Intelligent Control Method and System for Aquaculture Wastewater Based on Multi-Parameter Feedback

Abstract

A multi-parameter feedback-based intelligent control method and system for aquaculture wastewater, relating to the field of wastewater treatment and control technology, includes: acquiring key water quality monitoring indicators for each treatment subsequence; real-time monitoring of each treatment subsequence; critical change analysis of historical monitoring periods marked as abnormal for the treatment subsequences, obtaining critical feature sets for historical monitoring periods, comparing significant differences in the critical feature sets, and constructing a critical correlation feature network of abnormal indicators for abnormal historical monitoring periods; pseudo-correlation analysis of the critical correlation feature network, obtaining pseudo-correlation nodes and interference nodes, and constructing a critical directed acyclic graph and critical interference link graph of abnormal indicators; constructing a multi-device collaborative control model based on the critical directed acyclic graph and critical interference link graph of several abnormal indicators, outputting specific control parameters of the associated devices for each treatment subsequence and performing preventive control to achieve efficient, economical, and stable treatment of aquaculture wastewater.

Description

Technical Field

[0001] This invention relates to the field of wastewater treatment and control technology, specifically to an intelligent control method and system for aquaculture wastewater based on multi-parameter feedback. Background Technology

[0002] Chinese patent publication number CN118449981A discloses a water quality monitoring system for aquaculture, comprising: a control terminal, which is the main control terminal of the system and is used to issue execution commands; a construction module, which is used to construct a three-dimensional model of the aquaculture area; a data acquisition module, which is used to acquire water quality parameters in the aquaculture area; and a receiving module, which is used to set the parameter receiving period and receive the water quality parameters in the aquaculture area acquired by the data acquisition module based on the parameter receiving period.

[0003] Chinese patent CN113066278B discloses a method and system for monitoring water quality in aquaculture, including: initializing a wireless sensor network; collecting water information in the water through various sensor nodes of the wireless sensor network; preprocessing the water information to obtain effective data of the water information; and issuing an early warning of algal blooms based on the effective data of the water information.

[0004] Aquaculture wastewater contains a large amount of pollutants such as uneaten feed, feces, and nitrogen and phosphorus compounds. If discharged directly, it can easily cause eutrophication of the receiving water body and disrupt the ecological balance. Therefore, efficient treatment of wastewater has become a key link in the sustainable development of aquaculture.

[0005] The aforementioned traditional methods for controlling aquaculture wastewater treatment largely rely on human experience or simple parametric correlation analysis to formulate control strategies. However, these methods have many insurmountable drawbacks in practical applications:

[0006] First, traditional methods struggle to distinguish between the "correlation" and "causation" between wastewater parameters, often misjudging spurious correlations as genuine causal relationships, leading to fundamental errors in control direction. Parameter relationships in aquaculture wastewater systems are complex, and many seemingly synchronous changes do not stem from direct causal effects. For example, monitoring data may show a high correlation between "increased water temperature" and "decreased ammonia nitrogen," but in reality, water temperature indirectly affects ammonia nitrogen degradation by influencing microbial activity, not directly causing a decrease in ammonia nitrogen. If a "heated wastewater" control method is adopted based on this, in scenarios with already high ambient temperatures, such as summer, the increased water temperature will reduce dissolved oxygen solubility, inhibiting microbial activity and causing a significant decrease in ammonia nitrogen degradation efficiency, thus achieving the opposite of the intended control effect.

[0007] Secondly, control strategies based on spurious correlations often lead to increased ineffective energy consumption and significantly higher treatment costs. Traditional methods cannot accurately identify direct causal links, easily leading to the allocation of resources to the control of non-critical parameters. For example, the correlation between "light intensity" and "total phosphorus reduction" is actually due to phosphorus absorption indirectly achieved by light promoting algal growth. If these two are mistakenly considered a direct causal relationship and the power of lighting equipment is blindly increased, the contribution of simply increasing light intensity to total phosphorus removal is negligible, as algal growth is also constrained by more critical factors such as nitrogen concentration and pH. However, this can lead to a significant increase in energy consumption of 20%-30%, resulting in serious waste of resources.

[0008] Furthermore, spurious correlations can interfere with the speed and accuracy of response to sudden changes in water quality, increasing the risk of ammonia nitrogen exceeding standards in aquaculture wastewater. The quality of aquaculture wastewater often changes drastically due to sudden events such as heavy rain or overfeeding. In such cases, spurious correlations can easily obscure the true causal mechanism. For example, when heavy rain causes a sudden increase in influent flow, a correlation may appear between "sudden increase in flow" and "temporary decrease in ammonia nitrogen." This is actually a result of the dilution effect, while the true causal chain is "increased flow → flushing of uneaten feed and feces → subsequent rebound of ammonia nitrogen due to microbial decomposition." If the decrease in ammonia nitrogen is mistakenly attributed directly to the increased flow, and aeration intensity is reduced, the system will exceed standards due to insufficient aeration when ammonia nitrogen rebounds, severely impacting the treatment effect.

[0009] Finally, spurious correlations reduce the generalization ability of intelligent control models, hindering the system's self-learning and scenario adaptation. Intelligent control relies on the model's adaptability to different aquaculture scenarios, but spurious correlations can cause the model to learn incorrect patterns. For example, in a pond aquaculture scenario, the correlation between "pH increase" and "nitrite decrease" stems from the common cause of "algal photosynthesis." If the model misjudges this as a direct causal relationship, the pH regulation strategy will completely fail when transferred to a factory-style aquaculture scenario without algae. This model defect caused by spurious correlations severely restricts the widespread application of intelligent control systems in different aquaculture scenarios.

[0010] In summary, traditional methods for treating and controlling aquaculture wastewater have significant shortcomings in terms of precision control, energy consumption control, risk avoidance, and scenario adaptability because they cannot effectively distinguish between spurious correlations and true causal relationships. There is an urgent need for an intelligent control method that can accurately identify the true interaction mechanism between parameters in order to overcome existing technological bottlenecks and achieve efficient, economical, and stable treatment of aquaculture wastewater. Summary of the Invention

[0011] To address the aforementioned technical problems, the present invention aims to provide an intelligent control method for aquaculture wastewater based on multi-parameter feedback, comprising the following steps:

[0012] Step s1: Obtain the current aquaculture wastewater treatment process information, set up multi-source monitoring points according to the process information, and obtain the key water quality monitoring indicators for each treatment subsequence;

[0013] Step s2: Monitor each processing subsequence in real time, and classify the processing subsequence into normal or abnormal states based on the monitoring results;

[0014] Step s3: Perform critical change analysis on the historical monitoring periods in which the processed subsequences are marked as abnormal, obtain the critical feature set of the historical monitoring periods, compare the significant differences of the critical feature set, and construct the critical correlation feature network of abnormal indicators of the abnormal historical monitoring periods.

[0015] Step s4: Perform pseudo-correlation analysis on the critical correlation feature network of each abnormal indicator to obtain pseudo-correlation nodes and interference nodes, and construct the critical directed acyclic graph and critical interference link graph of the abnormal indicators.

[0016] Step s5: Construct a multi-device collaborative control model based on the critical directed acyclic graph and critical interference link graph of several abnormal indicators, output the specific control parameters of the associated devices of each processing sub-sequence and perform preventive control.

[0017] Furthermore, the process of obtaining current aquaculture wastewater treatment process information, setting up multi-source monitoring points based on the process information, and obtaining key water quality monitoring indicators for each treatment subsequence includes:

[0018] Obtain the process characteristics of the aquaculture wastewater treatment facility, and divide the aquaculture wastewater treatment process into several treatment sub-sequences based on the process characteristics.

[0019] Multi-source monitoring points were set up in several treatment subsequences. Based on the functional characteristics of the process flow of the corresponding treatment subsequence, the monitoring indicators of each multi-source monitoring point were obtained by data retrieval (including water quality parameters, environmental parameters, and process parameters; the water quality parameters include ammonia nitrogen, nitrite, total nitrogen, total phosphorus, COD, pH value, dissolved oxygen, and turbidity; the environmental parameters include light intensity, atmospheric temperature, and aquaculture water flow rate; the process parameters include aeration intensity, chemical dosage, filter flow rate, and wetland water level).

[0020] Multi-source monitoring points acquire multi-source data in real time based on monitoring indicators (including turbidity and flow rate detected by physical sensors; ammonia nitrogen, nitrite, pH value, and dissolved oxygen detected by electrochemical sensors; and biotoxicity detected by biosensors), mark the monitoring time, and set the monitoring cycle.

[0021] Furthermore, the process of real-time monitoring of each processing subsequence and classifying the processing subsequence into normal or abnormal states based on the monitoring results includes:

[0022] Preset standard threshold ranges for each type of monitoring indicator at each multi-source monitoring point, and extract the numerical time series of each type of monitoring indicator from the multi-source data of each multi-source monitoring point;

[0023] The numerical time series of each type of monitoring indicator at multiple monitoring points is compared with the corresponding standard threshold interval. The cumulative time of each type of monitoring indicator that is not within the corresponding standard threshold interval is obtained. An error threshold is preset. If there is a monitoring indicator whose cumulative time is greater than the error threshold, the monitoring indicator is marked as an abnormal indicator and the processing subsequence is marked as an abnormal state. If the cumulative time of each type of monitoring indicator is less than the error threshold, the processing subsequence is marked as a normal state.

[0024] Furthermore, the process of performing critical change analysis on historical monitoring periods in which the processed subsequences were marked as abnormal states, and obtaining the critical feature set of historical monitoring periods, includes:

[0025] Extract several historical monitoring periods marked as abnormal states for each processing subsequence, and mark these historical monitoring periods as abnormal historical monitoring periods. Obtain the times when the monitoring indicators are marked as abnormal indicators within these several abnormal historical monitoring periods, and mark these times as critical timestamps. Extract the numerical time series sequences corresponding to each type of monitoring indicator within the start timestamp and critical timestamp of these several abnormal historical monitoring periods. Perform critical change analysis on the numerical time series sequences corresponding to each type of monitoring indicator to obtain the critical feature set between each type of monitoring indicator (including the average standard deviation of each type of monitoring indicator and the average Pearson correlation coefficient between each type of monitoring indicator).

[0026] Furthermore, the process of constructing a critical correlation feature network of abnormal indicators for the historical monitoring period by comparing the significant differences of the critical feature set includes:

[0027] Extract the historical monitoring periods that are marked as normal for each processing subsequence, and mark the historical monitoring periods as normal historical monitoring periods;

[0028] For each processing subsequence, the next historical monitoring period of the normal historical monitoring period is determined. If the next historical monitoring period of the normal historical monitoring period is an abnormal historical monitoring period, the conventional feature set (including the mean standard deviation of each type of monitoring indicator and the mean Pearson correlation coefficient between each type of monitoring indicator) is obtained within the normal historical monitoring period. The conventional feature set within the normal historical monitoring period is compared with the critical feature set of the abnormal historical monitoring period to obtain the absolute value of the difference between the mean standard deviation of each type of monitoring indicator in the critical feature set and the mean standard deviation of each type of monitoring indicator in the conventional feature set. Monitoring indicators whose absolute value of the difference between the two is less than a preset threshold are removed. Subsequently, the average Pearson correlation coefficient between each type of monitoring indicator in the critical feature set and the average Pearson correlation coefficient between each type of monitoring indicator in the conventional feature set are compared. The average Pearson correlation coefficients are compared for significance, and the absolute value of the difference in average Pearson correlation coefficients between various types of monitoring indicators is obtained. For each type of monitoring indicator, those whose absolute value of the difference in average Pearson correlation coefficients is greater than a preset threshold are identified. These indicators are then marked as key monitoring indicators. Abnormal indicators are extracted from these key indicators (abnormal indicators in the historical abnormal monitoring period are necessarily key monitoring indicators). Each key monitoring indicator and abnormal indicator is used as a node, and these nodes are interconnected to construct a critical correlation feature network of abnormal indicators in the historical abnormal monitoring period. Edge weight coefficients are set between nodes in the critical correlation feature network (based on the absolute value of the difference in average Pearson correlation coefficients between nodes; the larger the absolute value of the difference in average Pearson correlation coefficients, the larger the edge weight coefficient).

[0029] Furthermore, the process of obtaining pseudo-correlated nodes and constructing a critical directed acyclic graph of anomaly indicators includes:

[0030] Extract the edge weight coefficients between each node and the abnormal indicator in the critical correlation feature network, preset the edge weight coefficient threshold, extract the nodes whose edge weight coefficients with the abnormal indicator are greater than the edge weight coefficient threshold, and mark the nodes as first nodes. Perform pseudo-correlation analysis on each first node to obtain the independent correlation coefficients between each first node and the abnormal indicator, preset the correlation significance threshold, mark the first nodes whose independent correlation coefficients with the abnormal indicator are greater than or equal to the correlation significance threshold as master nodes, and mark the first nodes whose independent correlation coefficients with the abnormal indicator are less than the correlation significance threshold as pseudo-correlation nodes.

[0031] Construct a critical directed acyclic graph (DAG) of the abnormal indicators based on the abnormal indicators and master nodes. The abnormal indicators and master nodes are used as nodes in the critical DAG. Each node corresponding to a master node in the critical DAG points to the node corresponding to the abnormal indicator, clearly showing the causal direction and dependency relationship between variables, while strictly excluding "causal loops" (i.e., there is no path that starts from a node, goes through a series of directed edges, and returns to itself).

[0032] Furthermore, the process of acquiring interference nodes and constructing a critical interference link diagram for abnormal indicators includes:

[0033] Pseudo-correlation analysis is performed between pseudo-correlation nodes and principal nodes. Pseudo-correlation nodes with an independent correlation coefficient greater than or equal to the correlation significance threshold with the principal node are marked as interference nodes, and pseudo-correlation nodes with an independent correlation coefficient less than the correlation significance threshold with the principal node are marked as irrelevant nodes.

[0034] Pseudo-correlation analysis is performed between irrelevant and interfering nodes. Irrelevant nodes with independent correlation coefficients greater than or equal to the significance threshold with interfering nodes are marked as secondary interfering nodes. A critical interference link diagram for the abnormal index is constructed based on the primary node, interfering nodes, and secondary interfering nodes (secondary interfering node → interfering node → primary node → abnormal index). For example, the critical interference link diagram is: water temperature ↑ → microbial activity ↑ → ammonia nitrogen degradation rate ↑ (but excessively high water temperature will lead to dissolved oxygen solubility ↓, forming a negative feedback).

[0035] Furthermore, the process of constructing a multi-device collaborative control model based on the critical directed acyclic graph and critical interference link graph of several abnormal indicators includes:

[0036] A multi-device collaborative control model is constructed based on a hybrid intelligent algorithm. The numerical time series sequences corresponding to various types of monitoring indicators in several historical monitoring periods of each processing sub-sequence, the critical directed acyclic graphs of abnormal indicators, and the critical interference link graphs are extracted as training sets and test sets. The training sets are input into the multi-device collaborative control model for training until the loss function is stable and the model parameters are saved. The multi-device collaborative control model is tested through the test set until it meets the preset requirements, and then the multi-device collaborative control model is output.

[0037] Furthermore, the process of outputting the specific control parameters of each processing sub-sequence associated device and performing preventative control includes:

[0038] Use the end timestamp of the monitoring period as the control time point;

[0039] When the processing subsequence is in a normal state, at the control time point, the multi-source data of the monitoring cycle is input into the multi-device collaborative control model. Based on the multi-device collaborative control model, the specific control parameters of the associated devices of each processing subsequence in the next acquisition cycle are output. Based on the specific control parameters of the associated devices of each processing subsequence in the next acquisition cycle, the associated devices of each processing subsequence are controlled.

[0040] The multi-device collaborative control model can accurately identify core causal links such as "aeration intensity → dissolved oxygen → ammonia nitrogen" from multi-parameter data of aquaculture wastewater, avoiding interference from spurious correlations such as "water temperature and ammonia nitrogen", and providing clear control logic for intelligent control. For example, when ammonia nitrogen exceeds the standard, the multi-device collaborative control model will prioritize increasing aeration intensity (direct causality) rather than blindly adjusting water temperature (spurious correlation), thereby improving control efficiency and reducing energy consumption.

[0041] The intelligent control system for aquaculture wastewater based on multi-parameter feedback includes a control center, which is connected to a data acquisition module, a real-time monitoring module, a feature extraction module, a pseudo-correlation analysis module, and a collaborative regulation module.

[0042] The data acquisition module is used to obtain information on the current aquaculture wastewater treatment process, set up multi-source monitoring points based on the process information, and obtain key water quality monitoring indicators for each treatment subsequence.

[0043] The real-time monitoring module is used to monitor each processing subsequence in real time and classify the processing subsequence into normal or abnormal states based on the monitoring results.

[0044] The feature extraction module is used to perform critical change analysis on the historical monitoring periods in which the processed subsequences are marked as abnormal, obtain the critical feature set of the historical monitoring period, compare the significant differences of the critical feature set, and construct the critical correlation feature network of abnormal indicators of the abnormal historical monitoring period.

[0045] The pseudo-correlation analysis module is used to perform pseudo-correlation analysis on the critical correlation feature network of various abnormal indicators, obtain pseudo-correlation nodes and interference nodes, and construct the critical directed acyclic graph and critical interference link graph of abnormal indicators.

[0046] The collaborative control module is used to construct a multi-device collaborative control model based on critical directed acyclic graphs and critical interference link graphs of several abnormal indicators, output the specific control parameters of the devices associated with each processing sub-sequence, and perform preventive control.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] 1. By eliminating spurious correlations through causal reasoning, the system can focus on direct causal links such as "carbon source addition → denitrification → total nitrogen reduction", concentrating resources on core control measures and significantly reducing ineffective energy consumption.

[0049] 2. After identifying spurious correlations through causal reasoning, the system can predict potential risks in advance (such as the causal chain of "residual bait input → delayed increase in ammonia nitrogen"), and maintain aeration intensity when the flow rate increases sharply to avoid exceeding the standard.

[0050] 3. After eliminating spurious correlations through causal reasoning, the model can capture universal causal chains such as "aeration → dissolved oxygen → nitrite oxidation", maintaining stable regulatory effects in different scenarios and laying the foundation for system self-learning (optimizing parameters based on real causality).

[0051] This invention distinguishes pseudo-correlation, allowing subsequent multi-device collaborative control models to "see clearly" the true interaction mechanism between tailwater parameters. This enables precise targeting of control points, optimization of resource input, and avoidance of potential risks, ultimately achieving the goal of "achieving the best treatment effect with the least energy consumption." This is also a key prerequisite for moving from "experience-based control" to "precise intelligent control."

[0052] The multi-device collaborative control model can accurately identify core causal links such as "aeration intensity → dissolved oxygen → ammonia nitrogen" from multi-parameter data of aquaculture wastewater, avoiding interference from spurious correlations such as "water temperature and ammonia nitrogen", and providing clear control logic for intelligent control. For example, when ammonia nitrogen exceeds the standard, the multi-device collaborative control model will prioritize increasing aeration intensity (direct causality) rather than blindly adjusting water temperature (spurious correlation), thereby improving control efficiency and reducing energy consumption. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the intelligent control method for aquaculture wastewater based on multi-parameter feedback, according to an embodiment of this application.

[0054] Figure 2 This is a schematic diagram of the intelligent control system for aquaculture wastewater based on multi-parameter feedback, according to an embodiment of this application. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] like Figure 1 As shown, the intelligent control method for aquaculture wastewater based on multi-parameter feedback includes the following steps:

[0057] Step s1: Obtain the current aquaculture wastewater treatment process information, set up multi-source monitoring points according to the process information, and obtain the key water quality monitoring indicators for each treatment subsequence;

[0058] Step s2: Monitor each processing subsequence in real time, and classify the processing subsequence into normal or abnormal states based on the monitoring results;

[0059] Step s3: Perform critical change analysis on the historical monitoring periods in which the processed subsequences are marked as abnormal, obtain the critical feature set of the historical monitoring periods, compare the significant differences of the critical feature set, and construct the critical correlation feature network of abnormal indicators of the abnormal historical monitoring periods.

[0060] Step s4: Perform pseudo-correlation analysis on the critical correlation feature network of each abnormal indicator to obtain pseudo-correlation nodes and interference nodes, and construct the critical directed acyclic graph and critical interference link graph of the abnormal indicators.

[0061] Step s5: Construct a multi-device collaborative control model based on the critical directed acyclic graph and critical interference link graph of several abnormal indicators, output the specific control parameters of the associated devices of each processing sub-sequence and perform preventive control.

[0062] It should be further explained that, in the specific implementation process, the process of obtaining information on the current aquaculture wastewater treatment process, setting up multi-source monitoring points based on the process information, and obtaining key water quality monitoring indicators for each treatment subsequence includes:

[0063] Obtain the process characteristics of the aquaculture wastewater treatment facility, and divide the aquaculture wastewater treatment process into several treatment sub-sequences based on the process characteristics.

[0064] Multi-source monitoring points were set up in several treatment subsequences. Based on the functional characteristics of the process flow of the corresponding treatment subsequence, the monitoring indicators of each multi-source monitoring point were obtained by data retrieval (including water quality parameters, environmental parameters, and process parameters; the water quality parameters include ammonia nitrogen, nitrite, total nitrogen, total phosphorus, COD, pH value, dissolved oxygen, and turbidity; the environmental parameters include light intensity, atmospheric temperature, and aquaculture water flow rate; the process parameters include aeration intensity, chemical dosage, filter flow rate, and wetland water level).

[0065] Multi-source monitoring points acquire multi-source data in real time based on monitoring indicators (including turbidity and flow rate detected by physical sensors; ammonia nitrogen, nitrite, pH value, and dissolved oxygen detected by electrochemical sensors; and biotoxicity detected by biosensors), mark the monitoring time, and set the monitoring cycle.

[0066] It should be further explained that, in the specific implementation process, the process of real-time monitoring of each processing subsequence and classifying the processing subsequence into normal or abnormal states based on the monitoring results includes:

[0067] Preset standard threshold ranges for each type of monitoring indicator at each multi-source monitoring point, and extract the numerical time series of each type of monitoring indicator from the multi-source data of each multi-source monitoring point;

[0068] The numerical time series of each type of monitoring indicator at multiple monitoring points is compared with the corresponding standard threshold interval. The cumulative time of each type of monitoring indicator that is not within the corresponding standard threshold interval is obtained. An error threshold is preset. If there is a monitoring indicator whose cumulative time is greater than the error threshold, the monitoring indicator is marked as an abnormal indicator and the processing subsequence is marked as an abnormal state. If the cumulative time of each type of monitoring indicator is less than the error threshold, the processing subsequence is marked as a normal state.

[0069] It should be further explained that, in the specific implementation process, the process of performing critical change analysis on the historical monitoring periods in which the processed subsequences were marked as abnormal, and obtaining the critical feature set of the historical monitoring periods, includes:

[0070] Extract several historical monitoring periods marked as abnormal states for each processing subsequence, and mark these historical monitoring periods as abnormal historical monitoring periods. Obtain the times when the monitoring indicators are marked as abnormal indicators within these several abnormal historical monitoring periods, and mark these times as critical timestamps. Extract the numerical time series sequences corresponding to each type of monitoring indicator within the start timestamp and critical timestamp of these several abnormal historical monitoring periods. Perform critical change analysis on the numerical time series sequences corresponding to each type of monitoring indicator to obtain the critical feature set between each type of monitoring indicator (including the average standard deviation of each type of monitoring indicator and the average Pearson correlation coefficient between each type of monitoring indicator).

[0071] It should be further explained that, in the specific implementation process, the process of comparing significant differences in the critical feature set and constructing the critical correlation feature network of abnormal indicators for the historical monitoring period includes:

[0072] Extract the historical monitoring periods that are marked as normal for each processing subsequence, and mark the historical monitoring periods as normal historical monitoring periods;

[0073] For each processing subsequence, the next historical monitoring period of the normal historical monitoring period is determined. If the next historical monitoring period of the normal historical monitoring period is an abnormal historical monitoring period, the conventional feature set (including the mean standard deviation of each type of monitoring indicator and the mean Pearson correlation coefficient between each type of monitoring indicator) is obtained within the normal historical monitoring period. The conventional feature set within the normal historical monitoring period is compared with the critical feature set of the abnormal historical monitoring period to obtain the absolute value of the difference between the mean standard deviation of each type of monitoring indicator in the critical feature set and the mean standard deviation of each type of monitoring indicator in the conventional feature set. Monitoring indicators whose absolute value of the difference between the two is less than a preset threshold are removed. Subsequently, the average Pearson correlation coefficient between each type of monitoring indicator in the critical feature set and the average Pearson correlation coefficient between each type of monitoring indicator in the conventional feature set are compared. The average Pearson correlation coefficients are compared for significance, and the absolute value of the difference in average Pearson correlation coefficients between various types of monitoring indicators is obtained. For each type of monitoring indicator, those whose absolute value of the difference in average Pearson correlation coefficients is greater than a preset threshold are identified. These indicators are then marked as key monitoring indicators. Abnormal indicators are extracted from these key indicators (abnormal indicators in the historical abnormal monitoring period are necessarily key monitoring indicators). Each key monitoring indicator and abnormal indicator is used as a node, and these nodes are interconnected to construct a critical correlation feature network of abnormal indicators in the historical abnormal monitoring period. Edge weight coefficients are set between nodes in the critical correlation feature network (based on the absolute value of the difference in average Pearson correlation coefficients between nodes; the larger the absolute value of the difference in average Pearson correlation coefficients, the larger the edge weight coefficient).

[0074] It should be further explained that, in the specific implementation process, the process of obtaining pseudo-correlation nodes and constructing the critical directed acyclic graph of abnormal indicators includes:

[0075] Extract the edge weight coefficients between each node and the abnormal indicator in the critical correlation feature network, preset the edge weight coefficient threshold, extract the nodes whose edge weight coefficients with the abnormal indicator are greater than the edge weight coefficient threshold, and mark the nodes as first nodes. Perform pseudo-correlation analysis on each first node to obtain the independent correlation coefficients between each first node and the abnormal indicator, preset the correlation significance threshold, mark the first nodes whose independent correlation coefficients with the abnormal indicator are greater than or equal to the correlation significance threshold as master nodes, and mark the first nodes whose independent correlation coefficients with the abnormal indicator are less than the correlation significance threshold as pseudo-correlation nodes.

[0076] The specific process of performing pseudo-correlation analysis on each first node includes:

[0077] By keeping other first-node parameters constant and continuously adjusting the parameters of the first node, the absolute value of the difference in the average Pearson correlation coefficient between the abnormal indicator and the first node is obtained. This absolute value is then standardized to convert it into an independent correlation coefficient. First nodes with an independent correlation coefficient greater than or equal to the significance threshold are marked as master nodes, while those with a coefficient less than the significance threshold are marked as pseudo-correlation nodes. "If two parameters X and Y remain correlated after controlling for other variables (such as Z), then X and Y have a direct causal relationship." For example, to verify the causal relationship of "aeration intensity → dissolved oxygen": with fixed water temperature and pH, the aeration intensity is increased from 2 m³ / h to 4 m³ / h. If dissolved oxygen increases from 3 mg / L to 6 mg / L (p < 0.01), then the causal relationship is confirmed.

[0078] Meanwhile, the correlation between water temperature (X) and ammonia nitrogen (Y) may be a spurious correlation because the true causal relationship is "increased water temperature → enhanced microbial activity (Z) → accelerated ammonia nitrogen degradation (Y)". After controlling Z, the correlation between X and Y will disappear, thus eliminating the direct causal relationship between X and Y. To verify the spurious correlation between water temperature and ammonia nitrogen: control microbial activity (e.g., maintain constant DO and pH), and increase the water temperature from 20℃ to 30℃. If there is no significant change in ammonia nitrogen concentration (p>0.05), then the correlation between the two is confirmed to be spurious.

[0079] Construct a critical directed acyclic graph (DAG) of the abnormal indicators based on the abnormal indicators and master nodes. The abnormal indicators and master nodes are used as nodes in the critical DAG. Each node corresponding to a master node in the critical DAG points to the node corresponding to the abnormal indicator, clearly showing the causal direction and dependency relationship between variables, while strictly excluding "causal loops" (i.e., there is no path that starts from a node, goes through a series of directed edges, and returns to itself).

[0080] It should be further explained that, in the specific implementation process, the process of acquiring interference nodes and constructing a critical interference link diagram of abnormal indicators includes:

[0081] Pseudo-correlation analysis is performed between pseudo-correlation nodes and principal nodes. Pseudo-correlation nodes with an independent correlation coefficient greater than or equal to the correlation significance threshold with the principal node are marked as interference nodes, and pseudo-correlation nodes with an independent correlation coefficient less than the correlation significance threshold with the principal node are marked as irrelevant nodes.

[0082] Pseudo-correlation analysis is performed between irrelevant and interfering nodes. Irrelevant nodes with independent correlation coefficients greater than or equal to the significance threshold with interfering nodes are marked as secondary interfering nodes. A critical interference link diagram for the abnormal index is constructed based on the primary node, interfering nodes, and secondary interfering nodes (secondary interfering node → interfering node → primary node → abnormal index). For example, the critical interference link diagram is: water temperature ↑ → microbial activity ↑ → ammonia nitrogen degradation rate ↑ (but excessively high water temperature will lead to dissolved oxygen solubility ↓, forming a negative feedback).

[0083] It should be further explained that, in the specific implementation process, the process of constructing a multi-device collaborative control model based on the critical directed acyclic graph and critical interference link graph of several abnormal indicators includes:

[0084] A multi-device collaborative control model is constructed based on a hybrid intelligent algorithm. Numerical time-series sequences of various monitoring indicators, critical directed acyclic graphs (DAGs) of abnormal indicators, and critical interference link graphs from several historical monitoring periods for each processing sub-sequence are extracted as training and testing sets. The training sets are input into the multi-device collaborative control model for training until the loss function stabilizes, and the model parameters are saved. The multi-device collaborative control model is then tested using the testing set until it meets preset requirements, at which point the multi-device collaborative control model is output. It should be further noted that every week, the numerical time-series sequences of various monitoring indicators, critical DAGs of abnormal indicators, and critical interference link graphs from several historical monitoring periods for each processing sub-sequence are re-acquired as training and testing sets to retrain the multi-device collaborative control model.

[0085] The process of constructing a multi-device collaborative control model based on hybrid intelligent algorithms is the core step in transforming the multi-parameter time-series characteristics, causal relationships (DAG), and disturbance patterns of the entire aquaculture wastewater treatment process into precise control strategies. This process requires the integration of multiple algorithms, including time-series prediction, graph neural networks (GNN), and reinforcement learning (RL), to achieve integrated decision-making based on "data-driven, causal constraints, and device collaboration." The following is a detailed supplementary explanation of this process:

[0086] Three types of time series are extracted from the numerical time series of various monitoring indicators corresponding to various types of monitoring indicators within several historical monitoring periods of each processing subsequence, and used as the basic input of the model:

[0087] Normal time series: The processing subsequence contains the values ​​of monitoring indicators under normal conditions (such as the time series of DO fluctuations in a biofilter between 3-6 mg / L) to learn the equipment synergy patterns under steady-state conditions;

[0088] Abnormal time series: Processing segments of subsequences with abrupt changes in water quality parameters under abnormal conditions (such as 1-hour data of sedimentation tank turbidity jumping from 20 NTU to 100 NTU after a rainstorm) to learn abnormal response strategies;

[0089] Equipment status time sequence: The operating parameters of the equipment associated with each treatment subsequence (such as the change of aeration intensity from 0 to 5 m³ / h, and the change of dosing pump frequency from 10 to 50 Hz) are used to associate the mapping relationship between "equipment operation and water quality change".

[0090] For water quality anomalies (such as excessive ammonia nitrogen and excessive total phosphorus) observed in historical data, the critical directed acyclic graphs of the corresponding anomaly indicators are extracted and used as the "causal constraints" input for the model.

[0091] The critical directed acyclic graph of anomalies is transformed into an adjacency matrix (nodes are parameters / devices, and matrix elements represent causal weights) as the input format for a graph neural network (GNN).

[0092] The interference link diagram is used as the input for the model's "interference adaptation," describing the impact of non-causal interference factors (such as sudden environmental changes or equipment failures) on water quality.

[0093] The interference link diagram of abnormal indicators is transformed into a time-series feature vector (such as interference intensity, propagation velocity, and influence range), and then fused with the numerical time-series sequence.

[0094] The multi-device collaborative control model adopts an architecture of "multimodal input - hierarchical feature fusion - collaborative decision output", which integrates three types of algorithms: LSTM (temporal prediction), GNN (causal reasoning), and deep reinforcement learning (DRL, device collaboration).

[0095] Input layer:

[0096] The time-series input branch receives the numerical time-series sequences of each processing subsequence (such as the turbidity time-series of the sedimentation tank and the DO time-series of the biofilter), and extracts time-series features (such as trend terms, periodic terms, and mutation terms) through an LSTM network; the DAG input branch receives the DAG adjacency matrix of the abnormal indicators, and extracts causal features (such as core control nodes and causal transmission paths) through a GNN (such as GraphSAGE); the interference input branch receives the feature vector of the interference link graph, and extracts interference features (such as interference intensity and duration of influence) through a fully connected network.

[0097] Feature fusion layer: Cross-modal feature integration:

[0098] An attention mechanism is used to fuse three types of features: dynamic weights are assigned to time-series features (e.g., the weight of DO time-series on ammonia nitrogen regulation is higher than that of water temperature time-series); causal confidence weights are assigned to causal features (e.g., the causal weight of aeration intensity → DO is 0.9, which is given a high weight); and disturbance urgency weights are assigned to disturbance features (e.g., the disturbance weight of equipment failure is higher than that of natural fluctuations). The output is a fused feature vector (128 dimensions, containing comprehensive information on time-series trends, causal constraints, and disturbance adaptation).

[0099] Decision layer: Deep reinforcement learning (DRL) generates collaborative policies.

[0100] Using the fused feature vector as state input, the DRL agent generates multi-device collaborative control commands: State space (S): fused feature vector + current operating status of each device (e.g., aeration intensity, dosage); Action space (A): combination of control parameters for each device (e.g., aeration intensity adjustment range 0-5 m³ / h, dosage adjustment step size 0.1 L / h); Reward function (R): comprehensive evaluation of control effect, formula: R = Water quality compliance rate - Total energy consumption - Equipment coordination deviation; where, water quality compliance rate = (1 - number of parameters exceeding the standard / total number of parameters) × 100%; total energy consumption = power of each device × sum of operating time; equipment coordination deviation = Euclidean distance between actual control parameters and theoretical coordination parameters (such as the matching degree between aeration intensity and drug dosage); α, β, γ are weighting coefficients (set according to the aquaculture scenario, such as α=0.6, β=0.3, γ=0.1 for pond aquaculture).

[0101] Output layer: Outputs the specific control parameters of each treatment sub-sequence associated equipment, for example: biological filter: aeration intensity adjusted to 3.2 m³ / h, carbon source dosage adjusted to 0.8 L / h; sedimentation tank: inlet valve opening adjusted to 60%, sludge pump start frequency adjusted to 1 time / 30 minutes; constructed wetland: water level regulator lowered to 0.4 m, circulating water pump power adjusted to 5 kW.

[0102] After determining the model architecture, the next step is to define the loss function. In this invention, a "multi-objective composite loss" is adopted, which simultaneously constrains the control accuracy, energy consumption, and stability: Loss = Loss_quality + λ × Loss_energy + μ × Loss_stability, where Loss_quality is the mean square error (MSE) between the predicted water quality and the target water quality, such as the deviation between the predicted ammonia nitrogen value and the target value (0.5 mg / L), Loss_energy is the deviation between the actual energy consumption and the theoretical minimum energy consumption, Loss_stability is the fluctuation range of continuous control commands (to avoid frequent start-ups and shutdowns of equipment, such as the difference between adjacent adjustments of aeration intensity), and λ and μ are preset weights (λ = 0.3, μ = 0.2). The above formulas are all calculated by removing the dimensions and taking their numerical values. The formulas are obtained by software simulation based on a large amount of data, and are the closest to the real situation. The preset parameters and preset thresholds in the formulas are set by those skilled in the art according to the actual situation or obtained by simulation based on a large amount of data. Then, the prepared training set is input into the selected deep learning model to start training. During training, the weights are continuously updated using the backpropagation algorithm, causing the loss function to gradually decrease until it reaches a stable state. During this period, techniques such as early stopping are used to avoid overfitting. In addition to the basic training process, various model parameters, including the learning rate, batch size, and regularization coefficient, are fine-tuned using grid search.

[0103] Once the model training is complete and the parameters have been tuned, a final evaluation is performed using a test set to obtain the model's evaluation results. These results include classification metrics such as accuracy, recall, and F1 score. Based on the evaluation results on the test set, it is determined whether the model has met the expected standards. If the requirements are met, the model parameters are saved and deployment is prepared; otherwise, it is necessary to return to a previous stage to re-examine issues such as data quality, model structure, or training strategy.

[0104] It should be further explained that, in the specific implementation process, the process of outputting the specific control parameters of each processing sub-sequence associated device and performing preventive control includes:

[0105] Use the end timestamp of the monitoring period as the control time point;

[0106] When the processing subsequence is in a normal state, at the control time point, the multi-source data of the monitoring cycle is input into the multi-device collaborative control model. Based on the multi-device collaborative control model, the specific control parameters of the associated devices of each processing subsequence in the next acquisition cycle are output. Based on the specific control parameters of the associated devices of each processing subsequence in the next acquisition cycle, the associated devices of each processing subsequence are controlled.

[0107] The multi-device collaborative control model can accurately identify core causal links such as "aeration intensity → dissolved oxygen → ammonia nitrogen" from multi-parameter data of aquaculture wastewater, avoiding interference from spurious correlations such as "water temperature and ammonia nitrogen", and providing clear control logic for intelligent control. For example, when ammonia nitrogen exceeds the standard, the multi-device collaborative control model will prioritize increasing aeration intensity (direct causality) rather than blindly adjusting water temperature (spurious correlation), thereby improving control efficiency and reducing energy consumption.

[0108] like Figure 2 As shown, the intelligent control system for aquaculture wastewater based on multi-parameter feedback includes a control center, which is connected to a data acquisition module, a real-time monitoring module, a feature extraction module, a pseudo-correlation analysis module, and a collaborative control module.

[0109] The data acquisition module is used to obtain information on the current aquaculture wastewater treatment process, set up multi-source monitoring points based on the process information, and obtain key water quality monitoring indicators for each treatment subsequence.

[0110] The real-time monitoring module is used to monitor each processing subsequence in real time and classify the processing subsequence into normal or abnormal states based on the monitoring results.

[0111] The feature extraction module is used to perform critical change analysis on the historical monitoring periods in which the processed subsequences are marked as abnormal, obtain the critical feature set of the historical monitoring period, compare the significant differences of the critical feature set, and construct the critical correlation feature network of abnormal indicators of the abnormal historical monitoring period.

[0112] The pseudo-correlation analysis module is used to perform pseudo-correlation analysis on the critical correlation feature network of various abnormal indicators, obtain pseudo-correlation nodes and interference nodes, and construct the critical directed acyclic graph and critical interference link graph of abnormal indicators.

[0113] The collaborative control module is used to construct a multi-device collaborative control model based on critical directed acyclic graphs and critical interference link graphs of several abnormal indicators, output the specific control parameters of the devices associated with each processing sub-sequence, and perform preventive control.

[0114] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An intelligent control method for aquaculture tail water based on multi-parameter feedback, characterized in that, Comprising the following steps: Step s1: Obtain current aquaculture tail water treatment process information, set multiple source monitoring points according to the process information, and obtain key water quality monitoring indicators of each processing sub-sequence; Step s2: Real-time monitoring of each processing sub-sequence, and dividing the processing sub-sequence into normal state or abnormal state according to the monitoring result; Step s3: Critical change analysis on the historical monitoring period of the processing sub-sequence marked as abnormal state, obtaining the critical feature set of the historical monitoring period, comparing the significant difference of the critical feature set, and constructing the critical correlation feature network of the abnormal index of the abnormal historical monitoring period; Step s4: Extract the edge weight coefficient between each node in the critical correlation feature network and the abnormal index, mark the node with the edge weight coefficient greater than the preset edge weight coefficient threshold as the first node, perform pseudo correlation analysis on each first node, obtain the independent correlation coefficient between each first node and the abnormal index, mark the first node with the independent correlation coefficient greater than or equal to the preset correlation significance threshold as the main node, and mark the first node with the independent correlation coefficient less than the preset correlation significance threshold as the pseudo correlation node; Construct a critical directed acyclic graph of the abnormal index according to the abnormal index and the main node; Pseudo correlation analysis between the pseudo correlation node and the main node, mark the pseudo correlation node with the independent correlation coefficient greater than or equal to the correlation significance threshold as the interference node, and mark the pseudo correlation node with the independent correlation coefficient less than the correlation significance threshold as the irrelevant node; Pseudo correlation analysis between the irrelevant node and the interference node, mark the irrelevant node with the independent correlation coefficient greater than or equal to the correlation significance threshold as the secondary interference node, and construct a critical interference link graph of the abnormal index according to the main node, the interference node and the secondary interference node; Step s5: Based on the critical directed acyclic graph and the critical interference link graph of several abnormal indexes, a multi-device collaborative control model is constructed, and the specific control parameters of the associated devices of each processing sub-sequence are output and preventive control is performed.

2. The multi-parameter feedback based intelligent control method of aquaculture effluent according to claim 1, characterized in that, The process of obtaining the current aquaculture tail water treatment process information, setting multiple source monitoring points according to the process information, and obtaining the key water quality monitoring indicators of each processing sub-sequence includes: Obtain the process flow characteristics of the aquaculture tail water treatment facility, and divide the aquaculture tail water treatment process into several processing sub-sequences according to the process flow characteristics; Set multiple source monitoring points in several processing sub-sequences, and obtain monitoring indicators of each multiple source monitoring point by data retrieval according to the functional characteristics in the process flow characteristics of the corresponding processing sub-sequence, the multiple source monitoring points collect multiple source data and mark the monitoring time according to the monitoring indicators, and set the monitoring period.

3. The multi-parameter feedback based intelligent control method of aquaculture effluent according to claim 2, wherein, The process of real-time monitoring of each processing sub-sequence includes: The preset standard threshold interval corresponding to each type of monitoring index of each multi-source monitoring point is extracted from the multi-source data of each multi-source monitoring point. The value time sequence corresponding to each type of monitoring index of the multi-source monitoring point is compared with the corresponding standard threshold interval, and the cumulative time of each type of monitoring index not located in the corresponding standard threshold interval is obtained. A preset error threshold is set. If there is a corresponding cumulative time greater than the error threshold, the monitoring index is marked as an abnormal index, and the processing subsequence is marked as an abnormal state. If the cumulative time of each type of monitoring index is less than the error threshold, the processing subsequence is marked as a normal state.

4. The multi-parameter feedback based intelligent control method of aquaculture effluent according to claim 3, characterized in that, The process of critical change analysis on the historical monitoring period in which the processing subsequence is marked as an abnormal state includes: The historical monitoring period in which each processing subsequence is marked as an abnormal state is marked as an abnormal historical monitoring period. The time when the monitoring index in the abnormal historical monitoring period is marked as an abnormal index is marked as a critical timestamp. The value time sequence corresponding to each type of monitoring index within the start timestamp and the critical timestamp of the abnormal historical monitoring period is extracted. The value time sequence corresponding to each type of monitoring index is subjected to critical change analysis, and a critical feature set between each type of monitoring index is obtained.

5. The multi-parameter feedback based intelligent control method of aquaculture effluent according to claim 4, wherein, The process of significant difference comparison on the critical feature set to construct the critical correlation feature network of the abnormal index of the abnormal historical monitoring period includes: The historical monitoring period in which each processing subsequence is marked as a normal state is marked as a normal historical monitoring period. It is determined whether the next historical monitoring period of the normal historical monitoring period is an abnormal historical monitoring period. If so, the regular feature set between each type of monitoring index in the normal historical monitoring period is obtained. The regular feature set is compared with the critical feature set of the abnormal historical monitoring period to obtain a key monitoring index, and the abnormal index in the key monitoring index is extracted. Each key monitoring index and abnormal index is used as a node, and each node is interconnected to construct a critical correlation feature network of the abnormal index of the abnormal historical monitoring period. The edge weight coefficients between the nodes in the critical correlation feature network are set.

6. The multi-parameter feedback based intelligent control method of aquaculture effluent according to claim 5, wherein, The process of constructing a multi-device collaborative control model based on the critical directed acyclic graph and the critical interference link graph of a plurality of abnormal indexes includes: A multi-device collaborative control model is constructed. The value time sequence corresponding to each type of monitoring index in each historical monitoring period of each processing subsequence, the critical directed acyclic graph and the critical interference link graph of the abnormal index are extracted as training data. The multi-device collaborative control model is trained using the training data to obtain a trained multi-device collaborative control model.

7. The multi-parameter feedback based intelligent control method of aquaculture effluent according to claim 6, wherein, The process of preventive control includes: The end timestamp of the monitoring period is used as a control time point; When the processing sub-sequence is in a normal state, the multi-source data of the monitoring period is input into the multi-device collaborative regulation model at the regulation time point, specific regulation parameters of the associated devices of each processing sub-sequence in the next collection period are output according to the multi-device collaborative regulation model, and the associated devices of each processing sub-sequence are regulated according to the specific regulation parameters of the associated devices of each processing sub-sequence in the next collection period.

8. The intelligent control system for aquaculture tail water based on multi-parameter feedback, particularly applied to the intelligent control method for aquaculture tail water based on multi-parameter feedback according to any one of claims 1 to 7, characterized in that, The control center is communicatively connected with a data acquisition module, a real-time monitoring module, a feature extraction module, a pseudo-correlation analysis module, and a collaborative regulation module; The data acquisition module is configured to obtain current aquaculture tail water treatment process information, set multi-source monitoring points according to the process information, and obtain key water quality monitoring indexes of each processing sub-sequence. The real-time monitoring module is configured to perform real-time monitoring on each processing sub-sequence, and divide the processing sub-sequences into normal states or abnormal states according to the monitoring results. The feature extraction module is configured to perform critical change analysis on historical monitoring periods of the processing sub-sequences marked as abnormal states, obtain a critical feature set of the historical monitoring periods, perform significant difference comparison on the critical feature set, and construct a critical correlation feature network of abnormal indexes of abnormal historical monitoring periods. The pseudo-correlation analysis module is configured to perform pseudo-correlation analysis on the critical correlation feature network of each abnormal index, obtain pseudo-correlation nodes and interference nodes, construct a critical directed acyclic graph and a critical interference link graph of the abnormal indexes. The collaborative regulation module is configured to construct a multi-device collaborative regulation model based on the critical directed acyclic graph and the critical interference link graph of the abnormal indexes, output specific regulation parameters of the associated devices of each processing sub-sequence, and perform preventive regulation.

Images