents the normalized weight of the i-th evaluation indicator. represents the normalized score or state value of the i-th evaluation indicator, and n represents the total number of evaluation indicators.

[0194] The experiment selected 270 samples with calculable health level labels, including 252 healthy samples and 18 sub-healthy samples. Through scoring consistency verification, the accuracy of the system in continuous scoring output can be evaluated, providing a quantitative basis for subsequent health level determination and risk attribution.

[0195] Figure 10 The results demonstrate the consistency verification of the ecological health score. Experimental results show that the average absolute error between the system's predicted score and the actual score is 5.18, indicating that the method of this invention can accurately reflect the quantitative level of ecosystem health status. From the score distribution results, the overall trend of the system's predicted score distribution is consistent with that of the actual evaluation score distribution, indicating that after initial evaluation using a large language model and supervised correction using an expert small model, the system can reduce the scoring bias that may arise from solely relying on the large language model for inference.

[0196] To verify the system's ability to identify ecological health levels, a further hierarchical consistency verification experiment was conducted. Hierarchical consistency was evaluated using accuracy as the metric, expressed as:

[0197] ;

[0198] in, This represents the ecological health level of the i-th sample predicted by the system. The table represents the true ecological health level of the corresponding sample. This is an indicator function; it takes a value of 1 when the predicted level matches the actual level, and a value of 0 otherwise. Accuracy represents the proportion of samples where the predicted ecological health level matches the actual level.

[0199] The data sources for the grading consistency verification are consistent with those for the scoring consistency verification. Based on the grading rules in GB / T 42631—2023 "Guidelines for Assessment of Nearshore Marine Ecological Health", the health status of marine ecosystems is divided into three levels: unhealthy, sub-healthy, and healthy. When the comprehensive ecological health index is in the range [0, 43), it is judged as unhealthy; in the range [43, 80), it is judged as sub-healthy; and in the range [80, 100], it is judged as healthy.

[0200] Figure 11 The results of the consistency verification of the ecological health level classification are presented. The consistency indicators include: Accuracy, which represents the accuracy of level prediction; F1, which represents the classification evaluation index that comprehensively considers precision and recall; Kappa, which represents the consistency coefficient between the predicted level and the actual level; and TP, TN, FP, and FN, which represent true positives, true negatives, false positives, and false negatives, respectively. The sample size used was 270.

[0201] in, Figure 11 (a) For consistency verification of grade levels: distribution comparison plot, used to compare the difference between the actual sample grade distribution and the system predicted grade distribution. The results show that in the actual sample, grade 0 accounts for 93.3% and grade 1 accounts for 6.7%, while in the system prediction results, grade 0 accounts for 90.7% and grade 1 accounts for 9.3%. The overall distribution of the two grades is relatively close, indicating that the system predicted grade distribution has good consistency with the actual grade distribution.

[0202] Figure 11 (b) Validation of hierarchical consistency: A normalized confusion matrix is ​​used to show the proportion of samples with different true levels that are predicted by the system to be each level. The results show that 96.0% of the true level 0 samples were correctly predicted as level 0, and 4.0% were predicted as level 1; 83.3% of the true level 1 samples were correctly predicted as level 1, and 16.7% were predicted as level 0. Combined with the results of Accuracy=0.952, F1=0.698, Kappa=0.672, and TP=15, TN=242, FP=10, FN=3, it is shown that there is good consistency between the predicted level results and the true level results.

[0203] In conclusion, Figure 11 The results show that the state supervision and correction mechanism proposed in this invention can not only improve the accuracy of continuous scoring results, but also improve the reliability of health level judgment.

[0204] To verify the stability of the system's repeated inference under the same input conditions, 10 representative typical samples were selected from the historical evaluation samples and uniformly numbered as Sample 1 to Sample 10. While keeping the input data, evaluation task, evaluation rules, model parameters, and calling conditions consistent, 10 repeated inferences were performed on each sample group, and the ecological health scores and health levels obtained from each inference were compared.

[0205] Figure 12 The results demonstrate the score changes in regions 1 through 10 across multiple inference rounds. Experimental results show that while the ecological health scores output by the system fluctuate somewhat between different inference rounds, the overall fluctuation range is small. The score changes in each region consistently hover around a relatively stable central value, without any obvious abnormal jumps or extreme deviations. The score fluctuation range for most regions is between 5 and 10 points, indicating that in evaluation processes involving large language model inference and multi-agent collaborative analysis, the state agent supervision and correction mechanism can effectively constrain continuous scoring results.

[0206] Figure 13 The results demonstrate the consistency of health levels across samples 1 through 10 during repeated inference. Experimental results show that the health levels obtained by each group of samples remained consistent after 10 repeated inferences, achieving a 100% consistency rate. This result indicates that the system of this invention not only maintains minimal fluctuations at the continuous scoring level but also maintains stable output at the health level determination level, avoiding changes in evaluation levels due to differences in single inference iterations.

[0207] The experimental results above demonstrate that this invention, by combining the state semantic understanding capability of a large language model with the numerical supervision capability of a multilayer perceptron, maintains good stability in continuous scoring and health level determination. The multilayer perceptron does not replace the large language model but provides an independent data-driven reference for the initial scores and levels obtained from the large language model. When there are differences between the outputs of the two channels, the system verifies the results using evaluation rules, historical examples, and state data, and generates a corrected state S-dimensional evaluation result.

[0208] In summary, this invention constructs a complete intelligent assessment process for marine ecological health through a multi-spatial-temporal scale organization mechanism, a DPSIR agent collaborative evaluation and causal consistency detection mechanism, and a state agent supervised correction mechanism. Multi-spatial-temporal scale experiments verified the system's adaptability to assessment tasks at different spatial levels and temporal scales; DPSIR agent and causal consistency experiments verified the system's ability to conduct cross-dimensional correlation analysis and causal explanation; S-state supervised correction experiments verified the system's stability in continuous scoring, health level determination, and repeated inference; and typical anomaly event experiments further verified the system's ability to conduct pre-risk perception based on changes in previous state indicators.

[0209] Unless otherwise defined, the technical or scientific terms used in this invention shall have the general meaning understood by one of ordinary skill in the art to which this invention pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this invention do not indicate quantitative limitations and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this invention are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “A plurality” used in this invention refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. In this invention, the terms "first," "second," "third," etc., are used only to distinguish similar objects and do not represent a specific order of objects.

[0210] Obviously, the accompanying drawings are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, certain design, manufacturing, or production modifications made according to the technical content disclosed in the present invention are merely conventional technical means and should not be considered as insufficient disclosure of the present invention.

[0211] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A nearshore ecological health assessment method based on DPSIR intelligent agents, characterized in that, Includes the following steps: S1. Multi-temporal and spatial scale task analysis: The evaluation task input by the user is obtained, the evaluation object, evaluation target and scale constraint contained in the evaluation task are parsed, and spatial scale selection and time scale selection are performed respectively. The selected spatial scale and time scale are combined into a target evaluation unit for the DPSIR agent to call. S2, DPSIR agent collaborative evaluation and data-driven supervised correction of state agents: The evaluation is carried out collaboratively by DPSIR agents, which include a coordinating agent, a driving force agent, a stress agent, a state agent, an influencing agent, and a response agent. Each agent performs a multi-dimensional evaluation based on the evaluation objective within the spatial range and time span defined by the target evaluation unit, and obtains the evaluation results for each dimension. The state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervision and correction channel. The output results of the two channels are judged for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent. S3. Causality Consistency Detection: According to the adjacent causal relationships between the driving force (D) dimension and the pressure (P) dimension, the pressure (P) dimension and the state (S) dimension, the state (S) dimension and the influence (I) dimension, and the influence (I) dimension and the response (R) dimension, causal consistency checks are performed sequentially. After the consistency check of each causal edge passes, the coordinating agent summarizes the evaluation results of each dimension to form a comprehensive evaluation result.

2. The method according to claim 1, characterized in that, The selection of spatial scale and time scale includes: In terms of spatial scale, a multi-level spatial scale organization method from coarse to fine is adopted. The multi-level spatial scale organization method includes the overall nearshore zone, typical marine units, regional unit nearshore zones, and functional shorelines. Among them, the overall nearshore zone is used for ecological health assessment of the entire area, typical marine units are used for assessment of key or representative marine areas, regional unit nearshore zones are used for regional differentiated assessment, and functional shorelines are used for refined assessment of different shoreline functional types such as ports, estuaries, industrial towns, aquaculture, and natural ecology. In terms of time scale, a multi-level time scale organization method is adopted, from long to short. The multi-level time scale organization method includes annual scale, quarterly scale and monthly scale. Among them, the annual scale is used for long-term trend and annual comprehensive evaluation, the quarterly scale is used for seasonal process and stage risk identification, and the monthly scale is used for short-term dynamics, abnormal state capture and early warning clue extraction.

3. The method according to claim 2, characterized in that, Each agent, within the spatial range and temporal span defined by the target evaluation unit, performs a multi-dimensional evaluation based on the evaluation target to obtain evaluation results for each dimension, including: The driving force agent receives driving force indicators, which include per capita disposable income, gross domestic product, marine catch, and number of tourists; these indicators are used to identify the preceding driving factors that affect the ecological environment from socio-economic activities, resource development, and human activities. The pressure agent receives pressure indicators, including river runoff, land-based total phosphorus emissions, land-based ammonia nitrogen emissions, and pesticide and fertilizer application rates; these indicators are used to identify the environmental pressures directly affecting the marine ecosystem caused by land-based pollution inputs, nutrient pressures, and hydrodynamic changes. The state agent receives state indicators, including phosphate, chemical oxygen demand, inorganic nitrogen, dissolved oxygen, nitrogen-to-phosphorus ratio, chlorophyll a, salinity, and water temperature, which are used to evaluate the current water environment and ecological status of the target assessment unit. The influence agent receives red tide events to identify ecological risks or abnormal events caused by state changes. The response agent receives wastewater treatment volume information to evaluate the pollution control response level of the target assessment unit.

4. The method according to claim 1, characterized in that, The state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervised correction channel. The outputs of the two channels are then compared for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent, including: The state agent receives the state indicators of the target evaluation unit, and simultaneously receives the spatial scale, time scale and data quality information corresponding to the current evaluation task, which are respectively input into the large language model initial evaluation channel and the multilayer perceptron data-driven supervision and correction channel. The initial evaluation channel of the large language model, based on state indicators, evaluation tasks, time scales, spatial scales, evaluation criteria, historical evaluation examples, and the current task context, performs state semantic understanding, abnormal pattern recognition, and evaluation rule matching on the state indicators and their evaluation contexts. It generates an initial score for the S-dimensional state and determines the initial evaluation state level according to preset level classification rules, outputting the initial evaluation result of the large language model. The initial evaluation result of the large language model includes the initial score for the S-dimensional state, the initial evaluation state level, and an explanation of abnormal states. The evaluation criteria include ecological health evaluation-related standards, indicator threshold rules, and grading rules; historical evaluation examples include historical samples that have completed ecological health evaluations and have scores, grades, and anomaly explanations; the current task context includes user-input evaluation objects, evaluation objectives, spatial scale, time scale, and objective assessment unit information. The multilayer perceptron data-driven supervised correction channel constructs a state feature vector from the state indicators, time scale, spatial scale, and data quality information of the target evaluation unit, and performs data preprocessing on the state feature vector; the preprocessed state feature vector is input into a pre-trained multilayer perceptron, which outputs a state S-dimensional reference score and determines a reference state level according to a preset level classification rule; wherein, the state S-dimensional reference score and the reference state level are the output results of the multilayer perceptron. After the initial evaluation results of the large language model and the output results of the multilayer perceptron are processed through a consistency discrimination operation, the final state S-dimensional evaluation results are fused and corrected.

5. The method according to claim 4, characterized in that, The initial evaluation results of the large language model and the output results of the multilayer perceptron are then calibrated and fused after a consistency discrimination operation to output the final state S-dimensional evaluation results, including: Consistency determination operation: Compare the score difference between the initial score of state S dimension and the reference score of state S dimension; When the score difference does not exceed the preset threshold and the initial evaluation status level is consistent with the reference status level, the status S dimension evaluation result is directly formed. When the score difference exceeds the preset threshold, or the initial evaluation status level is inconsistent with the reference status level, the review and correction process includes: reviewing and locating the reasons for the difference between the initial evaluation channel of the large language model and the multilayer perceptron data-driven supervision correction channel, and taking corresponding correction strategies respectively, and completing the correction by numerical correction and abnormal state interpretation regeneration or correction followed by re-prediction and re-comparison or re-checking and re-correction. Correction and Fusion: Re-execute the consistency discrimination operation to output the consistency discrimination result. When the consistency discrimination result is consistent, retain the initial score, initial evaluation state level, and abnormal state explanation of the initial evaluation channel of the large language model to form the final state S dimension evaluation result. When the consistency discrimination result is inconsistent, locate the cause of conflict through rule review, sample comparison, and data backtracking, and correct the evaluation result of the corresponding channel according to the cause of conflict determined after review to obtain the final state S dimension evaluation result.

6. The method according to claim 5, characterized in that, The reasons for the discrepancies between the initial evaluation channel of the large language model and the multilayer perceptron data-driven supervised correction channel are explained, and corresponding correction strategies are adopted. Correction is achieved through numerical correction and abnormal state interpretation, regeneration, or re-prediction and re-comparison, or re-checking and re-correction. These strategies include: The initial evaluation results of the large language model and the output results of the multilayer perceptron are sequentially reviewed through evaluation criteria and level thresholds, comparison with historical similar evaluation samples, backtracking of state index data, and checking of data preprocessing results, and the output review results are then output. Based on the review results, determine the reasons for the differences between the initial evaluation channel of the large language model and the data-driven supervision and correction channel of the multilayer perceptron. When the review results indicate that the large language model has errors in calling evaluation rules, misunderstanding of indicator thresholds, or numerical judgment, the reference score for the state S dimension output by the multilayer perceptron and the reference state level corresponding to the state S dimension reference score are used as the basis for numerical correction. After determining the corresponding state level according to the preset level division rules, the large language model combines the state indicators, evaluation criteria, and historical evaluation examples to regenerate the abnormal state explanation; wherein the abnormal state explanation matches the state S dimension reference score and the reference state level corresponding to the state S dimension reference score. When the review results show that the status indicators have missing measurements, outliers, inconsistent dimensions, or insufficient data quality, the status indicators and preprocessed status feature vectors are corrected, the multilayer perceptron prediction is re-executed, the multilayer perceptron output results are output, and the scoring differences and status levels of the large language model initial evaluation channel and the multilayer perceptron data-driven supervision and correction channel are compared again. The multilayer perceptron prediction includes inputting the preprocessed state feature vector into a pre-trained multilayer perceptron, outputting a state S-dimensional reference score, and determining the reference state level according to a preset level division rule. When the review results show that the current multilayer perceptron output is significantly inconsistent with historical similar evaluation examples or evaluation criteria, the state feature vector of the model input, the data preprocessing process, and the model configuration called are re-examined. After the correction is completed, the state S-dimensional reference score is regenerated, and the reference state level is determined according to the preset level division rules.

7. The method according to claim 1, characterized in that, Based on the adjacent causal relationships between the driving force (D) dimension and the pressure (P) dimension, the pressure (P) dimension and the state (S) dimension, the state (S) dimension and the influence (I) dimension, and the influence (I) dimension and the response (R) dimension, causal consistency checks are performed sequentially, including: The elements to be verified are extracted from the dimensional evaluation results of the preceding agents and the subsequent agents at both ends of the current adjacent causal edge. The elements to be verified at the preceding agent end include the preceding dimensional score, the preceding causal explanation, and the preceding key element set. The elements to be verified at the subsequent agent end include the subsequent dimensional score, the subsequent causal explanation, and the subsequent type set. Invoke the causal mapping rules corresponding to the current adjacent causal edge, match the key elements of the preceding agent with the set of possible subsequent agent types, and then perform strength consistency detection, type consistency detection and coverage consistency detection in parallel. Among them, intensity consistency detection is used to determine whether the direction of change pointed to by the preceding dimension matches the score or level of the subsequent dimension; type consistency detection is used to determine whether the type output by the subsequent agent belongs to the set of allowed types obtained by the causal mapping rule of the key elements of the preceding agent; coverage consistency detection is used to determine whether the causal explanation of the subsequent agent covers the main key elements identified by the preceding agent and their corresponding subsequent types. When the strength consistency test, type consistency test, and coverage consistency test all pass, the current adjacent causal edge is determined to be consistent, and the evaluation results of subsequent agents are allowed to continue to be passed to the consistency test of the next adjacent causal edge. If any of the strength consistency test, type consistency test, or coverage consistency test fails, the coordinating agent locates the adjacent causal edge that is in conflict and the agents at both ends of it. After conflict review and result correction, the causal consistency test of the current adjacent causal edge is re-executed, and the result of the test that passes is passed and then passed on.

8. The method according to claim 7, characterized in that, The conflict review and result correction process includes: Re-examine the evaluation rules and relevant basis, verify the monitoring data with historical evaluation examples, and regenerate or revise the causal explanations of the relevant intelligent agents.

9. A nearshore ecological health assessment system based on DPSIR intelligent agents, characterized in that, include: The multi-spatiotemporal scale task parsing module is used for multi-spatiotemporal scale task parsing: acquiring the evaluation task input by the user, parsing the evaluation object, evaluation target and scale constraints contained in the evaluation task, performing spatial scale selection and temporal scale selection respectively, and combining the selected spatial scale and temporal scale into a target evaluation unit for the DPSIR agent to call. The DPSIR agent collaborative evaluation and state agent data-driven supervision and correction module is used for DPSIR agent collaborative evaluation and state agent data-driven supervision and correction: the evaluation is performed collaboratively by DPSIR agents, which include a coordinating agent, a driving force agent, a stress agent, a state agent, an influencing agent, and a response agent. Each agent performs a multi-dimensional evaluation according to the evaluation target within the spatial range and time span defined by the target evaluation unit, and obtains the evaluation results of each dimension. The state agent receives state indicators and evaluates them in parallel through a large language model initial evaluation channel and a multilayer perceptron data-driven supervision and correction channel. The output results of the two channels are judged for consistency, and the final state S-dimensional evaluation result is transmitted to the influencing agent. The causal consistency detection module is used for causal consistency detection: according to the adjacent causal relationships between the driving force (D dimension) and the pressure (P dimension), the pressure (P dimension) and the state (S dimension), the state (S dimension) and the influence (I dimension), and the influence (I dimension) and the response (R dimension), causal consistency detection is performed sequentially; after the consistency detection of each causal edge passes, the overall planning agent summarizes the evaluation results of each dimension to form a comprehensive evaluation result.

Images