Multi-agent collaborative decision and regulation system for environmental protection equipment autonomous intelligence

Abstract

The application relates to the technical field of intelligent equipment control and discloses a multi-agent collaborative decision-making and regulation system for the autonomous intelligence of environmental protection equipment, which comprises a routing agent, a knowledge question and answer agent, a data retrieval and generation agent, a water environment analysis agent, an intelligent equipment execution agent and a result processing module. The routing agent performs context completion and intent recognition based on multi-round historical dialogues, and distributes tasks to corresponding agents; the data retrieval and generation agent realizes high-reliability structured query generation and error correction through parallel retrieval of three types of indexes; the water environment analysis agent aggregates multi-source evidence to output traceable research and judgment results; and the intelligent equipment execution agent forms a safety control closed loop through entity uniqueness verification and feedback. The application solves the problems of limited natural language interaction understanding, unreliable query generation, non-uniform statistical caliber and lack of control closed loop, and realizes full-process intelligentization collaboration from perception to execution.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment control technology, specifically to a multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment. Background Technology

[0002] With the continuous improvement of the intelligence and networking of water environment monitoring and treatment equipment, the current mainstream water environment management system generally adopts a layered architecture of "Internet of Things sensing layer + data platform + application layer". By deploying online monitoring instruments for multiple parameters such as chemical oxygen demand (COD), ammonia nitrogen, and total phosphorus, and combining them with a data acquisition and monitoring control (SCADA) system, remote data acquisition and basic control of key environmental protection equipment such as aerators, dosing pumps, and booster pump stations can be achieved.

[0003] However, existing technologies face the following significant technical bottlenecks in the integration of intelligent interaction, data analysis, and decision execution for complex operation and maintenance scenarios: 1. Weak Natural Language Interaction Understanding Capability: Existing systems primarily rely on graphical user interfaces (GUIs) and predefined operation templates, lacking a deep understanding of natural language commands from operations and maintenance personnel, especially in multi-turn dialogues involving referential resolution, omission completion, and complex compound intentions. This makes it difficult for users to perform complex data queries (such as cross-time period multi-indicator comparative analysis) or issue compound control commands using natural and coherent language, resulting in low operational efficiency and slow emergency response.

[0004] 2. Structured query generation is unreliable and insecure: Directly applying Text-to-SQL technology based on the General Large Language Model (LLM) is prone to generating SQL statements with syntax errors, logical flaws, or bypassing access controls when dealing with the time-series data table structure, complex multi-table join logic, and specific aggregate functions unique to the water environment domain. Existing solutions lack deep adaptation and strong constraints to the domain database schema, and also lack a pre-screening mechanism for semantic security and operational risks of the generated SQL.

[0005] 3. Inconsistent Data Analysis and Statistical Standards: Water quality assessment, pollution load calculation, and equipment operating efficiency statistics must strictly adhere to national or industry standards (such as GB3838-2002 "Surface Water Environmental Quality Standard"). Current systems often lack embedded, standardized calculation rule bases, leading to discrepancies in calculation results for the same indicator across different modules and time points. This seriously affects the consistency and authority of regulatory decisions and makes cross-regional and cross-period comparative analysis difficult.

[0006] 4. Insufficient ability to fuse multi-source heterogeneous data and perform causal reasoning: Accurately judging abnormal water environment events (such as water quality exceeding standards) requires comprehensive analysis of multi-source heterogeneous information such as water quality time-series data, equipment operating status, process parameters, alarm records, and even meteorological data. Most existing systems remain at the level of data display and simple alarms, lacking mechanisms for automatic correlation, evidence chain construction, and root cause reasoning. Fault diagnosis heavily relies on expert experience, resulting in low efficiency and difficulty in standardization and traceability.

[0007] 5. Lack of a secure closed-loop system for remote control of intelligent equipment: Remote control commands for environmental protection equipment (such as pumps, valves, and fans) are typically issued through industrial protocols. Existing methods either rely on fixed scripts or lack accurate parsing and semantic verification of natural language control commands. More importantly, there is a lack of uniqueness verification of the controlled object (equipment entity), dynamic verification of operating permissions, confirmation feedback of command execution results (receipts), and auditing of complete operation logs. This poses a safety hazard that could lead to equipment damage or secondary environmental risks due to misoperation.

[0008] In summary, existing technologies for intelligent operation and maintenance of the water environment suffer from core deficiencies, including unintelligent interaction, insecure queries, inconsistent analysis, lack of basis for decision-making, and lack of closed-loop control. Therefore, there is an urgent need to construct a comprehensive, intelligent, collaborative, and secure solution that connects the entire process of "perception-cognition-decision-execution." Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment.

[0010] To achieve the above-mentioned objectives, the present invention provides the following technical solution: A multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment, the system's working method includes the following steps: S1. Session Management and Intent Routing: Receives user natural language input and user identifier, obtains multi-turn historical dialogue information corresponding to the user identifier; based on the multi-turn historical dialogue information, performs context semantic completion and entity normalization processing on the current input, identifies user intent, and outputs the service type. S2. Collaborative execution of business intelligent agents: Based on the business type, tasks are distributed to the corresponding business intelligent agent modules for execution. The business intelligent agent modules include at least one of the following: knowledge question answering intelligent agent, data retrieval and generation intelligent agent, water environment analysis intelligent agent, and intelligent equipment execution intelligent agent. S3. Result Feedback and Status Update: Receive the execution result returned by the business intelligence module, format the execution result and output it to the user, and update the multi-round historical dialogue information.

[0011] As a further description of the technical solution of the present invention, the contextual semantic completion and entity normalization processing of the current input in step S1 specifically includes: extracting and completing the entities of site, project, region, device, time range and indicator based on the multi-turn historical dialogue information; normalizing the synonyms, aliases and indicator descriptions in the entities; the business type includes at least one of knowledge question answering, structured data query, data analysis, prediction, anomaly detection, alarm management, water quality query, water quality level assessment, cause analysis and equipment control instructions.

[0012] As a further description of the technical solution of the present invention, the operations performed by the data retrieval and generation agent in step S2 include: S21. Enhanced Parallel Search: Parallel search of the database description index, data definition index, and question-answer example index to obtain search fragments related to the current input; S22. Context constraint generation: The search segments are reordered based on relevance, and the top K segments are selected to form a constraint context. S23. Structured Query Generation and Review: Based on the generation constraint context, drive the generation model to generate structured query statements, and review the security, syntax, and execution cost of the structured query statements. S24. Execution and Error Correction: Execute the structured query statement that has passed the review; if an error or timeout occurs during execution, generate an error correction prompt based on the error or timeout information to drive the generation model to generate a corrected structured query statement and retry execution until successful or the preset number of retries is reached.

[0013] As a further description of the technical solution of the present invention, the database description index is used to store the business meaning, unit and statistical caliber information of the data table fields; the data definition index is used to store data structure definition information; and the question and answer example index is used to store historical questions and corresponding example structured query statements.

[0014] As a further description of the technical solution of the present invention, when the business type is data analysis, the method further includes statistical calculation following a preset indicator calculation rule; the indicator calculation rule includes at least one of the following: For cumulative indicators, the actual consumption is calculated by the difference between the maximum and minimum values ​​of the corresponding fields of the cumulative indicator within the statistical period. For instantaneous flow rate indicators, when converting them into periodic processing volume, the average flow rate and the preset periodic time coefficient are used for conversion. For year-on-year or month-on-month analysis, calculate the statistical values ​​of the current period and the comparison period, and output the rate of change based on this; For statistics on equipment runtime or operating rate, a dual-judgment logic is used to determine the effective runtime based on the time difference between adjacent time series records and the equipment's online heartbeat records.

[0015] As a further description of the technical solution of the present invention, when the business type is alarm management, water quality level assessment or cause analysis, the operations performed by the water environment analysis intelligent agent in step S2 include: Multi-source evidence acquisition: Based on the aforementioned business type, acquire at least two of the following: alarm records, water quality monitoring time-series data, equipment status data, statistical summary results, and grade assessment results; Structured aggregation and analysis: The acquired multi-source evidence is aggregated in a structured manner to form a structured evidence set; based on the structured evidence set, analysis and analysis results including cause analysis, risk assessment and disposal recommendations are output; When conducting water quality level assessment, the original monitoring data is obtained, and the program calculates and determines the level based on a preset standard threshold table.

[0016] As a further description of the technical solution of the present invention, the operation performed by the intelligent device and the intelligent agent in step S2 includes: Instruction structured parsing: Parsing the natural language input into a structured set of execution parameters, the structured set of execution parameters including at least device identifier, action type, and action parameters; Entity uniqueness verification: Map the device identifier to verify whether it can uniquely identify the target environmental protection equipment; if not, stop the process and output a prompt message. Security verification and command issuance: The physical range and operation permission of the action parameters are verified; after the verification is passed, control commands are issued through the device control interface. Receipt Receipt Reception and Closed-Loop Feedback: Receive the execution receipt information returned by the target environmental protection equipment, generate a structured execution result based on the execution receipt information and feed it back, and write the instructions and receipt information into the audit log.

[0017] As a further description of the technical solution of the present invention, the method further includes a step of verification before performing prediction or anomaly detection, the verification including: Site whitelist verification: Based on historical data quality and model performance, verify whether the target site is allowed to perform prediction or anomaly detection; Parameter range verification: Verify whether the input parameters conform to physical boundaries, historical statistical distribution patterns, and logical relationship constraints among multiple indicators.

[0018] As a further description of the technical solution of the present invention, the prediction is performed based on a fusion prediction model, which adopts parallel modeling of Transformer path and LSTM path, and dynamically adjusts the output weights through attention weighted fusion strategy; the anomaly detection is performed based on an improved isolated forest algorithm and configured with a multi-level anomaly classification system.

[0019] A multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment, the system comprising: The routing agent module is used to execute step S1 to achieve session management, intent recognition, and task routing. The knowledge question answering intelligent agent module is used to execute knowledge questions and answers based on retrieval enhancement when the business type is knowledge question answering; The data retrieval and generation intelligent agent module is used to perform enhanced structured query generation, review, execution, and error correction when the business type is structured data query, data analysis, prediction, or anomaly detection. The intelligent agent module for water environment analysis is used to perform structured analysis and judgment based on multi-source evidence when the business type is alarm management, water quality query, water quality level assessment or cause analysis. The intelligent equipment executes an intelligent agent module, which is used to perform instruction parsing, security verification, closed-loop control and feedback when the business type is equipment control instruction; The result processing and interaction module is used to execute step S3 to achieve formatted output of the results and update the session state.

[0020] The beneficial effects of this invention are: 1) Improve the reliability and security of structured queries: Reduce the error rate of generative SQL in terms of syntax, fields and filtering conditions by parallel retrieval evidence constraints, SQL review and error correction retry mechanism.

[0021] (2) Improve the consistency of statistical standards: By pre-setting the standard rules of indicators and calculating on the program side, reduce the standard deviation in the statistics of cumulative amount, conversion type indicators and running time.

[0022] (3) Improve the traceability of analysis: By gathering structured evidence and outputting in a fixed format, the analysis conclusions can correspond to the data sources.

[0023] (4) Enhance control security: Reduce the risk of misoperation and improve the verifiability of execution through entity uniqueness verification and feedback mechanism.

[0024] (5) Form a collaborative closed loop: realize the closed-loop linkage of monitoring, query, analysis, decision-making and execution, and reduce manual switching and repeated verification. Attached Figure Description

[0025] The present invention will now be further described with reference to the accompanying drawings.

[0026] Figure 1 This is an end-to-end overall flowchart of a multi-agent cooperative system provided in an embodiment of the present invention.

[0027] Figure 2 The flowchart of the knowledge question answering intelligent agent module provided in the embodiment of the present invention.

[0028] Figure 3 The flowchart of the data retrieval and generation intelligent agent module (structured query) provided in the embodiments of the present invention is shown.

[0029] Figure 4 The flowchart of the data analysis and Kanban intelligent agent module (data analysis) provided in the embodiments of the present invention.

[0030] Figure 5 The flowchart illustrates the closed-loop control process of the intelligent equipment execution intelligent agent module provided in this embodiment of the invention.

[0031] Figure 6 The flowchart of the data analysis and dashboard intelligent agent module (operation and maintenance report generation) provided in the embodiment of the present invention.

[0032] Figure 7 The flowchart of the data analysis and Kanban intelligent agent module (prediction and anomaly detection) provided in the embodiments of the present invention is shown.

[0033] Figure 8 The flowchart of the data analysis and dashboard intelligent agent module (performance management) provided in the embodiments of the present invention.

[0034] Figure 9 The flowchart of the data analysis and dashboard intelligent agent module (real-time statistics and threshold discrimination) provided in the embodiment of the present invention.

[0035] Figure 10 The flowchart of the data analysis and dashboard intelligent agent module (device detection) provided in the embodiment of the present invention.

[0036] Figure 11 The flowchart of the water environment analysis intelligent agent module (water quality level assessment) provided in the embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Please see Figure 1 As shown, a multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment includes the following steps in its operation: S1. Session Management and Intent Routing: Receives user natural language input and user identifier, obtains multi-turn historical dialogue information corresponding to the user identifier; based on the multi-turn historical dialogue information, performs context semantic completion and entity normalization processing on the current input, identifies user intent, and outputs the service type. The contextual semantic completion and entity normalization processing of the current input in step S1 specifically includes: extracting and completing the entities of site, project, region, device, time range and indicator based on the multi-turn historical dialogue information; normalizing the synonyms, aliases and indicator descriptions in the entities; the business type includes at least one of knowledge question answering, structured data query, data analysis, prediction, anomaly detection, alarm management, water quality query, water quality level assessment, cause analysis and equipment control instructions.

[0039] Contextual completion: When user input contains omissions (such as "What was the COD of yesterday's influent?") or references (such as "Turn it off"), the system automatically completes the input by combining the entities that are already clearly defined in the previous dialogue (such as "Sewage Treatment Plant A" and "Influent Pump Station No. 1 Pump"). Entity normalization: This involves standardizing the mapping of the completed entities, for example, mapping "COD", "chemical oxygen demand", and "COD" to the standard mapping. cr "Unified mapping to the standard indicator code "I001"; "Pump No. 1", "Pump #1", and "Inlet Pumping Station Booster Pump" are unified into the equipment ID "PUMP-A-01" through the asset database; Intent recognition and routing: Based on the processed input, the routing agent recognizes the user's intent and outputs the determined service type. At the same time, it encapsulates the parsed structured parameters (such as {site: Plant A, indicator: COD of influent, time: yesterday}) and passes them to the downstream link.

[0040] S2. Collaborative execution of business intelligent agents: Based on the business type, tasks are distributed to the corresponding business intelligent agent modules for execution. The business intelligent agent modules include at least one of the following: knowledge question answering intelligent agent, data retrieval and generation intelligent agent, water environment analysis intelligent agent, and intelligent equipment execution intelligent agent. Please see Figure 2 As shown, the operations performed by the knowledge-answering agent in step S2 include: This module targets unstructured knowledge such as procedures and standards for the operation and maintenance of water environment and intelligent equipment, key points for fault handling, equipment maintenance manuals, and domain knowledge items, completing the retrieval and traceable question-and-answer output. Its internal processing flow is as follows: Hybrid retrieval unit: Vectorizes user questions and performs keyword retrieval (based on BM25 algorithm) and semantic retrieval (based on embedding vector similarity) in the knowledge base to obtain a set of candidate knowledge fragments; Re-ranking unit: The candidate knowledge fragments are relevance scored, and the search results are refined and re-ranked using a cross-encoder model. The top K most relevant fragments (K value is preset according to the application scenario, usually 3-5) are selected to form the evidence context. Evidence Judgment Unit: Determines whether the relevance of the evidence context to the user's question reaches a preset threshold (e.g., 0.7). If not, it outputs a prompt that additional information is needed or an explanation that the evidence cannot be matched. Response generation unit: Under the constraints of the evidence context, it drives the large language model to generate natural language responses and marks the sources cited in the responses (such as Chapter 3, Section 2 of the "Technical Manual for Operation and Maintenance of Urban Wastewater Treatment Plants"). Session writing unit: Writes the generated response and its reference information into the session record to ensure knowledge traceability.

[0041] Please see Figure 3 As shown, the operations performed by the data retrieval and generation agent in step S2 include: S21. Enhanced Parallel Search: Parallel search of the database description index, data definition index, and question-answer example index to obtain search fragments related to the current input; Database descriptive index: Stores the business meaning, unit, and statistical caliber information of data table fields. For example, "the field total_power means 'cumulative meter reading', the unit is kWh, and the difference between the maximum and minimum values ​​within the day should be used when calculating daily electricity consumption"; Data definition index: Stores data structure definition information, including table structure, field types, primary and foreign key relationships, index information, and other DDL content; Question and Answer Example Index: Stores historical questions and corresponding example structured query statements, such as "Question: 'Power consumption of Factory A last month', SQL: 'SELECT MAX(total_power)-MIN(total_power)FROM meterWHERE site_id='A'ANDtimeBETWEEN…'.

[0042] S22. Context constraint generation: The search segments are reordered based on relevance, and the top K segments are selected to form a constraint context. S23. Structured Query Generation and Review: Based on the generation constraint context, drive the generation model to generate structured query statements, and review the security, syntax, and execution cost of the structured query statements. Security audit: Sensitive field filtering, prohibiting or restricting deletion operations (DROP, DELETE), prohibiting or restricting structural change operations (ALTER, CREATE); Syntax review: Checks the correctness of SQL syntax, identifying unclosed quotation marks, missing keywords, etc. Execution cost review: Estimate the amount of data involved in the query and the expected execution time, and intercept or issue warnings for high-cost operations such as full table scans without time range restrictions and multi-table Cartesian products.

[0043] S24. Execution and Error Correction: Execute the structured query statement that has passed the review; if an error or timeout occurs during execution, generate an error correction prompt based on the error or timeout information to drive the generation model to generate a corrected structured query statement and retry execution until successful or the preset number of retries is reached.

[0044] The database description index is used to store the business meaning, unit, and statistical caliber information of the data table fields; the data definition index is used to store data structure definition information; and the question-and-answer example index is used to store historical questions and corresponding example structured query statements.

[0045] Please see Figure 4 As shown, when the business type is data analysis, the method further includes statistical calculations following preset indicator calculation rules; the indicator calculation rules include at least one of the following: For cumulative indicators, the actual consumption is calculated by the difference between the maximum and minimum values ​​of the corresponding fields of the cumulative indicator within the statistical period. For example, when calculating daily power consumption, the formula is Daily Power Consumption = MAX(total_power) - MIN(total_power) to avoid calculation errors caused by resetting the meter reading. For instantaneous flow rate indicators, when converting them to periodic processing volume, the average flow rate and the preset periodic time coefficient are used for conversion; for example, when calculating daily processing volume, daily processing volume = average flow rate × 24 is used. For year-on-year or month-on-month analysis, calculate the statistical values ​​of the current period and the comparison period and output the rate of change based on this; the output rate of change is: rate of change = (current value - comparison value) / comparison value × 100%, and the trend is determined according to the preset thresholds (±5% normal, ±15% significant, ±30% abnormal).

[0046] For equipment runtime or uptime statistics, a dual-judgment logic is used to determine effective runtime based on the time difference between adjacent time series records and the equipment's online heartbeat records. Effective runtime is determined when the time difference between adjacent records is less than 30 minutes and the equipment is online; shutdown is determined when the time difference exceeds 60 minutes or there is no heartbeat record.

[0047] Please see Figure 6 As shown, when the business type is operation and maintenance report generation, the report generation unit aggregates information such as water quality monitoring, alarm handling, work orders, and assets according to a unified standard to form auditable operation and maintenance statistics: Report element extraction unit: Extracts elements such as project / region / site and time range; Range verification unit: Determines whether the element can uniquely determine the statistical range. If it cannot be uniquely determined (e.g., there are multiple sites with the same name), it prompts the user to refine the range. Multi-source statistical unit: aggregates and statistically analyzes data from alarms, water quality, work orders, assets, etc. Completeness determination unit: Determines whether key data items are missing. If missing items are found, they are marked or a fault-tolerant output strategy is adopted. Template generation unit: Generates and outputs structured report content based on preset templates (such as daily, weekly, monthly, and annual reports).

[0048] Please see Figure 7 As shown, when the business type is prediction and anomaly detection processing, the prediction and anomaly detection unit is used to predict the trend of water quality indicators or to provide alerts for abnormal discharge risks: Parameter extraction unit: Extracts parameters such as site, indicator, and prediction step size from user input; Whitelist and Range Verification Unit: Determines whether a site is within the allowed range and whether the parameters meet preset thresholds. The whitelist is dynamically maintained based on historical data quality and model performance; sites with high data missing rates and large historical prediction errors will be removed from the whitelist. Data preparation unit: acquires historical monitoring data and determines whether the sample size meets the modeling requirements (e.g., at least 90 days of data are required). If not, it prompts for expanding the time range or performs a downgraded output. Prediction Unit: Outputs prediction results for the next preset number of days based on the fusion prediction model. The fusion prediction model employs a dual-path parallel modeling approach using both Transformer and LSTM paths: the Transformer path is responsible for long-term trend prediction, employing a multi-head self-attention mechanism with 8 attention heads, each with 64 dimensions, for a total of 512 dimensions, combined with sine and cosine positional encoding and learnable timestamp encoding; the LSTM path is responsible for short-term fluctuation modeling, employing an adaptive forgetting mechanism; the two models dynamically adjust the output weights through an attention-weighted fusion strategy.

[0049] Anomaly detection unit: Based on the improved isolated forest algorithm, configured with key parameters such as n_estimators=100, max_samples=0.9, contamination=0.1, and establishes a four-level anomaly classification system: normal (anomaly score ≥ 0), mild anomaly (-0.1≤anomaly score < 0), moderate anomaly (-0.2≤anomaly score < -0.1), and severe anomaly (anomaly score < -0.2).

[0050] Please see Figure 8 As shown, when the business type is performance evaluation and management processing, the performance evaluation and management unit meets the statistical needs for compliance status, abnormal status, and ranking comparison in the water environment governance and operation and maintenance supervision scenario: Object identification unit: Identifies the assessment object, assessment cycle, and assessment criteria; Data validation unit: Determines whether the assessment object and its data exist; if not, returns an explanation and supplementary suggestions. Rule calculation unit: Calculates assessment indicators and achievement results based on preset assessment rules (such as pass rate, number of anomalies, and response time); Output unit: Performs multi-object comparison ranking and outputs results when needed.

[0051] Please see Figure 9 As shown, when the business type is real-time statistics and threshold discrimination processing, the real-time statistics and threshold discrimination unit is used to output real-time statistical results for water quality or equipment operation indicators, and to perform threshold discrimination when necessary: Statistical element analysis unit: analyzes statistical indicators, time windows (such as the most recent 1 hour, the most recent 24 hours) and aggregation methods (mean, maximum value, minimum value); Real-time data acquisition unit: Acquires the latest monitoring data; if it fails, it returns the reason for the failure and suggestions. Aggregation calculation unit: Performs mean, extreme value, total, or other statistical aggregations on the data; Threshold discrimination unit: When the activation conditions are met, the statistical results are compared with thresholds (such as pH value 6.0-9.0, dissolved oxygen 0-15mg / L), and anomaly markers are output; Output Unit: Outputs structured statistical results and writes them to the session log.

[0052] Please see Figure 10 As shown, when the business type is equipment detection processing, the equipment detection unit combines temporal continuity and online evidence to determine the equipment status in the intelligent equipment operation and maintenance scenario, reducing misjudgments caused by missing data. Timing data acquisition unit: Acquires device data timing and necessary online / heartbeat information; Time difference calculation unit: Calculates the time difference between adjacent records and determines whether it is less than or equal to a preset threshold (e.g., 30 minutes). Online evidence verification unit: When the time difference is greater than the preset threshold, it determines whether there is a heartbeat or online record in the corresponding time period to distinguish between data missing and offline / fault status; Status determination unit: Outputs device status determination results (continuous operation, online but data not uploaded, offline / fault); Alarm triggering unit: When conditions are met (such as no heartbeat for 30 consecutive minutes), trigger an equipment abnormality alarm and output handling suggestions.

[0053] Please see Figure 11 As shown, the operations performed by the water environment analysis agent in step S2 include: Multi-source evidence acquisition: Based on the aforementioned business type, acquire at least two of the following: alarm records, water quality monitoring time-series data, equipment status data, statistical summary results, and grade assessment results; Structured aggregation and analysis: The acquired multi-source evidence is aggregated in a structured manner to form a structured evidence set; based on the structured evidence set, analysis and analysis results including cause analysis, risk assessment and disposal recommendations are output; When conducting water quality level assessment, the original monitoring data is obtained, and the program side performs calculations and level determination based on a preset standard threshold table. The core of water quality assessment lies in the principle of "programmatic calculation": Element analysis unit: analysis site, indicator set, and evaluation time range; Raw data retrieval unit: Generates raw monitoring data query requests, restricts the use of aggregate functions in the query statement, and only retrieves raw monitoring records to avoid introducing unnecessary aggregation bias during the data retrieval stage; Data sufficiency assessment unit: Determines whether the number of sampling points, coverage area, etc. meet the assessment requirements (e.g., the calculation of daily average value requires at least 18 hours of valid data). If not, it prompts to expand the time range or supplement the data. Standard threshold rating unit: On the program side, the statistical quantity is compared with the national or industry standard threshold table (such as GB3838-2002 "Surface Water Environmental Quality Standard"), the single factor level and comprehensive level are output, and the main influencing factors are identified; Results output unit: Outputs evaluation conclusions and key evidence, and updates session records.

[0054] Please see Figure 5 As shown, the operations performed by the intelligent device and the intelligent agent in step S2 include: Command structured parsing: The natural language input is parsed into a set of structured execution parameters, which includes at least device identifier, action type (such as turn on, turn off, adjust), action parameters (such as frequency 40Hz, opening degree 50%), duration, cycle, and effective time. Entity uniqueness verification: Map the equipment identifier to verify whether it can uniquely identify the target environmental protection equipment; if not, stop the process and output a prompt message; if multiple devices with the same name are found (e.g., there is one "Pump No. 1" in the inlet pump room and one in the return pump room), stop the process immediately and output a prompt message: "Multiple 'Pump No. 1' have been found. Please specify the process unit (e.g., 'Pump No. 1 in the inlet pump room')". Security verification and command issuance: The action parameters are verified for physical range (e.g., whether the frequency is within the 20-50Hz range) and operation permission (whether the current user has the right to operate this device); after the verification is passed, control commands are issued through the device control interface (e.g., Modbus TCP, OPC UA, MQTT); Receipt Receipt Reception and Closed-Loop Feedback: Receive the execution receipt information returned by the target environmental protection equipment, generate a structured execution result based on the execution receipt information and feed it back, and write the instructions and receipt information into the audit log.

[0055] As a further description of the technical solution of the present invention, the method further includes a step of verification before performing prediction or anomaly detection, the verification including: Site whitelist verification: Based on historical data quality and model performance, verify whether the target site is allowed to perform prediction or anomaly detection; Parameter range verification: Verify whether the input parameters conform to physical boundaries, historical statistical distribution patterns, and logical relationship constraints among multiple indicators.

[0056] As a further description of the technical solution of the present invention, the prediction is performed based on a fusion prediction model, which adopts parallel modeling of Transformer path and LSTM path, and dynamically adjusts the output weights through attention weighted fusion strategy; the anomaly detection is performed based on an improved isolated forest algorithm and configured with a multi-level anomaly classification system.

[0057] S3. Result Feedback and Status Update: Receive the execution result returned by the business intelligence module, format the execution result and output it to the user, and update the multi-round historical dialogue information.

[0058] A multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment, the system comprising: The routing agent module is used to execute step S1 to achieve session management, intent recognition, and task routing. The knowledge question answering intelligent agent module is used to execute knowledge questions and answers based on retrieval enhancement when the business type is knowledge question answering; The data retrieval and generation intelligent agent module is used to perform enhanced structured query generation, review, execution, and error correction when the business type is structured data query, data analysis, prediction, or anomaly detection. The intelligent agent module for water environment analysis is used to perform structured analysis and judgment based on multi-source evidence when the business type is alarm management, water quality query, water quality level assessment or cause analysis. The intelligent equipment executes an intelligent agent module, which is used to perform instruction parsing, security verification, closed-loop control and feedback when the business type is equipment control instruction; The result processing and interaction module is used to execute step S3 to achieve formatted output of the results and update the session state.

[0059] To facilitate understanding of the overall collaborative logic and practical application effects of this invention, a complete demonstration is provided below using a typical operation and maintenance scenario as an example: Scenario: Xiao Wang, an operations and maintenance worker at a water utility group, received an alarm text message stating that "ammonia nitrogen levels in the effluent from Wastewater Treatment Plant A exceeded the standard."

[0060] Interaction process: Xiao Wang typed: "Check the ammonia nitrogen levels in the effluent from Plant A today." The routing agent identified the query as "water quality query," completed the site as "Plant A," the time as "today," and the indicator as "effluent ammonia nitrogen." The data retrieval agent generates SQL queries to obtain hourly ammonia nitrogen data for today's effluent. The report stated: "The average ammonia nitrogen level in Plant A's effluent today was 4.2 mg / L, exceeding the discharge standard (3.0 mg / L). The exceedances occurred primarily between 02:00-06:00 and 14:00-16:00." Xiao Wang pressed further, "Why is it exceeding the standard? Is it related to the equipment?" The routing agent identifies the issue as "cause analysis" and automatically triggers multi-source evidence aggregation. The intelligent water environment analysis system acquires: aeration fan operating status, dissolved oxygen curve, influent water quality data, and sludge concentration data during periods of exceeding standards. Analysis revealed that: between 02:00 and 06:00, the aeration blower operated at a reduced frequency due to a frequency converter malfunction, resulting in dissolved oxygen in the biological treatment tank falling below 1.0 mg / L; between 14:00 and 16:00, the influent COD load suddenly increased. Output: "Possible causes: 1) The exceedance of the standard from 02:00 to 06:00 was due to a malfunction of the No. 1 aerator (log check: 01:55 reported 'inverter overheating'), inhibiting nitrification; 2) The exceedance of the standard from 14:00 to 16:00 was due to an influent load shock (COD suddenly increased from 320mg / L to 580mg / L). Recommended actions: 1) Immediately repair the No. 1 aerator; 2) It is recommended to increase the carbon source dosage during the load shock period." Xiao Wang's decision: "Increase the frequency of fan #2 by 5Hz to compensate for the aeration volume." The routing agent identifies the command as a "device control command". Intelligent equipment execution agent analysis: Equipment identifier "2# Fan", action type "Frequency Adjustment", action parameter "+5Hz" Entity uniqueness verification passed (No. 2 wind turbine is unique). Safety check: The current frequency is 35Hz, which will increase by 5Hz to 40Hz, within the allowable range of 20-50Hz. The instruction was sent to the frequency converter, and the response was: "Instruction received, current frequency 40Hz". Feedback: "Fan #2 has been upgraded from 35Hz to 40Hz and is operating stably." The system automatically updates: The complete interaction information of this round is recorded in the conversation history to provide context for subsequent dialogues.

[0061] As can be seen from the above embodiments, the present invention achieves a fully intelligent closed loop from problem discovery and cause analysis to decision execution through a multi-agent collaborative architecture, which significantly improves the efficiency and safety of environmental protection equipment operation and maintenance.

[0062] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A multi-agent collaborative decision-making and control system for autonomous intelligence in environmental protection equipment, characterized in that: The system operation method includes the following steps: S1. Session Management and Intent Routing: Receive user natural language input and user identifier, and obtain multi-turn historical dialogue information corresponding to the user identifier; Based on the multi-turn historical dialogue information, the current input is processed for contextual semantic completion and entity normalization to identify user intent and output business type. S2. Collaborative execution of business intelligent agents: Based on the business type, tasks are distributed to the corresponding business intelligent agent modules for execution. The business intelligent agent modules include at least one of the following: knowledge question answering intelligent agent, data retrieval and generation intelligent agent, water environment analysis intelligent agent, and intelligent equipment execution intelligent agent. S3. Result Feedback and Status Update: Receive the execution result returned by the business intelligence module, format the execution result and output it to the user, and update the multi-round historical dialogue information.

2. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 1, characterized in that, The contextual semantic completion and entity normalization processing of the current input in step S1 specifically includes: extracting and completing the entities of site, project, region, device, time range and indicator based on the multi-turn historical dialogue information; normalizing the synonyms, aliases and indicator descriptions in the entities; the business type includes at least one of knowledge question answering, structured data query, data analysis, prediction, anomaly detection, alarm management, water quality query, water quality level assessment, cause analysis and equipment control instructions.

3. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 1, characterized in that, The operations performed by the data retrieval and generation agent in step S2 include: S21. Enhanced Parallel Search: Parallel search of the database description index, data definition index, and question-answer example index to obtain search fragments related to the current input; S22. Context constraint generation: The search segments are reordered based on relevance, and the top K segments are selected to form a constraint context. S23. Structured Query Generation and Review: Based on the generation constraint context, drive the generation model to generate structured query statements, and review the security, syntax, and execution cost of the structured query statements. S24. Execution and Error Correction: Execute the structured query statement that has passed the review; if an error or timeout occurs during execution, generate an error correction prompt based on the error or timeout information to drive the generation model to generate a corrected structured query statement and retry execution until successful or the preset number of retries is reached.

4. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 3, characterized in that, The database description index is used to store the business meaning, unit, and statistical caliber information of the data table fields; the data definition index is used to store data structure definition information; and the question-and-answer example index is used to store historical questions and corresponding example structured query statements.

5. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 1, characterized in that, When the business type is data analysis, the method further includes statistical calculations following preset indicator calculation rules; the indicator calculation rules include at least one of the following: For cumulative indicators, the actual consumption is calculated by the difference between the maximum and minimum values ​​of the corresponding fields of the cumulative indicator within the statistical period. For instantaneous flow rate indicators, when converting them into periodic processing volume, the average flow rate and the preset periodic time coefficient are used for conversion. For year-on-year or month-on-month analysis, calculate the statistical values ​​of the current period and the comparison period, and output the rate of change based on this; For statistics on equipment runtime or operating rate, a dual-judgment logic is used to determine the effective runtime based on the time difference between adjacent time series records and the equipment's online heartbeat records.

6. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 1, characterized in that, When the business type is alarm management, water quality level assessment, or cause analysis, the operations performed by the water environment analysis agent in step S2 include: Multi-source evidence acquisition: Based on the aforementioned business type, acquire at least two of the following: alarm records, water quality monitoring time-series data, equipment status data, statistical summary results, and grade assessment results; Structured aggregation and analysis: The acquired multi-source evidence is aggregated in a structured manner to form a structured evidence set; based on the structured evidence set, analysis and analysis results including cause analysis, risk assessment and disposal recommendations are output; When conducting water quality level assessment, the original monitoring data is obtained, and the program calculates and determines the level based on a preset standard threshold table.

7. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 1, characterized in that, The operations performed by the intelligent device and the intelligent agent in step S2 include: Instruction structured parsing: Parsing the natural language input into a structured set of execution parameters, the structured set of execution parameters including at least device identifier, action type, and action parameters; Entity uniqueness verification: Map the device identifier to verify whether it can uniquely identify the target environmental protection equipment; if not, stop the process and output a prompt message. Security verification and command issuance: The physical range and operation permission of the action parameters are verified; after the verification is passed, control commands are issued through the device control interface. Receipt Receipt Reception and Closed-Loop Feedback: Receive the execution receipt information returned by the target environmental protection equipment, generate a structured execution result based on the execution receipt information and feed it back, and write the instructions and receipt information into the audit log.

8. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 1, characterized in that, The method further includes a verification step before performing prediction or anomaly detection, the verification including: Site whitelist verification: Based on historical data quality and model performance, verify whether the target site is allowed to perform prediction or anomaly detection; Parameter range verification: Verify whether the input parameters conform to physical boundaries, historical statistical distribution patterns, and logical relationship constraints among multiple indicators.

9. The multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment according to claim 8, characterized in that, The prediction is performed based on a fusion prediction model, which uses a dual-path parallel modeling of Transformer and LSTM paths and dynamically adjusts the output weights through an attention-weighted fusion strategy; the anomaly detection is performed based on an improved isolated forest algorithm and configured with a multi-level anomaly classification system.

10. A multi-agent collaborative decision-making and control system for autonomous intelligence of environmental protection equipment, characterized in that, The system for implementing the method as described in any one of claims 1 to 9 comprises: The routing agent module is used to execute step S1 to achieve session management, intent recognition, and task routing. The knowledge question answering intelligent agent module is used to execute knowledge questions and answers based on retrieval enhancement when the business type is knowledge question answering; The data retrieval and generation intelligent agent module is used to perform enhanced structured query generation, review, execution, and error correction when the business type is structured data query, data analysis, prediction, or anomaly detection. The intelligent agent module for water environment analysis is used to perform structured analysis and judgment based on multi-source evidence when the business type is alarm management, water quality query, water quality level assessment or cause analysis. The intelligent equipment executes an intelligent agent module, which is used to perform instruction parsing, security verification, closed-loop control and feedback when the business type is equipment control instruction; The result processing and interaction module is used to execute step S3 to achieve formatted output of the results and update the session state.

Images