Characterization monitoring system for saxitoxin risk formation chain

Abstract

The invention relates to the technical field of aquatic product quality safety, in particular to a saxitoxin risk formation chain characterization monitoring system, which comprises a data acquisition module, a data processing module, a data processing module and a display module, the risk chain association module is used for constructing a multi-factor toxin association mechanism model, analyzing and quantifying influence weights of environmental factors through a causal inference algorithm and a random forest weight, and identifying key regulation points and threshold values of toxin synthesis; the risk chain mechanism analysis module carries out deduction through a digital twin simulation platform and a dynamic Bayesian network, simulates the dynamic change of a toxin anabolism pathway under different environment scenes, and obtains a risk formation chain and a risk change rule thereof; and the risk monitoring and intervention module constructs a global risk prediction model to obtain a risk level, obtains an emergency scheme from the emergency decision library according to the risk level and the toxin type, and performs intervention according to the emergency scheme. Therefore, the problems of monitoring lag, unknown causality, low intervention efficiency and the like are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aquatic product quality and safety, and particularly relates to a characterization monitoring system for a risk formation chain of shellfish toxins. BACKGROUND

[0002] Shellfish is an important aquatic product, and its food safety is closely related to public health and the development of the aquaculture industry. Shellfish toxins are metabolized by toxic algae and can be enriched in shellfish through the food chain. The shellfish toxins have the characteristics of strong concealment and high toxicity, and are difficult to eliminate by conventional cooking methods, which can easily cause poisoning of consumers and pose a serious challenge to the supervision of the aquatic product market and the aquaculture industry.

[0003] Existing shellfish toxin monitoring mainly adopts a terminal detection mode, which focuses on the detection of toxins in terminal shellfish products. This mode is difficult to capture the process of the formation of shellfish toxins, and problems are often found after the shellfish toxins exceed the standard or even poisoning occurs, which has obvious lag. The formation of shellfish toxins is a complex process, which is influenced by multiple environmental factors and has significant temporal variability. Conventional statistical-based analysis is difficult to distinguish the causal relationship between environmental factors and toxins, cannot accurately quantify the contribution of each factor to different stages of risk formation, and is also difficult to identify the key control points that trigger the risk change, resulting in unclear risk traceability, weak pertinence of management measures, and limited intervention effect. SUMMARY

[0004] The present application provides a characterization monitoring system for a risk formation chain of shellfish toxins to solve the problems of lagging monitoring, unclear causality, and low intervention efficiency in the prior art.

[0005] The first aspect embodiment of the present application provides a shellfish toxin risk formation chain characterization monitoring system, comprising: a data acquisition module, a risk chain association module, a risk chain mechanism analysis module, a risk monitoring and intervention module, wherein the data acquisition module is configured to collect environmental data and obtain shellfish toxin detection results, wherein the environmental data includes hydrological data, biological data, and pollutant data, and the detection results include toxin types and toxin contents; the risk chain association module is configured to construct a multi-factor toxin association mechanism model, take key nodes as core links, and according to the environmental data and the detection results, quantize the influence weight of environmental factors on each stage of toxin formation through a causal inference algorithm and a random forest weight analysis, identify key regulation points and threshold values of toxin synthesis, wherein the key nodes include toxic algal population dynamics, environmental trigger factors, toxin synthesis metabolism, and shellfish accumulation; the risk chain mechanism analysis module is configured to deduce through a digital twin simulation platform and a dynamic Bayesian network according to the factor influence weight, the key regulation points and the threshold values quantized by the risk chain association module, simulate the dynamic changes of the toxin synthesis metabolic pathway under different environmental scenarios, and obtain a risk formation chain and its risk change rule; and the risk monitoring and intervention module is configured to construct a global risk prediction model, obtain a risk formation chain and its risk change rule according to the detection results, input the environmental data, the detection results, the risk formation chain and its risk change rule into the global risk prediction model to obtain a risk level, obtain an emergency scheme from an emergency decision library according to the risk level and the toxin type, and intervene according to the emergency scheme.

[0006] Optionally, the data acquisition module comprises an environmental data acquisition unit, an environmental data standardization processing unit, and a shellfish toxin detection result acquisition unit, wherein the environmental data acquisition unit is configured to obtain hydrological data through a buoy equipped with a multi-parameter water quality sensor, obtain biological data through a drone equipped with a hyperspectral imaging and eDNA sampling system, and obtain pollutant data through a mass spectrometer; the environmental data standardization processing unit is configured to filter out inconsistent data according to a predefined reasonable range, and then perform interpolation on missing values, and perform spatio-temporal alignment on the processed hydrological data, biological data, and pollutant data; and the shellfish toxin detection result acquisition unit is configured to obtain shellfish toxin detection results through a standardized interface, wherein the detection results include toxin types and toxin contents.

[0007] Optionally, the risk chain association module comprises a causal relationship mining unit, an influence weight quantification unit, and a regulation threshold identification unit. The causal relationship mining unit is configured to construct an initial causal graph based on shellfish toxin field knowledge and historical monitoring data, take environmental factors as cause variables, and take core parameters of each key node as result variables, mark direct association, indirect association, and confounding factors in the initial causal graph, control and remove confounding factors in the initial causal graph using a backdoor adjustment strategy of Do-Calculus algorithm, and output a causal relationship list. The influence weight quantification unit is configured to receive the causal relationship list, construct independent random forest regression models for four key nodes respectively, and perform interpretability analysis through SHAP to quantify the contribution of each environmental factor to each key node, and output a multi-dimensional factor influence weight matrix. The regulation threshold identification unit is configured to filter combinations of factors and key nodes with weights greater than a target threshold as key regulation points based on the multi-dimensional factor influence weight matrix, fit the relationship between factor changes and toxins in historical monitoring data using Logistic regression and piecewise linear fitting methods, and determine threshold parameters.

[0008] Optionally, the risk chain mechanism analysis module comprises a metabolic pathway model construction unit and an inference engine. The metabolic pathway model construction unit is configured to construct a virtual simulation scene based on a digital twin simulation platform, embed key nodes in the virtual simulation scene, assign each node with a factor influence weight, a key regulation point, and a threshold parameter, and construct different physical mapping models of toxin risk formation chains for different toxin types. The inference engine is configured to construct an initial Bayesian network with key nodes as network nodes and factor influence weights as connection strength between nodes, convert regulation threshold parameters into transition probabilities of node states, calculate risk state probabilities of each node of a toxin formation chain at each time node based on a dynamic Bayesian network inference algorithm by inputting different environmental scenario parameters, and output a risk probability change curve of node state change over time.

[0009] Optionally, the risk monitoring and intervention module comprises a risk level classification module, an emergency scheme acquisition module, and an intervention module, wherein the risk level classification module is configured to construct a global risk prediction model based on an attention mechanism and an LSTM network, obtain a risk formation chain and its risk change law of a corresponding toxin type according to a toxin type of the detection result, input the environmental data, the detection result, the risk formation chain, and its risk change law into the global risk prediction model to obtain a risk level, and the risk level is low risk, medium risk, high risk, or extremely high risk; the emergency scheme acquisition module is configured to acquire an emergency scheme from an emergency decision library according to the risk level and the toxin type, the emergency decision library is preconfigured with a two-dimensional response rule of the risk level and the toxin type, and the two-dimensional response rule comprises a responsible subject, a disposal time limit, and core measures; and the intervention module is configured to intervene according to the emergency scheme.

[0010] Optionally, the risk chain association module further comprises: after the threshold parameter is determined, the threshold parameter is dynamically verified according to a risk level output by the global risk prediction model, and when a risk level change trend deviates from the threshold parameter significantly, the threshold parameter is modified.

[0011] The second aspect embodiment of the present application provides a method for characterizing and monitoring a risk formation chain of a shellfish toxin, comprising the following steps: collecting environmental data and obtaining a detection result of a shellfish toxin, wherein the environmental data comprises hydrological data, biological data, and pollutant data, and the detection result comprises a toxin type and a toxin content; constructing a multi-factor toxin association mechanism model, taking a key node as a core link, quantifying an influence weight of an environmental factor on each stage of toxin formation according to the environmental data and the detection result through a causal inference algorithm and a random forest weight analysis, identifying a key regulation point and a threshold value of toxin synthesis, wherein the key node comprises a toxic algal population dynamics, an environmental trigger factor, a toxin synthesis metabolism, and a shellfish accumulation; inferring, through a digital twin simulation platform and a dynamic Bayesian network, a dynamic change of a toxin synthesis metabolism pathway under different environmental scenarios according to a factor influence weight, a key regulation point, and a threshold parameter quantified by the risk chain association module, obtaining a risk formation chain and its risk change law; constructing a global risk prediction model, obtaining a risk formation chain and its risk change law according to the detection result, inputting the environmental data, the detection result, the risk formation chain, and its risk change law into the global risk prediction model to obtain a risk level, and acquiring an emergency scheme from an emergency decision library according to the risk level and the toxin type; and intervening according to the emergency scheme.

[0012] The third aspect of the present application provides an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to perform a shellfish toxin risk formation chain characterization monitoring method as described in the above embodiments.

[0013] The fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement a shellfish toxin risk formation chain characterization monitoring method as described in the above embodiments.

[0014] The fifth aspect of the present application provides a computer program product, the computer program product storing a computer program, the program being executed by a processor to implement a shellfish toxin risk formation chain characterization monitoring method as described in the above embodiments.

[0015] The beneficial effects of the present application are as follows:

[0016] The embodiments of the present application build a multi-factor toxin correlation mechanism model, model the risk formation chain with four key nodes as core links, realize systematic correlation analysis of the whole process of shellfish toxins, combine the causal inference algorithm with the random forest weight analysis, not only identify the statistical correlation between environmental factors and toxin formation, but also strip the influence of confounding factors, establish the causal driving relationship, and quantify the contribution weight of each environmental factor in different stages of risk formation, providing a targeted basis for precise intervention. By identifying the key control points and threshold parameters, the key points of toxin risk jump can be locked, and quantitative judgment standards for risk state probability changes are provided. Through the digital twin simulation platform and the dynamic Bayesian network, a virtual scene that can simulate the dynamic changes of the toxin metabolic pathway under different environmental scenarios is constructed, which can reproduce and deduce the dynamic changes of the toxin synthesis and metabolism pathway under different environmental scenarios in the virtual environment, summarize the quantifiable risk formation chain and the risk state probability change rule, provide mechanism support and scenario priori for the global risk prediction model, make the risk level judgment more in line with the actual ecological process, and reduce the probability of false and missed judgments. The global risk prediction model integrates real-time environmental data, toxin detection results and the internal law of the risk formation chain, can output multi-dimensional and graded risk assessment results, and intelligently matches targeted emergency plans from the emergency decision library, improving the pertinence and efficiency of intervention. Thus, the problems of monitoring lag, unknown causality and low efficiency of intervention in the prior art are solved.

[0017] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description, taken in conjunction with the following drawings of exemplary embodiments of the present application, and where:

[0019] Figure 1 A structural schematic diagram of a shellfish toxin risk formation chain characterization monitoring system according to an embodiment of the present application;

[0020] Figure 2 A flowchart of a shellfish toxin risk formation chain characterization monitoring method according to an embodiment of the present application;

[0021] Figure 3 A structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] Embodiments of the present application are described in detail below with reference to the attached drawing figures, wherein the same or like component have the same or similar designations. The embodiments described below are exemplary and intended to be illustrative of the present application, and are not to be understood as limiting of the present application.

[0023] A shellfish toxin risk formation chain characterization monitoring system of an embodiment of the present application is described below with reference to the accompanying drawings. In view of the problems of monitoring lag, unclear causality, and inefficient intervention mentioned in the background art, the present application provides a shellfish toxin risk formation chain characterization monitoring system, in which an embodiment of the present application builds a multi-factor toxin correlation mechanism model, models the risk formation chain with four key nodes as core links, realizes systematic correlation analysis of the whole process of shellfish toxins, identifies the statistical correlation between environmental factors and toxin formation by combining causal inference algorithms with random forest weight analysis, further strips the influence of confounding factors, establishes causal driving relationships, and quantifies the contribution weights of each environmental factor at different stages of risk formation, thereby providing a targeted basis for precise intervention. By identifying key control points and threshold parameters, the key points of toxin risk jump can be locked, and quantitative judgment criteria for risk state probability changes can be provided. Through a digital twin simulation platform and a dynamic Bayesian network, a virtual scenario that can simulate the dynamic changes of toxin metabolic pathways under different environmental scenarios is constructed, which can reproduce and deduce the dynamic changes of toxin synthesis metabolic pathways under different environmental scenarios in a virtual environment, summarize quantifiable risk formation chains and risk state probability change rules, provide mechanism support and scenario priori for global risk prediction models, make risk level determination more consistent with actual ecological processes, and reduce the probability of false positives and false negatives. The global risk prediction model integrates real-time environmental data, toxin detection results, and the internal laws of the risk formation chain, can output multi-dimensional and graded risk assessment results, and intelligently matches targeted emergency plans from an emergency decision library, thereby improving the targeting and efficiency of intervention. Thus, the problems of monitoring lag, unclear causality, and inefficient intervention in the prior art are solved.

[0024] Specifically, Figure 1 A structural diagram of a shellfish toxin risk formation chain characterization monitoring system provided by an embodiment of the present application.

[0025] As Figure 1 shown, the shellfish toxin risk formation chain characterization monitoring system 10 includes a data acquisition module 100, a risk chain correlation module 200, a risk chain mechanism analysis module 300, and a risk monitoring and intervention module 400.

[0026] The data acquisition module 100 is configured to collect environmental data and obtain detection results of shellfish toxins, wherein the environmental data includes hydrological data, biological data, and pollutant data, and the detection results include toxin types and toxin contents; the risk chain association module 200 is configured to construct a multi-factor toxin association mechanism model, take key nodes as core links, and quantize the influence weight of environmental factors on each stage of toxin formation by a causal inference algorithm and a random forest weight analysis according to the environmental data and the detection results, and identify key regulation points and threshold values of toxin synthesis, wherein the key nodes include toxic algal population dynamics, environmental trigger factors, toxin synthesis metabolism, and shellfish accumulation; the risk chain mechanism analysis module 300 is configured to deduce, by a digital twin simulation platform and a dynamic Bayesian network, the dynamic changes of the toxin synthesis metabolic pathway under different environmental scenarios according to the factor influence weight, the key regulation points and the threshold value parameters quantized by the risk chain association module, obtain a risk formation chain and a risk change rule, and construct a global risk prediction model; the risk monitoring and intervention module 400 is configured to obtain a risk level by inputting the environmental data, the detection results, the risk formation chain and the risk change rule into the global risk prediction model, obtain an emergency scheme from an emergency decision library according to the risk level and the toxin type, and intervene according to the emergency scheme.

[0027] It can be understood that, by constructing a multi-factor toxin association mechanism model and taking four key nodes as core links to model a risk formation chain, the embodiment of the application realizes systematic correlation analysis of the whole process of shellfish toxins, identifies statistical correlations between environmental factors and toxin formation by combining a causal inference algorithm with a random forest weight analysis, further strips the influence of confounding factors, establishes a causal driving relationship, quantizes the contribution weight of each environmental factor at different stages of risk formation, and provides a targeted basis for precise intervention. By identifying key regulation points and threshold parameters, the key points of toxin risk jump can be locked, and quantitative determination standards for risk state probability changes are provided. By a digital twin simulation platform and a dynamic Bayesian network, a virtual scenario that can simulate the dynamic changes of the toxin metabolic pathway under different environmental scenarios is constructed, the dynamic changes of the toxin synthesis metabolic pathway under different environmental scenarios can be reproduced and deduced in a virtual environment, and quantifiable risk formation chains and risk state probability change rules are summarized, mechanism support and scenario priori are provided for a global risk prediction model, the risk level determination is more in line with the actual ecological process, and the probability of false negatives and false positives is reduced. The global risk prediction model fuses real-time environmental data, toxin detection results and the internal rules of the risk formation chain, can output multi-dimensional and graded risk assessment results, intelligently matches a targeted emergency scheme from an emergency decision library, and improves the pertinence and efficiency of intervention.

[0028] In the embodiment of the present application, the data acquisition module comprises: an environmental data acquisition unit, an environmental data standardization processing unit, and a shellfish toxin detection result acquisition unit.

[0029] The environmental data acquisition unit is configured to acquire hydrological data through buoys carrying multi-parameter water quality sensors, acquire biological data through unmanned aerial vehicles carrying hyperspectral imaging and eDNA sampling systems, and acquire pollutant data through a mass spectrometer; the environmental data standardization processing unit is configured to filter out data that does not conform to a predefined reasonable range according to the acquired hydrological data, biological data, and pollutant data, and then perform interpolation on missing values, and perform spatio-temporal alignment on the processed hydrological data, biological data, and pollutant data; and the shellfish toxin detection result acquisition unit is configured to acquire shellfish toxin detection results through a standardized interface, and the detection results include toxin types and toxin contents.

[0030] Specifically, the culture area is deployed in a grid manner, the culture areas in each grid are numbered, and a plurality of sampling points are arranged in each culture area, and the ecological characteristics of the sampling points are as different as possible.

[0031] The hydrological data is acquired through buoys carrying multi-parameter water quality sensors arranged in the culture area, one buoy is configured for each sampling point, and the buoy is bound to the corresponding culture area and sampling point, and the acquisition frequency is once every 10 minutes, wherein the hydrological data includes water temperature, salinity, pH value, dissolved oxygen concentration, turbidity, flow rate, and nutrient salt concentration. The collected data is generated in a structured record in the format of culture area number, sampling point number, and time.

[0032] The biological data is acquired through unmanned aerial vehicles carrying hyperspectral imaging and eDNA sampling systems arranged in the culture area, the unmanned aerial vehicles perform inspection and mobile sampling according to a preset flight route, the hyperspectral imaging is used to identify algal community structure and distribution, and the eDNA sampling system is used to capture genetic material of toxin-producing algae in water, and species identification and quantitative analysis are performed through real-time fluorescent quantitative PCR, wherein the biological data includes toxin-producing algal density and community structure, zooplankton abundance, and physiological state indicators. The collected data is generated in a structured record in the format of culture area number, sampling point number, and time.

[0033] The water quality of the sampling points is periodically collected through a sampling station, and the pollutant data is analyzed through a mass spectrometer, wherein the pollutant data includes heavy metal (cadmium, lead, and mercury) concentration, organic pollutant (polycyclic aromatic hydrocarbon and polychlorinated biphenyl) concentration, and new pollutant (microplastic and antibiotic) concentration. The collected data is generated in a structured record in the format of culture area number, sampling point number, and time.

[0034] According to the predefined reasonable range of each index, the abnormal values exceeding the range are removed by using the 3 sigma principle, and for the missing values caused by sampling omission or removal, the missing values are completed by using the space-time interpolation algorithm, the linear interpolation is used for filling the missing values in the time dimension, and the Kriging interpolation is used for filling the local area missing values in the space dimension, so as to ensure the continuity of the data.

[0035] The hydrological data (10 minutes / time), biological data (1 day / time) and pollutant data (1 week / time) are matched in time and space by taking the aquaculture area number, sampling point number and time as the identifier to obtain the environmental data.

[0036] The RESTful API standardized interface is connected with the detection system, the detection results of the shellfish toxin within 10 minutes of the sampling point are obtained by the aquaculture area number, sampling point number and time, the detection results include the toxin type and toxin content, the toxin type includes paralytic shellfish toxin STX, diarrhetic shellfish toxin OA and microcystin MCs, and the unit of the toxin content is μg / kg.

[0037] It can be understood that the embodiment of the application realizes the accurate acquisition of the hydrological data, biological data, pollutant data and toxin data by the mode of gridding partition, differential sampling point layout and multi-method collaborative collection, the integrity, consistency and continuity of the data are ensured by the standardized processing of the environmental data and the space-time alignment of the data, and the accurate association of the environmental data and the toxin detection results is ensured by the structured data record format and the standardized interface design, thereby providing high-quality and high-credibility data support for subsequent risk chain association analysis, mechanism analysis and accurate intervention.

[0038] In the embodiment of the application, the risk chain association module includes a causal relationship mining unit, an influence weight quantification unit and a regulation threshold identification unit.

[0039] The cause-effect mining unit is used to construct an initial cause-effect diagram based on shellfish toxin field knowledge and historical monitoring data, taking environmental factors as cause variables and core parameters of each key node as result variables, marking direct association, indirect association and confounding factors in the initial cause-effect diagram, controlling and removing confounding factors in the initial cause-effect diagram by using a backdoor adjustment strategy of Do-Calculus algorithm, and outputting a cause-effect relationship list; the influence weight quantification unit is used to receive the cause-effect relationship list, construct independent random forest regression models for the four key nodes respectively, and perform interpretability analysis by SHAP to quantify the contribution of each environmental factor to each key node, and output a multi-dimensional factor influence weight matrix; the regulation threshold identification unit is used to filter combinations of factors and key nodes with weights greater than a target threshold as key regulation points based on the multi-dimensional factor influence weight matrix, and use Logistic regression and piecewise linear fitting methods to fit the relationship between factor changes and toxins in historical monitoring data to determine threshold parameters.

[0040] Specifically, environmental factors are taken as cause variables, and the environmental factors include water temperature, salinity, pH value, dissolved oxygen concentration, turbidity, flow rate, and nutrient salt concentration. The key nodes are "toxic algae population dynamics, environmental trigger factors, toxin synthesis and metabolism, and shellfish accumulation", and preset core parameters are set for each key node, wherein the core parameters of toxic algae population dynamics are algae type, algae density and community structure, zooplankton abundance, and physiological state indicators, the core parameters of environmental trigger factors are temperature-nutrient salt synergy index, nutrient salt stoichiometric ratio, and water stability index, the core parameters of toxin synthesis and metabolism are algae type and toxin type, and the core parameters of shellfish accumulation are toxin type and toxin content. The core parameters of each key node are taken as result variables. An initial cause-effect diagram is constructed based on shellfish toxin field knowledge, the initial cause-effect diagram is a directed acyclic graph, the cause variables and the result variables are mapped to the nodes of the cause-effect diagram, the cause-effect logic of the field consensus is used to draw directed edges, the assumed cause-effect relationship is indicated by arrows, and direct association, indirect association and confounding factors are marked.

[0041] The backdoor adjustment strategy of Do-Calculus algorithm is to identify and block the backdoor path from the cause variable to the result variable to remove the interference of the confounding factor on the true cause-effect. The backdoor path is a non-causal path from the cause variable to the result variable. All paths connecting the cause variable and the result variable but not passing through the causal direction are found in the cause-effect diagram, and if these paths are opened by confounding factors, they are judged as backdoor paths. After identifying the effective backdoor path, the interference is removed by using the backdoor adjustment formula, and the removed cause-effect relationship is output as the cause-effect relationship list. The backdoor adjustment formula is:

[0042]

[0043] Where X is the causal variable, and x is the specific value of the causal variable; Y is the result variable, and y is the specific value of the result variable; Z is the adjustment set, which is a set of variables that can block all backdoor paths, and the variables do not contain descendants of the causal variable. For specific numerical values; For example, the term "cause variable" indicates an operation that forces the causal variable X to be fixed to the value x. The intervention distribution represents the probability that Y takes the value y when X=x is forcibly set, i.e., the causal effect; This represents the conditional probability that Y=y when X=x and Z=z; Indicates the adjustment variable is The probability of.

[0044] Independent random forest regression models were constructed for each of the four key nodes. Environmental factors with direct causal relationships to each node were selected from a causal relationship list as input features, and the core parameters of each node were used as the model output targets. Historical monitoring data were divided into training and testing sets in a 7:3 ratio. The random forest regression model was trained on the training set, and the model accuracy and overfitting risk were balanced by adjusting core parameters such as the number of decision trees and the maximum depth. The model performance was then validated using the testing set.

[0045] After model training is complete, the SHAP method is introduced for interpretability analysis to accurately quantify the contribution of environmental factors. Based on cooperative game theory, SHAP calculates the marginal contribution of each environmental factor to the model's prediction results. By decomposing the output of each prediction sample, it measures the impact of a single factor from its missing state to its presence on the predicted value. First, the trained random forest model is parsed using the SHAP interpreter to calculate the SHAP value of each environmental factor in all training samples. Then, the global contribution of each environmental factor to the node is obtained through the average absolute SHAP value; the larger the value, the more significant the impact of the environmental factor on the node's core parameters.

[0046] The environmental factor contributions of each node are standardized so that the sum of the contributions of all factors within a single node is 1. This is then used to generate a multi-dimensional factor influence weight matrix, with the dimensions of "key node, environmental factor, contribution".

[0047] Based on the multi-dimensional factor influence weight matrix, combinations of factors with weights greater than the target threshold and key nodes are selected as key control points.

[0048] Threshold parameters include the critical threshold for a single factor and the threshold for the rate of change of a factor.

[0049] Logistic regression is used to identify the abrupt inflection point in the probability of toxin exceeding the standard when the factor value crosses a certain critical point. Using the probability of exceeding the toxin standard as the dependent variable and the factor value at the key control point as the independent variable, it finds the inflection point where the probability rises rapidly by fitting the relationship between the binary risk state and the continuous factor value. The formula is as follows:

[0050]

[0051] in, P represents the factor value; Indicates the probability of exceeding the toxin limit; , is the regression coefficient.

[0052] The x-values ​​corresponding to P values ​​greater than the target threshold are used as the critical thresholds for a single factor.

[0053] Piecewise linear fitting is used to capture the continuous nonlinear relationship between factors and toxins, identifying points of slope change. Inflection points are roughly determined based on historical data, and segmentation points are set based on business experience. Linear regression is performed on each segment, and the slope and goodness of fit of each segment are calculated. By comparing the differences in slopes across different segments, the intensity of the factor's impact on toxin risk can be clearly identified; the points of slope change are the thresholds for the factor's rate of change.

[0054] It is understood that the embodiments of this application construct an initial causal graph based on domain knowledge and historical data through a causal relationship mining unit, and use a Do-Calculus backdoor adjustment strategy to remove confounding factors, outputting a reliable list of causal relationships to ensure the authenticity and scientific validity of the association between factors and risk nodes. An influence weight quantification unit establishes random forest regression models for the four key nodes respectively, and introduces SHAP interpretability analysis to accurately quantify the contribution of each environmental factor at different stages, forming a multi-dimensional factor influence weight matrix, making the weight results transparent and traceable. A control threshold identification unit selects combinations of high-weight factors and nodes as key control points, and combines Logistic regression and piecewise linear fitting to fit the relationship between factor changes and toxins from historical data to determine scientific and interpretable threshold parameters. This provides a scientifically quantified decision-making basis for subsequent risk prediction and targeted intervention.

[0055] In this embodiment, the risk chain mechanism analysis module includes: a metabolic pathway model construction unit and a deduction engine.

[0056] The metabolic pathway model construction unit is used to build virtual simulation scenarios based on a digital twin simulation platform. Key nodes are embedded into the virtual simulation scenario, and each node is assigned factor influence weights, key control points, and threshold parameters. Physical mapping models of different toxin risk formation chains are constructed for different toxin types. The inference engine is used to construct an initial Bayesian network with key nodes as network nodes and factor influence weights as the connection strength between nodes. The control threshold parameters are converted into the transition probabilities of node states. By inputting different environmental scenario parameters, the risk state probability of each node in the toxin formation chain at each time node is calculated based on the dynamic Bayesian network inference algorithm, and the risk probability change curve of the node state over time is output.

[0057] Specifically, geographical and ecological basic data of the target aquaculture area are collected. In the digital twin simulation platform, the physical space of the aquaculture area is restored at a 1:1 scale using a 3D modeling tool. A hydrological dynamic simulation module is configured to simulate the changes in environmental factors such as temperature and salinity. At the same time, a data interface is set up to support the real-time input of subsequent environmental scenario parameters, ensuring that the ecological characteristics of the virtual scene are consistent with those of the real aquaculture environment.

[0058] Key nodes are embedded into the virtual simulation scene. The "environmental trigger factor" node is bound to the hydrological monitoring module of the virtual scene, with its data input directly connected to the platform's real-time monitoring parameters such as water temperature and salinity. The "toxic algae population dynamics" node is bound to the biodistribution module, associating it with the algae growth area in the virtual waters. The "toxin synthesis and metabolism" node is embedded as a sub-module of the algae population into the algae physiological activity simulation unit. The "shellfish accumulation" node is bound to the cage aquaculture module, associating it with the filter feeding and metabolic processes of individual shellfish. Directional transmission links between nodes are built using the platform's visual programming tools. For each type of toxin, a unidirectional transmission rule of "environmental trigger factor → toxic algae population dynamics → toxin synthesis and metabolism → shellfish accumulation" is set. The corresponding multi-dimensional factor influence weight matrix is ​​imported, and the weight values ​​are assigned to the transmission links between nodes. At the same time, the trigger conditions of the links are configured. Signals can only be transmitted to downstream nodes when the current node's state meets the threshold requirements, ensuring that the interaction logic between nodes is consistent with the real ecological chain.

[0059] The inference engine constructs an initial Bayesian network, using four key nodes—environmental triggering factors, toxic algal population dynamics, toxin synthesis and metabolism, and shellfish accumulation—as the core nodes. The hierarchical relationships and action paths of these nodes are clearly defined, and a directed topology is built following the ecological logic of "environmental triggering → algal response → toxin synthesis → shellfish enrichment." A multi-dimensional factor influence weight matrix is ​​imported, converting the contribution values ​​of "factor-node" in the matrix into connection strength weights between nodes. Higher weight values ​​indicate a more significant impact of changes in the state of upstream nodes on downstream nodes. For each key node's threshold parameters, the frequency of node states within different threshold intervals is statistically analyzed using historical monitoring data, and this frequency is converted into a conditional probability table for the Bayesian network.

[0060] The initial Bayesian network is extended into a dynamic Bayesian network, with the inference period divided into time slices, and state transition rules for nodes within adjacent time slices are defined. Different environmental scenario parameters are input, and these parameters are mapped to the node state probability distribution of the initial time slice. Based on the forward inference algorithm of the dynamic Bayesian network, the risk state probability of each node within each time slice is calculated sequentially. The node risk probability data from each time slice are integrated, and a "time-risk probability" risk probability change curve is generated for each key node. Simultaneously, key feature points are extracted from the curve, and the curve is output through the visualization interface of the digital twin platform.

[0061] It is understood that the embodiments of this application utilize digital twin technology to achieve virtual replication and dynamic simulation of the shellfish toxin risk formation chain. A highly accurate physical mapping model of toxin formation is constructed using parameters from real-world aquaculture scenarios and quantified factor weights. Through a dynamic Bayesian network algorithm, threshold parameters are transformed into node state transition probabilities. This supports simulation of the temporal changes in risk probabilities at key nodes under different environmental disturbance scenarios, outputting intuitive risk probability change curves. This not only accurately traces the causes and transmission paths of risks but also identifies critical time windows for risk escalation in advance, providing a scientific and visualized decision-making basis for targeted early warning and precise intervention of shellfish toxin risks.

[0062] In this embodiment, the risk monitoring and intervention module includes: a risk level classification module, an emergency plan acquisition module, and an intervention module.

[0063] The risk level classification module is used to build a global risk prediction model based on the attention mechanism and LSTM network. It obtains the risk formation chain and risk change law of the corresponding toxin type according to the toxin type detected. The environmental data, detection results, risk formation chain and risk change law are input into the global risk prediction model to obtain the risk level, which is low risk, medium risk, high risk and very high risk. The emergency plan acquisition module is used to obtain emergency plans from the emergency decision database according to the risk level and toxin type. The emergency decision database is preset with two-dimensional response rules for risk level and toxin type, including the responsible party, the time limit for disposal and the core measures. The intervention module is used to intervene according to the emergency plan.

[0064] Specifically, the global risk prediction model includes a feature encoding layer, an LSTM temporal modeling layer, an attention mechanism layer, and a classification layer. The feature encoding layer unifies multi-dimensional data—environmental data, detection results, risk formation chains, and risk change patterns—into a unified dimension through a fully connected layer. The LSTM temporal modeling layer inputs the encoded feature sequence, arranged according to time steps, into an LSTM network containing 128 hidden units to capture the temporal dependencies of environmental data. The attention mechanism layer calculates the attention weights of features at each time step, weighted and fused to obtain a context vector, highlighting key factors and time points influencing risk. The classification layer inputs the context vector into the fully connected layer and a Softmax classifier, outputting a probability distribution of four risk levels. The level corresponding to the highest probability is taken as the final prediction result.

[0065] Based on the risk level and toxin type, emergency plans are retrieved from the emergency decision-making database. These plans are assigned a plan number, which provides specific emergency measures, including the responsible party, response timeframe, and core measures. The responsible party clearly identifies the responsible units and individuals at each level of response. The response timeframe sets the time nodes for tiered responses. The core measures involve developing targeted interventions for different scenarios.

[0066] It is understood that the embodiments of this application, through the constructed global risk prediction model, achieve in-depth fusion analysis of multi-dimensional environmental data, toxin detection results, and risk formation patterns, enabling the identification of four risk levels and addressing the pain points of traditional risk assessments that rely on experience and lack accuracy. Simultaneously, relying on a pre-set emergency decision-making database, it quickly matches standardized emergency plans containing responsible parties, response timelines, and core measures, avoiding the blindness and lag in emergency response. Through intervention modules, it implements tiered control measures, forming an integrated management system from risk warning to response execution, significantly improving the rapid response capability and precise control level of shellfish toxin risks.

[0067] In this embodiment of the application, the risk chain association module further includes: after determining the threshold parameter, dynamically verifying the threshold parameter according to the risk level output by the global risk prediction model, and correcting the threshold parameter when the trend of risk level change deviates significantly from the threshold parameter.

[0068] Specifically, a real-time verification data link is established to connect the initial threshold parameters output by the threshold identification unit and the output results of the global risk prediction model of the risk level classification module in real time. The system extracts environmental monitoring data, toxin detection results, and risk level change curves within the verification window period according to a preset cycle, and compares the consistency between the risk status predicted based on the initial threshold parameters and the actual risk level output by the global risk prediction model. If the predicted risk status does not match the actual risk level output by the model, it is determined that the risk level change trend has deviated significantly from the threshold parameters.

[0069] For cases deemed significantly deviating, an iterative parameter optimization process is initiated. First, the core cause of the deviation is identified, distinguishing between an unreasonable threshold setting for a single factor and the need to adjust the threshold combination resulting from the synergistic effect of multiple factors. If the deviation is due to an unreasonable threshold setting for a single factor, attention weight data from the global risk prediction model is introduced to extract the core environmental factor with the greatest impact on risk level. Combined with historical data within the validation window, Logistic regression and piecewise linear fitting methods are re-applied to calculate the optimal threshold range for that factor. If the deviation is due to synergistic effects of multiple factors, the threshold combination rules for multiple factors are adjusted based on the factor influence weight matrix. The corrected threshold parameters are then input again into the control threshold identification unit and simulated using the inference engine of a digital twin simulation platform. The consistency between the risk evolution curve under the corrected threshold and the model prediction results is observed. Once verification is successful, the threshold parameter library is officially updated, completing one closed-loop optimization.

[0070] It is understood that the embodiments of this application break through the limitations of static setting and long-term fixation of threshold parameters, and can continuously verify the matching degree between threshold parameters and real risk scenarios based on the real-time risk level change trend output by the global risk prediction model. When the two deviate significantly, the threshold parameters are iteratively optimized in a timely manner to ensure that the threshold always conforms to the dynamic change law of the aquaculture environment, effectively avoiding the threshold failure problem caused by fluctuations in environmental factors and adjustments in ecological structure.

[0071] The following is a detailed description of a characterization and monitoring system for the formation chain of shellfish toxin risks through specific embodiments:

[0072] Taking a certain aquaculture area as the target scenario, the aquaculture area has a total area of ​​15 square kilometers and is divided into 42 grid units of 600m×600m, numbered from FJ-01 to FJ-42. Three differentiated sampling points (intertidal shallow water area, subtidal aquaculture area, and open sea confluence area) are set up in each grid, numbered CY-01, CY-02, and CY-03.

[0073] Data acquisition module: Each sampling point's buoy collects hydrological data every 10 minutes. At 2 PM on July 18th, the real-time data for sampling point CY-02 in grid FJ-27 were: water temperature 28.5℃, salinity 23.2‰, pH 8.0, dissolved oxygen 5.8 mg / L, turbidity 18 NTU, flow velocity 0.45 m / s, and nutrient (phosphorus) 0.8 mg / L. A drone patrols daily at 10 AM along a preset route. Hyperspectral imaging identified *Aureobasidium* as the dominant algae at this sampling point. eDNA sampling combined with quantitative real-time PCR detection yielded a *Aureobasidium* density of 5.6 × 10⁻⁶. 4 The abundance of zooplankton was 1.2 × 10³ ind / m³. Water samples were collected every Friday at the sampling station. Mass spectrometry analysis showed that the levels of heavy metal lead, polycyclic aromatic hydrocarbons (PAHs), and microplastics at this sampling point were 0.003 mg / L, 0.015 mg / L, and 0.4 mg / L, all within the safe baseline range. Outliers were removed using the 3σ principle. For example, the water temperature of 35.1℃ (outside the reasonable range of -2 to 30℃) recorded by the buoy at sampling point CY-02 in grid FJ-28 on July 18 was removed. The time dimension was completed by linear interpolation to fill in the data for this period, and the spatial dimension was filled by Kriging interpolation to fill in the local data gaps in grid FJ-28. Finally, the data were aligned in time and space using the unified identifier "FJ-27, CY-02, 2025-07-18 14:00" to complete the spatiotemporal alignment of hydrological, biological, and pollutant data. By connecting to the shellfish toxin detection system via RESTful API, the test results for this sampling point at 14:00 on July 18 were obtained: the toxin type was diarrhetic shellfish toxin (OA), and the toxin content was 1.8 mg / kg, which exceeded the food safety limit of 1.6 mg / kg.

[0074] The risk chain association module, specifically the causal relationship mining unit, uses seven environmental factors, including water temperature and salinity, as causal variables and core parameters from four key nodes—environmental triggering factors, toxic algal population dynamics, toxin synthesis and metabolism, and shellfish accumulation—as outcome variables. Based on domain knowledge of OA toxin formation, an initial causal graph is constructed, labeling "water temperature → Acerola densitifera density" as a direct correlation, "nutrient concentration → Acerola densitifera density → toxin synthesis rate" as an indirect correlation, and "water turbidity" as a confounding factor. A backdoor adjustment strategy using the Do-Calculus algorithm is employed, setting water turbidity as the adjustment set Z. The intervention distribution is calculated using the backdoor adjustment formula, and after removing confounding factors, a causal relationship list is output, clearly identifying water temperature and nutrient concentration as the core driving factors influencing Acerola densitifera proliferation and OA synthesis. The influence weight quantification unit constructed random forest regression models for the four key nodes. Historical monitoring data from 2023-2024 were divided into a training set (1820 records) and a test set (780 records) in a 7:3 ratio. When training the "toxic algae population dynamics" node model, the number of decision trees was set to 120, with a maximum depth of 10, and the R² of the test set reached 0.82. The SHAP method was introduced for interpretability analysis. The calculated average absolute SHAP value of water temperature on the density of *Alternaria cerevisiae* was 0.71, significantly higher than that of nutrient concentration (0.45) and salinity (0.22). After standardization, a multi-dimensional factor influence weight matrix was generated, with a weight of 0.71 for "water temperature-toxic algae population dynamics" and 0.83 for "*Alternaria cerevisiae* density-toxin synthesis metabolism". The regulation threshold identification unit set a target weight of 0.5 and selected the above two combinations as key regulation points. A model was constructed using Logistic regression. When the probability of toxin exceeding the standard P > 80%, the corresponding water temperature is ≥ 28.2℃, and the critical water temperature threshold is determined to be 28.2℃. Through piecewise linear fitting, it was found that when the daily increase in water temperature is > 1.5℃, the fitting slope jumps from 0.4 to 3.1, and the threshold for the rate of change of water temperature is determined to be 1.5℃ / d.

[0075] Risk Chain Mechanism Analysis Module: The digital twin platform replicates the geographical topography, hydrodynamics, and aquaculture layout of the aquaculture area at a 1:1 scale. Four key nodes are embedded into corresponding virtual modules: "Environmental Triggering Factors" are bound to the hydrological monitoring module; "Toxic Algae Population Dynamics" to the biological distribution module; "Toxin Synthesis and Metabolism" to the algal physiological activity unit; and "Shellfish Accumulation" to the oyster farming cage unit. A customized physical mapping model is developed for OA toxins, importing factor influence weight matrices and threshold parameters. A transmission chain of "water temperature increase → acuminate algae proliferation → OA synthesis → oyster accumulation" is established, with the trigger condition configured as "when water temperature ≥ 28.2℃ and daily increase > 1.5℃, downstream nodes initiate risk transmission." The inference engine constructs an initial Bayesian network centered on the four key nodes, converting the weight of "water temperature - toxic algae population dynamics" (0.71) into the connection strength between nodes. Combining historical data, the probability of water temperature ≥ 28.2℃ being converted into a high-risk state of "environmental triggering factors" is 88%, generating a conditional probability table. The initial Bayesian network was extended into a dynamic Bayesian network, with a 72-hour extrapolation period set at 6-hour time slices. The environmental scenario parameters "starting from July 19th, affected by the subtropical high pressure, the water temperature rises to 29.1℃, with a daily increase of 1.8℃" were input. Based on the forward inference algorithm, the following calculations were made: after 18 hours, the probability of high risk of Alternaria acuminate density rises to 75%; after 36 hours, the probability of high risk of toxin synthesis and metabolism reaches 86%; and after 60 hours, the probability of OA content exceeding the standard in oysters exceeds 90%. The curves of risk probability changes over time at each node were output, clearly showing the complete transmission path of OA toxin risk from triggering factors to terminal outbreak.

[0076] The risk monitoring and intervention module: The global risk prediction model takes the environmental data from July 19th (water temperature 29.1℃, daily increase 1.8℃), OA detection results (1.8mg / kg), and the inferred risk change pattern as input. The feature encoding layer unifies the multi-dimensional data into a 256-dimensional feature space through a fully connected layer. The LSTM time series modeling layer captures the continuous trend of water temperature rise at the 10-minute level. The attention mechanism layer automatically strengthens the weight of the key feature "daily increase of water temperature 1.8℃". The classification layer outputs the risk level probability distribution through the Softmax function, where the high risk probability reaches 92%, and finally determines that the grid is at a high risk level. The emergency plan acquisition module matches the two-dimensional response rule of "high risk - OA toxin", retrieves the plan number OA-G-02 from the emergency decision database, clarifies that the responsible entity is the local marine and fisheries bureau to take the lead, with the cooperation of aquaculture enterprises, the disposal time limit is 8 hours, and the core measures are "physical temperature control + algae suppression + suspension of oyster sales". The intervention module was executed rapidly according to the plan: within 8 hours, the aquaculture enterprises were organized to start the circulating water cooling system to reduce the water temperature in the FJ-27 grid aquaculture area to 27.5℃; modified clay flocculant was added to reduce the density of acervulae to 1.2×10⁻⁶. 4Cells / L; A notice was issued to suspend the sale of the oysters from this area, and the harvested oyster products were sealed and tested. Monitoring on July 22 showed that the water temperature at the sampling point remained stable at 27.3℃, the OA toxin content dropped to 1.1 mg / kg, and the risk level returned to low risk.

[0077] In summary, this application's embodiments accurately captured key risk factors such as water temperature and algae density through gridded sampling, multi-technology collaborative monitoring, and standardized data processing. By using causal inference to eliminate confounding factors, and combining random forest and SHAP analysis, water temperature was identified as the core driving factor. Logistic regression and piecewise linear fitting were used to pinpoint the critical threshold of 28.2℃ and the rate of change threshold of 1.5℃ / d, overcoming the limitations of empirical threshold setting. Based on digital twins and dynamic Bayesian network inference, a 90% probability of OA exceeding the standard was predicted in advance within 72 hours, and an emergency plan was matched. Targeted interventions such as temperature control, algae suppression, and product sealing were completed within 8 hours. The toxin content was reduced from 1.8 mg / kg to 1.1 mg / kg in just 4 days, and the risk level returned to low risk. This significantly improved the accuracy, foresight, and efficiency of shellfish toxin risk management, effectively ensuring the quality and safety of aquatic products and the stable development of the aquaculture industry.

[0078] Next, referring to the accompanying drawings, a method for characterizing and monitoring the risk formation chain of shellfish toxins according to an embodiment of this application is described.

[0079] Specifically, Figure 2 This is a flowchart illustrating a method for characterizing and monitoring the risk formation chain of shellfish toxins provided in an embodiment of this application.

[0080] like Figure 2 As shown, the method for characterizing and monitoring the risk formation chain of shellfish toxins includes the following steps:

[0081] In step S101, environmental data is collected and the detection results of shellfish toxins are obtained. The environmental data includes hydrological data, biological data, and pollutant data, and the detection results include toxin type and toxin content.

[0082] Specifically, hydrological data is acquired via buoys equipped with multi-parameter water quality sensors, biological data is acquired via drones equipped with hyperspectral imaging and eDNA sampling systems, and pollutant data is acquired via mass spectrometry. The acquired hydrological, biological, and pollutant data are filtered out according to predefined reasonable ranges, and missing values ​​are imputed. The processed hydrological, biological, and pollutant data are then spatiotemporally aligned. Shellfish toxin detection results, including toxin type and concentration, are obtained through a standardized interface.

[0083] It is understood that the embodiments of this application ensure the consistency and comparability of data from different sources and of different types by using multiple means to collect data collaboratively, standardize data processing, and align multidimensional data in time and space. This provides high-quality and standardized data support for core steps such as risk chain correlation analysis and global risk prediction model training.

[0084] In step S102, a multi-factor toxin association mechanism model is constructed, with key nodes as the core links. Based on environmental data and detection results, the influence weight of environmental factors on each stage of toxin formation is quantified through causal inference algorithm and random forest weight analysis. Key regulatory points and thresholds for toxin synthesis are identified. Among them, key nodes include toxic algal population dynamics, environmental triggering factors, toxin synthesis metabolism, and shellfish accumulation.

[0085] Specifically, based on knowledge of shellfish toxins and historical monitoring data, an initial causal graph is constructed, using environmental factors as causal variables and core parameters of each key node as outcome variables. Direct, indirect, and confounding factors are marked in this initial causal graph. A backdoor adjustment strategy using the Do-Calculus algorithm is employed to control and remove confounding factors, outputting a list of causal relationships. Receiving this list, independent random forest regression models are constructed for each of the four key nodes, and interpretability analysis is performed using SHAP to quantify the contribution of each environmental factor to each key node, outputting a multi-dimensional factor influence weight matrix. Based on this multi-dimensional factor influence weight matrix, combinations of factors with weights greater than a target threshold and key nodes are selected as key control points. Logistic regression and piecewise linear fitting methods are used to fit the relationship between factor changes and toxins in historical monitoring data to determine the threshold parameters.

[0086] Understandably, this application's embodiments construct an analytical framework based on four key nodes as the core link. By employing the Do-Calculus algorithm to remove confounding factors, the true causal relationships between environmental factors and each stage of toxin formation are obtained. Through node-specific modeling and interpretability analysis, the contribution of each environmental factor to the toxin formation chain is accurately quantified, identifying core regulatory points. Regression fitting clarifies the threshold parameters for factor-triggered risks, providing a scientific and quantitative core basis for source identification, targeted monitoring, and precise intervention of shellfish toxin risks.

[0087] In step S103, based on the factor influence weights, key control points and threshold parameters quantified by the risk chain association module, the dynamic changes of toxin synthesis and metabolism pathways under different environmental scenarios are simulated through a digital twin simulation platform and a dynamic Bayesian network to obtain the risk formation chain and its risk change law.

[0088] Specifically, a virtual simulation scenario is constructed based on a digital twin simulation platform. Key nodes are embedded into the virtual simulation scenario, and each node is assigned factor influence weights, key control points, and threshold parameters. Different physical mapping models of toxin risk formation chains are constructed for different toxin types. Using key nodes as network nodes and factor influence weights as the connection strength between nodes, an initial Bayesian network is constructed. The control threshold parameters are transformed into the transition probabilities of node states. By inputting different environmental scenario parameters, the risk state probability of each node in the toxin formation chain at each time point is calculated based on a dynamic Bayesian network inference algorithm, and the risk probability change curve of the node state over time is output.

[0089] It is understood that the embodiments of this application, through the collaboration of a digital twin simulation platform and a dynamic Bayesian network, transform the quantified factor influence weights, key control points and threshold parameters into a visualized and deducible physical mapping model of toxin risk. This breaks through the limitations of traditional risk analysis, which can only perform static assessments and cannot reproduce dynamic evolution. Furthermore, by inputting different environmental scenario parameters, it can accurately simulate the temporal changes of risk probability at each key node, output intuitive risk probability change curves, and clearly present the complete transmission path from environmental factor disturbances to toxin accumulation.

[0090] In step S104, a global risk prediction model is constructed. The risk formation chain and its risk change pattern are obtained based on the detection results. Environmental data, detection results, risk formation chain and its risk change pattern are input into the global risk prediction model to obtain the risk level. Based on the risk level and toxin type, emergency plans are obtained from the emergency decision-making database, and intervention is carried out according to the emergency plan.

[0091] Specifically, a global risk prediction model is constructed based on an attention mechanism and an LSTM network. The model obtains the risk formation chain and risk change patterns corresponding to the toxin type detected. Environmental data, detection results, risk formation chains, and risk change patterns are input into the global risk prediction model to obtain risk levels: low, medium, high, and extremely high. Based on the risk level and toxin type, emergency response plans are retrieved from an emergency decision-making database. This database pre-defines two-dimensional response rules for risk level and toxin type, including the responsible party, response timeframe, and core measures. Intervention is then implemented according to the emergency response plan.

[0092] It is understood that the embodiments of this application, by constructing a global risk prediction model combining an attention mechanism and an LSTM network, achieve deep fusion analysis of environmental data, toxin detection results, and risk formation patterns. This enables four levels of risk assessment. Through a pre-set emergency decision-making database, standardized emergency plans, including responsible parties, response timelines, and core measures, can be quickly matched based on toxin type and risk level, avoiding blind and delayed emergency responses. Targeted control based on the emergency plan improves the accuracy of shellfish toxin risk warnings and the efficiency of emergency response.

[0093] The following section will elaborate on a method for characterizing and monitoring the risk formation chain of shellfish toxins:

[0094] A certain aquaculture area was divided into 36 grids of 500m × 500m. Each grid had three sampling points: nearshore, mid-water, and offshore. Hydrological data were collected every 10 minutes using buoys. At 3 PM on August 5th, the hydrological data at the mid-water sampling point in grid DY-18 were: water temperature 26.8℃, salinity 27.5‰, pH 8.2, dissolved oxygen 6.5mg / L, and nutrient (nitrogen) 1.5mg / L. Daily drone inspections identified the dominant algae at this sampling point as *Alexandrium tamarense* using hyperspectral imaging. eDNA detection combined with qPCR quantification yielded a density of 4.2 × 10⁻⁶. 4 The water quality samples were collected weekly at the sampling station. Mass spectrometry analysis revealed the following concentrations: mercury 0.001 mg / L, polycyclic aromatic hydrocarbons 0.012 mg / L, and microplastics 0.25 mg / L. Outliers were removed using the 3σ principle. Missing temporal data were filled using linear interpolation, and spatial data were filled using Kriging interpolation. The final spatiotemporal alignment was completed using the identifier "DY-18, Central Sampling Point, 2025-08-05 15:00". The sampling point's results at 15:00 on August 5th were obtained via a RESTful API connection to the shellfish toxin monitoring system: the toxin type was paralytic shellfish toxin (STX), with a concentration of 58 μg / kg, exceeding the safety limit of 50 μg / kg.

[0095] Using environmental factors such as water temperature and salinity as causal variables, and four key nodes—environmental triggering factors, toxic algal population dynamics, toxin synthesis and metabolism, and shellfish accumulation—as outcome variables, an initial causal graph was constructed based on STX domain knowledge. The graph labeled "salinity → Alexandrium tamarense density" as a direct correlation, "nutrient concentration → algal density → STX synthesis rate" as an indirect correlation, and "zooplankton abundance" as a confounding factor. A Do-Calculus algorithm backdoor adjustment strategy was employed, setting zooplankton abundance as the adjustment set. After removing confounding interference, a list of causal relationships was output, clearly identifying salinity and water temperature as the core driving factors. Random forest regression models were constructed for the four nodes. 2400 historical data points from 2023-2024 were divided into training and testing sets in a 7:3 ratio. When training the model for the "toxic algae population dynamics" node, 110 decision trees were set with a maximum depth of 9, achieving an R² of 0.80 on the testing set. SHAP analysis revealed that the average absolute SHAP value of salinity for this node was 0.68. After standardization, a weight matrix was generated, with a weight of 0.68 for "salinity-toxic algae population dynamics" and 0.79 for "algae density-toxin synthesis and metabolism." A target weight of 0.5 was set, and the above two groups were selected as key control points. A logistic regression model was constructed. When the exceedance probability P > 80%, the corresponding salinity was ≤ 26.8‰, which was determined to be the critical threshold. Piecewise linear fitting revealed that when salinity dropped sharply by > 5.0‰ over 24 hours, the slope jumped from 0.35 to 3.0, determining the rate of change threshold to be 5.0‰ / 24 hours.

[0096] Based on a digital twin platform, the geographical topography and hydrodynamics of the Daya Bay aquaculture area were replicated at a 1:1 scale. Four key nodes were embedded into corresponding virtual modules: "environmental triggering factors" were bound to the hydrological monitoring module, "toxic algae population dynamics" were bound to the biological distribution module, "toxin synthesis and metabolism" were embedded into the algal physiological unit, and "shellfish accumulation" was bound to the scallop aquaculture cage unit. A dedicated physical mapping model was constructed for STX, and weight matrices and threshold parameters were imported to set the transmission link of "sudden drop in salinity → Alexandrium proliferation → STX synthesis → scallop accumulation". The inference engine constructs an initial Bayesian network with four core nodes. The weight of "salinity-toxic algae population dynamics" (0.68) is converted into the connection strength between nodes. Combining historical data, salinity ≤26.8‰ is converted into a high-risk state transition probability of 86% for "environmental triggering factors," generating a conditional probability table. The network is then expanded into a dynamic Bayesian network, with a 72-hour inference period set at 6-hour time slices. The scenario parameter "salinity drops sharply by 5.5‰ to 22.0‰ in 24 hours due to rainfall on August 6" is input. Based on the forward inference algorithm, it is found that: after 24 hours, the high-risk probability of Alexandrium density rises to 78%; after 48 hours, the high-risk probability of toxin synthesis metabolism reaches 89%; and after 72 hours, the probability of scallop STX exceeding the standard exceeds 92%. The risk probability change curves of each node are output.

[0097] A global risk prediction model was constructed based on an attention mechanism and an LSTM network. Environmental data from August 6th, STX detection results, and risk probability change curves were input into the model. The feature encoding layer unified the multi-dimensional data into a 256-dimensional feature space. The LSTM layer captured the temporal trend of the sudden drop in salinity. The attention mechanism layer reinforced the key feature of a "5.5‰ drop in salinity over 24 hours." The classification layer output the probability distribution of risk levels, with an extremely high risk probability of 95%, classifying the grid as extremely high risk. The emergency response module matched the "extremely high risk - STX" two-dimensional rule, retrieving the plan number STX-E-01 from the decision database. The responsible entity was clearly identified as the local marine and fisheries bureau, with the disease control center providing support. The response timeframe was 4 hours, and the core measures were "comprehensive lockdown + emergency purification + harmless disposal of products." The intervention module executed the plan rapidly: within 4 hours, grid DY-18 was designated as a lockdown area, suspending scallop harvesting and distribution; a salinity buffer was added to raise the salinity to 27.0℃; and an algicidal agent was added to reduce the density of Alexandrium algae to 0.8 × 10⁻⁶. 4 cells / L; Harmless disposal was carried out on the harvested scallops. Monitoring on August 10 showed that the STX content at this sampling point dropped to 22 μg / kg, and the risk level decreased to low risk.

[0098] In summary, this application's embodiments utilize technologies such as buoys, drones, and mass spectrometers to achieve standardized collection of environmental data and toxin detection results. By employing the Do-Calculus algorithm and random forest model, it accurately identifies salinity as the core driving factor, critical threshold, and rate of change threshold. Through digital twins and dynamic Bayesian networks, it predicts 92% of exceedance risks 72 hours in advance. Based on the risk level and toxin type, it matches an "extremely high risk - STX" emergency plan, completing targeted measures such as containment, salinity stabilization, and algae suppression within 4 hours, reversing the risk situation. This significantly improves the accuracy, foresight, and efficiency of shellfish toxin risk management, effectively ensuring the quality and safety of aquaculture products and the stability of the industry.

[0099] The method for characterizing and monitoring the risk formation chain of shellfish toxins proposed in this application constructs a multi-factor toxin association mechanism model. Using four key nodes as the core links, it models the risk formation chain, achieving a systematic association analysis of the entire shellfish toxin process. By combining causal inference algorithms with random forest weight analysis, it not only identifies the statistical association between environmental factors and toxin formation but also removes the influence of confounding factors, establishing a causal driving relationship and quantifying the contribution weight of each environmental factor at different stages of risk formation, providing a targeted basis for precise intervention. By identifying key control points and threshold parameters, it can pinpoint the key points of toxin risk escalation, providing a quantitative judgment standard for changes in the probability of risk status. Through a digital twin simulation platform and dynamic Bayesian networks, a virtual scenario capable of simulating the dynamic changes of toxin metabolic pathways under different environmental scenarios is constructed. This virtual environment can reproduce and deduce the dynamic changes of toxin anabolic pathways under different environmental scenarios, summarizing quantifiable risk formation chains and the probability changes of risk status, providing mechanistic support and scenario priors for global risk prediction models, making risk level judgments more consistent with actual ecological processes and reducing the probability of misjudgment and omission. The global risk prediction model integrates real-time environmental data, toxin detection results, and the inherent patterns of risk formation chains. It can output multi-dimensional, tiered risk assessment results and intelligently match targeted emergency response plans from an emergency decision-making database, improving the targeting and efficiency of interventions. This solves the problems of lagging monitoring, unclear causality, and inefficient interventions inherent in existing technologies.

[0100] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0101] The memory 301, the processor 302, and the computer program stored on the memory 301 and capable of running on the processor 302.

[0102] When the processor 302 executes the program, it implements a method for characterizing and monitoring the risk formation chain of shellfish toxins provided in the above embodiments.

[0103] Furthermore, electronic devices also include:

[0104] Communication interface 303 is used for communication between memory 301 and processor 302.

[0105] The memory 301 is used to store computer programs that can run on the processor 302.

[0106] The memory 301 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0107] If the memory 301, processor 302, and communication interface 303 are implemented independently, then the communication interface 303, memory 301, and processor 302 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0108] Optionally, in a specific implementation, if the memory 301, processor 302, and communication interface 303 are integrated on a single chip, then the memory 301, processor 302, and communication interface 303 can communicate with each other through an internal interface.

[0109] Processor 302 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0110] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for characterizing and monitoring the risk formation chain of shellfish toxins.

[0111] This application also provides a computer program product that stores a computer program that, when executed by a processor, implements the above-described method for characterizing and monitoring the formation chain of shellfish toxin risks.

[0112] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0113] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0114] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0115] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0116] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A characterization and monitoring system for the formation chain of shellfish toxin risks, characterized in that, include: The module includes a data acquisition module, a risk chain correlation module, a risk chain mechanism analysis module, and a risk monitoring and intervention module. The data acquisition module is used to collect environmental data and obtain the detection results of shellfish toxins. The environmental data includes hydrological data, biological data, and pollutant data, and the detection results include toxin type and toxin content. The risk chain association module is used to construct a multi-factor toxin association mechanism model. With key nodes as the core link, the model quantifies the influence weight of environmental factors on each stage of toxin formation based on the environmental data and detection results through causal inference algorithms and random forest weight analysis, and identifies key regulatory points and thresholds for toxin synthesis. The key nodes include toxic algal population dynamics, environmental triggering factors, toxin synthesis metabolism, and shellfish accumulation. The risk chain mechanism analysis module is used to simulate the dynamic changes of toxin synthesis and metabolism pathways under different environmental scenarios by using a digital twin simulation platform and a dynamic Bayesian network to extrapolate the factor influence weights, key control points and threshold parameters quantified by the risk chain association module, thereby obtaining the risk formation chain and its risk change patterns. The risk monitoring and intervention module is used to construct a global risk prediction model, obtain the risk formation chain and its risk change pattern based on the detection results, input the environmental data, detection results, risk formation chain and its risk change pattern into the global risk prediction model to obtain the risk level, obtain emergency plans from the emergency decision-making database based on the risk level and toxin type, and intervene according to the emergency plan.

2. The characterization and monitoring system for the formation chain of shellfish toxins according to claim 1, characterized in that, The data acquisition module includes: an environmental data acquisition unit, an environmental data standardization processing unit, and a shellfish toxin detection result acquisition unit, wherein... The environmental data acquisition unit is used to acquire hydrological data through a buoy equipped with a multi-parameter water quality sensor, acquire biological data through a drone equipped with a hyperspectral imaging and eDNA sampling system, and acquire pollutant data through a mass spectrometer. The environmental data standardization processing unit is used to filter out non-compliant data from the acquired hydrological data, biological data, and pollutant data according to a predefined reasonable range, then imput missing values, and perform spatiotemporal alignment on the processed hydrological data, biological data, and pollutant data. The shellfish toxin detection result acquisition unit is used to acquire shellfish toxin detection results through a standardized interface. The detection results include toxin type and toxin content.

3. The characterization and monitoring system for the formation chain of shellfish toxins according to claim 1, characterized in that, The risk chain association module includes: a causal relationship mining unit, an influence weight quantification unit, and a control threshold identification unit, wherein... The causal relationship mining unit is used to construct an initial causal graph based on knowledge of shellfish toxins and historical monitoring data, taking environmental factors as causal variables and the core parameters of each key node as outcome variables. In the initial causal graph, direct correlation, indirect correlation and confounding factors are marked. The backdoor adjustment strategy of the Do-Calculus algorithm is used to control and remove confounding factors in the initial causal graph and output a list of causal relationships. The influence weight quantification unit is used to receive the causal relationship list, construct independent random forest regression models for the four key nodes, and perform interpretability analysis through SHAP to quantify the contribution of each environmental factor to each key node and output a multi-dimensional factor influence weight matrix. The control threshold identification unit is used to select combinations of factors with weights greater than the target threshold and key nodes as key control points based on the multi-dimensional factor influence weight matrix. Logistic regression and piecewise linear fitting methods are used to fit the relationship between factor changes and toxins in historical monitoring data to determine the threshold parameters.

4. The characterization and monitoring system for the formation chain of shellfish toxins according to claim 1, characterized in that, The risk chain mechanism analysis module includes: a metabolic pathway model construction unit and a deduction engine, wherein... The metabolic pathway model building unit is used to build a virtual simulation scene based on the digital twin simulation platform, embed key nodes into the virtual simulation scene, and assign factor influence weights, key regulatory points and threshold parameters to each node, and build physical mapping models of different toxin risk formation chains for different toxin types. The inference engine is used to construct an initial Bayesian network with key nodes as network nodes and factor influence weights as the connection strength between nodes. It converts the control threshold parameters into the transition probability of node states. By inputting different environmental scenario parameters, it calculates the risk state probability of each node in the toxin formation chain at each time node based on the dynamic Bayesian network inference algorithm, and outputs the risk probability change curve of node state over time.

5. The characterization and monitoring system for the formation chain of shellfish toxins according to claim 1, characterized in that, The risk monitoring and intervention module includes: a risk level classification module, an emergency response plan acquisition module, and an intervention module, wherein... The risk level classification module is used to construct a global risk prediction model based on the attention mechanism and LSTM network. It obtains the risk formation chain and risk change law of the corresponding toxin type according to the toxin type of the detection results. The environmental data, detection results, risk formation chain and risk change law are input into the global risk prediction model to obtain the risk level, which is low risk, medium risk, high risk and extremely high risk. The emergency response plan acquisition module is used to acquire emergency response plans from the emergency decision database based on the risk level and toxin type. The emergency decision database is pre-set with two-dimensional response rules for risk level and toxin type, including the responsible party, the time limit for handling, and the core measures. The intervention module is used to intervene according to the emergency plan.

6. The characterization and monitoring system for the formation chain of shellfish toxins according to claim 3, characterized in that, The risk chain association module further includes: after determining the threshold parameter, dynamically verifying the threshold parameter based on the risk level output by the global risk prediction model, and correcting the threshold parameter when the trend of risk level change deviates significantly from the threshold parameter.

7. A method for characterizing and monitoring a shellfish toxin risk formation chain according to any one of claims 1-6, characterized in that, Includes the following steps: Collect environmental data and obtain detection results for shellfish toxins, wherein the environmental data includes hydrological data, biological data, and pollutant data, and the detection results include toxin type and toxin content; A multi-factor toxin association mechanism model was constructed, with key nodes as the core links. Based on the environmental data and detection results, the influence weight of environmental factors on each stage of toxin formation was quantified through causal inference algorithm and random forest weight analysis. Key regulatory points and thresholds of toxin synthesis were identified. The key nodes include toxic algal population dynamics, environmental triggering factors, toxin synthesis metabolism, and shellfish accumulation. Based on the factor influence weights, key control points and threshold parameters quantified by the risk chain association module, the dynamic changes of toxin synthesis and metabolism pathways under different environmental scenarios are simulated through a digital twin simulation platform and a dynamic Bayesian network to obtain the risk formation chain and its risk change law. A global risk prediction model is constructed. Based on the detection results, the risk formation chain and its risk change pattern are obtained. The environmental data, detection results, risk formation chain and its risk change pattern are input into the global risk prediction model to obtain the risk level. Based on the risk level and toxin type, an emergency plan is obtained from the emergency decision database, and intervention is carried out according to the emergency plan.

8. An electronic device, characterized in that, include: The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement a method for characterizing and monitoring a shellfish toxin risk formation chain as described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement a method for characterizing and monitoring the risk formation chain of shellfish toxins as described in claim 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the characterization and monitoring method for the formation chain of shellfish toxins as described in claim 7.

Images