A big data-based monitoring and aquaculture water pollution early warning system

Abstract

The application discloses a kind of based on big data's monitoring aquaculture water pollution early warning system, belong to aquaculture water quality monitoring technical field.The system includes data acquisition module, data preprocessing module, time series prediction early warning module, pollution cause tracing module and intelligent disposal decision module;With environmental data correction water quality time series data, breeding production data is split into time series and discrete type respectively standardized;Using pre-trained time series prediction AI model predicts water quality pollution index, realizes pollution risk grade classification early warning;After triggering early warning, the working condition comprehensive feature vector is obtained by fusing first fusion feature set and second fusion feature set, and is matched with the pollution factor correlation strength matrix constructed based on historical pollution data set Inference, output pollution cause;Based on pollution risk grade and pollution cause matching knowledge base output disposal scheme.The application can improve the accuracy of pollution early warning, and realize accurate tracing and targeted disposal, suitable for large-scale aquaculture water pollution prevention and control.

Description

Technical Field

[0001] This invention relates to the field of aquaculture water quality monitoring technology, specifically to a big data-based monitoring and early warning system for aquaculture water pollution. Background Technology

[0002] With the rapid development of my country's aquaculture industry towards large-scale and intensive operations, the risk of pollution in aquaculture water bodies continues to intensify. Endogenous pollution caused by the accumulation of uneaten feed and excrement from farmed organisms, as well as sudden pollution from surrounding agricultural non-point source pollution and external water sources, can easily lead to outbreaks of diseases and large-scale mortality in farmed organisms, causing huge economic losses to the industry. Traditional aquaculture water quality monitoring relies heavily on manual sampling and laboratory testing, which suffers from poor timeliness and limited coverage. Conventional online monitoring systems only trigger alarms based on fixed thresholds, failing to capture the dynamic evolution of water quality parameters, resulting in strong warning delays and high false alarm rates, making it difficult to meet the precise water pollution control needs of large-scale aquaculture.

[0003] With the widespread application of big data and IoT technologies in water environment management, existing technical solutions have constructed water pollution risk early warning systems based on big data processing. For example, patent application CN122200940A discloses a water pollution management risk early warning method based on big data processing. This method collects multi-source monitoring data on water quality, pollution sources, hydrology, and meteorology, completes data structuring and feature fusion based on a stream processing framework, combines water pollution propagation relationship diagrams and trend prediction models to output pollution risk prediction results, and identifies abnormal events through mutation point detection, ultimately achieving hierarchical dynamic early warning. This solution improves the efficiency of water environment data processing and the foresight of risk early warning, and is suitable for water environment management scenarios in open watersheds.

[0004] However, there are still obvious shortcomings when applying the above technologies to aquaculture scenarios: First, it does not distinguish between endogenous and exogenous pollution due to the relatively closed water body characteristics of aquaculture, and lacks the ability to trace the source of pollution. After the warning, manual investigation is still required, and the targeted treatment is insufficient. Second, it does not separate the natural fluctuations in water quality parameters caused by environmental factors such as temperature, air pressure, and light, and it is easy to misjudge non-pollution fluctuations as pollution risks, thus limiting the accuracy of the warning. Third, it does not include aquaculture production data such as feeding amount, stocking density, and water exchange operations, nor does it construct differentiated cross-features for discrete production events and time-series water quality data, making it impossible to explore the correlation between aquaculture activities and water pollution. Fourth, it lacks intelligent decision support based on historical treatment data, and only outputs the risk level, which cannot form a closed-loop management from warning to source tracing and then to treatment, making it difficult to give full play to the prevention and control value of the warning. Summary of the Invention

[0005] This invention provides a big data-based monitoring system for early warning of water pollution in aquaculture to address the problems of existing aquaculture water quality early warning systems, which are susceptible to false alarms due to environmental interference, lack of integration with aquaculture production data, lack of automatic pollution source tracing function, and inability to intelligently match treatment solutions, making it difficult to achieve precise and closed-loop prevention and control of water pollution.

[0006] To address the aforementioned problems, according to one aspect of the present invention, a big data-based monitoring and early warning system for aquaculture water pollution is disclosed, comprising a data acquisition module, a data preprocessing module, a time-series prediction and early warning module, a pollution cause tracing module, and an intelligent disposal decision-making module.

[0007] The data acquisition module is used to collect water quality time-series data, environmental data, and aquaculture production data in real time during the aquaculture process, and to store historical water pollution event datasets with pollution cause annotations. The data preprocessing module is used to correct environmental interference in water quality time series data based on environmental data, and to split aquaculture production data into time series production data and discrete production data, and to perform standardization processing on each of them. The time-series prediction and early warning module is used to input the water quality time-series data after environmental interference correction and environmental data into the pre-trained time-series prediction AI model, output the water quality pollution index prediction results, and complete the pollution risk level determination and early warning triggering according to the graded early warning rules. When the pollution cause tracing module is used to trigger an early warning, it fuses standardized time-series production data with water quality time-series data and environmental data after environmental interference correction to obtain a first fusion feature set. It then fuses standardized discrete production data with water quality time-series data and environmental data after environmental interference correction to obtain a second fusion feature set. After performing feature comprehensive screening on the first and second fusion feature sets to obtain a comprehensive feature vector of operating conditions, it performs matching and reasoning with the pollution factor correlation intensity matrix pre-constructed based on historical water pollution event datasets to output the cause of pollution. The intelligent disposal decision-making module is used to combine the pollution risk level and the cause of pollution, match it with a knowledge base pre-built based on historical disposal big data, and output a pollution early warning and disposal plan.

[0008] Furthermore, the environmental interference correction is as follows: For target water quality indicators in water quality time series data that are susceptible to environmental influences, a multivariate nonlinear regression function is fitted between each target water quality indicator and environmental impact factors based on a subset of normal operating conditions in the historical water pollution event dataset. The target water quality indicators are dissolved oxygen, pH value or ammonia nitrogen, and the environmental impact factors include air temperature, air pressure, water temperature and light intensity in the environmental data. The original measured values ​​of the target water quality indicators in the water quality time series data are substituted into the corresponding multivariate nonlinear regression function to calculate the natural fluctuation caused by changes in environmental impact factors. The original measured values ​​of the target water quality indicators are subtracted from the natural fluctuation amount, and then the benchmark mean values ​​of the corresponding target water quality indicators under standard environmental conditions are superimposed to obtain the water quality time series data after environmental disturbance correction.

[0009] Furthermore, the standardization process is as follows: The water quality time series production data is timestamped to maintain the same time granularity as the water quality time series data after environmental interference correction. At the same time, the missing values ​​are filled by linear interpolation and the units are unified by min-max normalization. One-hot encoding is used for categorical variables in discrete production data, time distance encoding is used for event variables to represent the time difference between the event occurrence time and the endpoint time, and min-max normalization is used for numerical discrete variables to achieve dimensional unification. The endpoint time is the time node when the warning is triggered.

[0010] Furthermore, the time-series prediction and early warning module concatenates water quality time-series data corrected for environmental interference within a preset time period with environmental data and inputs it into a pre-trained time-series prediction AI model. The model outputs hourly predicted values ​​of ammonia nitrogen, nitrite nitrogen, and dissolved oxygen pollution indicators within a preset prediction time period, i.e., the water quality pollution indicator prediction results. The time-series prediction AI model is a Seq2Seq-LSTM model with an attention mechanism.

[0011] Furthermore, the tiered early warning rule is as follows: The water quality safety thresholds for ammonia nitrogen, nitrite and dissolved oxygen pollution indicators are determined based on the corresponding aquaculture species. The hourly predicted values ​​of each pollution indicator within the preset prediction period are compared with the corresponding water quality safety thresholds one by one. The pollution risk level is determined by combining the time-series change trend of the pollution indicators. If any level of pollution risk is triggered, a water pollution early warning signal corresponding to the pollution risk level will be generated, and the pollution cause tracing module will be activated. The pollution risk levels include Level 1 pollution risk, Level 2 pollution risk, Level 3 pollution risk, and Level 4 pollution risk.

[0012] Furthermore, the method for time-series feature fusion is as follows: Using the time node when the warning is triggered as the endpoint, the time-series production data after standardization, the water quality time-series data after environmental interference correction, and the environmental data within a preset time period before the endpoint are respectively extracted, and the time-domain features corresponding to each type of extracted time-series data are extracted. The time-domain features include the time period mean, extreme value, hourly rate of change, volatility, and trend slope. The first fused feature set is obtained by weighting and fusing the temporal features corresponding to various types of extracted time-series data using a channel attention mechanism.

[0013] Furthermore, the method for fusing discrete operating condition features is as follows: Using the time node when the warning is triggered as the endpoint, an analysis window of a preset duration is extracted. Baseline statistical features are extracted from the water quality time series data and environmental data after environmental interference correction within the analysis window to obtain baseline statistical features including water quality baseline statistical features and environmental baseline statistical features. The standardized discrete production data is divided into categorical discrete features, event-based discrete features, and numerical discrete features according to data type. These are then subjected to differential cross-operations with the baseline statistical features to obtain cross features. The standardized discrete production data is concatenated with cross features to obtain an initial high-dimensional discrete feature set. The variance threshold method is then used to filter redundant features from the initial high-dimensional discrete feature set to obtain a second fused feature set.

[0014] Furthermore, the method for obtaining the cause of pollution is as follows: The first and second fusion feature sets are concatenated, and the top K features with the highest mutual information coefficient and the correlation with pollution causes are selected to obtain the comprehensive feature vector of the working condition. Multiply the corresponding pollution-causing weight vectors of each feature in the comprehensive feature vector of the working condition with the single-factor pollution-causing weight vectors in the correlation strength matrix of pollution factors, calculate the single-factor contribution of each factor, and filter them in descending order to obtain the top 3 candidate factors. Based on the two-factor synergy coefficient matrix in the pollution factor correlation strength matrix, the synergistic pollution effect of the pairwise combination of candidate factors is verified. If the synergistic contribution of the combination exceeds the set threshold, it is determined to be a common pollution factor. Output the 1 to 2 pollution causes with the highest overall contribution, and simultaneously determine the pollution type as pollution from aquaculture itself, external pollution, or mixed pollution; The pollution factor association strength matrix includes a single-factor pollution weight vector and a two-factor synergy coefficient matrix, which is constructed offline based on historical water pollution event datasets using the maximum mutual information coefficient method and the SHAP interaction effect analysis method.

[0015] Furthermore, the intelligent disposal decision module uses pollution risk level and pollution cause as search conditions to retrieve multiple sets of candidate disposal measures from the knowledge base. It prioritizes the candidate disposal measures based on water quality recovery speed, aquaculture loss rate and disposal cost as comprehensive evaluation indicators, and determines the pollution early warning disposal plan.

[0016] Compared with the prior art, the present invention has the following advantages: First, by using multivariate nonlinear regression fitting to isolate the natural fluctuations in water quality caused by environmental factors such as air temperature, water temperature, and light intensity, environmental interference correction is completed, effectively reducing false alarms caused by non-pollution fluctuations and improving the accuracy of early warnings.

[0017] Second, the aquaculture production data is divided into time-series and discrete types and standardized in a differentiated manner. Time-series fusion features and discrete working condition cross features are constructed respectively to fully explore the intrinsic relationship between aquaculture activities and water pollution, making up for the lack of production dimension data in the existing scheme.

[0018] Third, a Seq2Seq-LSTM model with attention mechanism is used to achieve forward prediction of water pollution indicators, and a four-level graded early warning rule is provided to solve the problem of strong lag in traditional threshold alarms.

[0019] Fourth, a pollution factor association strength matrix is ​​constructed based on historical labeled datasets. Through multi-feature fusion screening and matching reasoning, the source of pollution can be traced automatically, and the pollution from the aquaculture itself, external pollution and mixed pollution can be accurately distinguished. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the big data-based monitoring and early warning system for aquaculture water pollution. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. This embodiment uses intensive, contiguous aquaculture ponds as the application carrier, and relies on IoT sensing devices and a cloud-based big data platform to build the system of this invention. The invention adopts a dual-track operation mode of offline big data modeling and online real-time closed-loop control. In the offline big data modeling stage, based on labeled historical water pollution event datasets, the system completes multivariate nonlinear regression function fitting, time-series prediction AI model training, pollution factor correlation strength matrix construction, knowledge base mining and construction on the cloud-based big data platform, and then distributes the data. In the online real-time closed-loop control stage, the system relies on the big data-based monitoring aquaculture water pollution early warning system of this invention to complete the entire process of real-time operation from data collection, data preprocessing, water quality prediction and early warning, pollution cause tracing to intelligent disposal decision-making. It should be understood that the specific embodiments described herein are only for explaining the invention and do not limit the application scenarios of the invention.

[0022] The big data-based monitoring and early warning system for aquaculture water pollution applied in this embodiment of the invention includes a data acquisition module, a data preprocessing module, a time-series prediction and early warning module, a pollution cause tracing module, and an intelligent treatment decision-making module. The specific implementation methods of each module are as follows: The data acquisition module collects real-time water quality time-series data, environmental data, and aquaculture production data during the aquaculture process, and stores historical water pollution event datasets labeled with pollution causes. In specific implementation, multi-parameter water quality sensor probes are deployed in each aquaculture pond, with one set of multi-parameter water quality sensors submerged in the middle layer of the water. These probes collect water quality parameters such as dissolved oxygen, pH, ammonia nitrogen, nitrite nitrogen, total nitrogen, total phosphorus, and turbidity at fixed time intervals and are linked to a unified timestamp, forming time-series water quality data. Simultaneously, meteorological and hydrological monitoring units are set up in the aquaculture areas of the ponds to synchronously collect environmental factors such as air temperature, atmospheric pressure, light intensity, rainfall, wind speed, water level, and external inflow, forming environmental data. Production data is recorded through a production data entry terminal, recording information on aquaculture activities such as feeding, oxygenation, water exchange, medication, stocking density, and aquaculture species, forming aquaculture production data. This production data record covers the entire aquaculture cycle, from seedling stocking, daily feeding, water quality control to pond cleaning and disinfection. In this embodiment, the historical water pollution event dataset is constructed by labeling historical water pollution events. A water pollution event is defined as a pollution index exceeding the water quality safety threshold of the corresponding aquaculture species within a continuous preset time period. The pollution causes of each water pollution event are manually reviewed and labeled. Water quality time series data, environmental data, and aquaculture production data corresponding to the preset time period before the pollution occurred are extracted as positive samples. Normal operating condition data of the same period are extracted as negative samples at a positive-to-negative sample ratio of 1:2. The normal operating condition data is the data corresponding to the period when there is no water quality exceeding the standard and the aquaculture status is stable. All positive and negative samples are persistently stored in a distributed file system after being unified in time granularity and field structure, which serves as the underlying data support for the subsequent training of time series prediction AI models and the construction of the pollution factor correlation strength matrix.

[0023] The data preprocessing module corrects for environmental interference in water quality time series data based on environmental data. At the same time, it splits aquaculture production data into time series production data and discrete production data, and performs standardization processing on the time series production data and discrete production data respectively. The environmental interference correction is performed on three target water quality indicators that are significantly affected by the environment: dissolved oxygen, pH, and ammonia nitrogen. First, based on normal operating condition samples from the historical water pollution event dataset (which are the negative samples in the aforementioned historical water pollution event dataset), a multivariate nonlinear regression function is fitted between each target water quality indicator and environmental impact factors such as air temperature, air pressure, water temperature, and light intensity. Then, environmental impact factors from real-time collected environmental data are substituted into the corresponding multivariate nonlinear regression function to calculate the natural fluctuation of the corresponding target water quality indicator caused by environmental impact factors. This natural fluctuation reflects the normal fluctuation range of water quality parameters caused by non-pollution factors. Finally, the original measured values ​​of the target water quality indicators are subtracted from the natural fluctuation, and then the baseline mean under standard environmental conditions is superimposed to obtain the corrected water quality time series data after eliminating environmental interference. The baseline mean is the steady-state average concentration of each target water quality indicator under normal operating conditions monitored over a long period of time in a standard aquaculture environment (generally air temperature 25℃, standard atmospheric pressure, moderate light intensity, and water temperature 22℃).

[0024] Aquaculture production data was split into time-series and discrete data based on its temporal characteristics and then standardized. Time-series data was timestamped to maintain the same temporal granularity as the environmentally corrected water quality time-series data. Missing values ​​were filled using linear interpolation, and dimensional unification was achieved through min-max normalization, resulting in standardized time-series production data. Discrete data was further categorized into categorical variables, event-based variables, and numerical discrete variables. Categorical variables included species and growth stages of aquaculture, while event-based variables included water change operations. The data includes events such as pesticide application and disinfection, and numerical discrete variables such as fish stocking density. Categorical variables are converted into binary feature vectors using one-hot encoding. Event-type variables are represented by time distance encoding to indicate the time difference between the event occurrence and the current time. The time difference between the event occurrence and the current time is the number of hours between the event occurrence and the warning trigger time, which is used to characterize the timeliness of the event's impact on water quality. Min-max normalization is performed on numerical discrete variables. Finally, the three types of encoded discrete features are integrated to output a standardized discrete feature vector, which is the standardized discrete production data.

[0025] The time-series prediction and early warning module inputs environmentally corrected water quality time-series data into a pre-trained time-series prediction AI model, combining it with environmental data. It outputs predictions of water pollution indicators for a preset time period and determines pollution risk levels and triggers warnings based on tiered early warning rules. The time-series prediction AI model is a Seq2Seq-LSTM (Sequence-to-Sequence Long Short-Term Memory) model with an attention mechanism. This model is trained offline on a historical water pollution event dataset, and its parameters are iteratively optimized periodically as the dataset is updated. During real-time prediction, the water quality time-series data corrected for environmental interference from the past 24 hours is concatenated with the corresponding environmental data to form a two-dimensional input sequence. This sequence is then input into the time-series prediction AI model. The model extracts the temporal latent features of the two-dimensional input sequence through a built-in encoder and automatically assigns feature weights to each time step using an attention mechanism, highlighting the characteristics of periods with significant pollution trend changes. Finally, the decoder outputs hourly predictions of ammonia nitrogen, nitrite nitrogen, and dissolved oxygen for the next 6 hours, representing the water pollution indicator prediction results.

[0026] This embodiment of the graded early warning rule adopts a four-level graded judgment logic. When the prediction result meets any risk level judgment condition, a corresponding early warning signal is immediately generated and the pollution cause tracing module is activated. The risk level judgment conditions in this embodiment are as follows: Level 1 pollution risk (mild warning): The predicted value of any one of the indicators, such as ammonia nitrogen, nitrite nitrogen, and dissolved oxygen, reaches 95%-100% of the water quality safety threshold of the aquaculture species within the next 6 hours, and the time-series trend of the indicator continues to rise in the direction of exceeding the standard; Level 2 pollution risk (moderate warning): The predicted value of any pollution indicator reaches the water quality safety threshold within the next 6 hours. Level 1: 100%, or two or more indicators simultaneously reaching 95% or above, indicating a clear trend of continuous deterioration in water quality, triggering a moderate warning; Level 2: Pollution risk (severe warning): At least one indicator's predicted value exceeds the water quality safety threshold within the next 6 hours, and the cumulative duration of exceeding the standard is ≥2 hours, indicating a high risk of water pollution and a high likelihood of triggering aquatic stress, triggering a severe warning; Level 3: Pollution risk (emergency warning): A single indicator significantly exceeds the standard (exceeding the water quality safety threshold by 120%) within the next 6 hours, or two or more indicators simultaneously exceed the water quality safety threshold with a duration of exceeding the standard ≥4 hours, indicating a risk of large-scale disease outbreaks and deaths in fish and shrimp, triggering an emergency warning.

[0027] When the pollution cause tracing module is used to trigger an early warning, it fuses standardized time-series production data with environmental disturbance-corrected water quality time-series data and environmental data to obtain a first fused feature set. It then fuses standardized discrete production data with environmental disturbance-corrected water quality time-series data and environmental data with discrete operating condition features to obtain a second fused feature set. After performing feature synthesis and screening on the first and second fused feature sets to obtain an operating condition comprehensive feature vector, it performs matching and reasoning with the pollution factor correlation intensity matrix pre-constructed based on historical water pollution event datasets to output the cause of pollution. Specifically, the time series feature fusion takes the time node when the warning is triggered as the endpoint, and extracts the standardized time series production data, environmental interference corrected water quality time series data, and environmental data within 6 hours before the endpoint. It then extracts the time domain features corresponding to each type of extracted time series data, including the time period mean, extreme values, hourly rate of change, volatility, and trend slope. Each type of extracted time series data is a time domain feature set corresponding to the three types of time series data: time series production data, corrected water quality time series data, and environmental data. The three types of time domain features are then weighted and allocated through a channel attention mechanism, and the first fused feature set is obtained after feature splicing.

[0028] The discrete operating condition feature fusion is performed with the time node when the early warning is triggered as the endpoint. An analysis window of a preset duration is extracted, and baseline statistical features are extracted from the water quality time series data and environmental data after environmental interference correction within the analysis window. Statistical quantities such as mean, median, and cumulative change within the analysis window are extracted to obtain baseline statistical features including water quality baseline statistical features and environmental baseline statistical features. Then, the standardized discrete production data is divided into categorical discrete features, event-type discrete features, and numerical discrete features according to data type. Differential cross operations are performed with the baseline statistical features to obtain cross features. Specifically, categorical discrete features and baseline statistical features undergo a product-type cross operation, event-type discrete features and baseline statistical feature change rate undergo a decay-weighted cross operation, and numerical discrete features and baseline statistical features undergo a proportional cross operation. The combined features generated by these three types of operations are the cross features. After constructing the cross features, the standardized discrete production data is concatenated with all the cross features to obtain an initial high-dimensional feature set. Then, low-variance redundant features are removed by using the variance threshold method. Low-variance redundant features are those that have very small differences in values ​​among samples of different pollution causes and do not make a significant contribution to the distinction of causes. The effective features are retained to form the second fusion feature set.

[0029] In this embodiment, the pollution factor correlation strength matrix includes a single-factor pollution weight vector and a two-factor synergy coefficient matrix, constructed offline based on a historical water pollution event dataset. The single-factor pollution weight vector is obtained by calculating the nonlinear correlation between each factor and the water pollution event using the maximum mutual information coefficient method; the factors referred to here are the pollution causes. The two-factor synergy coefficient matrix is ​​obtained by mining the synergistic pollution gain of pairwise combinations of factors through SHAP interaction effect analysis. During online matching and inference, the first and second fusion feature sets are first concatenated. Based on the maximum mutual information coefficient, features with the highest correlation to pollution causes are selected to obtain the comprehensive feature vector of the working condition. Then, the feature values ​​of each feature in the comprehensive feature vector of the working condition are multiplied by the corresponding single-factor pollution weight vector to calculate the single-factor contribution of each factor and filter them in descending order to obtain the Top 3 candidate factors. Subsequently, the two-factor synergy coefficient matrix is ​​retrieved to verify the synergistic pollution effect of pairwise combinations of candidate factors. If the combined synergistic contribution exceeds a preset threshold, it is determined to be a synergistic pollution gain. The pollution causes are considered together; the comprehensive contribution is obtained by weighted summation of the contribution of single factors and the synergistic contribution of two factors, and the 1-2 pollution causes with the highest comprehensive contribution are output. The pollution type is determined simultaneously according to the attribute category of the pollution cause, namely, pollution from aquaculture itself, pollution from external sources, or mixed pollution. If the pollution cause is excessive feeding, excessive stocking density, or deterioration of bottom sediment, it is determined to be pollution from aquaculture itself. If the pollution cause is all external factors such as the inflow of external water bodies or non-point source pollution from surrounding areas, it is determined to be pollution from external sources. If both internal and external factors are included, it is determined to be mixed pollution.

[0030] The intelligent disposal decision-making module combines pollution risk level and pollution cause, matches it with a knowledge base built based on historical disposal big data, and outputs a pollution early warning and disposal plan. In this embodiment, the knowledge base is a pollution disposal knowledge graph. The pollution disposal knowledge graph is built offline based on big data of disposal measures and disposal effects of historical water pollution events. The pollution disposal knowledge graph stores entities and relationships in the form of a four-tuple of <early warning level, pollution cause, disposal measure, disposal effect>. The disposal effect dimension includes three types of quantitative indicators: water quality recovery time, aquaculture loss rate, and disposal cost, which are stored in a graph database to support fast retrieval. When generating pollution early warning and response plans online, the current pollution risk level and the causes of pollution identified through source tracing are used as joint search criteria to retrieve pollution response knowledge graphs and match multiple sets of candidate response measures. Then, the candidate response measures are prioritized and ranked using water quality recovery speed, aquaculture loss rate, and response cost as comprehensive evaluation indicators. Finally, the optimal response plan with the highest ranking is output as the pollution early warning and response plan. The pollution early warning and response plan includes immediate emergency response measures for the water pollution event that is about to occur, as well as medium- and long-term aquaculture regulation suggestions for the causes of pollution, so as to achieve early prevention and targeted treatment of water pollution events.

[0031] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A big data-based monitoring and early warning system for aquaculture water pollution, characterized in that, include: The data acquisition module is used to collect water quality time-series data, environmental data, and aquaculture production data in real time during the aquaculture process, and to store historical water pollution event datasets with pollution cause annotations. The data preprocessing module is used to correct environmental interference in water quality time series data based on environmental data, and to split aquaculture production data into time series production data and discrete production data, and to perform standardization processing on each of them. The time-series prediction and early warning module is used to input the environmental disturbance-corrected water quality time-series data and environmental data into the pre-trained time-series prediction AI model, output the water pollution index prediction results, and complete the pollution risk level determination and early warning triggering according to the graded early warning rules. The pollution cause tracing module is used to fuse standardized time-series production data with environmental disturbance-corrected water quality time-series data and environmental data to obtain a first fusion feature set when an early warning is triggered. It also fuses standardized discrete production data with environmental disturbance-corrected water quality time-series data and environmental data to obtain a second fusion feature set. After comprehensively filtering the first and second fusion feature sets to obtain a comprehensive feature vector of operating conditions, it performs matching and reasoning with the pollution factor correlation intensity matrix pre-constructed based on historical water pollution event datasets to output the cause of pollution. The intelligent disposal decision-making module is used to combine the pollution risk level and the cause of pollution, match it with a knowledge base pre-built based on historical disposal big data, and output a pollution early warning and disposal plan.

2. The monitoring and early warning system for aquaculture water pollution according to claim 1, characterized in that, The environmental interference correction is as follows: For target water quality indicators in water quality time series data that are susceptible to environmental influences, a multivariate nonlinear regression function is fitted between each target water quality indicator and environmental impact factors based on a subset of normal operating conditions in the historical water pollution event dataset. The target water quality indicators are dissolved oxygen, pH value or ammonia nitrogen, and the environmental impact factors include air temperature, air pressure, water temperature and light intensity in the environmental data. The original measured values ​​of the target water quality indicators in the water quality time series data are substituted into the corresponding multivariate nonlinear regression function to calculate the natural fluctuation caused by changes in environmental impact factors. The original measured values ​​of the target water quality indicators are subtracted from the natural fluctuation amount, and then the benchmark mean values ​​of the corresponding target water quality indicators under standard environmental conditions are superimposed to obtain the water quality time series data after environmental disturbance correction.

3. The monitoring and early warning system for aquaculture water pollution according to claim 1, characterized in that, The standardization process is as follows: The water quality time series production data is timestamped to maintain the same time granularity as the water quality time series data after environmental interference correction. At the same time, the missing values ​​are filled by linear interpolation and the units are unified by min-max normalization. One-hot encoding is used for categorical variables in discrete production data, time distance encoding is used for event variables to represent the time difference between the event occurrence time and the endpoint time, and min-max normalization is used for numerical discrete variables to achieve dimensional unification. The endpoint time is the time node when the warning is triggered.

4. The monitoring and early warning system for aquaculture water pollution according to claim 1, characterized in that, The The time-series prediction and early warning module concatenates water quality time-series data corrected for environmental interference within a preset time period with environmental data and inputs it into a pre-trained time-series prediction AI model. It then outputs hourly predicted values ​​of ammonia nitrogen, nitrite nitrogen, and dissolved oxygen pollution indicators within a preset prediction time period, i.e., the water quality pollution indicator prediction results. The time-series prediction AI model is a Seq2Seq-LSTM model with an attention mechanism.

5. The monitoring and early warning system for aquaculture water pollution according to claim 4, characterized in that, The tiered early warning rules are as follows: The water quality safety thresholds for ammonia nitrogen, nitrite and dissolved oxygen pollution indicators are determined based on the corresponding aquaculture species. The hourly predicted values ​​of each pollution indicator within the preset prediction period are compared with the corresponding water quality safety thresholds one by one. The pollution risk level is determined by combining the time-series change trend of the pollution indicators. If any level of pollution risk is triggered, a water pollution early warning signal corresponding to the pollution risk level will be generated, and the pollution cause tracing module will be activated. The pollution risk levels include Level 1 pollution risk, Level 2 pollution risk, Level 3 pollution risk, and Level 4 pollution risk.

6. The monitoring and early warning system for aquaculture water pollution according to claim 1, characterized in that, The method for time-series feature fusion is as follows: Using the time node when the warning is triggered as the endpoint, the time-series production data after standardization, the water quality time-series data after environmental interference correction, and the environmental data within a preset time period before the endpoint are respectively extracted, and the time-domain features corresponding to each type of extracted time-series data are extracted. The time-domain features include the time period mean, extreme value, hourly rate of change, volatility, and trend slope. The first fused feature set is obtained by weighting and fusing the temporal features corresponding to various types of extracted time-series data using a channel attention mechanism.

7. The monitoring and early warning system for aquaculture water pollution according to claim 1, characterized in that, The method for fusing discrete operating condition features is as follows: Using the time node when the warning is triggered as the endpoint, an analysis window of a preset duration is extracted. Baseline statistical features are extracted from the water quality time series data and environmental data after environmental interference correction within the analysis window to obtain baseline statistical features including water quality baseline statistical features and environmental baseline statistical features. The standardized discrete production data is divided into categorical discrete features, event-based discrete features, and numerical discrete features according to data type. These are then subjected to differential cross-operations with the baseline statistical features to obtain cross features. The standardized discrete production data is concatenated with cross features to obtain an initial high-dimensional discrete feature set. The variance threshold method is then used to filter redundant features from the initial high-dimensional discrete feature set to obtain a second fused feature set.

8. The monitoring and early warning system for aquaculture water pollution according to claim 6 or 7, characterized in that, The method for obtaining the cause of the pollution is as follows: The first and second fusion feature sets are concatenated, and the top K features with the highest mutual information coefficient and the correlation with pollution causes are selected to obtain the comprehensive feature vector of the working condition. Multiply the corresponding pollution weight vectors of each feature in the comprehensive feature vector of the working condition with the single-factor pollution weight vectors in the correlation strength matrix of pollution factors, calculate the single-factor contribution of each factor, and filter them in descending order to obtain the top 3 candidate factors. Based on the two-factor synergy coefficient matrix in the pollution factor correlation strength matrix, the synergistic pollution effect of the pairwise combination of candidate factors is verified. If the synergistic contribution of the combination exceeds the set threshold, it is determined to be a common pollution factor. Output the 1 to 2 pollution causes with the highest overall contribution, and simultaneously determine the pollution type as pollution from aquaculture itself, external pollution, or mixed pollution; The pollution factor association strength matrix includes a single-factor pollution weight vector and a two-factor synergy coefficient matrix, which is constructed offline based on historical water pollution event datasets using the maximum mutual information coefficient method and the SHAP interaction effect analysis method.

9. The monitoring and early warning system for aquaculture water pollution according to claim 1, characterized in that, The intelligent disposal decision-making module uses pollution risk level and pollution cause as search conditions to retrieve multiple sets of candidate disposal measures from the knowledge base. It prioritizes the candidate disposal measures based on water quality recovery speed, aquaculture loss rate and disposal cost as comprehensive evaluation indicators, and determines the pollution early warning disposal plan.

Images