Method and system for dynamic evaluation of aquaculture capacity

Abstract

This invention provides a method and system for dynamic assessment of aquaculture capacity, relating to the field of data processing technology. The method includes: generating multiple state evolution trajectories within a preset future time window based on dynamic causal fingerprints; obtaining a feasible solution set for aquaculture density regulation based on the disease probability and mortality distribution predicted by the state evolution trajectories; dynamically calibrating and correcting prediction biases of the state evolution trajectories based on the latest real-time monitoring data stream to obtain calibrated results; updating a preset structural causal model using the calibrated results to obtain a feasible solution set for aquaculture density regulation; and selecting and outputting recommended dynamic adjustment amounts for aquaculture capacity based on the feasible solution set for aquaculture density regulation. This invention can deeply integrate real-time monitoring data and has the ability to perform counterfactual inference and dynamic calibration, enabling accurate and adaptive dynamic assessment of aquaculture capacity.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for dynamic assessment of aquaculture capacity. Background Technology

[0002] Aquaculture, especially closed recirculating aquaculture systems, is a highly intensive production model. In this model, stocking capacity (i.e., reasonable stocking density) is a core variable affecting economic benefits and ecological risks. Excessively high stocking densities exacerbate water pollution, increase environmental stress, and significantly increase the risk of disease outbreaks; while excessively low densities lead to low facility and resource utilization rates, harming economic benefits.

[0003] Currently, the assessment of aquaculture capacity largely relies on static empirical models or statistical analysis based on historical data. Common methods include calculations based on empirical formulas for carrying capacity using key water quality parameters such as dissolved oxygen and ammonia nitrogen, or establishing statistical relationships between historical density and disease incidence using regression models. However, these traditional methods have some drawbacks:

[0004] Aquaculture systems are dynamic systems with highly coupled and real-time changes in multiple variables (water temperature, dissolved oxygen, pH, ammonia nitrogen, fish behavior, pathogen numbers, etc.). Some existing models struggle to characterize the complex nonlinear interactions and time lag effects among these variables. When environmental factors fluctuate in real time, the models cannot adjust the capacity assessment results in a timely manner, resulting in recommended stocking densities that are either too conservative, failing to fully utilize the system's potential, or too optimistic, ignoring the potential accumulation of risks.

[0005] In recent years, although some studies have attempted to introduce causal inference models (such as the CausalTune-Fish framework) to improve the interpretability of disease causes and have made progress in identifying high-dimensional confounding variables, their training data are still limited to records of diseases that have already occurred. When faced with counterfactual inference tasks, i.e., simulating the evolution of a system under intervention conditions that have not occurred (such as changing density), such models are still limited by methodological and data foundations. Some cannot effectively construct and extrapolate potential high-risk states, making it difficult to improve the accuracy of dynamic capacity assessment. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a dynamic assessment method and system for aquaculture capacity, which can deeply integrate real-time monitoring data and has the ability to perform counterfactual inference and dynamic calibration, thereby achieving accurate and adaptive dynamic assessment of aquaculture capacity.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0008] Firstly, a method for dynamically assessing aquaculture capacity, the method comprising:

[0009] Step 1: Collect real-time water quality sensor data and fish behavior and physiological status monitoring data of the closed recirculating aquaculture system, and integrate historical disease outbreak records. Based on a predefined knowledge base of environmental stress factor association rules, perform time-series alignment and feature fusion of multi-source heterogeneous data to construct a feature matrix.

[0010] Step 2: Extract core causal feature variables based on the feature matrix;

[0011] Step 3: Based on the core causal characteristic variables and the preset structural causal model, generate a dynamic causal fingerprint by calculating the counterfactual distribution of the variables under the preset intervention conditions;

[0012] Step 4: Based on the dynamic causal fingerprint, generate multiple state evolution trajectories within a future preset time window; based on the disease probability and mortality distribution predicted by the state evolution trajectories, solve for the feasible solution set for breeding density regulation.

[0013] Step 5: Dynamically calibrate and correct prediction deviations of the state evolution trajectory based on the latest real-time monitoring data stream to obtain calibrated results; update the preset structural causal model using the calibrated results to obtain a feasible solution set for stocking density regulation; and select and output the recommended dynamic adjustment amount of stocking capacity based on the feasible solution set for stocking density regulation.

[0014] Secondly, a dynamic assessment system for aquaculture capacity includes:

[0015] The data acquisition module is used to collect real-time water quality sensor data and fish behavior and physiological status monitoring data of the closed recirculating aquaculture system, and integrate historical disease outbreak records. Based on a predefined knowledge base of environmental stress factor association rules, it performs time-series alignment and feature fusion of multi-source heterogeneous data to construct a feature matrix.

[0016] The extraction module is used to extract core causal feature variables based on the feature matrix;

[0017] The generation module is used to generate dynamic causal fingerprints based on core causal feature variables and a preset structural causal model by calculating the counterfactual distribution of the set of variables under preset intervention conditions.

[0018] The processing module is used to generate multiple state evolution trajectories within a preset future time window based on dynamic causal fingerprints; and to obtain a feasible solution set for controlling breeding density based on the disease probability and mortality distribution predicted by the state evolution trajectories.

[0019] The determination module is used to dynamically calibrate and correct prediction deviations of the state evolution trajectory based on the latest real-time monitoring data stream to obtain the calibrated results; the calibrated results are used to update the preset structural causal model to obtain a feasible solution set for stocking density regulation; based on the feasible solution set for stocking density regulation, the recommended dynamic adjustment amount of stocking capacity is selected and output.

[0020] The above-described solution of the present invention has at least the following beneficial effects:

[0021] This invention integrates real-time water quality sensor data, fish behavior and physiological data, and historical disease records, combined with a knowledge base of environmental stress factor association rules, to fully capture multivariate coupling relationships. By calculating the counterfactual distribution under intervention conditions through a structural causal model, it accurately quantifies the lagged impact of different stocking density adjustments on the subsequent system state. In a dynamically changing aquaculture environment, it can provide a precise density range that fully utilizes the system's carrying capacity while avoiding disease risks.

[0022] This invention visualizes the abstract system risk state through dynamic causal fingerprints. As an intuitive vector encoding the contribution of risk factors and the real-time risk level, dynamic causal fingerprints can clearly inform users that at the current density, ammonia nitrogen accumulation and fish stress behavior are the main sources of risk, and that the core reason for the reduction in disease risk after adjusting the density is the increase in dissolved oxygen reserve redundancy. This can enhance the interpretability of decision-making suggestions, increase users' trust in the assessment results, and better adapt to the decision-making needs of actual aquaculture scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a dynamic assessment method for aquaculture capacity provided by an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of a dynamic assessment system for aquaculture capacity provided by an embodiment of the present invention. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] like Figure 1 As shown, an embodiment of the present invention proposes a method for dynamic assessment of aquaculture capacity, the method comprising the following steps:

[0027] Step 1: Collect real-time water quality sensor data and fish behavior and physiological status monitoring data of the closed recirculating aquaculture system, and integrate historical disease outbreak records. Based on a predefined knowledge base of environmental stress factor association rules, perform time-series alignment and feature fusion of multi-source heterogeneous data to construct a feature matrix.

[0028] Step 2: Extract core causal feature variables based on the feature matrix;

[0029] Step 3: Based on the core causal characteristic variables and the preset structural causal model, generate a dynamic causal fingerprint by calculating the counterfactual distribution of the variables under the preset intervention conditions;

[0030] Step 4: Based on the dynamic causal fingerprint, generate multiple state evolution trajectories within a future preset time window; based on the disease probability and mortality distribution predicted by the state evolution trajectories, solve for the feasible solution set for breeding density regulation.

[0031] Step 5: Dynamically calibrate and correct prediction deviations of the state evolution trajectory based on the latest real-time monitoring data stream to obtain calibrated results; update the preset structural causal model using the calibrated results to obtain a feasible solution set for stocking density regulation; and select and output the recommended dynamic adjustment amount of stocking capacity based on the feasible solution set for stocking density regulation.

[0032] In this embodiment, a combination of multi-source time-series data fusion, core causal feature extraction, and dynamic causal fingerprint generation achieves precise adaptation to dynamic systems. On the one hand, by integrating real-time water quality sensor data, fish behavior and physiological data, and historical disease records, and combining them with a knowledge base of environmental stress factor association rules, it breaks through the limitations of traditional models that rely solely on single or a few parameters such as dissolved oxygen and ammonia nitrogen, or are limited to historical data statistics, and fully captures multivariate coupling relationships. On the other hand, by calculating the counterfactual distribution under intervention conditions through a structural causal model, it accurately quantifies the lagged impact of different aquaculture density adjustments on the subsequent system state, avoiding the problems of conservative waste of facility potential and optimistic neglect of risk accumulation caused by the inability of traditional models to respond to real-time fluctuations. It can provide a precise density range that fully utilizes the system's carrying capacity and avoids disease risks in a dynamically changing aquaculture environment.

[0033] This invention utilizes dynamic causal fingerprinting and state evolution trajectory generation technology. Dynamic causal fingerprinting can encode the contribution and immediate risk level of various risk factors (water temperature, pathogen quantity, fish behavior, etc.). Based on this, it can extrapolate state trajectories within multiple future time windows, simulate the system evolution corresponding to the uninterrupted behavior of changing aquaculture density, accurately predict the probability of potential disease outbreaks and mortality distribution, and solve the shortcomings of traditional causal models that can only explain existing diseases and cannot predict risks that have not yet occurred, thereby improving the accuracy of dynamic capacity assessment.

[0034] This invention establishes a real-time closed-loop mechanism of monitoring-prediction-calibration-updating, forming dynamic adaptive capabilities. By acquiring the latest monitoring data stream in real time, the predicted state evolution trajectory is dynamically calibrated and biased, which can promptly offset prediction errors caused by sudden fluctuations in environmental factors (such as sudden changes in water temperature or sudden drops in dissolved oxygen). At the same time, the calibration results are used to update the parameters of the structural causal model online, enabling the model to continuously learn from new data and gradually optimize the modeling accuracy for high-dimensional, time-varying, coupled systems. This mechanism not only solves the static defects of traditional models that provide a one-time solution, but also specifically addresses the imbalance problem in aquaculture data, where there are few disease events and many normal state data. It improves the reliability of predicting low-probability disease events and ensures that capacity assessment recommendations are always synchronized with the actual state of the system, taking into account both timeliness and stability.

[0035] This invention visualizes the abstract system risk state through dynamic causal fingerprinting. As an intuitive vector encoding the contribution of risk factors and the real-time risk level, dynamic causal fingerprinting can clearly inform users that at the current density, ammonia nitrogen accumulation and fish stress behavior are the main sources of risk, and that the core reason for the reduction in disease risk after adjusting the density is the increase in dissolved oxygen reserve redundancy. This breaks through the limitations of traditional models that only output density values ​​and lack decision-making basis. This design upgrades capacity assessment from simply providing results to a complete solution that includes results and causes, enhances the interpretability of decision-making suggestions, increases users' trust in the assessment results, and is more suitable for the decision-making needs of actual aquaculture scenarios.

[0036] In a preferred embodiment of the present invention, step 1 includes:

[0037] Step 100 involves cleaning and interpolating missing values ​​in the real-time water quality sensor data to generate a standardized time-series sequence of first water quality parameters. This includes: collecting raw data from water quality sensors in a closed recirculating aquaculture system, including key water quality parameters such as dissolved oxygen, ammonia nitrogen, pH, water temperature, and nitrite concentration; cleaning the raw data to remove outliers caused by sensor malfunctions or signal interference (such as sudden increases / decreases exceeding the reasonable range for aquaculture); supplementing missing values ​​caused by equipment offline or transmission interruptions during data acquisition using linear interpolation, moving average interpolation, and other methods adapted to aquaculture water quality data, taking into account the time-series characteristics of water quality parameters (such as the gradual changes in water temperature and dissolved oxygen); and finally standardizing the cleaned and supplemented water quality data (such as unifying dimensions and mapping values ​​to the 0-1 range) to generate a standardized time-series sequence of first water quality parameters arranged in order of timestamps.

[0038] Step 101: Input the fish behavior and physiological state monitoring data into the preset behavior and physiological analysis model, process and transform it into a standardized fish state time series, including: organizing the raw monitoring data of fish behavior and physiological state, where the behavior data includes 1080P, 30 frames / second video stream data and motion sensor data with a sampling frequency of 10 times / minute, specifically including the fish swimming frequency (normal range 5-12 times / minute), school density (tails / m 2 The data is categorized into low density (≤5), medium density (6-15), and high density (≥16). Feeding activity and abnormal swimming patterns (such as belly-up or hitting the wall, judged by hitting the wall ≥3 times / hour or swimming speed ≤0.1m / s / ≥0.9m / s) are also considered. Physiological data includes sampling and testing data from 30 fish per pond every 24 hours, specifically covering fish group growth and weight, blood biochemical indicators (such as ALT normal range 20-40U / L, blood glucose 4.5-7.0mmol / L), stress hormone levels (cortisol normal range 0.05-0.2μg / dL), and surface health status. The raw data is preprocessed to remove redundant information (such as fluctuations in individual fish swimming direction, features with a correlation of less than 0.3 with health status), systematic errors in calibrating monitoring equipment (including high-definition cameras, motion sensors, and physiological detection equipment), and invalid data. Missing data is supplemented using the average of the same scene and time period, with the supplementation amount controlled to ≤3% to ensure data integrity and accuracy.

[0039] The preprocessed data is then input into a preset behavioral and physiological analysis model. This model is trained on 80,000 labeled samples covering 12 aquaculture bases, 30 aquaculture ponds (15 each of Litopenaeus vannamei and California bass) over a 12-month period. It adopts a multimodal fusion hierarchical architecture (including a data input layer, feature extraction layer, state quantification layer, and output standardization layer), which can accurately identify four states of fish populations: normal, mild stress, severe stress, and disease precursors. The accuracy rate of behavioral feature recognition reaches 89.5%, and the accuracy rate of physiological state assessment is ≥92%. It can effectively identify fish behavior patterns and quantify the degree of physiological stress. The model will transform unstructured video data and semi-structured sensor data into structured quantitative indicators, including feeding activity (0-10 points, below 2 points is severely abnormal, and above 6 points is normal), stress level (1-5 levels, corresponding to a standardized value range of 0.2-1.0), and abnormal behavior incidence (number of abnormal behaviors per hour / total number of behaviors × 100%, threshold 14%). At the same time, it will establish a correlation mapping between behavior and physiological characteristics (e.g., swimming speed ≤0.2m / s + cortisol concentration ≥0.3μg / dL corresponds to stress level 4).

[0040] Finally, these quantitative indicators are standardized, and all indicators are uniformly mapped to the 0-1 range (e.g., stress level 1 corresponds to 0.2, and stress level 5 corresponds to 1.0). The data are aligned according to the 1-hour time granularity to generate a standardized fish swarm status time series containing 19 features (8 behavioral features + 6 physiological features + 5 fusion indicators). This ensures that it is completely consistent with the first water quality parameter time series generated in step 100 in terms of time granularity and dimensional range, with a data missing rate of ≤1%. It can be directly used for subsequent time alignment and feature fusion.

[0041] The process of constructing the pre-defined behavioral and physiological analysis model is as follows:

[0042] For the raw fish behavior and physiological data in step 101, a full-process transformation from multi-source input to structured quantification to standardized output is achieved, outputting a fish state time series that is completely aligned with the water quality parameter time series. Specifically, behavioral data such as swimming frequency, grouping degree, feeding activity, and abnormal swimming mentioned in step 101, as well as physiological data such as body length and weight, blood biochemical indicators, stress hormone levels, and body surface health status, need to be converted into quantitative indicators such as feeding activity, stress level, and abnormal behavior incidence. This ensures that the state identification accuracy is not less than 92% and the quantification error is not more than 4%, adapting to the closed recirculating aquaculture scenario involved in step 101, focusing on two major species, Litopenaeus vannamei and California bass, covering the entire aquaculture stage from seedling to growth.

[0043] The input data is adapted to the monitoring type in step 101. Behavioral data includes a 1080P video stream at 30 frames per second, simultaneously collected by three cameras at the head, tail, and center of the pool, as well as motion sensor data sampled 10 times per minute. Physiological data includes laboratory test data every 24 hours, with 30 fish sampled from each pool each time, and real-time physiological sensor data every 15 minutes. The output time series has a uniform time granularity of 1 hour per data point, perfectly matching the first water quality parameter time series generated in step 100.

[0044] Each layer of the architecture is designed around the original data from step 101, enabling accurate data reception and transformation:

[0045] The data input layer is adapted to three main branches based on the data type in step 101. The behavioral video branch receives video data such as abnormal swimming and clustering degree from step 101, with a single segment duration of 1 hour. It supports simultaneous access of data from 3 cameras and automatically associates the corresponding identification information of the breeding pond. The behavioral sensor branch receives numerical data such as swimming frequency and feeding activity from step 101, reads in real time at a sampling frequency of 10 times per minute, and filters invalid sampling points. The physiological data branch is divided into two access points. The real-time branch receives stress hormone level data, mainly cortisol, while the offline branch receives laboratory data such as body length and weight, blood biochemical indicators, and body surface health score from step 101, and automatically matches the sampling time with the breeding pond information.

[0046] The feature extraction layer extracts core features from the data type in step 101 and removes redundant information. Behavioral characteristics were extracted from the video, including abnormal swimming behavior identified in step 101. Abnormal behavior was defined as swimming at least 3 times per hour, swimming speed not exceeding 0.1 m / s or not less than 0.9 m / s, and the degree of grouping, graded as no more than 5 fish per square meter, 6 to 15 fish per square meter, and no less than 16 fish per square meter. Swimming frequency was extracted from sensor data, with a normal range of 5 to 12 times per minute. The percentage of time spent feeding was also extracted, with a normal range of 20% to 40%. A total of 8 core behavioral characteristics were identified, and redundant information with a correlation less than 0.3 with health status, such as fluctuations in individual swimming direction, was removed. Physiological characteristics were extracted from step 101, including stress hormones (cortisol, normal range 0.05 to 0.2 μg / dL), blood biochemical indicators (ALT, normal range 20 to 40 IU / L, blood glucose, normal range 4.5 to 7.0 mmol / L), body length and weight growth rate, and body surface health score, totaling 6 core physiological characteristics. Systematic errors in laboratory testing equipment were calibrated, such as an ALT detection deviation not exceeding 2 IU / L.

[0047] The state quantification layer establishes a mapping between the behavior and physiological data from step 101, transforming them into structured indicators. Rules are set based on the characteristics of the data from step 101, such as a swimming frequency of no more than 3 times per minute and cortisol levels not lower than 0.3 micrograms per deciliter corresponding to stress level 4; a feeding time not exceeding 10% of the total time spent feeding and blood glucose levels not lower than 8.0 mmol / L corresponding to early signs of illness; and hitting the wall at least 3 times per hour and a body surface score not exceeding 2 points corresponding to severe stress. The quantification standards align with the requirements of step 101. Feeding activity is quantified from 0 to 10 points, with a score not exceeding 2 points indicating severe abnormality, adapting to the abnormal feeding judgment in step 101. The incidence of abnormal behavior is calculated by dividing the number of abnormal behaviors per hour by the total number of behaviors and then multiplying by 100%, with a threshold set at 14%, adapting to the abnormal swimming monitoring in step 101. The output standardization layer maps the quantitative indicators to the 0-1 range according to the alignment requirements of step 101 and the water quality time series. Stress level 1 corresponds to 0.2, level 5 corresponds to 1.0, and feeding activity 10 points corresponds to 1.0. The data is integrated according to the 1-hour time nodes to generate a single time series data with 19 features, including 8 behavioral features, 6 physiological features, and 5 fusion indicators. The data missing rate does not exceed 1%, ensuring that it can be directly aligned with the first water quality parameter time series of step 100 for time alignment and feature fusion.

[0048] The behavioral feature recognition module processes the behavioral monitoring data from step 101, primarily adapting to dual-source data from video and sensors. The video module accurately identifies abnormal swimming movements such as belly-up and wall-banging, with an error margin of no more than ±1 time per hour. The sensor module quantifies swimming frequency and feeding activity, precisely integrating with the pre-processed behavioral data from step 101 to output eight standardized behavioral features. The physiological state assessment module receives the physiological detection data from step 101, quickly assessing the rationality of body length and weight gain, the degree of abnormality in blood indicators, and stress hormone levels, outputting physiological assessment results within 10 minutes with an accuracy rate of no less than 92%. The quantification error for cortisol and ALT indicators is no more than 3%, meeting the detection accuracy requirements of physiological data in step 101. The cross-modal fusion and standardization module integrates the outputs of the first two modules, handling potential conflicts in the step 101 data, such as normal behavior but abnormal physiological indicators. It fuses data with a weight of 0.6 for physiological indicators and 0.4 for behavioral indicators, incorporating the trend correction results from the past 3 hours, ultimately outputting a standardized time-series sequence aligned with water quality data, fulfilling the core data transformation requirements of step 101.

[0049] Model training process, training dataset construction:

[0050] Multi-scenario data collection involved collecting 12-month full-cycle data from 12 closed recirculating aquaculture bases and 30 ponds, based on the monitoring data types outlined in step 101. This included 15 ponds each for Litopenaeus vannamei and California bass, covering all data dimensions involved in step 101. Behavioral data included over 5,000 hours of 1080P video, including abnormal scenarios such as belly-up and wall-banging; over 1 million motion sensor data points, including swimming frequency and cluster density; and over 80,000 feeding monitoring data points. Physiological data included over 20,000 blood biochemistry test reports, including ALT and blood glucose; over 50,000 real-time cortisol data points; and over 30,000 records of body length, weight, and body surface scores. Supporting data included over 1.2 million synchronized water quality data points and over 300 disease event records, ensuring a high degree of consistency between the dataset and the actual monitoring data from step 101.

[0051] Data preprocessing followed the preprocessing standards of step 101, removing over 200 hours of blurry video, over 50,000 sensor malfunction values, and over 300 contaminated physiological samples, with the total removal amount not exceeding 5%. Missing data was supplemented using the average of the same scene and time period, with the supplementation amount controlled to not exceed 3%, consistent with the original data preprocessing workflow of step 101. Five experts with over 10 years of experience were invited to label over 80,000 samples with status tags related to step 101, based on disease records. The labeling standards precisely corresponded to quantitative indicators: normal state was defined as feeding activity of 6 to 10 points, cortisol of 0.05 to 0.2 micrograms per deciliter, and no abnormal swimming; mild stress was defined as feeding activity of 4 to 6 points and cortisol of 0.2 to 0.3 micrograms per deciliter; severe stress was defined as feeding activity of 2 to 4 points and hitting the wall at least 3 times per hour; and disease precursors were defined as feeding activity of no more than 2 points and cortisol of no less than 0.5 micrograms per deciliter. The cross-verification pass rate is no less than 95%, and the data is divided into training set, validation set and test set in a ratio of 7:2:1, with more than 56,000 data entries in the training set, more than 16,000 data entries in the validation set and more than 8,000 data entries in the test set.

[0052] The initial parameters are based on industry benchmark values ​​from the data in step 101, tailored to the characteristics of the target species. The normal range for the swimming frequency in the behavior module is 5 to 12 times per minute, and the cluster density threshold is 16 tails per square meter. The benchmark values ​​for cortisol in the physiological module are 0.12 micrograms per deciliter and ALT are 30 units per liter, perfectly matching the normal ranges of the original data from step 101, ensuring the initial model is adapted to actual monitoring data.

[0053] The behavioral feature recognition module uses behavioral data from the training set as supervision signals, with a batch size of 32 and 100 iterations. It focuses on optimizing the recognition accuracy of abnormal swimming and feeding activity mentioned in step 101. Training stops after the accuracy rate stabilizes at or above 88% for 10 consecutive rounds, ultimately achieving a behavioral recognition accuracy of 89.5%, meeting the behavioral data recognition requirements of step 101. The physiological state assessment module uses physiological data from the training set as supervision signals, with a batch size of 32 and 80 iterations. It focuses on core indicators such as cortisol, ALT, body length, and weight from step 101. Training stops when the quantification error does not exceed 5%, ultimately achieving a physiological indicator quantification error of 4.2%, meeting the physiological data quantification requirements of step 101.

[0054] Cross-modal fusion training and end-to-end optimization:

[0055] Cross-modal fusion training simulates the dual-source data input scenario of step 101. The training module learns the correlation between abnormal swimming frequency and elevated cortisol, and between a sudden drop in feeding and abnormal blood glucose. When handling data conflicts, such as normal behavior but elevated cortisol, corrections are made according to the characteristics of the data in step 101. The consistency of the fused data reaches 93%, ensuring the accuracy of the dual-source data fusion. End-to-end full-process training integrates all modules, targeting the deviation between the output indicators and expert annotations, and iterates for 120 rounds, with each round improving the validation set accuracy by at least 0.2%. The final validation set state recognition accuracy reaches 91.2%, with a quantization error of 3.8%, and the generated time series fully meets the alignment requirements of step 101 and water quality data.

[0056] Testing was conducted using over 16,000 validation dataset entries, focusing on verifying the compatibility with the data from step 101. The overall state recognition accuracy reached 91.5%, with abnormal movement recognition accuracy at 92.8% and cortisol anomaly assessment accuracy at 93.1%. The average quantization error was 3.7%, all not exceeding 5%. Regarding temporal consistency, the match with actual data changes reached 92%, with no lag or lead exceeding one hour, meeting the temporal alignment requirements of step 101. To address the low accuracy of mild stress recognition, 2,000 additional scenario-specific data entries were added, aligning with the mild anomaly data characteristics of step 101. The weights of these features were strengthened, and the fusion module was retrained. After this retraining, the mild stress recognition accuracy improved to 91.7%, and the overall accuracy reached 92.3%. Simultaneously, the threshold for the occurrence rate of abnormal behavior was adjusted from 15% to 14% to reduce the misjudgment rate of the data in step 101. Parameters were fine-tuned using 2000 historical data points from the target aquaculture scenario to adapt to the characteristics of specific aquaculture species. The normal range for California bass swimming speed was adjusted to 0.15 to 0.85 meters per second, the normal range for cortisol was adjusted to 0.06 to 0.22 micrograms per deciliter, and the normal range for Litopenaeus vannamei cluster density was adjusted to 8 to 18 shrimp per square meter. This ensured that the model was adapted to the data characteristics of the actual aquaculture species in step 101. After fine-tuning and verification, the alignment error between the output time series and water quality data did not exceed 1 minute, and it could be directly used for time matching and feature fusion in subsequent steps.

[0057] The optimized parameters are solidified, and operating rules are set. A single processing run of 1000 data points of type 101 takes no more than 5 minutes, and parameters are automatically calibrated every 24 hours to meet the high-efficiency requirements of actual monitoring. Simultaneously, interpretation rules for indicators corresponding to the data from step 101 are established, such as stress level 3 being cortisol 0.3 to 0.4 micrograms per deciliter, feeding activity 4 to 5 points, and abnormal behavior incidence 15% to 20%, providing a clear basis for subsequent feature fusion and key factor identification in step 101.

[0058] Throughout the model's construction and training, precise adaptation to data dimensions, replication of preprocessing logic, and alignment with quantization standards ensure a complete correspondence between input data, model processing, and output sequence. The trained model efficiently transforms the multi-source monitoring data from step 101 into a standardized time-series sequence, achieving a state recognition accuracy of 92.3% and a quantization error of 3.7%. This not only meets the requirements for detail depth but also forms a highly closed loop with the preceding steps and data.

[0059] Step 102 involves matching and aligning the recorded time points of historical disease outbreaks with the time series of the first water quality parameters and the time series of fish population status. This ensures that the water quality and fish population status within the matched time period are labeled as contextual state data corresponding to the disease event. This includes: reviewing historical disease outbreak records to clarify the outbreak time / time period, disease type (e.g., bacterial gill rot, viral hemorrhagic disease), and scope of impact for each disease event; using time as the core benchmark, accurately matching the time nodes of disease events with the timestamps of the time series of the first water quality parameters and the time series of fish population status. For example, if a disease outbreak occurs at time T0, extract the water quality parameter time series data and fish population status time series data within a preset time period before and after T0 (e.g., from T0-72 hours to T0+24 hours); labeling the water quality and fish population data within these matched time periods as contextual state data corresponding to the disease event, establishing a one-to-one correspondence between disease events and time series state data, and ensuring that each disease record is supported by corresponding water quality and fish population status data.

[0060] Step 103: Match the context state data with a preset environmental stress factor association rule knowledge base to identify key factor combinations that appear before and after the disease event and conform to the stress rules. This includes: Step 1031: Input the context state data into the preset environmental stress factor association rule knowledge base for matching to obtain matching results, including:

[0061] A knowledge base of rules relating environmental stress factors was constructed, and its core components and rule system were defined. This knowledge base is built upon long-term aquaculture practice data from closed recirculating aquaculture systems, industry technical standards, publicly available research literature, and targeted experimental data, forming a standardized and callable set of rules. Its core components include three parts: First, a mapping table between disease types and stress factor combinations, covering prevalent bacterial, viral, parasitic, and stress-related diseases in closed recirculating aquaculture. Each type of disease corresponds to a unique combination of characteristic stress factors, which simultaneously includes subsets of water quality parameters and fish population status parameters. Among them, bacterial diseases correspond to two core characteristic stress factor combinations. Vibrio disease corresponds to the water quality parameter subset of persistently high ammonia nitrogen concentration, low dissolved oxygen concentration, and water temperature in the range of 25-30℃. Fish school status parameter subsets are feeding activity dropping below 6 points, stress level rising to level 3 or above, and abnormal behavior incidence exceeding 15%. Bacterial gill rot corresponds to the water quality parameter subset of pH value deviating from the normal range (6.5-8.5), hydrogen sulfide concentration exceeding the standard, and nitrite content rising. Fish school status parameter subsets are physiological health index dropping below 0.5, significantly increased incidence of abnormal swimming (frequent surfacing for air), and continuous decline in feeding activity. Among viral diseases, the water quality parameters corresponding to viral hemorrhagic disease are: sudden changes in water temperature exceeding 2℃ / 24 hours, and synergistic exceedance of ammonia nitrogen and nitrite. The fish population status parameters are: stress level rising to level 4 or above, abnormal behavior incidence exceeding 20%, and rapid decline in physiological health index. The water quality parameters corresponding to nerve necrosis virus disease are: drastic fluctuations in dissolved oxygen concentration, pH value remaining in the weakly acidic range for a long time, and fish population status parameters are: feeding activity dropping sharply to below 3 points, abnormal agglomeration (dispersed swimming), and persistently high stress level. In parasitic diseases, the subset of water quality parameters corresponding to Ichthyophthirius multifiliis (white spot disease) is water temperature of 18-22℃, increased water turbidity, and low dissolved oxygen content. The subset of fish status parameters is obvious stress response on the body surface, frequent abnormal rubbing against the pond wall, and a gradual decrease in feeding activity. The subset of water quality parameters corresponding to Trichodiniasis (wheelworm disease) is high nitrite concentration, stable water temperature of 20-28℃, and ammonia nitrogen concentration at the critical value. The subset of fish status parameters is low physiological health index, stress level 2-3, and increased frequency of abnormal swimming. The subset of water quality parameters corresponding to stress-related diseases is multi-factor synergistic fluctuation (such as water temperature, pH, and dissolved oxygen all deviating from the safe threshold simultaneously) and a sudden increase in ammonia nitrogen concentration. The subset of fish status parameters is a stress level of 3 or above, a significant decrease in feeding activity, and large fluctuations in physiological health index. Each parameter subset clearly identifies the core indicators and abnormal characteristics directly related to the outbreak of the corresponding disease, ensuring a precise and unique mapping relationship.Second, a single-factor safety threshold and a multi-factor synergistic threshold system are established. The single-factor safety threshold is calibrated separately according to the aquaculture species and growth stage, while the multi-factor synergistic threshold clarifies the safety boundary under the action of different parameter combinations, avoiding the limitations of single threshold judgment. Third, a time-related rule base is established, which clarifies the time interval range from the occurrence of abnormality to the induction of corresponding diseases by various stress factor combinations, the positive correlation between the duration of abnormality and the probability of disease outbreak, and the impact of the temporal sequence of factor abnormalities on disease type. All rule entries are marked with confidence levels, which are determined based on the corresponding data sample size and the validation pass rate.

[0062] The labeled disease context state data is standardized and preprocessed to form an input data format adapted to the knowledge base. The context state data corresponding to each disease event in step 102 is extracted, including time-series datasets of water quality parameters and fish population status parameters within a preset time period before and after the disease outbreak, as well as core information about the disease events. This core information covers the disease type, outbreak time, impact range, and severity. The extracted data undergoes a structured transformation, splitting it according to knowledge base rule entries. Each data entry is associated with a corresponding timestamp, parameter type, specific value, fluctuation range, and duration. Data noise and redundant information are removed, parameter representation standards and units of measurement are standardized, and a standardized data list is generated.

[0063] Based on a forward reasoning algorithm, the standardized preprocessed context state data list is matched against an environmental stress factor association rule knowledge base, one by one. Taking a single standardized data point as a unit, the system first identifies the candidate rule set corresponding to the disease type in the knowledge base based on the core information of the disease event. Then, it compares the water quality parameter values ​​in the data to see if they exceed the corresponding thresholds, and whether the fish population status parameters meet the abnormal characteristics, determining whether the parameter combinations are consistent with the characteristic stress factor combinations in the candidate rule set. Simultaneously, it compares the duration, time interval, and time sequence of parameter anomalies to see if they match the temporal correlation patterns in the candidate rules. For each candidate rule, the fit between the data and the rule is calculated. The fit is weighted by the number of parameter matches, the degree of threshold deviation, and the degree of temporal pattern consistency, forming a ranking result of the fit between a single data point and each candidate rule. A structured matching report is generated based on the matching comparison results, serving as the basis for subsequent processing steps. The report is categorized and summarized by disease events. The report content for each disease event includes: successfully matched rule entries and their corresponding confidence levels, the top three stress factor combinations with the highest matching degree, specific environmental parameters and fish population status indicators that meet the stress rules, the threshold deviation range and duration of each indicator, whether disease-related rules are triggered and the level of the triggered rules, with the rule level being classified based on a combination of matching degree and confidence level; at the same time, it clearly marks abnormal parameters that do not match the corresponding rules, rule entries with matching degree below the preset threshold and their cause analysis, ensuring that the matching results are traceable and verifiable, and providing complete data support for step 1032 to accurately identify key abnormal factors.

[0064] Step 1032: Based on the matching results, identify environmental parameters that continuously exceed the preset threshold in the time series of the first water quality parameter before and after the outbreak of the disease event, and abnormal indicators that show correlated abnormal fluctuations in the time series of fish population status. This includes: based on the matching results of step 1031, screening key abnormal factors before and after the outbreak of the disease event: on the one hand, identifying environmental parameters that exceed the preset safety threshold at multiple consecutive time points from the time series of the first water quality parameter (such as ammonia nitrogen concentration exceeding 0.2 mg / L for 12 consecutive hours and dissolved oxygen concentration below 5 mg / L for 8 consecutive hours); on the other hand, identifying indicators that show correlated abnormal fluctuations compared to the normal state from the time series of fish population status (such as a sudden drop of 50% in feeding activity and a sudden increase in stress hormone levels, and an abnormal increase in the degree of grouping accompanied by a sudden drop in swimming frequency).

[0065] Step 1033 involves combining environmental parameters with abnormal indicators to identify key factor combinations associated with disease outbreaks. This includes: based on the abnormal environmental parameters and abnormal fish population status indicators screened in Step 1032, preliminary correlation integration is performed according to the corresponding disease event types, forming one-to-one candidate combination pairs of abnormal environmental parameters and abnormal fish population status indicators. Subsequently, a time-series correlation analysis is conducted, using the disease outbreak time point as a benchmark to trace the occurrence time, duration, and fluctuation trends of various abnormal indicators. The focus is on verifying the temporal synchronicity and consistency of change trends between abnormal environmental parameters and abnormal fish population status indicators in the candidate combination pairs. Isolated indicators that only fluctuate once and have no time-series correlation or logical causal relationship with the disease outbreak are eliminated, such as non-core parameters whose time interval with the disease outbreak exceeds a preset threshold or fluctuates only instantaneously.Further combining the mapping relationships in the knowledge base of environmental stress factor association rules, the fit between candidate combinations and corresponding disease type characteristic stress factor combinations was verified. Factor combinations with clear temporal correlation, logical causal correlation, and synergistic relationship with the outbreak of this disease event were retained. Specifically, these include: in bacterial diseases, the combination corresponding to vibriosis with dissolved oxygen concentration continuously below 5 mg / L for 10 hours, ammonia nitrogen concentration above 0.3 mg / L for more than 8 hours, and abnormally increased fish schooling degree, stress hormone levels exceeding twice the normal value, and feeding activity reduced to below 6 points; and the combination corresponding to bacterial gill rot. The following combinations are considered indicative of fish disease: pH value deviating from the normal range of 6.5-8.5, hydrogen sulfide concentration exceeding 0.1 mg / L, and fish physiological health index dropping below 0.5, abnormal surfacing for air more than 5 times per hour, and feeding activity continuously declining for more than 3 hours; for viral diseases, viral hemorrhagic septicemia corresponds to a sudden change in water temperature exceeding 2℃ / 24 hours, synergistic exceedance of ammonia nitrogen and nitrite (ammonia nitrogen ≥ 0.2 mg / L and nitrite ≥ 0.1 mg / L), and fish stress level rising to level 4 or above, abnormal behavior incidence exceeding 20%, and physiological health index decreasing by more than 0.05 per hour. Nervous necrosis virus disease corresponds to a dissolved oxygen concentration fluctuation exceeding 2 mg / L / hour, a pH value consistently in the slightly acidic range of 6.0-6.5, a sudden drop in feeding activity to below 3 points, obvious abnormal scattered swimming, and a stress level consistently above 3. Among parasitic diseases, Ichthyophthirius multifiliis disease corresponds to a water temperature maintained at 18-22℃, increased water turbidity lasting for more than 6 hours, dissolved oxygen content below 4 mg / L, frequent rubbing of the pond walls by the fish (more than 8 times per hour), obvious body surface stress response, and a gradual decrease in feeding activity to below 4 points. Trichodiniasis corresponds to a nitrite concentration above 0. The following combinations were identified as key factors directly related to this disease outbreak: ammonia nitrogen concentration of 0.1 mg / L, stable water temperature of 20-28℃, ammonia nitrogen concentration at the critical value of 0.2 mg / L, fish physiological health index below 0.6, stress level 2-3, and abnormal swimming frequency exceeding 10 times per hour. Stress-induced diseases correspond to a combination where water temperature, pH, and dissolved oxygen all deviate from safe thresholds (water temperature fluctuation ±1.5℃, pH deviation 6.5-8.5, dissolved oxygen ≤4.5 mg / L), a sudden increase in ammonia nitrogen concentration to above 0.3 mg / L, fish stress level rising to level 3 or higher, feeding activity dropping below 5 points, and physiological health index fluctuation exceeding 0.1. These combinations, after multiple verifications, were ultimately determined to be the key factor combinations directly related to this disease outbreak. Each combination clearly defines the synergistic characteristics and quantitative standards of core environmental stress parameters and abnormal fish physiological behaviors, accurately reflecting the core driving factors of disease induction.

[0066] Step 104: Using the unified timeline after time alignment as a benchmark, the time series sequences of the first water quality parameters, the time series sequences of fish school states, and the key factor combinations are vector-concatenated and feature-fused to construct a feature matrix. This includes: establishing a unified timeline based on the smallest time granularity of 1 hour for all time series data; accurately mapping the time series sequences of the first water quality parameters generated in Step 100, the time series sequences of fish school states generated in Step 101, and the key factor combinations determined in Step 1033 to this unified timeline according to their corresponding timestamps, ensuring that the three types of data are completely synchronized and aligned in the time dimension, and eliminating discrepancies between time series data. The time granularity differences and misalignment biases were investigated. Subsequently, the time series of the first water quality parameter was transformed into a structured vector containing core indicators such as ammonia nitrogen, dissolved oxygen, pH, water temperature, nitrite, and hydrogen sulfide. Each indicator corresponds to an independent vector dimension, and the values ​​are all standardized. The time series of fish population status was transformed into a structured vector containing quantitative indicators such as feeding activity (0-10 standardized values), stress level (1-5 corresponding to 0.2-1.0 standardized values), abnormal behavior incidence (percentage standardized values), and physiological health index (0-1 interval). Similarly, each indicator corresponds to an independent dimension. Simultaneously, key factor combinations were labeled to the corresponding positions of the two types of vectors according to the corresponding time nodes, clarifying the synergistic characteristics of core environmental stress parameters and abnormal fish population indicators related to disease outbreaks at each time point. Furthermore, the labeling results corresponded to the characteristic stress factor combinations of the corresponding diseases in the environmental stress factor association rule knowledge base.

[0067] Finally, through feature fusion processing, the two types of structured vectors and key factor combination annotation information are concatenated point by point according to time nodes, preserving the temporal and causal relationships between each factor, especially the synergistic relationship between abnormal environmental parameters, fish population state fluctuations, and key factor combinations, ultimately constructing a feature matrix. This feature matrix uses each 1-hour time node as a row, with each row corresponding to the full feature data under a timestamp; the columns are divided into three main categories: water quality parameter columns, fish population state index columns, and key factor combination feature columns, with each main category further subdivided into specific sub-columns. The matrix is ​​divided into three sections: Water Quality Parameters, Fish Population Status, and Key Factor Combinations. The Water Quality Parameters section includes independent sub-columns for ammonia nitrogen, dissolved oxygen, pH, water temperature, nitrite, and hydrogen sulfide, each corresponding to a standardized quantified value for a single water quality indicator. Fish Population Status Index section includes independent sub-columns for feeding activity, stress level, abnormal behavior incidence, and physiological health index, each corresponding to a standardized quantified value for a single fish population status indicator. The Key Factor Combination Feature section includes disease type identifiers, core synergistic factor identifiers, and synergistic anomaly intensity columns. The disease type identifier column indicates the disease type corresponding to the key factor combination at that time point (e.g., vibrio disease, Ichthyophthirius multifiliis disease). The core synergistic factor identifier column indicates the core environmental parameters and fish population indicators in the combination (e.g., dissolved oxygen and stress hormones, pH and physiological health index). The synergistic anomaly intensity column indicates the degree of synergistic effect of the abnormal indicators in the combination (standardized values ​​quantified based on threshold deviation magnitude and duration). Each element in the matrix corresponds to a standardized quantified value or identifier for a specific feature at a specific time point, clearly presenting the synergistic characteristics of water quality, fish population status, and key factor combinations at each time point.

[0068] In a preferred embodiment of the present invention, step 2, extracting core causal feature variables based on the feature matrix, includes:

[0069] Step 200: Based on the time series of all aquaculture variables in the feature matrix, calculate the time-lag cross-correlation and covariance between each pair of variables to construct a weighted directed graph characterizing the strength of the association between variables, serving as the initial variable relationship graph. This includes: extracting the time series data of all aquaculture variables; based on the feature matrix constructed in step 104, extracting the time series of aquaculture variables corresponding to all column dimensions in the matrix, comprehensively covering three categories of variables: water quality parameter columns (including ammonia nitrogen, dissolved oxygen, pH value, water temperature, nitrite concentration, hydrogen sulfide concentration, etc.), fish school status index columns (including feeding activity, stress level, abnormal behavior incidence, physiological health index, swimming frequency, school density, etc.), and key factor combination feature columns (including the quantitative value corresponding to the core synergistic factor identifier, synergistic anomaly strength, etc.). Each variable corresponds to a complete time series data point, with the time granularity consistent with the feature matrix, uniformly set to 1 hour, and the time series length the same as the row dimension of the feature matrix, ensuring that the time series data of all variables are completely aligned on the time axis, laying the data foundation for subsequent pairwise calculations. The core parameters for time-lag calculation were set, taking into account the characteristics of closed recirculating aquaculture systems. Considering that fish typically require a certain period before exhibiting obvious physiological or behavioral responses after environmental stressors become abnormal, the time lag range was set from 0 to 72 hours, covering both immediate associations without time lag and lagged associations within the longest response period. Simultaneously, a time lag gradient of 1 hour was set to maintain consistency with the time granularity of the time-series data, ensuring the refinement and accuracy of the time-lag calculation. A significance level was set as the criterion for determining the validity of associations, used to subsequently screen statistically significant association results and avoid random associations interfering with the analysis.

[0070] The time-lag cross-correlation of each pair of variables is calculated using time-lag gradients to quantify the degree of correlation between variables as they change with time lag. The specific operation is as follows: For any two sets of variables to be analyzed, denoted as variable A and variable B respectively, starting from a time lag of 0 hours, the gradient is progressively increased by 1 hour, traversing all time-lag gradients up to 72 hours to complete the full-range calculation. Under each time-lag gradient, the time-series data of the variables are first shifted. If the current time lag is τ hours, the time-series data of variable A is shifted forward by τ hours, while keeping the time-series data of variable B unchanged, so that the state of variable A τ hours ago corresponds to the current state of variable B. Subsequently, the correlation between variable A after the shift and the original variable B is calculated to obtain the time-delay cross-correlation coefficient under this time-delay gradient. The coefficient ranges from -1 to 1. A positive number indicates a positive correlation, meaning that after variable A increases or decreases, variable B will increase or decrease synchronously after τ hours; a negative number indicates a negative correlation, meaning that after variable A increases, variable B will decrease after τ hours. The larger the absolute value of the coefficient, the stronger the correlation between the two variables under this time-delay gradient. After completing the calculation in one direction, a reverse verification is performed. The time series data of variable B is shifted forward by τ hours, while keeping the time series data of variable A unchanged. The above correlation calculation process is repeated to obtain the time-delay cross-correlation coefficient of variable B to variable A, avoiding misjudgment of the time series relationship caused by unidirectional calculation. Then, the significance of the two sets of coefficients obtained under each time-delay gradient is tested. The calculation results are compared with the preset significance level. If the results meet the significance requirements, the correlation between the variables under this time-delay gradient is determined to be statistically significant, and the corresponding coefficient is retained; if not, it is determined to be without significant correlation, the coefficient is set to 0, and invalid correlation data is excluded. After traversing all 73 time-delay gradients (0 to 72 hours), a set of time-delay cross-correlation coefficients between variables A and B under all effective time-delay gradients was obtained. Based on this set, the correlation characteristics were further quantified: First, the optimal time lag was extracted by selecting the time-delay gradient corresponding to the coefficient with the largest absolute value from all coefficients that passed the significance test, which was taken as the time lag duration for the strongest correlation between the two variables; Second, the correlation strength was divided into levels based on the absolute value of the coefficients and the actual characteristics of the correlation between aquaculture variables, the degree of correlation was divided into five levels: no correlation, weak correlation, moderate correlation, strong correlation, and extremely strong correlation. The smaller the absolute value of the coefficient, the weaker the correlation, and the larger the absolute value, the stronger the correlation; Third, the correlation direction was clarified by determining whether the correlation between variables was positive or negative based on the sign of the coefficients, and finally, a complete record of the time-delay correlation characteristics of each pair of variables was formed, clearly showing the correlation pattern of the two variables as time lag changes.

[0071] Simultaneously, the covariance of each pair of variables is calculated. Based on the original time-series data of variables A and B (without time shifting), their covariance is calculated to characterize the coordinated trend of numerical fluctuations of the two variables without time lag. The sign of the covariance is consistent with the sign of the aforementioned time-lag cross-correlation coefficient; a positive value indicates that the two variables fluctuate synchronously, and a negative value indicates that they fluctuate in opposite directions. The larger the absolute value of the covariance, the more obvious the degree of coordination or inversion of the numerical fluctuations of the two variables, providing supplementary basis for subsequent comprehensive judgment of the correlation strength. An initial weighted directed graph is constructed, with each aquaculture variable as an independent node in the graph, labeled with the variable name and its category (e.g., water quality, fish population status), making the node information clearly identifiable. The direction of the directed edges is determined based on the time-lag cross-correlation calculation results, pointing from the variable that changes first to the variable that responds later. That is, if the optimal time lag of variable A to variable B is greater than 0 and the corresponding coefficient passes the significance test, it indicates that the change of variable A precedes that of variable B, and the direction of the edge is set from variable A to variable B. The edge weights are calculated using a comprehensive quantification method, combining the absolute values ​​of the cross-correlation coefficients and covariances under the optimal time lag. The covariance is also normalized to eliminate dimensional differences. The time-lag cross-correlation coefficient carries a higher weight, with weights ranging from 0 to 1; larger values ​​indicate stronger correlations between the two variables. Furthermore, each directed edge is labeled with its optimal time lag duration and correlation direction (forward / backward). The resulting weighted directed graph comprehensively represents the correlation strength, temporal sequence, and time lag characteristics among all aquaculture variables.

[0072] Step 201 involves transforming the initial variable relationship graph into a geometric constraint graph model, including: preprocessing redundant edges in the initial weighted directed graph and removing invalid weakly correlated edges to eliminate noise interference; combining practical experience in the field of closed recirculating aquaculture, and verifying and calibrating based on historical aquaculture data (covering 12 months of monitoring data from 12 aquaculture bases and 30 aquaculture ponds), setting a weight threshold, and determining the threshold to be 0.2 (corresponding to the critical value of weak correlation in step 200), that is, only directed edges with a weight ≥ 0.2 are retained, and weakly correlated edges with a weight < 0.2 are removed. During the preprocessing process, validity verification must be carried out simultaneously: First, ensure that all core correlation edges between the core water quality parameters (ammonia nitrogen, dissolved oxygen, water temperature) and the key state indicators of the fish population (stress level, incidence of abnormal behavior) are retained without omission; second, eliminate redundant weak correlations between similar variables (such as low-weight cross-correlation between different water quality parameters) to reduce the computational load of the model; third, record the eliminated edges, label the variable pair name, corresponding weight, and reason for elimination, so as to facilitate subsequent traceability and parameter adjustment, and finally obtain a simplified weighted directed graph that retains only the core correlations between variables.

[0073] Subsequently, the simplified weighted directed graph is transformed into a geometrically constrained graph model, and the visualization and structured representation of variable relationships are achieved through geometric space mapping. The specific mapping rules are as follows:

[0074] The coordinate mapping of variable nodes selects either two-dimensional or three-dimensional geometric space (adapted to the total number of variables; two-dimensional space is used when the number of variables is ≤20 for easier visualization; three-dimensional space is used when the number of variables is >20 to avoid node overlap), mapping each aquaculture variable node to an independent coordinate point in the space. Coordinate allocation follows the principle of clustering similar variables and separating dissimilar ones; that is, coordinate points of variables of the same category (water quality parameters, fish population status indicators, key factor combination features) are concentrated in the same area, while coordinate areas of variables of different categories maintain a reasonable spacing. For example, water quality parameter nodes are mapped to the left side of the space, fish population status indicator nodes to the right side, and key factor combination feature nodes to the middle transition area, ensuring visual clarity and facilitating subsequent constraint relationship construction. Simultaneously, each coordinate point is labeled with its corresponding variable name, category, and core attributes to ensure rapid identification of node information.

[0075] Geometric constraint transformation of directed edges: The three core features of the simplified directed edges are transformed into constraint relationships between two points in geometric space:

[0076] Edge weights are transformed into distance weights: a mapping logic of higher weights corresponding to closer distances is adopted. That is, the greater the weight of a directed edge between two variable nodes, the closer the straight-line distance between the two points in the geometric space, intuitively representing a stronger correlation between variables. For example, the edge weight between dissolved oxygen and the incidence of abnormal behavior is 0.8 (extremely strong correlation), and the distance between their coordinate points is set to the minimum value; while the edge weight between pH value and cluster density is 0.3 (upper limit of weak correlation), and the distance between their coordinate points is set to a medium level. The correlation strength is quantified through the difference in distance. Edge directions are transformed into vector pointing, matching the temporal response relationship between variables. The direction of directed edges is directly converted into a vector pointing between two points in the geometric space, with the variable node that changes first pointing to the variable node that responds later, forming a unidirectional vector. For example, ammonia nitrogen anomalies precede the increase in fish stress levels, corresponding to a vector direction from the ammonia nitrogen node to the stress level node; insufficient dissolved oxygen triggers abnormal behavior first, with the vector direction from the dissolved oxygen node to the abnormal behavior incidence rate node. The temporal logic of variables is clearly defined through vector pointing. The optimal time delay duration is converted into a vector length parameter. The optimal time delay duration extracted in step 200 is converted into the corresponding vector length. The longer the time delay duration, the longer the vector length, and vice versa. A length benchmark is set: the vector length corresponding to the optimal time delay of 1 hour is used as the base unit. For each additional hour of time delay, the vector length increases by a fixed proportion (e.g., 0.1 units for each additional hour), ensuring that time delay differences can be intuitively distinguished by vector length. For example, the optimal time delay for the stress level of fish populations due to ammonia nitrogen anomalies is 6 hours, corresponding to a vector length of 6 base units; the optimal time delay for the feeding activity level due to sudden water temperature changes is 2 hours, corresponding to a vector length of 2 base units.

[0077] Finally, by combining the knowledge base of association rules for environmental stress factors, prior constraints are added to the geometric constraint graph model to ensure that the model conforms to the inherent logic between variables in a closed recirculating aquaculture system and avoids geometric relationships that violate common sense in aquaculture. The specific supplementary rules are as follows:

[0078] Temporal irreversibility constraint: For clearly defined cause-and-effect relationships in the knowledge base, an irreversible vector pointing constraint is set to prohibit the generation of reverse vectors. For example, it is clear that abnormal ammonia nitrogen leads to increased stress level; excessive nitrite leads to decreased physiological health index; and sudden changes in water temperature cause fluctuations in feeding activity. The vector pointing to these relationships is irreversible, and in geometric space, only pointing from the former node to the latter node is allowed, and reverse vectors cannot be generated. This aligns with the objective law that environmental stress factors cause abnormalities first, and fish respond subsequently.

[0079] For the core characteristic stress factor combinations of various diseases in the knowledge base, a lower limit for the distance weight of key variable pairs is set to ensure that the core association strength is not lower than the benchmark. For example, it is clear that insufficient dissolved oxygen leads to an increased incidence of abnormal behavior; excessive hydrogen sulfide leads to a decrease in body surface health score. The distance weight of these two variables must not be lower than 0.5 (corresponding to a strong association level). The distance between the nodes of these two variables in geometric space must be controlled within the corresponding benchmark range. If the distance exceeds the range after conversion, the coordinate point position is automatically adjusted to ensure that the core association strength meets the standard. For the inherent association logic of variables of the same category, clustering constraints are set to ensure that the geometric distribution of nodes of the same type of variable conforms to the actual association characteristics. For example, ammonia nitrogen, nitrite, and hydrogen sulfide in water quality parameters are all harmful pollutants. The coordinate points of the three variables must form a tight cluster (the distance weight between them is ≥0.4). Feeding activity, stress level, and physiological health index in fish school status indicators all reflect the overall status of the fish school. Similarly, they must maintain a clustered distribution to avoid the dispersion of nodes of the same type of variable and to conform to the inherent association logic of the variables.

[0080] For different disease-specific stress factor combinations, vector association constraints are set for corresponding variable nodes. For example, the characteristic stress factor combination for vibrio is increased ammonia nitrogen and decreased dissolved oxygen, increased stress level and increased abnormal behavior. In the geometric model, it is necessary to ensure that the ammonia nitrogen and dissolved oxygen nodes point to the stress level and abnormal behavior incidence nodes, and that the corresponding vector length matches the typical time lag (12-24 hours) of vibrio. For Ichthyophthirius multifiliis disease, the corresponding water temperature is 18-22℃ and turbidity increases, and friction against the pool wall increases. It is necessary to constrain the vector orientation and distance weights of the water temperature and turbidity nodes and the abnormal behavior incidence nodes to ensure that the model fits the factor association logic of the specific disease.

[0081] After supplementing all prior constraints, the geometric constraint graph model is validated for consistency. The distribution of coordinate points, vector orientation, and distance weights are checked to ensure they all conform to the constraint rules. If conflicts exist (such as contradictions between vector orientation and irreversible constraints, or distance weights being lower than the baseline), the coordinate point positions or vector parameters are adjusted based on the knowledge base of environmental stress factor association rules until the model is free of logical conflicts. This results in a constraint model that accurately represents variable associations using geometric relationships. It retains the core information of the initial variable relationships and provides structured support for the subsequent extraction of core causal feature variables through geometric and rule-based processing.

[0082] Step 202: Determine the mutually independent simplified constraint equation system based on the geometric constraint graph model, including: algebraic transformation of distance weight constraints (association strength):

[0083] Based on the distance weights (range 0-1, higher weights indicate stronger associations) between variable nodes in the geometric model, this is transformed into an algebraic constraint on the strength of variable associations, as shown in the formula: ,in, This represents the actual correlation strength between variables X and Y (corresponding to the absolute value of the time-delay cross-correlation coefficient that passed the significance test in step 200). This represents the lower limit of the association strength between variables X and Y (determined by the weights of the edges retained in the geometric model, i.e., values ​​above the weight threshold set in step 201, such as dissolved oxygen and the incidence of abnormal behavior). The constraint logic requires that the actual correlation strength between variables must be maintained above the lower limit set by the geometric model to ensure correlation stability. Example: For the strong correlation constraint between dissolved oxygen (X) and the incidence of abnormal behavior (Y), it is transformed into the equation... That is, the correlation between the two must be strong or above.

[0084] Algebraic transformation of vector pointing constraints (temporal causality):

[0085] Based on the vector orientation in the geometric model (representing the irreversible temporal logic of the preceding dependent variable on the subsequent variable), it is transformed into a one-way causal algebraic constraint, using subscripts to distinguish the direction of action, as shown in the formula: and ,in, This represents the positive effect coefficient of variable X on variable Y (a non-zero value, the absolute value of which corresponds to the strength of the association). This represents the coefficient of the inverse effect of variable Y on variable X; the constraint logic only allows the dependent variable to affect the subsequent variable, prohibiting causal reversal, which aligns with the objective law in aquaculture systems that "environmental factor anomalies precede fish responses." Example: The vector direction of ammonia nitrogen X in response to stress level Y is transformed into the equation... and It is clear that changes in ammonia nitrogen can affect stress levels, while the reverse effect is negligible.

[0086] Algebraic transformation of time delay length constraint (duration of action):

[0087] Combining the optimal time delay length corresponding to the vector length, this is transformed into an algebraic constraint on the effect time of the variable, clarifying the response time window of the consequential variable, as shown in the formula: ,in, This indicates the point in time when the dependent variable X first undergoes an abnormal change; This indicates the time point at which the consequent variable Y produces its corresponding response; The optimal time delay of variable X with respect to Y (determined by the vector length in the geometric model, i.e., the effective time delay extracted in step 200); This indicates the allowable fluctuation range of response time (set to 1-2 hours based on 1-hour time-series data granularity to adapt to the data precision in aquaculture scenarios); the constraint logic requires that the consequential variable must produce a response within the window corresponding to the optimal time lag; otherwise, the association is considered invalid. Example: The effect of a sudden change in water temperature (X) on feeding activity (Y). Hour, hours, transformed into equations This means that within 2-3 hours after a sudden change in water temperature, the feeding activity must show a corresponding fluctuation. The above three types of algebraic constraint equations are summarized, and the geometric constraint sources corresponding to each type of equation are labeled (such as distance weight constraints - dissolved oxygen and abnormal behavior incidence rate), forming an initial set of constraint equations to ensure a one-to-one correspondence between the algebraic expression and the geometric model, with no information omissions.

[0088] A dual approach, employing both the variance inflation factor (VIF) test and aquaculture logic verification, is used to identify linear dependencies, eliminate redundant equations, and retain mutually independent constraints.

[0089] Variance inflation factor (VIF) test (quantifying the degree of linear dependence):

[0090] Calculate the VIF value for each equation in the initial system of equations to determine its linear dependence on other equations. The formula is as follows: ;in, Let represent the variance inflation factor of the i-th equation; The coefficient of correlation between the variable in the i-th equation and the variables in the other equations in the system of equations; Criterion: When the equation is determined to have a strong linear dependence on other equations (which can be derived from other equations), it is considered a redundant equation. When the equation is independent, it is determined to be an independent equation.

[0091] Based on the VIF test results and the importance of aquaculture variables, priority is given to retaining variables and equations with more direct causal logic, while redundant equations that repeatedly represent the same association are eliminated:

[0092] Example 1: Ammonia nitrogen X corresponds to stress level Y ( (independent) and nitrite Z corresponding to stress level Y ( Both (strong dependence) characterize the effects of nitrogenous pollutants on fish stress, and ammonia nitrogen is a pollutant. After removing redundant equations for Z to Y, the corresponding equation is... and Remove them all;

[0093] Example 2: Two sets of independent equations (corresponding equations) for X versus Y and Y versus physiological health index (W). and , and ), and X versus W equation ( and After combining them, the logic of X to W can be derived from the first two sets of deductions. If X to W... If so, then the redundant equation is eliminated.

[0094] After removing redundant equations, the VIF values ​​of the remaining equations are recalculated to ensure that all equations are correct. Furthermore, any two equations cannot be derived from each other, forming a set of intermediate constraint equations without linear dependence. At the same time, the number, VIF value, and reason of the removed equation are recorded to facilitate subsequent traceability and adjustment.

[0095] Based on the physical meaning of the variables in the characteristic matrix and the operating rules of the closed recirculating aquaculture system, the equation set is further simplified by eliminating secondary constraints and merging redundant logic, while taking into account both the simplicity and relevance of the equation set:

[0096] For highly correlated variables within the same category that exhibit synchronized trends and consistent impact logic, retain the constraint equations corresponding to the variables, eliminate redundant constraints on secondary variables, and only provide supplementary explanations. Example: In water quality parameters, ammonia nitrogen (X) and nitrite (Z) are both nitrogenous pollutants. Their concentration changes are highly synchronized, and their stress mechanisms on fish populations are similar. Therefore, the constraint equation for variable X should be retained first. Eliminate duplicate equations of Z (including , and Only by adding the Z concentration after the equation, which fluctuates synchronously with X, can a synergistic stress effect be generated, reducing the number of equations without omitting correlation characteristics.

[0097] For variable pairs with a clear causal relationship and extremely weak inverse effects, retain the one-way causal equation, delete the two-way association statements, and optimize the equation structure. Example: An increase in stress level (Y) directly leads to a decrease in physiological health index (W), but a decrease in W is unlikely to quickly reverse and change Y; therefore, retain the one-way equation. and The redundant descriptions of the original bidirectional associations have been removed to better reflect the objective logic of the role of variables in the aquaculture system.

[0098] Constraint equations that exist only in mathematical derivation and lack support from actual aquaculture logic are eliminated. Example: The weak correlation equation between pH (P) and cluster density (Q) (the corresponding geometric edges have been eliminated in step 201), although... (Independent), but in actual aquaculture, pH value has no clear regulatory effect on cluster density, so they are directly removed (including) , and This ensures that each equation in the system corresponds to a real variable interaction relationship.

[0099] After simplification, the final system of equations undergoes triple verification to ensure it meets the requirements of independent, reasonable, and interconnected constraints:

[0100] Independence check, all equations Wireless dependencies and logical conflicts; correlation verification, each equation corresponds to constraints in the geometric constraint graph model, covering key variables and temporal characteristics of water quality and fish population status; rationality verification, equation parameter thresholds ( , , The rules are consistent with the knowledge base of association rules for environmental stress factors and conform to the operating rules of closed recirculating aquaculture systems. Ultimately, a set of simplified constraint equations that are mutually independent are obtained. This set of equations contains only non-redundant, strongly correlated aquaculture variables. Each equation precisely corresponds to geometric constraint information, while also taking into account the temporal logic and causal relationships between variables, providing concise and efficient algebraic support for the subsequent extraction of causal feature variables.

[0101] Step 203: From the solution space of the simplified constraint equation system, determine the core causal characteristic variables, including: refining the solution of the simplified constraint equation system, constructing an effective solution space according to the preprocessing-stage solution-integration and verification process, with each step combined with the practical application of the formulas:

[0102] Constraint preprocessing and priority definition: First, all equations in the simplified constraint equation set are classified and sorted to clarify three types of constraints and their corresponding core formulas. Simultaneously, a solution priority is set (time-series constraints > unidirectional causal constraints > correlation strength constraints) to ensure the solution logic is orderly. The classification results are as follows:

[0103] Formulas corresponding to time series and timeliness constraints: The response time range of core control variables; the corresponding formula for unidirectional causal constraints. and The core locks the direction of the variable's action; the corresponding formula for the correlation strength constraint. The core is to define the boundaries of the correlation strength between variables. At the same time, fixed parameters in each formula (such as...) are extracted. From step 201, the length of the geometric model vector, (From step 201 weight threshold), enter the parameter table uniformly to avoid parameter confusion during the solution process.

[0104] Iterative solution in stages:

[0105] The first step is to solve the time-series constraints. Using the time-series nodes of the feature matrix in one-hour units as a baseline, the time-series data for each pair of variables is extracted one by one. (Optimal time delay) and The (fluctuation range) parameter is substituted into the time series constraint formula to locate the response time window of the consequent variable. Specific operation: First, lock the time point when the dependent variable X experiences an abnormal change. Then, the result is obtained through formula calculation. The upper and lower limits are used to indicate that the time series data of the Y variable within this window are valid response data, while those outside the window are considered invalid responses and are temporarily stored (for later verification). For example, for water temperature (X) - feeding activity (Y), given... Hour, Hours, if If it is the 10th hour, then substituting into the formula yields... For each hour, only the feeding activity data from the 12th to 13th hour are retained as the time-series basis data for this variable pair.

[0106] The second step is to solve for the superposition of unidirectional causal constraints: based on the time series data obtained in the first step, and combined with the unidirectional causal constraint formula... and Remove inverse correlation data and lock unidirectional action paths. Specifically, perform direction verification on the time-series baseline data for each variable pair, retaining only... The non-zero corresponding correlation data (i.e., the data that drives the Y response to the X change) will Corresponding reverse data (data where X changes synchronously after Y fluctuations) are marked as incidental correlations and removed. For example, when analyzing ammonia nitrogen (X) - stress level (Y), if the stress level fluctuates at hour 15 and ammonia nitrogen changes at hour 16, this data set is considered an incidental correlation and is removed. Reverse correlation is used to directly eliminate data, retaining only data showing changes in ammonia nitrogen and stress levels within the window, further narrowing the range of variable values. The third step is a correlation strength constraint fusion solution: extracting the correlation strength constraints for each variable pair. The parameters are substituted into the correlation strength constraint formula to perform strength filtering on the valid data obtained in the first two steps. Specifically, the correlation strength of the valid variable pairs within the first step window and in the second step direction is calculated. Combine it with the formula Comparison, only retain The data was processed, and weakly correlated data below a threshold were removed. Simultaneously, the actual correlation of each variable pair was recorded. The range serves as the boundary of the correlation strength in the solution space. For example, dissolved oxygen (X) - the incidence of anomalous behavior (Y). Calculate the data within the window. If the data is below the threshold, that portion of the data will be removed, and only the data that is retained will be discarded. The data clearly defines the effective intensity range of fluctuations in the incidence of abnormal behavior corresponding to changes in dissolved oxygen.

[0107] Solution space fusion and invalid solution verification:

[0108] The first step is solution space fusion: Integrate the effective data obtained from solving all variable pairs in the three stages to construct a global solution space. The solution space needs to present four core pieces of information according to variable classification: the effective value range of a single variable (e.g., ammonia nitrogen 0.02-0.2 mg / L, stress level 1-5), the unidirectional action path of the variable pair (e.g., X to Y), the time series response window (e.g., [12,13] hours), and the correlation strength range (e.g., 0.5-0.8).

[0109] The second step involves two rounds of invalid solution elimination: First, parameter rationality verification, comparing the solution space against the safe thresholds for closed-loop aquaculture (e.g., the safe range of ammonia nitrogen is 0.01-0.25 mg / L, and the normal range of fish stress level is 1-3). Solutions that exceed the actual aquaculture parameter range are eliminated, for example, data with an ammonia nitrogen value of 0.3 mg / L are directly eliminated. Second, constraint logic consistency verification, checking whether solutions for different variable pairs conflict (e.g., the timing of the response window from X to Y is contradictory to the response window from Y to Z). Conflicting solutions are adjusted first based on the environmental stress factor knowledge base, and those that cannot be adjusted are directly eliminated, ultimately forming an effective solution space without logical conflicts and closely aligned with aquaculture practices.

[0110] Secondly, based on the effective solution space and combined with the formula constraint logic, core causal characteristic variables are selected according to three criteria:

[0111] Filtering by causal relationship: Based on the range of correlation strength in the solution space (corresponding to...) ) and unidirectional action path (corresponding ), prioritizing the selection of variables that directly drive disease outbreaks and fluctuations in fish population status. Specifically: retaining variables in the solution space Variables in the medium to high range (above 0.4) and directly affected by the pathway (such as ammonia nitrogen, dissolved oxygen, and water temperature) should be excluded. near Lower limits and secondary variables that only indirectly affect the core indicators (such as trace impurities in water quality and non-critical flow velocities) are selected to ensure that the selected variables conform to the causal logic of the formula constraints.

[0112] Screening by Independence: For variables initially screened, this is combined with the trend of values ​​in the solution space and... (Step 202 Independence Criterion) Conduct multicollinearity verification. Specifically: If the effective value ranges of two sets of variables fluctuate synchronously and their action paths completely overlap (such as ammonia nitrogen and nitrite), and both... If the underlying logic remains consistent, only the core variable (ammonia nitrogen) is retained, while secondary variables (nitrite) are removed, ensuring that each variable provides unique causal information without redundancy. The stability of variable constraint relationships is verified based on full-cycle time-series data and multi-scenario data from the solution space. Time dimension: Examining the variables at each stage of aquaculture. scope, To determine stability, transient variables that are only briefly correlated within a specific window (such as one-time water quality fluctuations caused by equipment failure) are excluded; Scenario dimension: Compare solution space data for different aquaculture ponds, fish species, and seasons to ensure the formula constraints on variable parameters are met. , There was no significant bias, and only variables that were stable across scenarios were retained. Finally, the variable set was optimized by combining key factor combinations for verification: The characteristic factor combinations of different diseases in step 1033 were correlated, and the variable action paths and formula constraint parameters in the solution space were compared to verify the fit between the selected variables and the core synergistic factors. For example, vibriosis corresponds to the combination of "increased ammonia nitrogen + decreased dissolved oxygen → increased stress level," and the corresponding values ​​for ammonia nitrogen, dissolved oxygen, and stress level were extracted from the solution space. Action path and Intensity data is used, retaining these three types of variables. For Ichthyophthirius multifiliis disease, which corresponds to combinations of water temperature fluctuations, increased turbidity, and increased abnormal behavior, water temperature, turbidity, and the incidence of abnormal behavior are retained, while variables without formulaic constraints (such as pH and population density) are removed. After validation, the variable set is integrated and optimized to ensure that core causal characteristic variables cover key water quality parameters and fish population status indicators, accurately match the temporal, causal, and intensity logic of formula constraints, and are free of redundancy, contradictions, and unstable variables. The final variable set provides core data support for subsequent dynamic causal fingerprint generation, ensuring that the fingerprint accurately reflects the key causal relationships in the aquaculture system.

[0113] In a preferred embodiment of the present invention, step 3, based on the core causal feature variables and a preset structural causal model, generates a dynamic causal fingerprint by calculating the counterfactual distribution of the set of variables under preset intervention conditions, including:

[0114] Step 300: In the pre-defined structural causal model, for each core causal feature variable, a pre-defined intervention condition representing an external regulatory operation is established. The conditional probability distribution of all core variables under this intervention condition is calculated to obtain the counterfactual distribution. This includes: adapting and calibrating the pre-defined structural causal model to ensure that the variable nodes and causal paths in the model completely correspond to the core causal feature variables determined in step 203. Simultaneously, parameter thresholds (such as water quality safety range and fish population status baseline values) from the environmental stress factor association rule knowledge base are loaded as the basis for intervention condition design and distribution verification, avoiding intervention scenarios that are divorced from actual aquaculture practices. For each core causal feature variable, a single and clear external regulatory intervention condition is designed based on the feasibility of regulation in aquaculture practice, ensuring that the intervention scenario is specific and simulable. The design is categorized by variable type as follows:

[0115] Water quality variables (ammonia nitrogen, dissolved oxygen, water temperature, etc.): The intervention conditions are set to deviate from the current actual value to the safe threshold boundary / abnormal threshold boundary. For example, if the current actual value of ammonia nitrogen is 0.1 mg / L, the intervention conditions are set to adjust to the upper limit of abnormality of 0.25 mg / L (simulating pollution stress) and to adjust to the lower limit of safety of 0.02 mg / L (simulating optimization control); if the current actual value of dissolved oxygen is 6 mg / L, the intervention conditions are set to reduce to the critical value of 3 mg / L (simulating hypoxia stress), to ensure that the intervention can trigger the response of subsequent variables.

[0116] Fish population status variables (stress level, incidence of abnormal behavior, etc.): These variables are mostly aftereffect variables. The intervention conditions are set to intervene indirectly by adjusting the preceding water quality variables. For example, when intervening in the stress level, the stress level is raised from the current level 2 to level 4 (simulating stress response) by adjusting the preceding variables such as ammonia nitrogen and dissolved oxygen, so as to avoid directly intervening in the aftereffect variables, which violates the causal logic of aquaculture.

[0117] Key synergistic factor variables: Intervention conditions are set to individually activate / inhibit a single synergistic factor. For example, for synergistic factors related to vibriosis, the intervention condition is set to activate only the ammonia nitrogen-increasing factor, while keeping other factors such as dissolved oxygen stable, to isolate the effect of a single factor. Simultaneously, all intervention conditions set a control baseline (i.e., the current actual value of the variable), and only one variable is intervened at a time, while other core variables maintain their initial association logic unchanged, avoiding effect confounding caused by the superposition of multiple variable interventions. The preset intervention conditions are substituted into a structural causal model, and by progressively simulating the causal transmission process between variables, the set of values ​​for all core variables under this intervention scenario is calculated, forming a counterfactual distribution. Specific process:

[0118] The first step is to fix the value of the intervention variable: The target intervention variable is locked at a preset intervention value. Based on the unidirectional causal path in the model, the impact of this variable change on directly related consequential variables is simulated. For example, after intervening in ammonia nitrogen to 0.25 mg / L, the change in the value of the immediately adjacent consequential variable, stress level, is simulated first. The second step is to simulate the cascading effects: Following the causal path, the changes in the values ​​of indirectly related variables are simulated sequentially. For example, after an increase in ammonia nitrogen leads to an increase in stress level, the chain reaction of stress level on physiological health index and abnormal behavior incidence is further simulated, ensuring that all core variables are included in the simulation. The third step is to construct and validate the distribution: All core variable values ​​obtained from the simulation are collected, and characteristics such as value range, fluctuation frequency, and mean are statistically analyzed according to variable type to construct a counterfactual distribution. Simultaneously, extreme invalid values ​​appearing during the simulation (such as stress levels exceeding the 1-5 level range) are removed by comparing with the actual parameters of aquaculture, ensuring that the counterfactual distribution has physical meaning. The above operations are repeated to complete the counterfactual distribution calculation for each core causal characteristic variable under intervention conditions, ultimately forming a set of correspondences between single-variable intervention and the counterfactual distributions of all core variables.

[0119] The construction and training process of the pre-defined structural causal model is as follows:

[0120] Based on the core causal characteristic variables identified in step 203, these variables are used as core nodes of the model. They are then categorized and sorted into water quality parameters (ammonia nitrogen, dissolved oxygen, water temperature, turbidity, etc.), fish population status (stress level, incidence of abnormal behavior, physiological health index, etc.), and key synergistic factors (corresponding variables to the combination of core stress factors) to ensure that the nodes are completely consistent with the screening results above, with no extra redundant variables or missing core variables.

[0121] Based on the knowledge base of environmental stress factor association rules, three main categories of core rules are extracted: variable causal logic rules (e.g., ammonia nitrogen abnormalities precede stress level increases, and insufficient dissolved oxygen directly drives an increase in abnormal behavior), aquaculture parameter threshold rules (e.g., safe ammonia nitrogen range, fish stress level grading standards), and disease association rules (e.g., feature factor combination logic corresponding to vibrio disease and Ichthyophthirius multifiliis disease). Simultaneously, the vector pointing and temporal delay information from the geometric constraint graph in step 201 are integrated as the core basis for model structure construction, avoiding deviation from actual aquaculture logic. Based on the feature matrix constructed in step 104, the time granularity (1 hour / node), value range, and basic association features between variables of the full-cycle time-series data are analyzed to determine the model data input format and adaptation range, ensuring accurate correspondence between subsequent training data and model nodes.

[0122] Using the categorized core variables as nodes, a unidirectional causal path is constructed based on the loaded prior logic and the preceding constraints. The path construction follows these principles: First, the time-priority principle strictly adheres to the vector pointing logic of the geometric constraint diagram in step 201, ensuring the path direction is consistent with the temporal relationship between the dependent and consequential variables. For example, only the ammonia nitrogen and stress level paths are constructed, and reverse paths are prohibited. Second, the direct association principle retains only direct causal paths between variables, eliminating indirect association paths (e.g., the direct path between ammonia nitrogen and physiological health index is not constructed; instead, the path is transmitted step-by-step through ammonia nitrogen, stress level, and physiological health index). Third, the disease adaptation principle strengthens the corresponding core paths for different disease characteristic factor combinations. For example, in the vibriosis scenario, the association between the ammonia nitrogen / dissolved oxygen and stress level / abnormal behavior incidence paths is strengthened. The model is divided into two layers: an environmental stress layer (water quality variables, synergistic factors) and a fish swarm response layer (fish swarm state variables). The environmental stress layer serves as the input layer, and the fish swarm response layer serves as the output layer. Causal transmission between layers is only allowed from top to bottom, and variables within a layer have no overlapping paths (to avoid redundant associations between similar variables). Simultaneously, initial association attributes are labeled for each path, including the time-series lag range (taken from the optimal lag in step 200 and the vector length in step 201) and the association strength threshold (taken from the weight threshold in step 201 and the association strength constraint in step 202), forming a preliminary path attribute table.

[0123] Key constraints in the steps are transformed into built-in rules in the model to ensure that the model's calculations do not deviate from the aquaculture logic. First, there are time-series lag constraints: a fixed time lag window is set for each causal path (e.g., a 2-3 hour response window is bound to the water temperature and feeding activity paths); variable associations exceeding this window are considered invalid. Second, there are association strength constraints: a path activation threshold is set; the subsequent variable response is only triggered when the change in the dependent variable reaches the threshold (e.g., ammonia nitrogen increases by more than 20% above the safe range). Third, there are parameter boundary constraints: value boundaries are set for each node variable (fitting the aquaculture safety threshold); values ​​exceeding the boundaries automatically trigger an anomaly warning and do not participate in normal causal transmission calculations. Practical characteristic rules of the closed recirculating aquaculture system are embedded, such as fish population state variables not being directly intervened in but only indirectly affected by water quality variables; and when similar water quality variables (ammonia nitrogen, nitrite) act synergistically, the core variable (ammonia nitrogen) has a higher weight than secondary variables. This avoids model calculations that violate aquaculture practices and lays the logical foundation for subsequent intervention simulations (step 300).

[0124] The model training process is as follows:

[0125] Two types of core data are integrated: first, the full-cycle time-series data of the feature matrix in step 104 (covering variable observations at different breeding stages and in different scenarios), serving as basic training data to support the model in learning the causal relationships of variables under natural conditions; second, simulated intervention data (a single-variable intervention-full-variable response dataset generated by the Monte Carlo simulation algorithm based on the intervention condition preset logic in step 300), serving as reinforcement training data to help the model grasp the causal transmission laws under intervention scenarios. Both types of data are aligned to a 1-hour time granularity, and the time dimension is calibrated using a timestamp synchronization algorithm to ensure that the variable data at each time-series node corresponds one-to-one, eliminating the impact of time deviation.

[0126] Three rounds of data cleaning were conducted on the dataset. The IQR quartile method was combined with dual verification of aquaculture parameter thresholds. First, the normal distribution range of the data was defined using the interquartile range. Then, the data was compared against the safe thresholds for variables such as ammonia nitrogen and dissolved oxygen (e.g., the safe range for ammonia nitrogen is 0.01-0.25 mg / L). Extreme values ​​caused by equipment malfunctions or acquisition errors (e.g., data where instantaneous ammonia nitrogen values ​​exceeded the safe range by 3 times) were removed. At the same time, reasonable outliers caused by aquaculture stress (e.g., sudden increases in stress levels) were retained to avoid erroneously deleting valid signals. For time-series nodes with missing durations ≤2 hours, a linear interpolation algorithm was used to fill in the gaps, generating intermediate values ​​based on the changing trends of adjacent valid data to match the time-series fluctuations of the variables. Data segments with missing durations exceeding 2 hours were directly removed because the variable correlation logic was easily broken. The location and reason for the missing data were recorded for bias analysis during subsequent model validation. The Min-Max normalization algorithm was used to standardize the value range according to variable type. Variables with different dimensions, such as stress level (1-5) and abnormal behavior incidence rate (0-100%), were all standardized to the 0-1 range to eliminate the interference of dimension differences on parameter calibration and ensure that the weights of each variable were balanced during training. After data cleaning, the data was divided into training and validation sets in a 7:3 ratio using a stratified sampling algorithm. The sampling preserved the data distribution characteristics of different breeding scenarios and disease occurrence cycles to avoid scenario bias between the training and validation sets and ensure the representativeness of the validation results (the training set was used for parameter calibration, and the validation set was used for effect verification).

[0127] Based on the loaded knowledge base of association rules for environmental stress factors and historical aquaculture data, the Analytic Hierarchy Process (AHP) is used to assign initial weights and response coefficients to each causal path, transforming qualitative associations into quantitative parameters:

[0128] Based on the correlation strength constraint results from step 202, a judgment matrix was constructed using the analytic hierarchy process (AHP) to quantify the priority of correlations between variables. For example, dissolved oxygen and the incidence of abnormal behavior were identified as a strong correlation path, with a weight of 0.8; pH and cluster density were identified as a weak correlation path, with a weight of 0.3, forming an initial weight matrix. Referring to practical experience in aquaculture and historical data patterns, a multiple linear regression algorithm was used to fit the response relationships between variables, determining the coefficient values. For instance, for every 0.1 mg / L increase in ammonia nitrogen, the stress level increased by 0.5 levels. This data was then integrated into a complete initial parameter matrix, laying the foundation for subsequent iterative calibration.

[0129] The batch gradient descent algorithm is used to input the training set data into the model in time-series batches. By comparing the deviation between the model output and the actual observed data, the parameters are iteratively adjusted until the prediction error is ≤10% (acceptable range).

[0130] The deviation between the predicted magnitude of the aftereffect variable change and the actual value is calculated. If the deviation is positive (e.g., the predicted stress level increases by 1 level, but the actual level increases by 2), the corresponding path weight is increased using the gradient descent algorithm to strengthen the transmission efficiency of that path. If the deviation is negative (the predicted magnitude exceeds the actual value), the path weight is reduced to suppress over-transmission. Based on the optimal time lag result from step 200, the difference between the model's predicted response time and the actual response time is verified using a time-series alignment algorithm. If the deviation exceeds 1 hour, the path time lag window boundary is adjusted. For example, if the original time lag window was 2-3 hours and the actual response time was 4 hours, the window is adjusted to 3-4 hours to ensure that the time-series logic aligns with actual aquaculture practices. To address the problem of frequent false triggering of path activation by the model (e.g., a slight fluctuation in ammonia nitrogen triggers a stress response), an adaptive threshold algorithm is used to increase the path activation threshold. Simultaneously, the threshold boundary is dynamically optimized by combining historical data statistics on the normal fluctuation range of variables, balancing model sensitivity and anti-interference capability.

[0131] Based on the validation set data, a multi-round iterative optimization strategy is adopted. After each iteration, the model fit is calculated using the coefficient of determination. If the fit is <0.8 (preset standard), the core parameters are adjusted backtrackingly to focus on solving two types of problems:

[0132] If the predictions from multiple paths highly overlap (e.g., the predictions for ammonia nitrogen and physiological health index are consistent with those for stress level and physiological health index), an L1 regularization algorithm is used to sparse the parameters, reduce the weight of secondary paths, and strengthen core paths (e.g., retaining the core paths for stress level and physiological health index while reducing the weight of the ammonia nitrogen path to below 0.1) to eliminate redundant correlations. If the predictions from different paths contradict each other (e.g., one path predicts an increase in stress level while another predicts a decrease), a causal structure learning algorithm (PC algorithm) and an environmental stress factor knowledge base are used to verify the causal rationality of the paths. The path direction is corrected based on the knowledge base first; if it cannot be corrected, conflicting paths are directly eliminated, and a causal link that conforms to the aquaculture logic is rebuilt.

[0133] Based on the intervention conditions preset in step 300, a reinforcement learning algorithm (Q-learning) is used to input simulated intervention data into the model, thereby enhancing the model's ability to respond to intervention scenarios.

[0134] Using single-variable intervention as the action and variable response as the reward, a reward and punishment mechanism is constructed. If the variable response magnitude and time lag output by the model are consistent with the simulation expectation (e.g., after ammonia nitrogen is regulated to the abnormal upper limit, the stress level rises to level 4 within 6 hours), a positive reward is given; if the deviation is large, a negative penalty is given, and the path parameters and constraint rules are iteratively optimized. In this way, it is ensured that the model can accurately simulate the causal transmission process under different intervention conditions, providing reliable support for the counterfactual distribution calculation in step 300.

[0135] Three types of algorithms were used to conduct 3D validation and comprehensively examine the model performance:

[0136] The model is tested using validation set data, and the goodness of fit is evaluated using the coefficient of determination (COP), which should be ≥0.85. The consistency between the predicted response time and magnitude of change of core variables and actual data is also verified, with deviations controlled within one hour and magnitude deviations ≤15%. New scenario data not used in training (such as aquaculture data for different fish species and seasons) is input, and K-fold cross-validation (K=5) is used to test the model's generalization ability, requiring a goodness of fit ≥0.8 and no significant scenario bias, ensuring stable operation of the model under diverse aquaculture scenarios. Three to five intervention conditions are randomly selected (such as dissolved oxygen dropping to a critical value or water temperature fluctuation regulation), and the propensity score matching (PSM) algorithm is used to compare the counterfactual distribution output by the model with the intervention effects in actual aquaculture operations, ensuring a deviation ≤10%, verifying the reliability of the intervention simulation. Based on the three-dimensional validation results, a final round of parameter fine-tuning is performed to correct deviations in a few scenarios, while simultaneously locking the model structure, parameters, and built-in rules to form the final pre-defined causal model. The model must meet two core characteristics: first, it must be fully compatible with the core variables, constraint logic, and intervention conditions mentioned above, without any logical conflicts; second, it must be able to accurately simulate the natural causal transmission and intervention response process between variables in the aquaculture system through algorithms, providing support for the generation of dynamic causal fingerprints.

[0137] Step 301: Based on the difference between the counterfactual distribution and the actual observed distribution of the core causal characteristic variables, quantify the marginal causal effect of each variable on the cumulative disease risk, including: for each intervention-control scenario, compare the difference between the counterfactual distribution (after intervention) and the actual observed distribution (before intervention) from three dimensions:

[0138] Differences in value ranges: Compare the direction and magnitude of the shift in the value range after intervention with that before intervention. For example, after ammonia nitrogen was intervened to 0.25 mg / L, the stress level value range shifted from level 1-2 to level 3-4. Record the direction of shift (positive increase) and magnitude (+2 levels).

[0139] Distribution trend differences are analyzed by examining changes in the fluctuation trends of variable values. For example, after intervention to reduce dissolved oxygen to 3 mg / L, the distribution of abnormal behavior incidence changed from being concentrated below 5% to being dispersed between 10% and 20%, and the recorded trend changed from stable low incidence to fluctuating high incidence. Mean differences are calculated by comparing the absolute difference and relative proportion of the mean values ​​of variables before and after intervention. For example, the mean of the physiological health index before intervention was 8.5 points, and after intervention it was 5.2 points, with an absolute difference of 3.3 points and a relative decrease of 38.8%, quantifying the overall shift in distribution. Cumulative disease risk association is established by linking the differences in bidistribution to cumulative disease risk using a knowledge base of association rules for environmental stress factors, clarifying the direction of the impact of variable changes on disease risk.

[0140] If the value of a variable shifts towards disease induction after intervention (e.g., increased ammonia nitrogen, decreased dissolved oxygen, increased stress level), then the change in that variable is considered to increase the cumulative risk of disease; if the value of a variable shifts towards disease suppression after intervention (e.g., decreased ammonia nitrogen, increased dissolved oxygen, decreased stress level), then the change in that variable is considered to decrease the cumulative risk of disease.

[0141] Marginal causality quantification: Based on the above difference analysis and risk association results, the marginal causal effect of each core variable on the cumulative disease risk is quantified:

[0142] Effect strength is indicated by the relative change in the distribution mean as the core indicator, combined with the direction of risk association, and the effect strength level is marked as weak, medium, or strong. For example, a relative change of risk-related indicators ≥30% after variable intervention is a strong effect, 10%-30% is a medium effect, and less than 10% is a weak effect.

[0143] The direction of the effect should be clearly identified as either positive (increasing disease risk) or negative (reducing disease risk). For example, increased ammonia nitrogen corresponds to a strong positive effect, while increased dissolved oxygen corresponds to a strong negative effect. The independence of the effect should be ensured by eliminating indirect effects caused by causal transmission between dependent variables, retaining only the direct marginal effects of the variable's own intervention. For instance, if the increase in stress level is partly caused by ammonia nitrogen intervention, the direct effect of ammonia nitrogen and the indirect effect of stress level need to be separately separated to ensure accurate quantification results. Finally, a list of marginal causal effects for each core causal characteristic variable should be compiled, clearly defining the direction, intensity, and independent contribution of the effect.

[0144] Step 302: Integrate the marginal causal effects of all variables and the current real-time values ​​of the core causal characteristic variables to generate a vector of immediate risk level and contribution of each risk factor as a dynamic causal fingerprint. This includes: collecting the current real-time values ​​of the core causal characteristic variables, comparing them with the parameter standards of the feature matrix in step 104 and the aquaculture safety threshold, and labeling the real-time status (safe, critical, abnormal) of each variable. For example, the real-time value of ammonia nitrogen is 0.18 mg / L (critical state), the real-time value of dissolved oxygen is 5.2 mg / L (safe state), and the stress level is 3 (abnormal state), providing a basis for risk level assessment.

[0145] Real-time risk level assessment, combining a list of marginal causal effects and the real-time status of variables, comprehensively calculates the current cumulative disease risk level:

[0146] First, for each abnormal / critical state variable, a risk weight is assigned according to its marginal effect strength: strong-effect variables account for 40%, medium-effect variables for 30%, weak-effect variables for 20%, and safe-state variables for 10% (basic correlation weight). Then, the weight coefficients are adjusted based on the direction of the variable's state deviation (the magnitude of deviation from the safe threshold). For example, ammonia nitrogen, an abnormal state with a strong positive effect, has its weight coefficient increased by 20%; stress levels, a critical state with a medium positive effect, have their weight coefficient increased by 10%. Finally, the immediate risk level is calculated through weighted summation and graded from 0 to 10 (0-3 for low risk, 4-6 for medium risk, and 7-10 for high risk), ensuring that the risk level is quantifiable and interpretable.

[0147] Risk factor contribution breakdown: Based on marginal causality, calculate the percentage contribution of each core variable to the current immediate risk level.

[0148] Positive effect variables (increasing risk): Contribution = Marginal effect strength of the variable × Percentage of its abnormal state amplitude. For example, ammonia nitrogen (strong positive effect 40%, abnormal amplitude 80%), contribution = 40% × 80% = 32%. Negative effect variables (reducing risk): Contribution = -(Marginal effect strength of the variable × Percentage of its safe state amplitude). For example, dissolved oxygen (strong negative effect 40%, safe amplitude 90%), contribution = -(40% × 90%) = -36%. The sum of the contributions of all variables corresponds to the immediate risk level, ensuring that the contribution breakdown accurately reflects the risk impact of each variable. The immediate risk level and the risk contribution of each variable are integrated to form a standardized vector as a dynamic causal fingerprint. The vector structure contains three core pieces of information:

[0149] The system includes basic identifiers such as fingerprint generation timestamp and corresponding aquaculture scenario (pond number, fish species, season); core indicators such as real-time risk level (0-10 points) and risk level description (low / medium / high risk); and factor details such as real-time values, status, marginal effect direction and intensity, and risk contribution percentage for each core causal characteristic variable. Simultaneously, a fingerprint update frequency is set (consistent with the temporal granularity of the feature matrix, updated every hour) to ensure that the fingerprint can dynamically track changes in aquaculture system variables and risk trends, providing core data support for subsequent disease early warning and control decisions.

[0150] In a preferred embodiment of the present invention, step 4 involves generating multiple state evolution trajectories within a preset future time window based on dynamic causal fingerprints; and solving for feasible solutions for stocking density regulation based on the disease probability and mortality distribution predicted by the state evolution trajectories, including:

[0151] Step 400: Initialize a dynamic simulator using the instantaneous state vector encoded in the dynamic causal fingerprint. This includes: importing the instantaneous state vector from the dynamic causal fingerprint into the simulator in its entirety. The vector contains real-time values ​​of core variables such as ammonia nitrogen, dissolved oxygen, water temperature, stress level, and abnormal behavior incidence rate (e.g., ammonia nitrogen 0.18 mg / L, dissolved oxygen 5.2 mg / L, stress level 3), state labels (safe / critical / abnormal), and risk contribution percentage. Simultaneously, load the causal path of the preset causal model (e.g., density, water quality, fish population status), time-series delay parameters (e.g., the effect of density change on ammonia nitrogen 2 hours later), and association strength rules to ensure that the simulator's computational logic is fully aligned with the preceding text. Simulator core parameter calibration: To meet the needs of state simulation, three core parameters are calibrated. The time granularity is maintained at 1 hour / step (consistent with time series data and the time lag range mentioned earlier, accurately capturing dynamic changes in variables); the variable transmission coefficient is matched with the strength of marginal causal effect (the coefficient for strong correlation paths is set at 0.7-0.9, and the coefficient for weak correlation is set at 0.2-0.4. For example, if dissolved oxygen and the incidence of abnormal behavior are strongly correlated, the coefficient is set to 0.8); the response trigger threshold is bound to aquaculture safety standards (such as ammonia nitrogen ≥0.2mg / L and dissolved oxygen ≤3.5mg / L triggering a stress response), ensuring that the parameters are adapted to the simulation accuracy and aquaculture practice.

[0152] Boundary and operational constraints are embedded: Two types of core boundary conditions are set: physical boundaries (water volume of the aquaculture pond, maximum oxygen supply capacity of the aeration equipment, maximum flow rate of the water exchange system) and biological boundaries (current stocking density of 20 fish / cubic meter, maximum system carrying capacity of 30 fish / cubic meter, and physiological tolerance threshold of fish species); at the same time, operational constraints are embedded (such as density adjustment rate ≤ 2 fish / cubic meter·hour to avoid sudden increases and decreases that may cause stress to the fish; fish state variables cannot be directly intervened in, but can only be indirectly controlled through water quality) to prevent abnormal states from occurring during the simulation that exceed the scope of actual operation.

[0153] A one-hour no-intervention simulation was conducted based on the initial state vector to verify whether the fluctuation of the core variable values ​​was ≤5% (the stable threshold under no-intervention conditions), whether the causal transmission was smooth (e.g., whether slight fluctuations in ammonia nitrogen affected the stress level according to the preset time lag), and whether there were no logical conflicts and the parameters were properly adapted before proceeding to the subsequent strategy generation stage.

[0154] The implementation process of the dynamic simulator is as follows:

[0155] A dynamic simulator is essentially a data-driven, rule-constrained, and iteratively validated dynamic numerical simulation system designed for closed recirculating aquaculture systems. Its core logic is to abstract the aquaculture system into a computable model of core variables, causal rules, and boundary constraints. Through standardized data parsing, rule parameterization, parameter calibration, constraint embedding, and pre-run validation, it achieves accurate replication of the aquaculture system's state evolution. The complete implementation process is described below, based on the initialization requirements of step 400:

[0156] The dynamics simulator automatically identifies and splits the three core data types of the real-time state vector, generating a structured index table. Real-time values ​​of variables such as ammonia nitrogen, dissolved oxygen, water temperature, stocking density, stress level, and abnormal behavior incidence rate are extracted (e.g., ammonia nitrogen 0.18 mg / L, dissolved oxygen 5.2 mg / L), converted to floating-point / integer values, and a basic index of variable name-value-unit is established. The state label (safe / critical / abnormal) of each variable is extracted and bound to the corresponding variable index, e.g., ammonia nitrogen - 0.18 mg / L - safe. The risk contribution percentage of each variable is extracted (e.g., ammonia nitrogen risk contribution 40%, dissolved oxygen 30%), converted to 0-1 weight coefficients, and added to the index table. The deconstructed data is then standardized to ensure no format conflicts and no invalid values. All variables are mapped to preset standard units (e.g., density = tails / cubic meter, water quality parameters = mg / L, stress level = 1-5). If the original data unit is incorrect (e.g., dissolved oxygen is labeled g / L), it is automatically converted to mg / L. For missing values ​​(e.g., the occurrence rate of abnormal behavior has no value), interpolation is used to complete them based on historical data from the same aquaculture scenario. For outliers (e.g., dissolved oxygen is negative, stress level > 5), they are reset according to common aquaculture knowledge (dissolved oxygen is reset to 0, stress level is reset to 5). At the same time, all data are bound to the timestamp of the initial time t0, generating an initial dataset with t0, variable index, standardized values, status labels, and risk weights.

[0157] The unidirectional causal path of density, water quality, and fish status is constructed as a directed weighted graph model. Nodes correspond to core variables (stocking density, ammonia nitrogen, dissolved oxygen, stress level, incidence of abnormal behavior, etc.); directed edges correspond to the causal transmission direction between variables, such as a positive edge from density to ammonia nitrogen, meaning that an increase in density will lead to an increase in ammonia nitrogen, and a negative edge from dissolved oxygen to the incidence of abnormal behavior, meaning that an increase in dissolved oxygen will reduce the incidence of abnormal behavior; the directed edges reserve assignment bits for association strength coefficients and time lag step sizes, providing an interface for subsequent parameter calibration.

[0158] The time-series delay rules are transformed into a mapping relationship between trigger variables, delay step sizes, and target variables for storage. Since the simulator's time granularity is 1 hour / step, the time unit needs to be converted to the corresponding step size. For example, if a change in density affects ammonia nitrogen 2 hours later, this is transformed into a mapping relationship where the trigger variable is stocking density, the delay step size is 2 steps, and the target variable is ammonia nitrogen. Similarly, if a change in ammonia nitrogen affects stress level 1 hour later, this is transformed into a mapping relationship where the trigger variable is ammonia nitrogen, the delay step size is 1 step, and the target variable is stress level. This method ensures that causal transmission strictly follows temporal logic, avoiding unrealistic situations of immediate response. Qualitative descriptions of strong / weak associations are transformed into quantitative transmission coefficient ranges, which are then bound to the weights of directed edges. The coefficient for strongly correlated paths (such as dissolved oxygen and the incidence of abnormal behavior) is set to 0.7-0.9, and in this example, the coefficient is set to 0.8. The coefficient for weakly correlated paths (such as water temperature and stress level) is set to 0.2-0.4. These coefficients are essentially the response magnitude of the target variable to the change of the trigger variable. For example, for every 1 mg / L decrease in dissolved oxygen, the incidence of abnormal behavior will increase by a coefficient of 0.8, that is, by 0.8 percentage points.

[0159] The dynamic simulator fixes the iteration step size at 1 hour / step. After each iteration of full-variable calculation, the virtual clock automatically advances by 1 hour, perfectly aligning with the time scale of time-series data and time-delay parameters, ensuring accurate capture of hourly dynamic changes in variables. Based on the strength of marginal causal effects, specific values ​​are assigned to the correlation strength coefficients of directed edges, rather than simply retaining intervals. First, the risk contribution weight in the dynamic causal fingerprint is retrieved, and the coefficients are weighted and corrected. For example, if the risk contribution of ammonia nitrogen is 40%, then its coefficient pointing to the stress level = 0.7 (lower limit of the strong correlation interval) × 1.2 (risk weighting coefficient) = 0.84. Then, fine-tuning is done in conjunction with the aquaculture scenario. For example, the coefficient pointing to dissolved oxygen consumption in the adult fish stage is 0.1 higher than that in the fry stage, ensuring that the coefficients closely match the actual correlation strength of the current aquaculture stage.

[0160] The aquaculture safety standards are converted into numerical threshold-trigger rules and stored. Threshold judgment and action triggering logic are set. For example, when ammonia nitrogen is ≥0.2mg / L, a stress level increase rule is triggered, and when dissolved oxygen is ≤3.5mg / L, an abnormal behavior incidence rate increase rule is triggered. At the same time, the thresholds are bound to the variable status labels (safe / critical / abnormal). For example, when ammonia nitrogen is 0.18mg / L (safe), no rule is triggered, when it is 0.21mg / L (critical), a low intensity response is triggered, and when it is 0.3mg / L (abnormal), a high intensity response is triggered.

[0161] The water volume of the aquaculture pond, the maximum oxygen supply capacity of the aeration equipment, and the maximum flow rate of the water exchange system are converted into upper and lower limits for variable values. For example, the maximum oxygen supply capacity of the aeration equipment corresponds to an upper limit of dissolved oxygen of 8 mg / L, and the upper limit of the water exchange flow rate corresponds to an ammonia nitrogen degradation rate of ≤0.05 mg / L / hour. If the variable value exceeds the upper or lower limit during the calculation, it will be automatically corrected to the boundary value. In terms of biological boundaries, the physiological tolerance threshold of fish species (such as water temperature 18-28℃) and the stocking density boundary (currently 20 fish / cubic meter, maximum 30 fish / cubic meter, minimum 12 fish / cubic meter) are stored. If the density value exceeds the range, it is directly judged as an invalid strategy. Regarding the density adjustment rate, a computational limit is set for the density variable to change by ≤2 fish / cubic meter·hour in a single iteration. If the density adjustment rate in the strategy exceeds this value, the simulator will automatically reduce the rate to 2 fish / cubic meter·hour. Regarding the intervention method, variable assignment permission restrictions are set so that fish swarm status variables (stress level, abnormal behavior incidence rate) can only be driven by water quality variables, and manual assignment or direct intervention is prohibited to ensure that the simulation conforms to the practical logic of regulating fish swarm status solely through water quality.

[0162] Using the standardized dataset at time t0 as the initial value, a virtual clock is started, and a 1-hour (1-step) uninterrupted iteration is performed. Only the natural fluctuations of variables (such as natural degradation of ammonia nitrogen and natural reoxygenation of dissolved oxygen) are simulated, without loading any density control strategy, and finally the variable values ​​at time t1 are generated. In terms of numerical stability verification, the relative fluctuation range of core variables (ammonia nitrogen, dissolved oxygen, stress level, etc.) at time t0-t1 is calculated. The fluctuation range is calculated as |(t1 value - t0 value) / t0 value|×100%. If the fluctuation of all variables is ≤5% (stability threshold), the value is considered stable. If it exceeds (e.g., ammonia nitrogen increases from 0.18 mg / L to 0.20 mg / L, with a fluctuation of 11.1%), the parameter calibration process is returned to adjust the natural degradation coefficient. In terms of causal transmission compliance verification, it is checked whether the changes in variables conform to the preset causal logic. For example, when ammonia nitrogen fluctuates slightly but does not reach the 0.2 mg / L threshold, the stress level remains unchanged. When dissolved oxygen increases slightly, the incidence of abnormal behavior decreases slightly. If there is a logical conflict such as ammonia nitrogen remaining unchanged but the stress level increasing, the triggering conditions of the causal rule are corrected. If the verification passes, the current parameters, rules, and constraints are locked, an initialization completion signal is generated, and the process is allowed to proceed to the breeding density control strategy generation stage in step 401. If the verification fails, specific anomalies are output (such as dissolved oxygen fluctuation of 6% or time lag rule not taking effect). After automatically adjusting parameters (such as reducing the natural fluctuation coefficient) or correcting the rules, a 1-hour no-intervention simulation is re-executed until the verification passes.

[0163] Step 401: In the dynamic simulator, a set of different stocking density control strategies are generated according to the preset stocking density adjustment range and adjustment step size. These strategies include: based on the biological boundary, physical boundary and economic needs of stocking embedded in step 400, the density adjustment range, step size, rhythm and duration parameters are defined one by one to ensure that the parameters are suitable for actual operation and cover the whole scenario: Adjustment range definition: based on the current stocking density of 20 fish / cubic meter as the benchmark, combined with the system's maximum carrying density of 30 fish / cubic meter and minimum economic stocking density of 12 fish / cubic meter, the comprehensive adjustment range is defined as 15-24 fish / cubic meter. The lower limit of 15 fish / cubic meter is higher than the minimum economic threshold of 12 fish / cubic meter to avoid the economic loss risk of low-density farming; the upper limit of 24 fish / cubic meter is lower than the maximum carrying capacity of the system of 30 fish / cubic meter, reserving redundancy for equipment control, and covering three core scenarios: density reduction (15-19 fish / cubic meter), maintenance (20 fish / cubic meter), and density increase (21-24 fish / cubic meter), eliminating ineffective ranges.

[0164] The adjustment step size is set to 1 fish / cubic meter. This is based on the precision requirements of density control in practical aquaculture. In a closed recirculating aquaculture system, the density fine-tuning range must be controlled within 1 fish / cubic meter to avoid exacerbating stress in the fish population. Simultaneously, it ensures sufficient granularity of the strategy, eliminating blind spots and enabling differentiated simulation of different density gradients. A two-stage rhythm of uniform adjustment and constant maintenance is uniformly adopted to adapt to the buffering and adaptation characteristics of the fish population and water quality. The adjustment rate is fixed at 1 fish / cubic meter·hour, strictly adhering to the practical constraint set in step 400: a density adjustment rate ≤ 2 fish / cubic meter·hour. This ensures rapid approach to the target density while avoiding sudden increases or decreases that could cause stress in the fish population. The adjustment duration dynamically changes with the difference between the target density and the current density. The absolute value of the difference is the adjustment duration (unit: hours). For example, if the difference between the target density of 21 fish / cubic meter and the current density of 20 fish / cubic meter is 1, the adjustment duration is 1 hour; if the difference between the target density of 16 fish / cubic meter and the current density is 4, the adjustment duration is 4 hours. The overall adjustment duration is controlled within the range of 1-4 hours. The maintenance period is uniformly set to 72 hours, starting from the moment the adjustment is completed. This covers the entire cycle after density control, from the establishment of water quality stability (usually 24-48 hours) to the complete response of the fish population (48-72 hours), ensuring that the simulation can capture the long-term effects after control.

[0165] Based on the principles of a single density target, a fixed adjustment rhythm, and other unchanged conditions, an initial control strategy is generated to ensure strategy independence and comparability.

[0166] Using a density range of 15-24 fish / m³ and a step size of 1 fish / m³, ten initial strategies were generated by successively substituting the target density value. Each strategy corresponds to a unique target density (15, 16, 17, 18, 19, 20, 21, 22, 23, 24 fish / m³). Each strategy clearly defines three core pieces of information, ensuring no key parameters are omitted: first, the target density adjustment value (i.e., the final maintained density); second, the adjustment duration (calculated based on the difference between the target density and 20 fish / m³); and third, the maintenance period (fixed at 72 hours, including the full response period after adjustment). To ensure the comparability of simulation results for different strategies, control rules were set with other conditions unchanged. Non-density parameters such as water temperature, water exchange frequency, basic power of aeration equipment, and feed input all used the initial settings from the dynamic causal fingerprint, only changing the stocking density-related parameters to eliminate interference from irrelevant variables in the simulation results.

[0167] Using the physical boundary set in step 400 as a rigid standard, the key focus is on verifying the equipment carrying capacity corresponding to the target density. The core includes the maximum oxygen supply capacity of the oxygenation equipment and the maximum flow rate of the water exchange system, while also considering the water quality load prediction results (such as whether the ammonia nitrogen generation and dissolved oxygen consumption exceed the equipment control limit).

[0168] Each strategy's target density was compared with the equipment's capacity threshold. For example, when the target density was 24 fish fry / cubic meter, the simulator's built-in load model predicted a 30% increase in ammonia nitrogen generation and a 25% increase in dissolved oxygen consumption. Calculations showed this was still within the maximum oxygen supply capacity of the aeration equipment (capable of supporting a 30% increase in dissolved oxygen consumption) and the degradation capacity of the water exchange system, thus deemed effective. If a strategy with a target density of 25 fish fry / cubic meter existed (exceeding the 15-24 range, used only as an example), its predicted 35% increase in dissolved oxygen consumption exceeded the aeration equipment's maximum capacity (30%), and was therefore eliminated. The 10 initial strategies generated in the 15-24 fish fry / cubic meter range were all fully verified and did not exceed the equipment's capacity and boundary constraints; therefore, all were retained to form a complete control strategy group.

[0169] Standardize effective strategies to provide a clear basis for subsequent state simulation and trajectory tracing:

[0170] Ten effective strategies are assigned unique numbers (Strategy 1 to Strategy 10) in ascending order of target density, corresponding to target densities of 15 fish / m³ to 24 fish / m³, facilitating rapid association between strategies and target densities. A structured archive table (in textual description form) is established to record the core parameters of each strategy, including strategy number, target density adjustment value, adjustment duration, adjustment rate (fixed at 1 fish / m³·hour), and maintenance period (72 hours). This is also linked to the boundary constraint parameters from step 400, forming a strategy-parameter-constraint correspondence, which is stored in the simulator strategy library. The predicted water quality load changes for each strategy are labeled. For example, Strategy 10 (target density 24 fish / m³) indicates a 30% increase in ammonia nitrogen generation and a 25% increase in dissolved oxygen consumption compared to the current level; Strategy 1 (target density 15 fish / m³) indicates a 25% decrease in ammonia nitrogen generation and a 20% decrease in dissolved oxygen consumption compared to the current level. This provides a clear load benchmark for the subsequent state simulation in step 402, assisting in the accurate extrapolation of water quality variable changes.

[0171] Step 402: For each stocking density control strategy, simulate the continuous state change process from the current moment to the end of the preset future time window under each strategy, and obtain the state evolution trajectory corresponding to each strategy. This includes: for the 10 effective stocking density control strategies (strategies 1 to 10, corresponding to 15-24 fish / cubic meter) archived in Step 401, based on the causal rules and constraint system of the dynamic simulator, and adhering to the principle of independent simulation of single strategies and comparison of control variables, stage-by-stage simulation of system state changes over 72 hours, generating a continuous state evolution trajectory specific to each strategy. The specific operation is as follows:

[0172] A unified 72-hour time window was preset for future simulations, with the immediate state of the dynamic causal fingerprint (at time t0, ammonia nitrogen 0.18 mg / L, dissolved oxygen 5.2 mg / L, stress level 3, etc.) serving as the starting point for simulations of all strategies. The controlled variable method was strictly implemented; non-density parameters such as water temperature, water change frequency (once every 4 hours), basic power of aeration equipment, and feed dosage (based on the current density baseline) remained unchanged from their initial settings. Only the density control parameters (target value, adjustment duration) for each strategy were differentiated in the simulations, ensuring that the simulation results of different strategies were driven solely by density changes and were comparable. Each strategy occupied a separate simulator thread, without interference. The entire simulation adhered to the unidirectional causal logic of density control, water quality changes, and fish population status responses, strictly following the time-series delay rules and variable correlation strength standards embedded in step 400.

[0173] The density is adjusted uniformly at a fixed rate of 1 fish / cubic meter / hour, with the current density value recorded at the end of each hour, until the target density is reached. For example, in Strategy 1 (target 15 fish / cubic meter, adjustment period 4 hours), the density is 19 fish / cubic meter at time t1 (1 hour later), 18 fish / cubic meter at t2, 17 fish / cubic meter at t3, 16 fish / cubic meter at t4, and maintained at 15 fish / cubic meter from t5 onwards. In Strategy 10 (target 24 fish / cubic meter, adjustment period 4 hours), the density rises to 24 fish / cubic meter at time t4, and then enters a maintenance phase. The density change trend is synchronized, combined with the water quality load prediction marked by the strategy, and the water quality variable changes are deduced according to the correlation strength rule, while embedding time-series lag rules. For example, in Strategy 10 (predicted load: ammonia nitrogen increases by 30%, dissolved oxygen decreases by 25%), the density increases hourly from t1 to t4. Since the time lag of density change on ammonia nitrogen is 2 hours, ammonia nitrogen still maintains its natural degradation trend from t1 to t2 (decreasing from 0.18 mg / L to 0.17 mg / L). At t3 (2 hours after the density change), the effect of density begins to appear, and ammonia nitrogen increases hourly according to a strong correlation. At t4, ammonia nitrogen rises to 0.20 mg / L (an increase of 11.1% from the initial level), which does not reach the predicted peak load. Dissolved oxygen is strongly correlated with density and has no time lag (respiratory consumption is reflected immediately). At t1, the density rises to 21 fish / m³, and dissolved oxygen immediately decreases from 5.2 mg / L to 5.0 mg / L. At t4, it decreases to 4.5 mg / L (a decrease of 13.5% from the initial level), which matches the phased changes in the predicted load. If the density reduction strategy is adopted (such as strategy 1), then the reverse deduction is performed: the density gradually decreases from t1 to t4 hours, the dissolved oxygen immediately rebounds, and the ammonia nitrogen begins to degrade rapidly after a 2-hour time lag. At t4, the ammonia nitrogen drops to 0.14 mg / L (a decrease of 22.2% from the initial value), which is close to the 25% reduction predicted by the load forecast.

[0174] Using changes in water quality variables as the transmission medium, stress levels and abnormal behavior rates are updated by combining time lags and correlation strength. For example, in Strategy 10, when ammonia nitrogen rises to 0.20 mg / L at time t3 (triggering the stress response threshold), because the time lag of ammonia nitrogen on stress levels is 1 hour, the stress level rises from level 3 to level 3.8 at time t4 (strong correlation, stress increases by 0.9 levels for every 0.1 mg / L increase in ammonia nitrogen), and the abnormal behavior rate increases from the initial 5% to 13% (dissolved oxygen decreases by 0.7 mg / L, which, based on a strong correlation coefficient of 0.8, increases by 5.6 percentage points, and after adding the effect of ammonia nitrogen, increases by a total of 8 percentage points). If there are fluctuations in weakly correlated variables (such as water temperature rising from 25℃ to 26℃ at time t2), the stress level only increases by 0.2 levels (weak correlation coefficient of 0.2), reflecting the difference in correlation strength. After updating every hour, the variable status is simultaneously labeled (such as labeling ammonia nitrogen of 0.20 mg / L as "critical") to ensure consistency of status.

[0175] Maintenance Phase (After Reaching Target Density): State Simulation and Steady-State Deduction

[0176] Starting from the end of the adjustment phase, with a fixed stocking density as the target value, the system steady-state establishment process before the end of the 72-hour time window is simulated, and the variable values ​​are updated every hour. The core manifestation is the change pattern of water quality from fluctuation to stability and the change of fish school status from adaptation to stability.

[0177] Following the logic of generation-consumption balance, the results gradually approach a steady-state value. Regarding ammonia nitrogen, the generation (fish excrement) corresponding to the target density balances with consumption (water exchange dilution, microbial decomposition). For example, in Strategy 10 (24 fish / m³), ammonia nitrogen rises slowly from t4 to t28 hours, reaching 0.23 mg / L at t28 (a 27.8% increase from the initial value, close to the predicted 30% peak load), and then fluctuates slightly (±0.01 mg / L). Regarding dissolved oxygen, the supply from aeration equipment balances with the fish's respiration and the water's natural reoxygenation. In Strategy 10, dissolved oxygen stabilizes at 4.0 mg / L at t24 (a 23.1% decrease from the initial value, close to the predicted 25% load load), while in Strategy 1 (15 fish / m³), dissolved oxygen stabilizes at 6.3 mg / L at t20 (a 21.2% increase from the initial value, close to the opposite of the predicted 20% load reduction). Water quality steady state is usually established 20-28 hours after the adjustment is completed, and thereafter it will only change slightly due to minor environmental fluctuations (such as water temperature and atmospheric oxygen partial pressure).

[0178] As water quality stabilizes, the fish population gradually adjusts to the corresponding level and remains stable. Regarding stress levels, Strategy 10 stabilizes at level 4.0 (critical) after water quality stabilizes, with an abnormal behavior rate stabilizing at 15%; Strategy 1 (low load) stabilizes at level 2.5 (safe), with an abnormal behavior rate stabilizing at 3%. The establishment of a stable fish population state lags behind water quality stabilizes by approximately 12-16 hours (the fish population's physiological adaptation cycle). For example, at water quality stabilization at t28, the fish population state stabilizes around t44 and remains at that level until t72 (the 72-hour endpoint) without significant fluctuations.

[0179] During the maintenance phase, the time lag and correlation strength rules are strictly adhered to. If temporary water quality fluctuations occur (such as a brief adjustment of the water exchange system at time t30 causing dissolved oxygen to rise to 4.2 mg / L), the incidence of abnormal behavior at time t31 will decrease to 13.2% based on the strong correlation coefficient (dissolved oxygen rises by 0.2 mg / L, corresponding to a decrease of 1.8 percentage points). Fluctuations in weakly correlated variables (such as water temperature dropping to 24℃ at time t40) will only cause the stress level to drop to level 3.8, without affecting the overall steady state.

[0180] During the simulation, the rules are not executed independently, but are deeply integrated into the two-stage variable update logic to ensure the rationality of causal transmission:

[0181] In addition to the 2-hour time lag between density and ammonia nitrogen during the adjustment phase, and the 1-hour time lag between ammonia nitrogen and stress level, the impact of minor water quality anomalies during the maintenance phase (such as a brief rise in ammonia nitrogen to 0.24 mg / L at t50) on the fish population status is also delayed by 1 hour (stress level rises to 4.1 at t51), avoiding unrealistic scenarios requiring immediate response. The entire process is quantified through the magnitude of variable changes. Strongly correlated paths (density-ammonia nitrogen, dissolved oxygen-abnormal behavior incidence, ammonia nitrogen-stress level) show changes within the set range, while weakly correlated paths (water temperature-stress level, turbidity-abnormal behavior incidence) have their changes controlled at low levels. For example, a 1 mg / L fluctuation in dissolved oxygen corresponds to a 15% fluctuation in the abnormal behavior incidence (strong correlation); a 1°C fluctuation in water temperature corresponds to only a 0.2 level fluctuation in stress level (weak correlation); and a 10 NTU increase in turbidity corresponds to only a 2% increase in the abnormal behavior incidence (weak correlation), ensuring consistent and effective implementation of the rules.

[0182] Using hourly time-series nodes (from t0 to t72, a total of 73 nodes), the full core data corresponding to each strategy is recorded in real time, forming a continuous and structured state evolution trajectory:

[0183] Each node includes a timestamp, stocking density value, water quality parameters (ammonia nitrogen, dissolved oxygen, water temperature) values ​​and status labels, fish population status indicators (stress level, abnormal behavior incidence rate) values ​​and status labels, and preliminary risk prediction values ​​(disease probability and mortality prediction based on the current status), with no key data missing. Node data is arranged in ascending order of timestamp, and each trajectory is bound to a corresponding strategy number, target density, and core parameters (adjustment duration, load prediction), forming a four-in-one trajectory data system of strategy, time series, status, and risk, which is stored in the simulator results library for easy subsequent aggregation and traceability.

[0184] After the single-strategy simulation is completed, three rounds of verification are carried out one by one to remove abnormal trajectories and correct and recalculate, ensuring that each trajectory can accurately replicate the system dynamics:

[0185] Referring to the physical and biological boundaries set in step 400, check whether the values ​​of all node variables are compliant. For example, ammonia nitrogen must not exceed 0.25 mg / L (the upper limit for equipment degradation), dissolved oxygen must not be lower than 3.5 mg / L (the lower limit of fish tolerance), and the stress level must not exceed level 5. If ammonia nitrogen suddenly rises to 0.3 mg / L or dissolved oxygen falls below 3.5 mg / L, it is judged as an abnormal trajectory. Check whether the changes in variables conform to unidirectional causal logic and linkage rules, and eliminate logical conflicts. For example, dissolved oxygen must not decrease synchronously when the density decreases, the stress level must not increase significantly when ammonia nitrogen does not change, and the state of the fish group must not fluctuate in the opposite direction after the water quality stabilizes. If logical contradictions occur, it is judged as an abnormal trajectory. A threshold for variable mutations was set (the hourly change of core variables must not exceed a preset range, such as stress level change ≤ 0.5 levels per hour, ammonia nitrogen change ≤ 0.02 mg / L per hour, and dissolved oxygen change ≤ 0.3 mg / L per hour). If abrupt changes occur, such as stress level increasing from level 2 to level 5 within one hour, or ammonia nitrogen increasing by 0.05 mg / L within one hour, the trajectory is considered abnormal. For trajectories deemed abnormal, the cause of the anomaly is traced (e.g., parameter conduction coefficient deviation, delay in triggering time-delay rules), and the corresponding parameters are fine-tuned (e.g., reducing the ammonia nitrogen conduction coefficient, correcting the time-delay step size). The 72-hour simulation of the strategy is then re-executed until a valid trajectory meeting the verification criteria is generated. All 10 strategies tested showed no anomalies, and their corresponding state evolution trajectories were all retained.

[0186] Step 403: Summarize the state evolution trajectories corresponding to all strategies, and output a set of multiple future state evolution trajectories within a preset time window. This includes: for the 10 strategy-specific state evolution trajectories that passed the verification in Step 402, the trajectories are summarized and optimized with standardized archiving, standardized format, and full-dimensional verification as the core, forming a set of future state evolution trajectories that are free of redundancy and comparable. This lays the foundation for subsequent risk indicator extraction and screening. The specific operation is as follows:

[0187] By associating each strategy number with its corresponding trajectory, precise strategy-trajectory binding is achieved. Simultaneously, identifier labeling and parameter association are completed, facilitating subsequent traceability and review. Based on the strategy numbers assigned in step 401 (Strategy 1 to Strategy 10, corresponding to target densities of 15-24 tails / cubic meter), an independent trajectory archiving directory is established. Each trajectory is stored separately and bound to a strategy number to prevent trajectory-strategy mismatch. Each trajectory is labeled with a unique identifier, with a unified identifier format of strategy number-time window-target density. For example, the trajectory identifier for Strategy 1 (target 15 tails / cubic meter) is Strategy 1-72 hours-15 tails / cubic meter, and the trajectory identifier for Strategy 10 (target 24 tails / cubic meter) is Strategy 10-72 hours-24 tails / cubic meter, clearly reflecting core information and avoiding confusion. In the archived information of each trajectory, the core parameters and simulation constraints of the corresponding strategy are synchronously associated. The core parameters include adjustment duration (1-4 hours, differentiated according to strategy), adjustment rate (fixed 1 fish / cubic meter·hour), and maintenance period (72 hours). The simulation constraints include the physical boundaries set in step 400 (such as dissolved oxygen upper limit 8 mg / L, ammonia nitrogen degradation rate upper limit 0.05 mg / L / hour), biological boundaries (fish species physiological tolerance threshold), and practical constraints (density adjustment rate limit, fish group status indirect control rules), forming a four-in-one archived system of identifier-trajectory-parameter-constraint, which can quickly retrieve complete associated information during subsequent tracing.

[0188] Based on the batch comparative analysis and subsequent step 404 risk indicator extraction requirements, the data standards of all trajectories are unified to ensure consistent structure, uniform accuracy, and complete dimensions.

[0189] Data format standardization: All trajectories adopt a structured format of time-series nodes, parameter items, standard values, and status labels, arranged in ascending order of timestamps (0-72 hours). The data of each time-series node (73 in total, including the initial time t0 and the final time t72) is in a separate row. The field order is uniformly set as "timestamp (hour), stocking density (tails / cubic meter), ammonia nitrogen (mg / L), dissolved oxygen (mg / L), water temperature (°C), stress level (1-5), abnormal behavior incidence (%), disease probability (%), and status label of each parameter (safe / critical / abnormal)," which facilitates batch reading and comparison.

[0190] The accuracy of each parameter is controlled according to preset standards: ammonia nitrogen is accurate to 0.01 mg / L (e.g., 0.18 mg / L, 0.20 mg / L, avoiding vague expressions such as 0.2 mg / L); dissolved oxygen is accurate to 0.1 mg / L (e.g., 5.2 mg / L, 4.5 mg / L); stress levels are labeled according to 1-5 levels (one decimal place can be retained, e.g., level 3.8, level 2.5, to ensure continuity of change and not exceed the level range); disease probability is accurate to 1% (e.g., 5%, 15%, generated based on real-time prediction of fish status); water temperature is accurate to 0.1℃; stocking density and abnormal behavior incidence are rounded to integers (density is fish / cubic meter, abnormal behavior incidence is %); and all trajectory accuracy standards are consistent to avoid affecting the comparison results due to accuracy differences. Each time-series node of each trajectory is checked for dimensionality to ensure that no dimension is missing. Each node must record all the above-mentioned field information completely. If there are missing predicted values ​​such as disease probability or abnormal behavior rate for individual nodes, they are interpolated and supplemented based on the fish population status and water quality parameters of the same node (e.g., when the stress level is 3.8 and the ammonia nitrogen is 0.20 mg / L, the disease probability is supplemented to 12%). At the same time, a supplementation mark is marked to ensure data integrity.

[0191] A comprehensive verification is conducted from three dimensions: coverage integrity, initial state consistency, and temporal node integrity, eliminating unqualified trajectories and ensuring the validity of the set. The number of trajectory sets is checked against the number of valid policies generated in step 401, ensuring that 10 policies correspond to 10 valid trajectories, with no omissions or redundant trajectories. Policy numbers (1-10) are compared one by one with the policy numbers in the trajectory identifiers to confirm that all valid policies have corresponding trajectories. If a trajectory is missing (e.g., the trajectory corresponding to policy 5 is missing), the simulation process in step 402 is retried, the trajectory for that policy is regenerated, and added to the set. Verify that the values ​​of all core parameters for each trajectory at time t0 (initial time) are completely consistent with the instantaneous state of the dynamic causal fingerprint. The ammonia nitrogen should be 0.18 mg / L, dissolved oxygen 5.2 mg / L, stress level 3, stocking density 20 fish / cubic meter, and water temperature 25℃ (example initial values). A small error of ±0.01 mg / L for ammonia nitrogen and ±0.1 mg / L for dissolved oxygen is allowed (within the allowable range of simulation accuracy). If the error exceeds the range, it is judged as an abnormal initial state. The trajectory generation process is backtracked, corrected, and then re-included into the set to ensure that the starting point of all trajectories is consistent and in accordance with the principle of controlled variables. Check that the time-series nodes of each trajectory cover the entire period from 0 to 72 hours, with no missing, duplicate, or disordered nodes. Ensure that each trajectory contains 73 time-series nodes (one per hour from t0 to t72). If there are missing nodes (e.g., missing data at time t25), complete them based on adjacent node data in a linear trend and mark the completion information. If there are duplicate nodes (e.g., duplicate records of data at time t30), remove redundant records. If there are disordered node time sequences (e.g., data at time t10 is arranged after time t11), adjust them to the correct position to ensure that the time sequence of all trajectories is consistent.

[0192] To address the potential for repetition in multi-strategy trajectories, we conduct repetition detection and optimization, retaining the optimal trajectory to ensure the set is free of redundancy.

[0193] Duplicate trajectory determination criteria: The determination is based on the consistency of core parameter values ​​across all time-series nodes. Specifically, if two trajectories using different strategies have completely identical core parameter values ​​(at uniform precision) across all time-series nodes from 0 to 72 hours for ammonia nitrogen, dissolved oxygen, stress level, and disease probability, and their status labels are identical, they are determined to be duplicate trajectories. Due to the varying target densities (15-24 fish / cubic meter) in this strategy, there are inherent differences in water quality load and fish population status responses. Theoretically, there should be no duplicate trajectories; these are only used as a fallback verification step. If duplicate trajectories appear, the one with the better target density is prioritized. The target density optimization standard balances aquaculture benefits and risks, meaning that the medium-density range (20-22 fish / cubic meter) is prioritized over high and low densities. Within the same density range, the trajectory with the lower disease probability is retained. After the above archiving, unification, verification, and removal processes, a structured, non-redundant, and comparable set of future state evolution trajectories containing 10 valid trajectories is formed. A summary list is generated simultaneously, labeling each trajectory with its identifier, corresponding strategy, core parameters, and verification results, and stored in a dedicated database.

[0194] Step 404: From the set of state evolution trajectories within multiple future preset time windows, extract the predicted disease probability value and predicted mortality value corresponding to the endpoint of each trajectory. This includes: using the t72 time (72-hour endpoint) data of each trajectory as the core, extracting the predicted disease probability value and predicted mortality value according to preset rules to ensure accurate extraction and logical derivation that fits the entire simulation process; relying on the standardized format of the trajectory set, quickly locating the path at t72 time by trajectory identifier, time sequence node sorting, and priority to retrieve the complete record with the timestamp marked as 72 hours in each trajectory; if there is an abnormal connection between the t72 time data and the t71 time data (such as parameter mutation), simultaneously retrieve the data of three consecutive nodes from t70 to t72 for cross-validation to ensure that the extracted endpoint data is the true value under steady state, rather than an abnormal fluctuation value; the core is based on the entire trajectory change of the fish population state variables, combined with the characteristics of high-risk diseases for comprehensive calculation. First, the steady-state values ​​of stress level and abnormal behavior incidence at time t72 are extracted. Then, the knowledge base of environmental stress factor association rules is retrieved to match the corresponding risk coefficients. For example, when the stress level is ≥4.0 (critical state) and the abnormal behavior incidence is ≥15%, the base probability of vibrio disease is 12% and the base probability of Ichthyophthirius multifiliis disease is 10%. When the stress level is 2.0-3.0 (safe state) and the abnormal behavior incidence is ≤5%, the base probabilities of both diseases are ≤3%. Subsequently, the overall trajectory trend is adjusted. If the average stress level is ≥3.5 and the number of critical water quality events is ≥3, the base probability is increased by 2-3 percentage points. If the water quality is safe and the fish population is stable throughout the entire process, the base probability is decreased by 1-2 percentage points. Finally, the highest probability of the two high-risk diseases is taken as the predicted disease probability value for this trajectory, rounded to the nearest 1% (e.g., 13.2% is rounded to 13%). Based on the 72-hour cumulative changes of physiological health index and abnormal behavior incidence, combined with the physiological characteristics of the fish species, the cumulative mortality risk throughout the entire process is reflected. First, extract the minimum, average, and t72 steady-state values ​​of the physiological health index (0-100 points, higher indicates better health) for each trajectory, along with the cumulative duration of abnormal behavior occurrence (e.g., the total duration when the abnormal behavior occurrence rate is ≥10%). Then, calculate the baseline mortality rate using the formula: Baseline Mortality Rate = (1 - Average Physiological Health Index / 100) × 5 + (Cumulative Duration of Abnormal Behavior / 72) × 3, with the coefficient calibrated based on historical aquaculture data (the physiological health index has a higher weight than abnormal behavior). If there are water quality exceedances (e.g., dissolved oxygen ≤3.5 mg / L, ammonia nitrogen ≥0.25 mg / L), add an additional 0.5 percentage points for each exceedance; if there are no water quality exceedances throughout the entire trajectory and the average physiological health index is ≥85 points, add 1 percentage point. The final mortality rate value is rounded down to an integer to ensure it is within a reasonable range.

[0195] By combining the knowledge base of association rules of environmental stress factors with simulation logic, the two extracted risk indicators are verified one by one, abnormal data are removed and the causes are traced to ensure that the indicators truly reflect the risks of the strategy.

[0196] Round 1: Range Verification (Hard Constraint Verification). Using 0-100% as the absolute reasonable range for both indicators, the indicator values ​​for each trajectory were checked one by one. If the disease probability was <0% or >100%, or the mortality rate was <0% or >100%, it was directly determined as an abnormal indicator. Abnormality Handling Method: First, the indicator derivation process was traced back to correct calculation deviations (such as incorrect coefficient application or errors in cumulative duration statistics), and the indicators were recalculated. If the values ​​still exceeded the range after correction, the trajectory data was determined to be abnormal, the corresponding trajectory was removed, and the cause of the abnormality was simultaneously marked (such as distortion of fish population status response due to deviation in simulated time delay parameters, or abnormal simulated values ​​of physiological health index), and recorded in the abnormality log. All 10 trajectory indicators extracted in this study were within the 0-100% range, and no such abnormalities were found.

[0197] Round Two: Trend Verification (Logical Consistency Verification). This round verifies the matching between core verification indicators and the overall trajectory state changes, establishing clear judgment rules to avoid discrepancies between indicators and actual simulated trends.

[0198] If the average stress level throughout the trajectory is ≥3.5, there are ≥5 critical / abnormal water quality nodes, and the cumulative duration of abnormal behavior is ≥20 hours, the predicted mortality rate should not be lower than 5% and the disease probability should not be lower than 12%. If it is lower than this standard, it is judged as an abnormally low indicator. For medium-risk trend assessment, if the average stress level throughout the trajectory is 2.5-3.5, there are 2-4 critical water quality nodes, and the cumulative duration of abnormal behavior is 10-20 hours, the mortality rate should be 3%-8% and the disease probability should be 8%-15%. If it exceeds this range, it is judged as abnormal. For low-risk trend assessment, if the average stress level throughout the trajectory is ≤2.5, there are no abnormal water quality nodes, and the cumulative duration of abnormal behavior is ≤10 hours, the mortality rate should not be higher than 3% and the disease probability should not be higher than 10%. If it is higher than this standard, it is judged as an abnormally high indicator. After verification, for abnormal trajectory of indicators (such as low risk throughout but 13% probability of disease), backtrack the simulation process, investigate the cause (such as deviation in the setting of correlation strength coefficient, untimely triggering of time delay rule), correct the parameters and regenerate the trajectory and indicators; if it cannot be corrected, remove the trajectory and mark the cause to ensure that the retained indicators are highly consistent with the simulation trend.

[0199] For the verified trajectories and indicators, a clear correlation is established to form a structured data base, which is adapted to the threshold comparison in the subsequent step 405.

[0200] The relationship between strategy number, target density, endpoint disease probability, and endpoint mortality rate is presented in a structured text format. Each record contains complete core information: strategy number (1-10), target density (15-24 fish / cubic meter, precisely corresponding), endpoint predicted disease probability (%, accurate to 1%), and endpoint predicted mortality rate (%, integer). Supplementary information is also provided (such as risk level prediction: low / medium / high risk, and whether it conforms to the trend). For example, strategy 1 (15 fish / cubic meter, low-risk trajectory) corresponds to strategy 1-15 fish / cubic meter -6%-2% (low risk, conforms to the trend); strategy 10 (24 fish / cubic meter, medium-risk trajectory) corresponds to strategy 10-24 fish / cubic meter -14%-7% (medium risk, conforms to the trend). Verify the consistency between the data in the association relationships and the trajectory endpoint data and verification results to avoid misaligned numbers and incorrect indicator entry. After verification, generate a standardized summary document, store it synchronously in the database, and establish an association index with the trajectory set in step 403 and the strategy parameters in step 401 to ensure that the entire chain of information from strategy to trajectory to indicator to verification results can be quickly retrieved during subsequent tracing, providing intuitive and accurate data support for the risk threshold comparison in step 405.

[0201] Step 405: Compare the predicted disease probability value and predicted mortality value with the preset disease risk safety threshold and mortality safety threshold respectively to obtain the comparison results, including: based on the preset risk safety thresholds (derived from aquaculture industry standards, fish physiological tolerance data, and a knowledge base of environmental stress factors), the disease probability safety threshold is ≤15% (corresponding to low disease risk, which can be maintained through routine prevention and control measures), and the mortality safety threshold is ≤5% (corresponding to an economically acceptable range, avoiding excessive losses in aquaculture); at the same time, warning thresholds (disease probability 15%-20%, mortality rate 5%-8%) and high-risk thresholds (disease probability >20%, mortality rate >8%) are set to assist in risk level determination and provide clear standards for screening. For the endpoint indicator of each trajectory, a one-to-one comparison is carried out with the two safety thresholds to form four types of comparison results:

[0202] Double compliance: Disease probability ≤15% and mortality rate ≤5% (risk fully controllable, meeting safety requirements); Single exceedance: Only disease probability exceeds the standard (≥15%) or only mortality rate exceeds the standard (≥5%) (partial risk out of control, strategy optimization required); Double exceedance: Both indicators exceed the standard (risk completely out of control, corresponding strategy directly eliminated); Warning range: Indicators within the warning threshold (strengthening prevention and control measures required, temporarily excluded from the safe range). Record and label the comparison results, detailing the comparison results for each trajectory, labeling the exceedance indicators and deviation range (e.g., disease probability 18%, exceeding the safety threshold by 3 percentage points), and simultaneously linking intermediate trajectory data (e.g., whether there was a momentary exceedance of disease probability within 48 hours, whether the water quality reached a critical state).

[0203] Step 406: Based on the comparison results, screen the state evolution trajectories where both the predicted disease probability value and the predicted mortality value are lower than their corresponding safety thresholds, and mark them as safe trajectories. This includes: prioritizing the screening of double-standard trajectories (disease probability ≤ 15% and mortality rate ≤ 5%), directly eliminating trajectories corresponding to single exceedances, double exceedances, and warning intervals, and initially forming a subset of safe trajectories. Intermediate state verification of the trajectories (core verification step): For the initially screened safe trajectories, trace back their complete state evolution process from 0 to 72 hours, focusing on verifying three aspects: first, no instantaneous risk exceedances (no disease probability ≥ 15% or mortality rate ≥ 5% at any given time); second, stable variable changes (the core variable value fluctuation range ≤ 10%, with no sudden changes); and third, consistent logical transmission (changes in density, water quality, and fish population status conform to the causal rules mentioned above, with no logical conflicts), eliminating trajectories with instantaneous exceedances, variable mutations, or logical anomalies. The safety trajectory is finally confirmed. The trajectory that passes the review is marked and its corresponding strategy number, target density and core advantages are clarified (such as low density strategy corresponding to low disease risk, medium density strategy taking into account the balance between breeding benefits and risks). The final set of safety trajectories is formed to ensure that each safety trajectory can achieve full risk control and no potential safety hazards.

[0204] Step 407: Based on the stocking density control strategy corresponding to each safe trajectory, extract the stocking density adjustment target value defined in the strategy, including: for each safe trajectory, trace its corresponding stocking density control strategy and extract the stocking density adjustment target value in the strategy, that is, the final density maintained after the adjustment is completed (ignore the intermediate density values ​​during the adjustment process, and only focus on the final stable density, such as 17 fish / cubic meter, 20 fish / cubic meter). The extracted target values ​​are deduplicated (e.g., if multiple safety tracks correspond to the same target value of 20 fish / cubic meter, only one target value is retained to avoid redundancy). At the same time, the number of corresponding safety tracks is associated (e.g., a target value of 17 fish / cubic meter corresponds to 3 safety tracks, and a target value of 20 fish / cubic meter corresponds to 2 safety tracks) to reflect the stability and adaptability of the target values ​​(the more tracks associated, the higher the safety and reliability of the target value). The deduplicated target values ​​are then verified again to ensure that they are all within the preset adjustment range (15-24 fish / cubic meter) and that there are no practical conflicts with the corresponding strategies (e.g., the target value of 24 fish / cubic meter does not exceed the oxygen supply limit of the aeration equipment and does not cause water quality variables to exceed the safe range). Finally, a list of safe density target values ​​is formed.

[0205] Step 408: The target values ​​for adjusting the stocking density corresponding to all safe trajectories are summarized and mapped to continuous numerical intervals to form a feasible solution set for stocking density control. This includes: arranging the target values ​​for safe density in ascending order (e.g., 15, 16, 17, 19, 20, 24 fish / cubic meter), merging adjacent continuous target values, and dividing them into several continuous numerical intervals, such as 15-17 fish / cubic meter (including 15, 16, 17, continuous without intervals, suitable for low-risk scenarios), 19-20 fish / cubic meter (including 19, 20, with an interval of 1 fish; since the middle 18 fish / cubic meter has no safe trajectory, it is divided separately, suitable for medium-risk scenarios), and 24 fish / cubic meter (a single value, with an interval of 4 fish from the previous interval, listed as a separate interval, suitable for low-risk scenarios with stable water quality).

[0206] To enhance practicality, three core attributes are added to each interval:

[0207] Basic attributes include the interval range, the number of target values ​​within the interval, and the number of corresponding safe trajectories. Applicable scenarios are defined by combining trajectory characteristics and simulation results, indicating the applicable scenarios for each interval (e.g., 15-17 fish / cubic meter is suitable for critical water quality and high disease susceptibility scenarios; 19-20 fish / cubic meter is suitable for stable water quality and moderate disease risk scenarios). The optimal target value is selected as the value within the interval that corresponds to the most safe trajectories and the lowest risk (minimum disease probability and mortality rate) (e.g., within the 15-17 fish / cubic meter interval, 17 fish / cubic meter corresponds to the lowest risk trajectory and is set as the optimal value). The feasible solution set is finally confirmed by verifying its completeness, reliability, and operability, ensuring that all strategies corresponding to target values ​​within the interval can generate safe trajectories, with no missing valid target values ​​or mixed-in target values ​​exceeding risk limits; the interval division conforms to aquaculture practice (adjusting the step size by 1 fish / cubic meter facilitates precise control); and the attribute annotations are clear, providing explicit guidance for aquaculture decision-making. Ultimately, a feasible solution set for controlling aquaculture density was formed, providing a precise and implementable basis for density control in subsequent aquaculture operations.

[0208] In a preferred embodiment of the present invention, step 5 involves dynamically calibrating and correcting the prediction deviation of the state evolution trajectory based on the latest real-time monitoring data stream to obtain the calibrated result; updating the preset structural causal model using the calibrated result to obtain a feasible solution set for stocking density regulation; and selecting and outputting the recommended dynamic adjustment amount of stocking capacity based on the feasible solution set for stocking density regulation, including:

[0209] Step 500: Obtain the latest monitoring data stream within a predetermined time window starting from the current moment, and convert it into a current actual observation state vector consistent with the dynamic causal fingerprint dimension. This includes: monitoring data acquisition and time window definition: Set the predetermined time window to 1 hour (consistent with the time series granularity and simulator iteration step size, taking into account both data timeliness and stability). Through the built-in sensors of the aquaculture system (water quality sensor, fish behavior monitoring equipment, density detector), collect continuous monitoring data streams within this time window in real time. The core data includes key variables such as ammonia nitrogen, dissolved oxygen, water temperature, stocking density, stress level, and abnormal behavior incidence rate. The data collection frequency is 5 minutes / time to ensure data continuity.

[0210] Data preprocessing and standardization: The raw data stream was cleaned to remove outliers caused by sensor malfunctions (such as dissolved oxygen sudden drops to 0 mg / L, stress levels exceeding the 1-5 range). Missing data was filled in using linear interpolation (if data was lost at a particular time point, it was filled in based on data from the adjacent 5 minutes before and after). Then, the data stream was standardized into an actual observation state vector of a single time node (the current time tcurrent) according to the dimensional standard of the dynamic causal fingerprint. The dimensions and accuracy were completely consistent with the dynamic causal fingerprint (ammonia nitrogen accurate to 0.01 mg / L, dissolved oxygen to 0.1 mg / L, and stress levels retained to one decimal place).

[0211] Vector dimension alignment and feature supplementation: Ensure that the actual observed state vector corresponds one-to-one with the core dimensions of the dynamic causal fingerprint. This includes not only the real-time values ​​of each variable, but also the state label (safe / critical / abnormal) and risk contribution ratio (recalculated based on real-time data, such as the real-time risk contribution of ammonia nitrogen rising to 45%). Finally, a complete vector of variable values, state labels and risk weights is formed, which can be directly compared with the predicted state points.

[0212] Step 501: From the set of multiple future preset time windows of state evolution trajectories, select the predicted state point that is close to the current actual observed state vector time point, and calculate the state deviation vector between the current actual observed state vector and the predicted state point. This includes: from the set of state evolution trajectories summarized in step 403, matching the predicted state point corresponding to time t_current (consistent with the end point of the latest monitoring time window) for each trajectory one by one to ensure that the time points are completely aligned (e.g., if the current time is t10, select the predicted state at time t10 from all trajectories); if t_current exceeds the adjustment stage (already entered the maintenance stage) of some trajectories, still select the predicted state point at the corresponding time to ensure matching consistency. Based on the actual observed state vector, calculate the difference between the predicted state point and the actual observed state vector variable by variable to generate the state deviation vector. The calculation rule is: deviation value = actual observed value - predicted state value, retaining the sign of the difference (positive deviation represents that the predicted value is lower than the actual value, and negative deviation represents that the predicted value is higher than the actual value). For example, the actual observed value of ammonia nitrogen was 0.20 mg / L, and the predicted value was 0.18 mg / L, with a deviation of +0.02 mg / L; the actual observed value of dissolved oxygen was 4.9 mg / L, and the predicted value was 5.0 mg / L, with a deviation of -0.1 mg / L. After calculating for each variable, the deviations were integrated into a complete deviation vector. Extreme deviations caused by temporary fluctuations in the sensor (such as a single variable deviation exceeding twice the normal fluctuation range of that variable, such as an ammonia nitrogen deviation of ±0.05 mg / L) were eliminated. The extreme deviation values ​​were corrected according to the steady-state fluctuation range of the variable to ensure that the deviation vector can truly reflect the difference between the prediction and the actual value, rather than accidental interference.

[0213] Step 502: Based on the state deviation vector, perform reverse gradient adjustment on the key model parameters of the dynamic simulator that generates the state evolution trajectory to obtain the adjusted dynamic simulator. This includes: locking in key parameters in the dynamic simulator that significantly affect the prediction results, including variable transmission coefficients (weights of strong / weak correlation paths), time-series lag step sizes (e.g., 2-hour lag for density and ammonia nitrogen), and response trigger thresholds (e.g., ammonia nitrogen stress trigger threshold of 0.2 mg / L). These parameters directly determine the trajectory prediction accuracy and are the core objects of adjustment. Gradient descent is used to adjust the parameters, with minimizing the absolute value of the deviation vector as the optimization objective. The learning rate is set to 0.05 (balancing adjustment magnitude and stability). The gradient direction is calculated for each parameter and iteratively adjusted. For example, if the ammonia nitrogen deviation is +0.02 mg / L (prediction is too low), it indicates that the density and ammonia nitrogen transmission coefficients are too small, so the coefficients are adjusted along the positive gradient direction (e.g., from 0.8 to 0.82); if the dissolved oxygen deviation is -0.1 mg / L (prediction is too high), the density and dissolved oxygen transmission coefficients are adjusted downwards (e.g., from 0.8 to 0.78). The time lag step size is only fine-tuned when the deviation persists (e.g., if three consecutive deviations show a lag in ammonia nitrogen response, adjust the density and ammonia nitrogen time lags from 2 hours to 2.5 hours). The trigger threshold is corrected according to the actual observed stress triggering situation (e.g., if stress is triggered only when the actual ammonia nitrogen level is 0.21 mg / L, adjust the threshold from 0.2 mg / L to 0.21 mg / L). The adjusted parameters are substituted into the simulator, and a 30-minute short-cycle no-intervention simulation is conducted. The deviation between the simulation results and the actual observed data is compared. If the mean absolute value of the deviation vector is ≤3% (preset accuracy threshold), the parameter adjustment is deemed effective. If the target is not met, the reverse gradient adjustment process is repeated until the accuracy requirements are met, and the adjusted kinetic simulator is locked.

[0214] Step 503: Using the adjusted dynamic simulator, starting from the current observed state vector, a new set of state evolution trajectories is generated as the calibrated result. This includes: using the current observed state vector generated in step 500 as the simulation starting point (replacing the original initial state of the dynamic causal fingerprint), setting the simulation time window to tcurrent to the 72-hour endpoint (e.g., t10, with 62 hours remaining in the simulation), and using the 10 stocking density control strategies from step 401 (target density and adjustment duration remain unchanged), while keeping the control variables (water temperature, water exchange frequency, etc.) consistent with real-time observations. Following the simulation logic of the adjustment and maintenance phases in step 402, the simulation is conducted based on the adjusted parameters, updating variable values ​​hourly, with a focus on optimizing prediction accuracy. The adjustment phase focuses on correcting the correlation response deviation between density and water quality, while the maintenance phase accurately simulates the adaptation process between water quality steady state and fish population state, while strictly adhering to the time lag and correlation strength rules (which have been optimized with parameter adjustments). The three-round verification criteria (boundary compliance, logical coherence, and no abrupt changes) of step 402 are reused to verify each of the 10 newly generated trajectories, eliminate abnormal trajectories and re-simulate them, and finally form a set of calibrated state evolution trajectories as the effective result after deviation correction.

[0215] Step 504: The calibrated results, i.e. the new state evolution trajectories, are used as new training samples and input into the preset structural causal model to update the functional relationship parameters between the core causal feature variables in the preset structural causal model, resulting in an updated structural causal model. This includes: splitting the 10 calibrated new trajectories into a sample set (1 sample per hour, for a total of 62 × 10 = 620 samples) according to time steps. Each sample contains the dependent variable value (density, water quality), the effect variable value (fish school state), and a time series label to ensure that the samples cover different density strategies and different state stages (adjustment / maintenance).

[0216] The sample set is input into a pre-defined causal model, with a focus on updating the functional relationship parameters between core causal feature variables. For Bayesian network models, the conditional probability distribution between nodes is updated (e.g., the probability of stress level 3.5 at a density of 22 fish / m³ and dissolved oxygen of 4.5 mg / L is updated from 20% to 28%). For linear correlation models, the regression coefficients between variables are corrected (e.g., the regression coefficient of ammonia nitrogen on stress level is adjusted from 0.9 to 0.92) to ensure that the functional relationship conforms to the calibrated trajectory. The updated model is validated through causal path consistency checks and sample fit checks to verify its effectiveness, ensuring that the unidirectional causal path of density, water quality, and fish status remains unchanged and without logical conflicts; the sample fit R0 is verified. 2 A value of ≥0.85 (preset threshold) indicates that the model can accurately capture the correlation patterns of variables and lock in the updated structural causal model.

[0217] Step 505: Using the updated structural causal model, based on the latest current observed state vector, generate an updated dynamic causal fingerprint. This includes: re-encoding core information based on the current observed state vector using the updated structural causal model, including the instantaneous state vector (real-time variable values, state labels, risk contribution), optimized causal path weights (e.g., density, the weight of ammonia nitrogen is adjusted from 0.75 to 0.78), corrected time-series delay parameters, and variable association strength coefficients, ensuring that the fingerprint information is fully aligned with the updated model. Using the dynamic causal fingerprint format from Step 3, integrate the extracted core information to form a structured fingerprint, annotating the update timestamp (tcurrent) and parameter adjustment records (e.g., details of the adjustment of the transmission coefficient and time delay step size) for easy subsequent tracing and review. Verify the consistency between the updated dynamic causal fingerprint and the current observed vector and the adjusted simulator parameters, ensuring no dimensional bias or parameter conflicts. If conflicts exist (e.g., the time delay parameters in the fingerprint are inconsistent with the adjusted simulator parameters), correct the fingerprint based on the simulator parameters, ensuring that the fingerprint can be directly used as a new initial condition.

[0218] Step 506: Using the updated dynamic causal fingerprint as the new initial condition, obtain the updated feasible solution set for stocking density control. This includes: following the process from steps 400 to 408, initializing the adjusted dynamic simulator with the updated dynamic causal fingerprint, regenerating 10 state evolution trajectories, extracting the endpoint disease probability and mortality rate, comparing with safety thresholds (disease probability ≤15%, mortality rate ≤5%), screening safe trajectories, extracting the corresponding density adjustment target values, and finally mapping them to continuous numerical intervals; comparing the feasible solution sets before and after the update, marking interval changes (e.g., the original solution set was 17-20 fish / cubic meter, the updated one is 16-19 fish / cubic meter), and the migration of the optimal target value (e.g., the original optimal one was 20 fish / cubic meter, the updated one is 18 fish / cubic meter), and simultaneously explaining the reasons for the changes (e.g., real-time data shows an increase in ammonia nitrogen risk, and the solution set shifts towards the low-density interval). Verify the operability of the updated solution set, ensuring that all target values ​​within the interval conform to the physical / biological boundaries (such as not exceeding the oxygenation equipment capacity and not falling below the minimum economic density), and that the simulation results of the corresponding trajectories are consistent with the real-time state trend, thus locking in the feasible solution set for the updated aquaculture density control.

[0219] Step 507: The updated feasible solution set for stocking density regulation is matched with the preset objective function to select the corresponding stocking density adjustment target value. This target value is used as the recommended dynamic adjustment amount for stocking capacity. The preset objective function is a dual-objective optimization of maximizing stocking benefits and minimizing risk, with weights allocated as benefit α and risk β (α + β = 1), where benefit weight α has a higher proportion, reflecting the principle of prioritizing benefits and controlling risk. Benefit indicators are correlated with stocking density, and a medium-density range [D] is set. - D+ ] (D - D is the lower limit of the interval. + The optimal benefit is achieved by taking the upper limit of the interval; the risk index is related to the endpoint disease probability and mortality rate, and the lower the values ​​of both, the lower the risk and the better the objective function performance.

[0220] The function expression is corrected to:

[0221] Objective function value = α × (1 - |D - D0| / D0) + β × (1 - (P + M) / R) max );

[0222] The function takes values ​​between 0 and 1. Higher values ​​indicate better overall benefits and risk management effectiveness. The meanings of each letter are defined as follows:

[0223] α represents the weight of the aquaculture benefit target (set to correspond to 60%), and β represents the weight of the aquaculture risk target (set to correspond to 40% in this case), with α + β = 1, used to balance the contributions of the two targets and achieve normalized weighted summation; D represents the target value for adjusting the aquaculture density to be evaluated, and D0 represents the medium-density range [D - D + The optimal benefit baseline density within [ ] (in this case, D0 corresponds to 19 fish / cubic meter), D - D + These represent the lower and upper limits of the optimal density range for economic benefits (corresponding to the original 18 fish / cubic meter and 20 fish / cubic meter, respectively); P is the predicted disease probability at the endpoint, and M is the predicted mortality rate at the endpoint, both expressed as percentages (%); R max The overall safety limit for risk indicators (R is set in this case) max The corresponding value of 20 (i.e., the upper limit of the safety threshold of P+M) is used to normalize the comprehensive risk value to the 0-1 range; the two 1s in the formula are general normalization benchmark constants, ensuring that the values ​​of the benefit sub-items and risk sub-items are mapped to the 0-1 range, ensuring the uniformity of the dual-objective scoring scale, and facilitating weighted superposition.

[0224] Matching calculation and optimal value selection: For each density target value in the updated feasible solution set, substitute it into the objective function to calculate the corresponding function value, and select the target value with the highest function value as the optimal adjustment target value; if multiple target values ​​have the same function value, prioritize the target value that is closest to the current density and has the shortest adjustment time (to reduce the risk of fish stress). For example, in the feasible solution set of 16-19 fish / cubic meter, 18 fish / cubic meter corresponds to a function value of 0.92 (the highest), which is selected as the optimal target value. Based on the optimal target value and the current actual density, calculate the dynamic adjustment amount of the aquaculture capacity: adjustment amount = optimal target value - current actual density (positive for increasing density, negative for decreasing density). Combined with the adjustment rate in step 401 (1 fish / cubic meter·hour), calculate the adjustment time (adjustment time = |adjustment amount|), and determine the maintenance period (until the 72-hour endpoint). The final output recommendation plan includes core information: dynamic adjustment amount (e.g., -2 fish / cubic meter, i.e., reducing density by 2 fish / cubic meter), adjustment duration (2 hours), adjustment rate (1 fish / cubic meter·hour), optimal target density (18 fish / cubic meter), risk warning (e.g., disease probability 8% and mortality rate 2% after adjustment, which is within the safe range), and also indicates precautions during the adjustment process (e.g., monitoring dissolved oxygen every 30 minutes during the adjustment period to ensure it is not lower than 3.5 mg / L), providing clear guidance for aquaculture practice.

[0225] A dynamic assessment system for aquaculture capacity, comprising the following components:

[0226] The data acquisition module is used to collect real-time water quality sensor data and fish behavior and physiological status monitoring data of the closed recirculating aquaculture system, and integrate historical disease outbreak records. Based on a predefined knowledge base of environmental stress factor association rules, it performs time-series alignment and feature fusion of multi-source heterogeneous data to construct a feature matrix.

[0227] The extraction module is used to extract core causal feature variables based on the feature matrix;

[0228] The generation module is used to generate dynamic causal fingerprints based on core causal feature variables and a preset structural causal model by calculating the counterfactual distribution of the set of variables under preset intervention conditions.

[0229] The processing module is used to generate multiple state evolution trajectories within a preset future time window based on dynamic causal fingerprints; and to obtain a feasible solution set for controlling breeding density based on the disease probability and mortality distribution predicted by the state evolution trajectories.

[0230] The determination module is used to dynamically calibrate and correct prediction deviations of the state evolution trajectory based on the latest real-time monitoring data stream to obtain the calibrated results; the calibrated results are used to update the preset structural causal model to obtain a feasible solution set for stocking density regulation; based on the feasible solution set for stocking density regulation, the recommended dynamic adjustment amount of stocking capacity is selected and output.

[0231] This method overcomes the limitations of traditional single-parameter dependence through multi-dimensional data integration and standardization, significantly improving data quality and feature completeness.

[0232] For water quality sensor data, the outlier removal rate was controlled within 5%, and the missing value supplementation was ≤3%. After linear interpolation and moving average interpolation, the data missing rate was ultimately ≤1%. The fish population status time series, after preprocessing, contained 19 features per sequence (8 behavioral features + 6 physiological features + 5 fusion indicators), perfectly aligned with the water quality data's time granularity (1 hour) and dimensional range (0-1 interval), laying the foundation for subsequent fusion analysis. The model was trained on over 80,000 labeled samples. For two species, Litopenaeus vannamei and California bass, the behavioral feature recognition accuracy reached 89.5%, and the physiological state assessment accuracy was ≥92%. The quantification error for core indicators such as cortisol and ALT was only 3.7%-4.2%. The accuracy of mild stress recognition was improved to 91.7% after optimization, accurately capturing early abnormal states in fish populations. By combining the knowledge base of association rules of environmental stress factors, the matching degree of the time series data before and after the disease is ≥93%, which can accurately identify the core factor combination corresponding to different diseases such as vibriosis and ichthyophthirius multifiliis. For example, the accuracy rate of identifying the combination of increased ammonia nitrogen, decreased dissolved oxygen and increased stress level associated with vibriosis is 92.8%, providing a precise variable set for causal analysis.

[0233] By employing time-delay correlation analysis and geometric constraint modeling, we can achieve accurate quantification and redundancy removal of variable relationships.

[0234] The time-delay cross-correlation calculation covers the entire response period from 0 to 72 hours, with the optimal time-delay identification error ≤ 1 hour, and the accuracy rate of pairwise variable correlation strength classification ≥ 90%. After redundant edge removal (retaining edges with a weight ≥ 0.2), the constructed weighted directed graph has no missing core correlation edges, and redundant correlations of similar variables are reduced by 40%, significantly reducing the computational load of the model. After variance inflation factor (VIF) test (VIF < 10) and verification with aquaculture logic, more than 30% of redundant equations are removed, and the final simplified constraint equation set has no linear dependence. The consistency between the equations and geometric constraints and the actual aquaculture logic reaches 95%, and the accuracy rate of core causal characteristic variable screening is ≥ 92%, effectively avoiding interference from secondary variables.

[0235] Based on structural causal models and dynamic simulations, we can achieve more accurate risk prediction and state projection.

[0236] Fingerprints can accurately encode the contribution and immediate risk level of risk factors such as water temperature, pathogens, and fish behavior. The risk factor quantification deviation is ≤4%, and it is updated hourly, enabling real-time tracking of system risk dynamics and addressing the pain point of traditional models' inability to predict risks that have not yet occurred. Within a 72-hour time window, the trajectory simulation matches the actual state changes by 92%, the prediction error for the water quality steady-state establishment time is ≤2 hours, and the prediction error for the fish state steady-state lag time is ≤1 hour. The hourly fluctuation range of core variables (ammonia nitrogen, dissolved oxygen, and stress level) is ≤5%, with no parameter mutations (fluctuations exceeding the preset range). The trajectory reliability is significantly better than that of traditional static models. The prediction error for the endpoint disease probability is ≤3 percentage points, and the prediction error for mortality is ≤1 percentage point. The prediction reliability for low-probability disease events (occurrence rate <5%) is improved by more than 40%, avoiding the neglect of risk accumulation by traditional models.

[0237] Through a closed-loop mechanism of monitoring-prediction-calibration-update, model adaptation and error cancellation are achieved.

[0238] Dynamic calibration based on real-time monitoring data can offset prediction errors caused by sudden fluctuations such as abrupt changes in water temperature and dissolved oxygen. After calibration, the mean absolute value of the prediction deviation vector is ≤3%, a reduction of more than 60% compared to before calibration. After updating the structural causal model using the calibration trajectory as a sample, the model fit R... 2 ≥0.85, after adjusting the core parameters (transmission coefficient, time delay step size), the adaptability to the actual scenario is improved by 20%, and it can be stably applied across breeding stages and fish species (there is no significant difference in adaptability between whiteleg shrimp and California bass).

[0239] Overall technical performance comparison and advantages:

[0240] (I) The core indicators are compared with those of traditional evaluation methods as shown in Table 1 below.

[0241] Table 1 Comparison of core indicators with traditional evaluation methods

[0242]

[0243] The Implementation Results of Core Technological Advantages:

[0244] By employing dynamic causal fingerprinting and counterfactual distribution calculations, the lagged effects of density adjustments are precisely quantified, avoiding the dual problems of conservative waste and overly optimistic approaches inherent in traditional models. After regulation, the mortality rate in aquaculture is stably controlled within the safe threshold of 5%, a reduction of 40%-50% compared to traditional methods. Under the premise of ensuring controllable risks, facility carrying capacity utilization is increased by 15%-20%, and the aquaculture efficiency per unit water volume increases by 10%-15%, achieving a balance between maximizing production capacity and minimizing risk. Dynamic causal fingerprinting clearly identifies the sources of risk (such as ammonia nitrogen accumulation and fish stress) and the causes of the regulation effects, overcoming the limitations of traditional models that only output numerical values ​​without decision-making basis. User trust in the assessment results is increased by over 80%, and the success rate of implementing regulation solutions is ≥90%. It is compatible with closed recirculating aquaculture systems, covering the entire stage from larvae to adulthood for species such as Litopenaeus vannamei and California bass. Facing complex scenarios such as water temperature fluctuations and sudden changes in dissolved oxygen, its stability is improved by 30% compared to traditional models.

[0245] Typical application scenario effect verification:

[0246] (a) High-risk scenarios for disease outbreaks (during periods of high water temperature in summer)

[0247] After applying this method, the early warning accuracy rate of high-incidence diseases such as vibriosis and ichthyophthirius multifiliis reached 91.2%. Through density regulation and feasible set screening (optimal range 17-19 fish / cubic meter), the incidence rate of diseases decreased from 18% under traditional regulation to 7.5%, the mortality rate decreased from 6.8% to 2.2%, and the yield per unit water body increased by 12% compared with traditional conservative regulation.

[0248] (ii) Scenario of sudden fluctuations in water quality (dissolved oxygen suddenly drops to 3.5 mg / L)

[0249] After real-time monitoring data triggers dynamic calibration, the model parameters are adjusted within 1 hour, the state trajectory prediction deviation is corrected from 8.2% to 2.9%, and through the updated feasible solution set (density reduced by 2 fish / cubic meter), dissolved oxygen is restored to a safe level within 24 hours, the fish stress level is reduced from level 4 to level 2.5, no large-scale mortality events occur, and losses are reduced by more than 70% compared to traditional models.

[0250] This dynamic capacity assessment method for aquaculture achieves its core objectives of accurate prediction, dynamic adaptation, and efficient regulation through a full-chain technical solution involving multi-source data fusion, causal modeling, and dynamic calibration. In terms of quantitative results, the accuracy of capacity assessment, the lead time for risk prediction, and facility utilization are all significantly improved compared to traditional methods. In practical applications, it can effectively balance aquaculture benefits and disease risks, adapt to complex and ever-changing aquaculture environments, and provide a scientific and feasible capacity regulation solution for closed recirculating aquaculture systems, possessing broad application value.

[0251] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

Claims

1. A method for dynamic assessment of aquaculture capacity, characterized in that, The method includes: Step 1: Collect real-time water quality sensor data and fish behavior and physiological status monitoring data of the closed recirculating aquaculture system, and integrate historical disease outbreak records. Based on a predefined knowledge base of environmental stress factor association rules, perform time-series alignment and feature fusion of multi-source heterogeneous data to construct a feature matrix. Step 2: Based on the feature matrix, extract core causal feature variables, including: calculating the time-lag cross-correlation and covariance between each pair of variables based on the time series of all aquaculture variables in the feature matrix to construct a weighted directed graph characterizing the strength of the association between variables, serving as the initial variable relationship graph; transforming the initial variable relationship graph into a geometric constraint graph model; determining mutually independent simplified constraint equation sets based on the geometric constraint graph model; determining core causal feature variables from the solution space of the simplified constraint equation sets; and constructing a predefined knowledge base of environmental stress factor association rules based on long-term aquaculture practice data of closed recirculating aquaculture systems, industry technical standards, publicly available research literature, and targeted experimental data, forming a standardized and callable set of rules, which includes three parts: first, a mapping table of disease types and stress factor combinations. The system covers several key aspects: First, it includes a comprehensive set of data. First, it addresses common bacterial, viral, parasitic, and stress-related diseases in closed-loop recirculating aquaculture systems. Each disease corresponds to a unique combination of characteristic stress factors, which includes subsets of water quality parameters and fish population status parameters. Second, it employs a single-factor safety threshold and a multi-factor synergistic threshold system. Single-factor safety thresholds are calibrated according to the aquaculture species and growth stage, while multi-factor synergistic thresholds define the safety boundaries under different parameter combinations. Third, it includes a time-related rule base that clarifies the time interval between the occurrence of abnormalities in various stress factor combinations and the induction of corresponding diseases, the positive correlation between the duration of abnormal states and the probability of disease outbreaks, and the impact of the temporal sequence of factor abnormalities on disease types. All rule entries are labeled with confidence levels, determined based on the corresponding data sample size and validation pass rate. Step 3: Based on the core causal characteristic variables and the preset structural causal model, a dynamic causal fingerprint is generated by calculating the counterfactual distribution of the variables under the preset intervention conditions. This includes: in the preset structural causal model, for each core causal characteristic variable, a preset intervention condition representing the external control operation is defined, and the conditional probability distribution of all core variables under the intervention condition is calculated to obtain the counterfactual distribution; based on the difference between the counterfactual distribution and the actual observed distribution of the core causal characteristic variables, the marginal causal effect of each variable on the cumulative disease risk is quantified; and by combining the marginal causal effects of all variables and the current real-time values ​​of the core causal characteristic variables, a vector of the immediate risk level and the contribution of each risk factor is generated as the dynamic causal fingerprint. Step 4: Based on the dynamic causal fingerprint, generate multiple state evolution trajectories within a future preset time window; based on the disease probability and mortality distribution predicted by the state evolution trajectories, solve for the feasible solution set for breeding density regulation. Step 5: Dynamically calibrate and correct prediction deviations of the state evolution trajectory based on the latest real-time monitoring data stream to obtain calibrated results; update the preset structural causal model using the calibrated results to obtain a feasible solution set for stocking density regulation; and select and output the recommended dynamic adjustment amount of stocking capacity based on the feasible solution set for stocking density regulation.

2. The method for dynamic assessment of aquaculture capacity according to claim 1, characterized in that, Step 1 includes: The real-time water quality sensor data is cleaned and missing value interpolation is performed to generate a standardized time series sequence of first water quality parameters. Fish behavior and physiological state monitoring data are input into a preset behavior and physiological analysis model, processed, and transformed into a standardized time series of fish state. The time points recorded in the history of disease outbreaks are matched and aligned with the time series of the first water quality parameter and the time series of fish status, so that the water quality and fish status within the matching period are marked as the context state data corresponding to the disease event. The context state data is matched with a pre-defined knowledge base of environmental stress factor association rules to identify key factor combinations that appear before and after disease events and conform to stress rules. Using a unified timeline after time alignment as a benchmark, the time series of the first water quality parameter, the time series of the fish population status, and the combination of key factors are vectorized and feature-fused to construct a feature matrix.

3. The method for dynamic assessment of aquaculture capacity according to claim 2, characterized in that, The contextual state data is matched with a pre-defined knowledge base of environmental stress factor association rules to identify key factor combinations that occur before and after disease events and conform to stress rules, including: The context state data is input into a pre-defined knowledge base of environmental stress factor association rules for matching to obtain the matching results; Based on the matching results, environmental parameters that continuously exceeded the preset threshold in the time series of the first water quality parameter before and after the outbreak of disease events, as well as abnormal indicators that showed correlated abnormal fluctuations in the time series of fish population status, were identified. By combining environmental parameters with abnormal indicators, key factor combinations associated with disease outbreaks are identified.

4. The method for dynamic assessment of aquaculture capacity according to claim 3, characterized in that, Based on dynamic causal fingerprints, multiple state evolution trajectories within a preset future time window are generated, including: A dynamic simulator is initialized using the instantaneous state vector encoded in the dynamic causal fingerprint; In the dynamic simulator, a set of different stocking density control strategies are generated based on the preset stocking density adjustment range and adjustment step size. For each stocking density control strategy, the continuous state change process from the current moment to the end of the preset time window under each strategy is simulated to obtain each state evolution trajectory corresponding to the strategy. Summarize the state evolution trajectories corresponding to all strategies, and output a set of state evolution trajectories within multiple future preset time windows.

5. The method for dynamic assessment of aquaculture capacity according to claim 4, characterized in that, Based on the disease probability and mortality distribution predicted by the state evolution trajectory, a feasible solution set for stocking density regulation is obtained, including: From the set of state evolution trajectories within multiple future preset time windows, extract the predicted disease probability value and predicted mortality value corresponding to the end time of each trajectory; The predicted disease probability value and predicted mortality value are compared with the preset disease risk safety threshold and mortality safety threshold respectively to obtain the comparison results. Based on the comparison results, the evolution trajectory of the state where the predicted disease probability value and the predicted mortality value are both lower than their corresponding safety thresholds is selected and marked as a safe trajectory. Based on the stocking density control strategy corresponding to each safe trajectory, extract the stocking density adjustment target value defined in the strategy; The target values ​​for adjusting the stocking density corresponding to all safe trajectories are aggregated and mapped to a continuous numerical range to form a feasible solution set for stocking density control.

6. The method for dynamic assessment of aquaculture capacity according to claim 5, characterized in that, The state evolution trajectory is dynamically calibrated and prediction bias corrected based on the latest real-time monitoring data stream to obtain the calibrated results, including: Obtain the latest monitoring data stream within a predetermined time window starting from the current moment, and transform it into the current actual observation state vector consistent with the dynamic causal fingerprint dimension; From the set of multiple future preset time windows of state evolution trajectory, select the predicted state point that is close to the time point of the current actual observed state vector, and calculate the state deviation vector between the current actual observed state vector and the predicted state point. Based on the state deviation vector, the key model parameters of the dynamic simulator that generates the state evolution trajectory are adjusted by reverse gradient to obtain the adjusted dynamic simulator. Using the adjusted dynamic simulator, a new set of state evolution trajectories is generated from the current actual observed state vector as the calibrated result.

7. The method for dynamic assessment of aquaculture capacity according to claim 6, characterized in that, The pre-set structural causal model is updated using the calibrated results to obtain a feasible solution set for stocking density regulation; Based on the feasible solution set for stocking density control, the recommended dynamic adjustment amounts for stocking capacity are screened and output, including: The calibrated result, i.e. the new state evolution trajectory, is used as a new training sample and input into the preset structural causal model to update the functional relationship parameters between the core causal feature variables in the preset structural causal model, thus obtaining the updated structural causal model. Using the updated structural causal model, an updated dynamic causal fingerprint is generated based on the latest current actual observation state vector; Using the updated dynamic causal fingerprint as the new initial condition, we obtain the updated feasible solution set for stocking density regulation. The updated feasible solution set for stocking density regulation is matched with the preset objective function to select the corresponding stocking density adjustment target value, and the stocking density adjustment target value is used as the recommended dynamic adjustment amount of stocking capacity.

8. A dynamic assessment system for aquaculture capacity, the system implementing the method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect real-time water quality sensor data and fish behavior and physiological status monitoring data of the closed recirculating aquaculture system, and integrate historical disease outbreak records. Based on a predefined knowledge base of environmental stress factor association rules, it performs time-series alignment and feature fusion of multi-source heterogeneous data to construct a feature matrix. The extraction module is used to extract core causal feature variables based on the feature matrix; The generation module is used to generate dynamic causal fingerprints based on core causal feature variables and a preset structural causal model by calculating the counterfactual distribution of the set of variables under preset intervention conditions. The processing module is used to generate multiple state evolution trajectories within a preset future time window based on dynamic causal fingerprints; and to obtain a feasible solution set for controlling breeding density based on the disease probability and mortality distribution predicted by the state evolution trajectories. The determination module is used to dynamically calibrate and correct prediction deviations of the state evolution trajectory based on the latest real-time monitoring data stream to obtain the calibrated results; the calibrated results are used to update the preset structural causal model to obtain a feasible solution set for aquaculture density regulation; Based on the feasible solution set for stocking density control, the recommended dynamic adjustment amount of stocking capacity is selected and output.

Images