Intelligent Agent collaborative decision-making system for environment monitoring

Abstract

The invention relates to the technical field of water quality monitoring and analysis, in particular to an intelligent Agent collaborative decision-making system oriented to environment monitoring. Comprising the steps that original data including water quality parameters and environment data are collected in real time; performing noise elimination, missing value filling and standardization processing on the collected original data, and outputting preprocessed data; the method comprises the following steps: integrating multi-source data including historical monitoring data, pollution source information and treatment cases, performing vectorization coding on the multi-source data by adopting a BERT model, and constructing a retrieval architecture based on hierarchical indexes and inverted indexes of a hierarchical navigable small world to obtain a dynamic water quality knowledge base; and on the basis of a retrieval enhancement generation technology, associating the preprocessed data with the dynamic water quality knowledge base, predicting a water quality change trend through a time sequence prediction model, and generating an abnormality diagnosis report and a closed-loop treatment scheme. According to the invention, analysis and early warning of water quality are realized through collaborative decision-making of the intelligent Agent.

Description

Technical Field

[0001] The present invention relates to the technical field of water quality monitoring and analysis, and in particular to an intelligent agent collaborative decision-making system for environmental monitoring. Background Art

[0002] With the rapid development of industry and cities, problems such as industrial wastewater discharge, agricultural non-point source pollution, and sudden environmental accidents have become increasingly prominent, leading to frequent water pollution incidents. Traditional environmental water quality monitoring methods rely primarily on manual sampling and laboratory analysis. Although this method offers high detection accuracy, it is inherently limited by its poor timeliness, often requiring hours or even days to obtain analysis results, making it difficult to meet the needs of real-time early warning and emergency response.

[0003] With the development of Internet of Things technology, sensor networks have been widely used in the field of water quality monitoring. They can obtain key water quality parameters such as pH, dissolved oxygen, turbidity, and conductivity in real time, improving the real-time nature of water quality monitoring.

[0004] However, most existing water quality monitoring systems still have significant shortcomings. They are often limited to processing a single data source and lack the ability to deeply integrate multi-source heterogeneous data. It is difficult to effectively integrate key auxiliary information such as meteorological data, geographic information, historical governance cases, and environmental regulations; resulting in inefficient data integration.

[0005] At the same time, in water quality forecasting and decision support, traditional prediction models, such as long-short-term memory networks and autoregressive integrated moving average models, have significant shortcomings in handling complex nonlinear relationships, capturing long-term dependencies, and conducting multivariate coupling analysis. This results in limited prediction accuracy, and is particularly prone to high false alarm rates during sudden pollution events.

[0006] Therefore, an intelligent agent collaborative decision-making system for environmental monitoring is proposed. Summary of the Invention

[0007] The purpose of the present invention is to provide an intelligent agent collaborative decision-making system for environmental monitoring, which can realize water quality analysis and early warning through intelligent agent collaborative decision-making.

[0008] To achieve the above object, the present invention provides the following technical solutions: An intelligent agent collaborative decision-making system for environmental monitoring, including: Data acquisition and preprocessing module: collects raw data in real time, including water quality parameters and environmental data; performs noise elimination, missing value filling and standardization on the collected raw data, and outputs preprocessed data; Dynamic Water Quality Knowledge Base Module: This module integrates multi-source data, including historical monitoring data, pollution source information, and governance case studies. It uses the BERT model to vectorize and encode the multi-source data and constructs a retrieval architecture based on a hierarchical navigable small-world hierarchical index and an inverted index to generate a dynamic water quality knowledge base. Agent analysis module: Based on retrieval enhancement generation technology, it associates pre-processed data with a dynamic water quality knowledge base, predicts water quality trends through a time series prediction model, and generates abnormal diagnosis reports and closed-loop treatment plans; Interaction and execution module: Provides early warning through abnormal diagnosis reports and issues equipment control instructions according to the closed-loop management plan.

[0009] The raw data collection includes placing sensors at key nodes in the target water area to collect data, and the collected data is transmitted to the edge server in real time via the Internet of Things protocol; The water quality parameters include pH value, dissolved oxygen, turbidity, conductivity, heavy metal concentration, organic matter content, and algae biomass; the environmental data include water temperature, flow rate, and meteorological data; The noise elimination includes using a sliding window method to preprocess and eliminate random noise, outliers and interference signals in the original data; using a moving average filtering method to replace the current data point with the average value of the data in the sliding window to weaken high-frequency noise.

[0010] The construction process of the dynamic water quality knowledge base includes: Integrate multi-source data including historical monitoring data, pollution source information, and governance cases; the historical monitoring data includes raw data collected during the historical period; the pollution source information includes industrial and agricultural emission sources; the governance cases include the time, plan, and governance effects of historical water quality governance; Perform data cleaning and format unification on multi-source data; For structured data, interpolation is used to fill missing values; For unstructured data, key information is extracted through text parsing and OCR technology, supplemented by manual annotation to enhance semantic labels; then, the text information in the unstructured data is mapped into high-dimensional vectors through a pre-trained language model.

[0011] The retrieval architecture of the dynamic water quality knowledge base module includes: A hierarchical navigable small-world algorithm is used to optimize high-dimensional vector retrieval, combined with an inverted index to accelerate tag filtering, forming a hybrid architecture of Elasticsearch and Faiss, and associating original text and vector data through unique IDs.

[0012] The dynamic water quality knowledge base includes a dynamic update mechanism and a performance optimization mechanism; The dynamic update mechanism synchronizes real-time sensor data incrementally and optimizes storage using a hot-cold separation strategy: hot data is stored in memory, and cold data is archived to distributed storage; The performance optimization mechanism rebuilds the Faiss index periodically to eliminate data distribution offsets and uses Redis to cache high-frequency search results, reducing response time to milliseconds.

[0013] The Agent analysis module includes: calculating the mean and standard deviation based on historical data; setting dynamic thresholds based on the mean and standard deviation, judging the pre-processed data, and identifying abnormal data; When abnormal data is detected, the Agent triggers RAG enhanced analysis; Abnormal data is vectorized using a pre-trained model, and then an approximate nearest neighbor search is performed in the dynamic water quality knowledge base using the Faiss index to recall similar historical cases. The search results are sorted by cosine similarity and filtered using an inverted index to narrow the search scope. The search results of similar historical cases are combined with real-time data to form prompt words to generate diagnostic reports and preliminary governance recommendations.

[0014] The time series prediction model is based on the Transformer time series model, which inputs past period data and predicts water quality changes in future periods; the model injects time series information through position coding; The self-attention mechanism of the time series prediction model captures long-term dependencies by calculating the attention weights of the query matrix, key matrix, and value matrix; The multi-head attention is expanded into multiple subspaces to enhance the expressive power, and each head calculates attention independently.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This method collects historical monitoring data, pollution source information, and remediation case studies to generate multi-source data. This multi-source data is cleaned and formatted to avoid format incompatibilities and uneven data quality across different data sources, ensuring the accuracy and integrity of the data in the knowledge base. Furthermore, a pre-trained language model maps text information into high-dimensional vectors, enabling semantic understanding of the text in the knowledge base and enhancing its intelligence and utility.

[0016] 2. This invention builds a hybrid search architecture, leveraging Elasticsearch (for inverted indexing and full-text search) and Faiss (for vector similarity search). Unique IDs are used to link raw text data with corresponding vector data. The knowledge base data is also partitioned and stored based on factors such as region and pollution type, further improving search efficiency.

[0017] 3. This invention incrementally synchronizes sensor data and optimizes storage through a hot-cold separation strategy, storing frequently accessed hot data in memory and archiving less frequently accessed cold data to distributed storage. This significantly improves the efficiency of storage resource utilization and significantly increases access speed to hot data, ensuring the system's rapid response to real-time queries and decision-making. Furthermore, this performance optimization mechanism effectively eliminates the potential for retrieval performance degradation caused by data distribution shifts, ensuring stable retrieval efficiency over the long term. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the data collection and preprocessing process of the present invention; Figure 2 A schematic diagram of the process of constructing the dynamic water quality knowledge base of the present invention; Figure 3 Schematic diagram of the intelligent agent workflow of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] Example 1: This application proposes an intelligent agent collaborative decision-making system for environmental monitoring, including: Data acquisition and preprocessing module: collects raw data in real time, including water quality parameters and environmental data; performs noise elimination, missing value filling and standardization on the collected raw data, and outputs preprocessed data; Dynamic Water Quality Knowledge Base Module: This module integrates multi-source data, including historical monitoring data, pollution source information, and governance case studies. It uses the BERT model to vectorize and encode the multi-source data and constructs a retrieval architecture based on a hierarchical navigable small-world hierarchical index and an inverted index to generate a dynamic water quality knowledge base. Agent analysis module: Based on retrieval enhancement generation technology, it associates pre-processed data with a dynamic water quality knowledge base, predicts water quality trends through a time series prediction model, and generates abnormal diagnosis reports and closed-loop treatment plans; Interaction and execution module: Provides early warning through abnormal diagnosis reports and issues equipment control instructions according to the closed-loop management plan.

[0021] The raw data collection includes placing sensors at key nodes in the target water area to collect data, and the collected data is transmitted to the edge server in real time via the Internet of Things protocol; The water quality parameters include pH value, dissolved oxygen, turbidity, conductivity, heavy metal concentration, organic matter content, and algae biomass; the environmental data include water temperature, flow rate, and meteorological data; The noise elimination includes using a sliding window method to preprocess and eliminate random noise, outliers and interference signals in the original data; using a moving average filtering method to replace the current data point with the average value of the data in the sliding window to weaken high-frequency noise.

[0022] The process of data collection and preprocessing is as follows Figure 1 shown.

[0023] Specifically, multi-type sensor arrays are deployed at key points in the target waters, such as river inlets, outlets, and near key sewage outlets, as well as in the central areas of lakes and reservoirs or at specific monitoring points. Sensors can be installed flexibly and diversely, using either buoy or fixed installations to adapt to different water environments. Data collection frequency can be dynamically adjusted based on actual needs, from once per minute to once per hour, to balance real-time data performance and system resource consumption.

[0024] Collected raw data is transmitted in real time to a cloud server or edge computing node via IoT protocols. All transmitted data is standardized in JSON or CSV format. Each data point includes information such as the timestamp, unique sensor identifier, parameter type, specific value, and unit.

[0025] In order to ensure data reliability and avoid being affected by random noise, instantaneous outliers or interference signals, the sliding window mean filtering method is used to eliminate noise; this method calculates the average value of all data points in the sliding window to replace the current data point, effectively weakening high-frequency noise and making the data curve smoother.

[0026] The sliding window method uses a sliding window with a length of N (for example, 3-7 sampling points) and replaces the current point by calculating the mean of the data in the window, thereby effectively weakening high-frequency noise.

[0027] The system uses linear interpolation to fill in missing values caused by data transmission or sensor failure. If data is missing at a particular point in time, the system uses linear extrapolation based on the values of the previous and next valid data points to estimate the value at the missing point. If data is missing for a long period of time, the system marks it as "unavailable" and automatically triggers a sensor self-test or maintenance alert.

[0028] The present invention collects data from target water areas through Internet of Things sensors to obtain water quality parameters; performs data cleaning and noise elimination on the water quality parameters to obtain preprocessed data; and effectively provides support for subsequent data identification.

[0029] After noise removal and missing value filling, preprocessed data is stored in the Faiss vector database. As an efficient similarity search library, Faiss supports the rapid storage and retrieval of high-dimensional vectors, providing efficient support for similar case retrieval in the subsequent intelligent agent analysis module. After normalization, structured data is also converted to vector form for storage.

[0030] The construction process of the dynamic water quality knowledge base is as follows: Figure 2 Shown, including: Integrate multi-source data including historical monitoring data, pollution source information, and governance cases; the historical monitoring data includes raw data collected during the historical period; the pollution source information includes industrial and agricultural emission sources; the governance cases include the time, plan, and governance effects of historical water quality governance; Perform data cleaning and format unification on multi-source data; For structured data, interpolation is used to fill missing values; For unstructured data, key information is extracted through text parsing and OCR technology, supplemented by manual annotation to enhance semantic labels; then, the text information in the unstructured data is mapped into high-dimensional vectors through a pre-trained language model.

[0031] Historical monitoring data includes raw historical data collected over time; unstructured text such as manual inspection records can also be used. Pollution source information includes the geographic location, emission type, primary pollutants, and historical emission data for industrial and agricultural sources. Remediation cases include water pollution remediation cases, including the type of pollution incident, time of occurrence, remediation plan, adopted technology, and remediation effectiveness evaluation.

[0032] Furthermore, environmental regulations and standards, as well as meteorological and geographic information, can be incorporated into the dynamic water quality knowledge base to enhance its data richness and ensure effective water quality analysis and prediction. These environmental regulations and standards include water quality standards, pollutant emission standards, and environmental laws and regulations, providing a legal basis and compliance verification for system decision-making. Meteorological and geographic information includes historical meteorological data, geographic information on rivers and lakes (watershed divisions, hydrological characteristics, topography), and ecological and environmental data (biodiversity, vegetation coverage).

[0033] The above data are obtained through channels such as scientific research institution databases, Internet of Things monitoring equipment, historical archives, and manual expert entry, forming a multi-dimensional and three-dimensional knowledge system.

[0034] Preferably, since the multi-source data contains data in various formats such as numerical data and text, it is necessary to perform data cleaning and format unification on the data.

[0035] Structured data processing: For structured data such as historical monitoring data, missing values are first filled (also using interpolation), and outliers are detected and eliminated to ensure the accuracy of the values.

[0036] Unstructured data processing: For unstructured data such as laboratory reports, inspection records, and historical case descriptions, text parsing technology and optical character recognition technology are used to extract key information; for example, fields such as "pollution type," "pollutant name," and "treatment measures" are extracted from pollution accident reports. The extracted information is supplemented by manual annotation to enhance semantic tags, improve the accuracy of information extraction, and improve the recall rate of subsequent retrieval.

[0037] For longer unstructured texts, the dynamic window method is used for cutting and semantic segmentation; a fixed-size window (such as 512 characters) is set and slides at a 20% overlap rate to divide the long text into semantically complete text blocks; multi-dimensional labels are also added to each text block, such as "pollution type = heavy metal", "region = Yellow River Basin", "event type = sudden pollution".

[0038] To enable efficient retrieval of text and multimodal data, pre-trained language models (such as BERT) are used to map processed text information into a high-dimensional vector space. The BERT model captures the deep semantics of text, mapping semantically similar text to closer locations in the vector space. For multimodal data that mixes numerical pollutant concentrations with textual descriptions, a joint embedding model is used to encode them into a unified vector space representation, ensuring comparability and relevance across data from different modalities.

[0039] This paper adopts the BERT-base-chinese model as the pre-trained language model and fine-tunes it based on professional corpus in the field of water quality monitoring to enhance the model's ability to understand semantic information in the water quality field.

[0040] This method collects historical monitoring data, pollution source information, and remediation case studies to generate multi-source data. This multi-source data is cleaned and formatted to avoid format incompatibilities and uneven data quality across different data sources, ensuring the accuracy and integrity of the data in the knowledge base. Furthermore, a pre-trained language model maps text information into high-dimensional vectors, enabling semantic understanding of the text in the knowledge base and enhancing its intelligence and utility.

[0041] The retrieval architecture of the dynamic water quality knowledge base module includes: A hierarchical navigable small-world algorithm is used to optimize high-dimensional vector retrieval, combined with an inverted index to accelerate tag filtering, forming a hybrid architecture of Elasticsearch and Faiss, and associating original text and vector data through unique IDs.

[0042] Hierarchical navigable small-world hierarchical indexing is implemented based on the hierarchical navigable small-world algorithm.

[0043] To achieve millisecond-level fast retrieval, the knowledge base adopts a hybrid architecture that combines the hierarchical navigable small world (HNSW) algorithm with an inverted index.

[0044] The hierarchical navigable small-world algorithm achieves the efficiency of approximate nearest neighbor (ANN) search for high-dimensional vector data by constructing a multi-layered graph structure. During search, the algorithm starts at the top level of the graph and gradually moves down to more refined levels, quickly finding the vectors most similar to the query vector, ensuring millisecond-level search latency.

[0045] The inverted index is used to accelerate filtering based on text tags or keywords. For example, users can quickly search for all cases where "ammonia nitrogen exceeds the standard" and the "region" is "Yellow River Basin".

[0046] When data is stored, the original text content is stored in Elasticsearch and an inverted index is created. Simultaneously, its corresponding BERT vector is stored in Faiss. The two are linked via a globally unique ID. During retrieval, for queries containing keywords and semantic similarity requirements, a preliminary screening is performed using Elasticsearch's inverted index to obtain the unique IDs of a batch of candidate documents. These IDs are then used to retrieve the corresponding vectors from Faiss. These vectors are then combined with the query vectors for a precise ANN search and cosine similarity sorting, ultimately returning the most relevant text content.

[0047] This paper builds a hybrid search architecture, leveraging Elasticsearch (for inverted indexing and full-text search) and Faiss (for vector similarity search). Unique IDs are used to link raw text data with corresponding vector data. The knowledge base is also partitioned and stored based on factors such as region and pollution type, further improving search efficiency.

[0048] The dynamic water quality knowledge base includes a dynamic update mechanism and a performance optimization mechanism; The dynamic update mechanism synchronizes real-time sensor data incrementally and optimizes storage using a hot-cold separation strategy: hot data is stored in memory, and cold data is archived to distributed storage; The performance optimization mechanism rebuilds the Faiss index periodically to eliminate data distribution offsets and uses Redis to cache high-frequency search results, reducing response time to milliseconds.

[0049] Sensor data is regularly updated to the knowledge base through incremental synchronization. To optimize storage and retrieval performance, a hot-cold separation strategy is adopted: frequently accessed "hot data" (such as the latest monitoring data and recent emergency plans) is stored in memory or cache, while infrequently accessed "cold data" (such as historical archives and outdated regulations) is archived to a distributed storage system.

[0050] The knowledge base also features a verification module that automatically detects data conflicts, triggers an audit process, and ensures data quality. For regulatory texts, the system adds a timeliness marker to indicate their expiration date, preventing the use of expired information. To maintain efficient retrieval performance, the system regularly optimizes performance, such as rebuilding the Faiss index monthly to eliminate potential retrieval performance degradation caused by data distribution shifts. Furthermore, high-frequency retrieval results are cached using in-memory databases such as Redis, reducing response times to milliseconds and ensuring that intelligent agents can quickly obtain the required information during decision-making.

[0051] This invention incrementally synchronizes sensor data and optimizes storage through a hot-cold separation strategy, storing frequently accessed hot data in memory and archiving less frequently accessed cold data to distributed storage. This significantly improves the efficiency of storage resource utilization and significantly increases access speed to hot data, ensuring the system's rapid response to real-time queries and decision-making. Furthermore, the performance optimization mechanism effectively eliminates the potential for retrieval performance degradation caused by data distribution offsets, ensuring the stability of retrieval efficiency over the long term.

[0052] The Agent analysis module includes: setting dynamic thresholds based on historical data, judging pre-processed data, and identifying abnormal data; dynamic threshold setting is based on the calculation of the mean of historical data. and standard deviation , if the real-time data point x satisfies , it is marked as a potential anomaly.

[0053] When abnormal data is detected, the Agent triggers RAG enhanced analysis; Abnormal data is vectorized using a pre-trained model, and then an approximate nearest neighbor search is performed in the dynamic water quality knowledge base using the Faiss index to recall similar historical cases. The search results are sorted by cosine similarity and filtered using an inverted index to narrow the search scope. The search results of similar historical cases are combined with real-time data to form prompt words to generate diagnostic reports and preliminary governance recommendations.

[0054] The workflow of intelligent agent is as follows Figure 3 As shown, the pre-processed data is judged based on preset or dynamically set thresholds to identify potential abnormalities. The dynamic threshold is set based on the mean and standard deviation of historical data.

[0055] When the system detects abnormal water quality, the intelligent agent will immediately trigger the retrieval enhancement generation (RAG) analysis process.

[0056] Abnormal water quality parameters or related text descriptions are vectorized using a pre-trained language model. Using the Faiss index, an approximate nearest neighbor search (ANN) is performed within the dynamic water quality knowledge base to quickly retrieve a specified number of historical cases that are most similar to the current anomaly. The results are sorted based on cosine similarity to determine their degree of similarity to the current event. Furthermore, an inverted index is used for label filtering. For example, if the current anomaly is "acid pollution," only historical cases with the label "acid pollution" are retrieved, further narrowing the search scope and improving search accuracy.

[0057] The large model generates a diagnostic report: This information is combined with historical recall cases, relevant regulations, real-time monitoring data, and preliminary anomaly information into a structured prompt word, which is then fed into the Large Language Model (LLM). Based on this information, the large model automatically generates a detailed diagnostic report. The report includes a detailed description of the anomaly, possible sources of contamination, potential impacts on the ecological environment, and preliminary remediation recommendations based on historical experience.

[0058] The present invention identifies abnormal data based on dynamic thresholds, vectorizes abnormal data and uses Faiss index to perform approximate nearest neighbor search to quickly recall the most similar historical cases; and combines inverted index with label filtering to greatly improve the accuracy and efficiency of diagnosis; then the retrieval results of similar historical cases and real-time data are spliced into prompt words, and the large model generates a diagnostic report and preliminary governance recommendations, which improves the intelligence level and reliability of decision-making.

[0059] The time series prediction model is based on the Transformer time series model, which inputs past period data and predicts water quality changes in future periods; the model injects time series information through position coding; The self-attention mechanism of the time series prediction model captures long-term dependencies by calculating the attention weights of the query matrix, key matrix, and value matrix; The multi-head attention is expanded into multiple subspaces to enhance the expressive power, and each head calculates attention independently.

[0060] While diagnosing, the intelligent agent also activates the water quality trend prediction function, using a time series model based on the Transformer architecture to capture complex time dependencies and nonlinear characteristics.

[0061] The time series prediction model adopts a Transformer Encoder-based architecture, consisting of four encoder layers, each with eight attention heads and a hidden layer dimension of 512. The position encoding uses standard sine-cosine position encoding. The model is trained on historical monitoring data, and the training dataset covers past water quality parameters and environmental data. The loss function uses mean square error (MSE), the optimizer is AdamW, and the learning rate uses cosine annealing scheduling. The training is conducted for a total of 100 epochs.

[0062] The time series forecasting model inputs preprocessed data from the past period (for example, the past 72 hours); through the position encoding mechanism, the position information of each data point in the time series is encoded into the input vector, so that the model can understand the temporal relationship and order of the data.

[0063] The time series forecasting model includes a self-attention mechanism, which allows the model to consider the importance of all other data points in the sequence when processing a data point in the sequence and assign different weights to them. This enables the model to capture long-term dependencies. For example, the current drop in dissolved oxygen may be related to the rainfall a few days ago.

[0064] Furthermore, to enhance the model's expressiveness and robustness, the system employs a multi-head attention mechanism. This involves running the self-attention mechanism multiple times in parallel, learning different attention weights in different subspaces. Each "head" independently computes attention, and ultimately the outputs of multiple heads are concatenated and linearly transformed, capturing complex patterns in the data from different perspectives.

[0065] The time series forecasting model predicts the water quality trend over a period of time (e.g., the next 48 hours), generating a forecast curve and predicted values for key indicators. For example, it can predict the trend of dissolved oxygen content over the next 48 hours.

[0066] Continuously monitor the accuracy of the time series forecasting model results; if the prediction error exceeds the preset threshold, the incremental learning mechanism will be triggered to update the model parameters using the latest real-time data, so that the model can dynamically adapt to environmental changes and improve the accuracy and robustness of the prediction.

[0067] When updating a time series forecasting model, we proactively filter real-time data with high information content for the update. For example, we select data points where the model's prediction uncertainty is high and the difference between the actual observed value and the predicted value is significant, rather than simply updating the model based on all new data. This prevents the model from wasting computing resources on large amounts of redundant or low-value data, improving learning efficiency.

[0068] This paper adopts a time prediction model based on the Transformer architecture to predict water quality change trends. Through the self-attention mechanism, the model learns complex patterns and features in the data from multiple perspectives in parallel, further improving the robustness and accuracy of the prediction model, and providing a reliable basis for future trends in water quality warnings and the formulation of treatment plans.

[0069] After completing anomaly diagnosis and trend prediction, the intelligent agent will comprehensively consider various factors and generate a detailed, executable closed-loop governance plan; this includes historical success cases retrieved from the knowledge base, current real-time environmental constraints (such as wind speed and water flow rate that may affect pollutant diffusion), and the latest environmental regulations. For various possible governance plans, the system will comprehensively score them based on multiple dimensions such as cost-effectiveness, timeliness, expected success rate, and secondary impact on the ecological environment, and recommend the optimal solution.

[0070] The remediation solution also includes control commands that can be directly issued to IoT devices. For example, in response to a sudden drop in dissolved oxygen, the agent might generate the command: "Start the aerator, adjust the power to 80%, and monitor the dissolved oxygen content once an hour for 24 hours." These commands are linked to IoT devices via APIs to achieve automated device control. This fully automated process, from water quality monitoring, anomaly diagnosis, trend prediction, to remediation solution generation and execution, improves the efficiency and accuracy of responding to sudden water quality issues.

[0071] Example 2: The present invention proposes an intelligent agent collaborative decision-making system for environmental monitoring, comprising: Data acquisition and preprocessing module: collects raw data in real time, including water quality parameters and environmental data; performs noise elimination, missing value filling and standardization on the collected raw data, and outputs preprocessed data; Dynamic Water Quality Knowledge Base Module: This module integrates multi-source data, including historical monitoring data, pollution source information, and governance case studies. It uses the BERT model to vectorize and encode the multi-source data and constructs a retrieval architecture based on a hierarchical navigable small-world hierarchical index and an inverted index to generate a dynamic water quality knowledge base. Agent analysis module: Based on retrieval enhancement generation technology, it associates pre-processed data with a dynamic water quality knowledge base, predicts water quality trends through a time series prediction model, and generates abnormal diagnosis reports and closed-loop treatment plans; Interaction and execution module: Provides early warning through abnormal diagnosis reports and issues equipment control instructions according to the closed-loop management plan.

[0072] Preferably, during the data collection and preprocessing stage, data is collected in real time at a frequency of once per minute by sensors deployed in the target waters; the data is transmitted to the cloud or edge server via the NB-IoT protocol.

[0073] The preprocessing module performs sliding window mean filtering on the received data to eliminate instantaneous noise; at the same time, if there is data missing, it is filled by linear interpolation, and the processed standardized DO data is stored in the Faiss vector database.

[0074] Based on retrieval enhancement generation technology, the pre-processed data is associated with the dynamic water quality knowledge base to analyze and identify anomalies. Taking dissolved oxygen data as an example, the intelligent agent monitors the dissolved oxygen data stream in real time. Normally, the dissolved oxygen data should be maintained at 5-8mg / L. The system detects that the dissolved oxygen data drops sharply from 6.5mg / L to 2.1mg / L in a short period of time. Based on the dynamic threshold (based on the mean and standard deviation of historical dissolved oxygen data), it is judged as a serious anomaly and triggers an early warning.

[0075] RAG Enhanced Analysis: The intelligent agent vectorized the abnormal parameter "dissolved oxygen data suddenly dropped to 2.1 mg / L." Using Faiss indexes and inverted indexes, the dynamic water quality knowledge base quickly retrieved historical cases related to tags such as "abnormal dissolved oxygen," "hypoxia," and "aquatic life death." The system recalled multiple similar cases, including one that indicated that "wastewater discharge from a factory caused algal blooms, a sharp increase in nighttime oxygen consumption, and a sudden drop in dissolved oxygen data."

[0076] Diagnostic report generation: The large language model combines retrieved historical case information and real-time dissolved oxygen data to generate a diagnostic report: "The preliminary judgment is that the water body is hypoxic, which may be caused by the death and decomposition of algae blooms upstream or the discharge of organic matter in industrial wastewater, which leads to a sharp increase in microbial oxygen consumption. Further investigation of upstream pollution sources is necessary."

[0077] A time-series prediction model inputs dissolved oxygen data from the past 72 hours and relevant meteorological data (e.g., no recent rainfall, slow water flow) to predict dissolved oxygen trends for the next 48 hours. The model predicts that without intervention, dissolved oxygen levels could drop further to 1.5 mg / L, potentially causing mass fish kills. The intelligent agent integrates diagnostic reports, prediction results, and governance cases and regulations (such as "Water Pollutant Emission Standards") in its knowledge base. The system then recommends the optimal solution based on cost, timeliness, and success rate.

[0078] A command was issued via the API to activate the buoy-type aerators near the river section, setting them to maximum power. The monitoring frequency of the dissolved oxygen sensors in that section was automatically adjusted from one minute to every 30 seconds to more closely monitor the recovery of dissolved oxygen data. The local environmental protection department was notified and dispatched personnel to investigate the upstream factory. Plans for the salvage and disposal of aquatic organisms were also recommended. A detailed PDF report was automatically generated, including a summary of the dissolved oxygen data anomaly, diagnostic evidence, a 48-hour forecast for dissolved oxygen data, and a detailed remediation plan. This report was then pushed to the mobile app of the environmental protection department and water quality management personnel via an encrypted link.

[0079] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent agent collaborative decision-making system for environmental monitoring, characterized by: include: Data acquisition and preprocessing module: real-time collection of raw data, including water quality parameters and environmental data; Perform noise elimination, missing value filling and standardization on the collected raw data, and output preprocessed data; Dynamic Water Quality Knowledge Base Module: This module integrates multi-source data, including historical monitoring data, pollution source information, and governance case studies. It uses the BERT model to vectorize and encode the multi-source data and constructs a retrieval architecture based on a hierarchical navigable small-world hierarchical index and an inverted index to generate a dynamic water quality knowledge base. Agent analysis module: Based on retrieval enhancement generation technology, it associates pre-processed data with a dynamic water quality knowledge base, predicts water quality trends through a time series prediction model, and generates abnormal diagnosis reports and closed-loop treatment plans; Interaction and execution module: Provides early warning through abnormal diagnosis reports and issues equipment control instructions according to the closed-loop management plan.

2. The intelligent agent collaborative decision-making system for environmental monitoring according to claim 1, characterized in that: The raw data collection includes placing sensors at key nodes in the target water area to collect data, and the collected data is transmitted to the edge server in real time via the Internet of Things protocol; The water quality parameters include pH value, dissolved oxygen, turbidity, conductivity, heavy metal concentration, organic matter content, and algae biomass; the environmental data include water temperature, flow rate, and meteorological data; The noise elimination includes using a sliding window method to preprocess and eliminate random noise, outliers and interference signals in the original data; using a moving average filtering method to replace the current data point with the average value of the data in the sliding window to weaken high-frequency noise.

3. The intelligent agent collaborative decision-making system for environmental monitoring according to claim 1 is characterized by: The construction process of the dynamic water quality knowledge base includes: Integrate multi-source data including historical monitoring data, pollution source information, and governance cases; the historical monitoring data includes raw data collected during the historical period; the pollution source information includes industrial and agricultural emission sources; the governance cases include the time, plan, and governance effects of historical water quality governance; Perform data cleaning and format unification on multi-source data; For structured data, interpolation is used to fill missing values; For unstructured data, key information is extracted through text parsing and OCR technology, supplemented by manual annotation to enhance semantic labels; then, the text information in the unstructured data is mapped into high-dimensional vectors through a pre-trained language model.

4. The intelligent agent collaborative decision-making system for environmental monitoring according to claim 1, characterized in that: The retrieval architecture of the dynamic water quality knowledge base module includes: A hierarchical navigable small-world algorithm is used to optimize high-dimensional vector retrieval, combined with an inverted index to accelerate tag filtering, forming a hybrid architecture of Elasticsearch and Faiss, and associating original text and vector data through unique IDs.

5. The intelligent agent collaborative decision-making system for environmental monitoring according to claim 1 is characterized by: The dynamic water quality knowledge base includes a dynamic update mechanism and a performance optimization mechanism; The dynamic update mechanism synchronizes real-time sensor data incrementally and optimizes storage using a hot-cold separation strategy: hot data is stored in memory, and cold data is archived to distributed storage; The performance optimization mechanism rebuilds the Faiss index periodically to eliminate data distribution offsets and uses Redis to cache high-frequency search results, reducing response time to milliseconds.

6. The intelligent agent collaborative decision-making system for environmental monitoring according to claim 1, characterized in that: The Agent analysis module includes: calculating the mean and standard deviation based on historical data; setting dynamic thresholds based on the mean and standard deviation, judging the pre-processed data, and identifying abnormal data; When abnormal data is detected, the Agent triggers RAG enhanced analysis; Abnormal data is vectorized using a pre-trained model, and then an approximate nearest neighbor search is performed in the dynamic water quality knowledge base using the Faiss index to recall similar historical cases. The search results are sorted by cosine similarity and filtered using an inverted index to narrow the search scope. The search results of similar historical cases are combined with real-time data to form prompt words to generate diagnostic reports and preliminary governance recommendations.

7. The intelligent agent collaborative decision-making system for environmental monitoring according to claim 1, characterized in that: The time series prediction model is based on the Transformer time series model, which inputs past period data and predicts water quality changes in future periods; the model injects time series information through position coding; The self-attention mechanism of the time series prediction model captures long-term dependencies by calculating the attention weights of the query matrix, key matrix, and value matrix; The multi-head attention is expanded into multiple subspaces, and each head calculates attention independently.

Images