Aquaculture product quality safety whole-process management method based on big data

Abstract

This invention discloses a method for the whole-process management of aquaculture product quality and safety based on big data, belonging to the field of big data technology. The method includes: acquiring input data, environmental parameter sequences, and product inspection results corresponding to multiple aquaculture batches throughout the entire aquaculture chain; performing feature binning based on density distribution to construct quality label vectors and risk factor association matrices; determining the risk transmission network based on the factor coupling paths in the risk factor association matrix, and using node centrality indicators to screen key control nodes and their control factor sets; using a perturbation propagation simulation strategy to evolve and deduce the quality label vectors to obtain predicted risk label sequences; generating quality offsets and performing risk level classification based on the deviation distribution between the predicted risk label sequences and product inspection results; mapping the risk level classification results and key control nodes to a management decision knowledge base to output a whole-process quality and safety control plan.

Description

Technical Field

[0001] This invention relates to the field of big data technology, specifically to a method for the whole-process management of the quality and safety of aquaculture products based on big data. Background Technology

[0002] In the field of aquaculture product quality and safety management, with the expansion of aquaculture scale and the extension of the industrial chain, the factors affecting product quality and safety are becoming increasingly complex. Existing full-process management methods mostly rely on isolated monitoring of single links, and the collected input data and environmental parameters cannot form effective correlation analysis across batches. Under this isolated monitoring model, the large amount of accumulated aquaculture data is only used for simple threshold comparisons and post-event statistics, and the deep coupling relationships and risk transmission patterns between data are not explored. Existing technical solutions have obvious defects. On the one hand, traditional methods lack a quantitative description of the coupling relationship of risk factors across aquaculture batches, making it difficult to identify the key aquaculture nodes and regulatory elements that play a key role in the spread of quality anomalies. Because each aquaculture batch is analyzed as an independent entity, the risk association between batches due to shared input sources or similar environmental patterns is ignored, resulting in the inability to construct a network structure that reflects the dynamic diffusion path of risks. Quality management measures can only take a broad, passive approach, unable to focus on bottlenecks or key links in the network. On the other hand, when performing quality risk prediction, existing technologies mainly rely on statistical regression or comparative analysis of historical data, and cannot simulate the dynamic spread of risks in the associated network. When the risk factors at a certain node change, the cascading effects of this change on related nodes and the subsequent evolution trend of quality labels are difficult to predict. This leaves quality management at the post-event response stage, lacking the ability to predict the direction and extent of potential risk spread. Therefore, a method is needed that can construct a risk factor correlation network from aquaculture data, identify key regulatory nodes, and simulate risk propagation on this network to predict quality evolution, addressing the problems of unclear risk coupling paths and lack of dynamic evolution trends in aquaculture quality and safety management. Summary of the Invention

[0003] This invention provides a big data-based method for the whole-process management of the quality and safety of aquaculture products. It aims to construct a transmission network reflecting the coupling relationship of risk factors between aquaculture batches from the data of the entire aquaculture chain, identify key control nodes with high bidirectional propagation intensity, and perform perturbation propagation evolution inference on this network to generate a predicted risk label sequence with a deviation distribution from the product inspection results, thereby completing the risk level classification and control plan output for aquaculture batches.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for the whole-process management of aquaculture product quality and safety based on big data, comprising: acquiring input data, environmental parameter sequences, and product inspection results corresponding to multiple aquaculture batches in the entire aquaculture chain, as a multidimensional initial dataset for quality and safety; performing feature binning based on density distribution on the multidimensional initial dataset for quality and safety, constructing a quality label vector and risk factor association matrix for each aquaculture batch; determining the risk transmission network for each aquaculture batch according to the factor coupling path in the risk factor association matrix, and using the node centrality index in the risk transmission network to determine the risk transmission network for each aquaculture batch. Key control nodes and their corresponding control factor sets are selected from each breeding batch. Based on the key control nodes and their corresponding control factor sets, a perturbation propagation simulation strategy is used to evolve and deduce the quality label vector to obtain the predicted risk label sequence for each breeding batch. According to the deviation distribution between the predicted risk label sequence and the product inspection results, the quality offset of each breeding batch is generated, and the risk level is classified according to the quality offset. The results of the risk level classification and the key control nodes are mapped to a preset management decision knowledge base to output the quality and safety full-process control plan for each breeding batch. This method employs density-aware binning encoding of multidimensional data across the entire chain, effectively extracting the correlation structure between quality labels and risk factors while preserving data distribution characteristics. This ensures that subsequent analysis is highly aligned with the actual data distribution. By utilizing directed risk transmission networks and bidirectional propagation intensity to identify core control nodes, it can accurately pinpoint links that play a crucial role in overall quality and safety, enhancing the targeted allocation of control resources. Through multi-scenario disturbance propagation simulation to deduce risk evolution paths and generate predicted risk labels, it significantly enhances the predictability and timeliness of response to hidden quality risks. By generating quality offsets based on the statistical characteristics of prediction deviations for risk level classification, it makes risk assessment results more objective and robust. Finally, by jointly mapping the level classification and key control nodes to a knowledge base, it automatically generates a comprehensive control plan that considers both risk levels and node linkages, achieving intelligent closed-loop management from data perception, risk positioning, evolution prediction to decision output.

[0005] As a technical solution of the present invention, the process of performing density distribution-based feature binning on the initial multidimensional dataset of quality and safety to construct the quality label vector and risk factor association matrix for each breeding batch specifically includes: calculating the numerical distribution interval for each dimension of data and dividing it into multiple density intervals according to the equal-density binning strategy, with each density interval containing the same number of data samples, thereby overcoming the problem of information loss in sparse data areas in traditional uniform binning; performing normalization mapping on the data samples of each density interval to obtain the binning coding sequence of that dimension of data, so that features of different dimensions and distributions are comparable under the same coding system; concatenating the binning coding sequences of all dimensions of the same breeding batch in dimensional order to generate an initial feature coding vector; extracting quality label sub-vectors from the initial feature coding vector according to a preset quality label lookup table, and constructing risk factor sub-vectors from the remaining dimension binning coding sequences to achieve decoupled representation of quality status and influencing factors; combining the quality label sub-vectors to obtain the quality label vector, and calculating the multidimensional spatial distance between the risk factor sub-vectors of any two breeding batches to construct the risk factor association matrix. This density distribution-based coding and distance calculation method effectively preserves the inherent distribution patterns of data in each dimension, making the correlation strength between risk factors more accurately reflect the true similarity and differences between batches.

[0006] Preferably, when calculating the multidimensional spatial distance, the absolute value of the difference in the same dimension of the risk factor sub-vectors of the two breeding batches is extracted to obtain multiple dimension differences, and the multiple dimension differences are weighted and summed. The weight corresponding to each dimension difference is positively correlated with the degree of data variation in that dimension, so that risk factors with high volatility occupy a more important position in the distance measurement and the sensitivity of the correlation matrix to key volatility factors is improved.

[0007] In the above technical solution, the process of determining the risk transmission network and screening key control nodes includes: using each breeding batch as a network node and elements in the risk factor association matrix that are greater than a preset association threshold as network edges to construct an initial risk transmission network; for each network edge, comparing the number of factor dimensions with consistent trends and the number of factor dimensions with opposite trends in the risk factor sub-vectors of the two connected breeding batches, when the consistent trend dimension is dominant, setting the direction of the edge from the source breeding batch to the target breeding batch, otherwise setting it in the opposite direction, thereby generating a directed risk transmission network. This direction determination method more reasonably describes the risk transmission logic from the perspective of factor change synergy; calculating the in-degree centrality and out-degree centrality of each network node in the directed risk transmission network, and using their product as the bidirectional propagation strength of the node; sorting the bidirectional propagation strength from largest to smallest, selecting a number of network nodes in the first preset proportion as key control nodes, and constructing the corresponding control factor set by all risk factors in the risk factor sub-vector of each key control node. By comprehensively measuring the dual role of nodes in risk convergence and diffusion using the product of in-degree and out-degree, the key control nodes selected have a stronger impact on the overall stability of the risk network. Implementing control measures based on this can achieve twice the result with half the effort.

[0008] Furthermore, based on the key control nodes and their corresponding control factor sets, the process of evolving and deducing the quality label vector using a perturbation propagation simulation strategy is as follows: For each control factor in the control factor set corresponding to each key control node, positive and negative perturbations of preset amplitudes are applied respectively to generate multiple perturbation simulation scenarios to fully cover the bidirectional impact that factor fluctuations may bring; for each perturbation simulation scenario, with the key control node as the propagation starting node, the reachable network nodes are determined step by step along the directed risk transmission network, and the perturbation amplitude is multiplied by the propagation attenuation coefficient of each directed edge according to the propagation level to calculate the change in risk factors for each other network node; the change in risk factors is superimposed on the original risk factor sub-vector to obtain the updated risk factor sub-vector, and then mapped to the simulated quality label vector through the risk mapping function; the simulated quality label vectors corresponding to multiple perturbation simulation scenarios are arranged in order to obtain the predicted risk label sequence corresponding to each aquaculture batch. This perturbation propagation simulation method fully considers the cascading amplification or attenuation effect of factor changes in the network, and can generate risk evolution maps with temporal depth and multiple possibilities, thereby improving the richness and foresight of prediction information.

[0009] In a preferred embodiment of the present invention, the process of generating a quality offset based on the deviation distribution between the predicted risk label sequence and the product inspection results is as follows: Each simulated quality label vector is compared with the product inspection results of the same breeding batch to obtain multiple predicted deviation values; the mean and standard deviation of these multiple predicted deviation values ​​are calculated as the quality offset for the breeding batch; a preset set of risk level classification intervals is obtained, and the quality offset is matched with these intervals to determine the risk level to which the breeding batch belongs. Using statistical features composed of the mean and standard deviation to characterize the offset reflects both the overall level of prediction error and the uncertainty information of prediction fluctuations, making the risk level classification results more statistically reliable.

[0010] When generating a comprehensive quality and safety control plan, based on the node identifiers of key control nodes in the directed risk transmission network, multiple candidate control measures associated with each node identifier are retrieved from a pre-set management decision knowledge base. This knowledge base pre-stores the mapping relationship between node identifiers and control measures. Target control measures matching the risk level classification results are selected from these candidate measures. For each key control node, the corresponding target control measures are prioritized according to the ratio of the in-degree centrality to the out-degree centrality of that key control node, generating a node control sequence. Then, the node control sequences of each key control node are concatenated and combined according to the bidirectional propagation strength from largest to smallest to generate the comprehensive quality and safety control plan. Using the in-degree to out-degree ratio to prioritize control measures allows for determining the order of implementation based on whether a node primarily acts as a risk receiver or risk transmitter in the network. This ensures that the internal logic of the control plan aligns with the network flow characteristics, strengthening systematic intervention in the risk transmission chain.

[0011] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0012] By calculating the multidimensional spatial distance between risk factor sub-vectors of any two farming batches, a risk factor correlation matrix is ​​constructed using the distance value as the basis for quantifying the coupling strength. Based on this, using farming batches as nodes and coupling relationships exceeding a preset correlation threshold as initial edges, factor dimensions with consistent and opposite trends in the source and target farming batches are further extracted. The directional attribute of the network edges is determined by comparing the number of the two types of dimension sets, generating a directed risk transmission network. In this network, the product of in-degree centrality and out-degree centrality is used as the bidirectional propagation strength, and several top-ranked network nodes are selected as key control nodes. This method organizes scattered cross-batch risk correlations into a directional network structure and uses bidirectional propagation strength to locate nodes with high influence in both risk reception and risk output directions. The selected key control nodes are at the hub of risk propagation; intervention at these nodes can simultaneously block the spread of upstream risks and cut off the source of downstream risks, allowing quality management resources to shrink from full coverage to precise control of a few core nodes in the network. After identifying key control nodes and their corresponding control factor sets, positive and negative preset amplitude perturbations are applied to each key control node to generate multiple perturbation simulation scenarios. For each perturbation simulation scenario, starting with the key control node, reachable network nodes are determined step-by-step along the direction of the directed edges. The perturbation amplitude is multiplied by the attenuation coefficient of the corresponding directed edge according to the propagation level, and the changes in risk factors of other network nodes at each level are accumulated. After superimposing the changes onto the original risk factor sub-vectors, they are converted into simulated quality label vectors through a mapping function, and the simulated quality label vectors under multiple scenarios are arranged in order to form a predicted risk label sequence. This process performs multi-scenario perturbation injection and step-by-step attenuation calculation on the directed risk propagation network, which can deduce the evolution trajectory of the quality label of the entire network over time or propagation depth after a local node is perturbed. Based on the deviation distribution between the predicted risk label sequence obtained by this deduction and the product inspection results, a quality offset reflecting the degree of prediction offset can be generated, so that the risk level classification is not only based on historical facts, but also integrates the multivariate simulation results after the risk dynamically spreads in the network. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a flowchart of a big data-based method for the whole-process management of aquaculture product quality and safety.

[0015] Figure 2 This is a flowchart of the construction process for the correlation matrix between quality label vectors and risk factors;

[0016] Figure 3 This is a flowchart of the risk transmission network construction and key control node selection process;

[0017] Figure 4 This is a flowchart of the process for generating predicted risk label sequences based on perturbation propagation simulation;

[0018] Figure 5 This is a flowchart for classifying the risk level of aquaculture batches based on quality offset;

[0019] Figure 6 This is a flowchart for generating a comprehensive quality and safety control plan. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figure 1 This invention provides a big data-based method for the whole-process management of aquaculture product quality and safety, comprising: acquiring input data, environmental parameter sequences, and product inspection results corresponding to multiple aquaculture batches in the entire aquaculture chain, as a multi-dimensional initial dataset for quality and safety; performing feature binning based on density distribution on the multi-dimensional initial dataset for quality and safety, constructing quality label vectors and risk factor association matrices for each aquaculture batch; determining the risk transmission network for each aquaculture batch based on the factor coupling paths in the risk factor association matrix, and using the node centrality index in the risk transmission network to screen out key control nodes and their corresponding control factor sets from each aquaculture batch; based on the key control nodes and their corresponding control factor sets, using a perturbation propagation simulation strategy to evolve and deduce the quality label vectors, obtaining the predicted risk label sequence for each aquaculture batch; generating the quality offset for each aquaculture batch based on the deviation distribution between the predicted risk label sequence and the product inspection results, and classifying the risk level of each aquaculture batch according to the quality offset; mapping the results of the risk level classification and the key control nodes to a preset management decision knowledge base, and outputting a whole-process quality and safety control plan for each aquaculture batch.

[0022] Example 1:

[0023] In specific implementation, please refer to Figure 2This study acquires input data, environmental parameter sequences, and product inspection results corresponding to multiple aquaculture batches across the entire aquaculture chain, serving as the initial multidimensional dataset for quality and safety. Each dimension of the initial multidimensional dataset corresponds to one input indicator, one environmental parameter indicator, or one product inspection indicator. Multiple aquaculture batches constitute the sample set in the initial multidimensional dataset. Density distribution-based feature binning is performed on the initial multidimensional dataset to construct the quality label vector and risk factor correlation matrix for each aquaculture batch.

[0024] For each dimension of the initial multidimensional dataset for quality and safety, the numerical distribution interval corresponding to that dimension is calculated. The calculation method is as follows: obtain all values ​​for that dimension across all breeding batches, arrange all values ​​in ascending order, use the minimum value after sorting as the lower bound of the numerical distribution interval, and use the maximum value after sorting as the upper bound. The numerical distribution interval is a continuous interval of real numbers from the lower bound to the upper bound. The numerical distribution interval is divided into multiple density intervals using an equal-density binning strategy, with each density interval containing the same number of data samples. The partitioning method is as follows: set the number of bins, and evenly divide the numerical distribution interval into multiple density interval boundary values ​​according to the number of bins. Each density interval boundary value corresponds to a specific quantile of all sorted values, ensuring that the number of data samples falling into each density interval is equal. When the total number of data samples is not divisible by the number of bins, the difference in the number of data samples falling into each density interval does not exceed 1.

[0025] For each density interval, data samples are normalized and mapped to obtain the binning code sequence corresponding to that dimension. The normalization mapping process involves assigning a unique interval identifier to each density interval, mapping data samples belonging to the same density interval to the same interval identifier, and ensuring that interval identifiers for different density intervals are distinct. The binning code sequences of all dimensions corresponding to the same breeding batch are concatenated in dimensional order to generate the initial feature encoding vector for that batch. The dimensional order follows the original arrangement of the dimensions in the initial multidimensional dataset for quality and safety.

[0026] Based on a pre-defined quality label lookup table, the quality label sub-vector corresponding to the breeding batch is extracted from the initial feature encoding vector, and the binning codes of the remaining dimensions are used to construct the risk factor sub-vector corresponding to the breeding batch. The pre-defined quality label lookup table records the index sets for the quality label dimension and the index sets for the risk factor dimension. The index set for the quality label dimension points to multiple dimensions related to product inspection results in the initial multidimensional dataset for quality and safety, while the index set for the risk factor dimension points to multiple dimensions related to input data and environmental parameter sequences in the initial multidimensional dataset for quality and safety. During extraction, based on the index set for the quality label dimension in the quality label lookup table, the binning codes at the corresponding index positions are extracted from the initial feature encoding vector and arranged in the original index order to form the quality label sub-vector; based on the index set for the risk factor dimension, the binning codes at the remaining index positions are extracted from the initial feature encoding vector and arranged in the original index order to form the risk factor sub-vector.

[0027] The quality label sub-vectors corresponding to each breeding batch are combined in batch order to obtain the quality label vector for each breeding batch. The method of combining in batch order is as follows: using the breeding batch number as the row identifier, the quality label sub-vectors of each breeding batch are stacked as row vectors to form the quality label vector.

[0028] Calculate the multidimensional spatial distance between risk factor sub-vectors of any two farming batches, and construct the risk factor correlation matrix for each farming batch based on the multidimensional spatial distance. Let any two farming batches be labeled as the first farming batch and the second farming batch, respectively. Extract the first risk factor sub-vector corresponding to the first farming batch and the second risk factor sub-vector corresponding to the second farming batch. Assume that both the first and second risk factor sub-vectors contain... One dimension, This is equal to the number of elements in the index set of the risk factor dimension. The first risk factor sub-vector is denoted as... ,in, For the first batch of aquaculture, in risk factor number [number missing] The binning code values ​​of each dimension, the second risk factor sub-vector is denoted as ,in, For the second batch of aquaculture, under risk factor number [number missing] The binning codes for each dimension are calculated. The absolute value of the difference between each element in the first risk factor sub-vector and each element in the same dimension of the second risk factor sub-vector is calculated to obtain the multi-dimensional differences. Dimensional differences of each dimension Calculated using the following formula: By weighted summing of the differences across multiple dimensions, the multidimensional spatial distance between the first and second farming batches is obtained. The calculation formula is:

[0029]

[0030] in, This represents the multidimensional spatial distance between the first and second breeding batches. This represents the total number of dimensions of the risk factor sub-vectors; The dimension index representing the risk factor sub-vector. The value range is from 1 to Integers; Indicates the risk factor number The weight coefficients corresponding to each dimension; Indicates the risk factor number The dimensional difference of each dimension.

[0031] Weighting coefficient The value of this is positively correlated with the degree of data variability in that dimension. The degree of data variability is represented by the ... The coefficient of variation of each dimension in the bin-coded values ​​across all farming batches is used to characterize the risk factor. The risk factor is obtained by calculating the ratio of the standard deviation to the mean of the bin-coded value sequence for each dimension across all farming batches. coefficient of variation in each dimension The coefficients of variation for all risk factor dimensions are summed and normalized to obtain the weighting coefficients. The calculation method is as follows:

[0032]

[0033] in, Indicates the risk factor number The coefficient of variation in each dimension Indicates the risk factor number The coefficient of variation in each dimension This represents the sum of the coefficients of variation across all risk factor dimensions.

[0034] Based on the multidimensional spatial distance between any two farming batches, a distance matrix with a dimension equal to the total number of farming batches is constructed as the risk factor correlation matrix. Let the total number of farming batches be... The distance matrix is ​​a The square formation, the first in the square formation Line number The element value of the column is the first The breeding batch and the first Multidimensional spatial distance between individual breeding batches and The value range is from 1 to Integers.

[0035] Example 2:

[0036] In specific implementation, please refer to Figure 3 Based on the factor coupling paths in the risk factor association matrix, the risk transmission network for each breeding batch is determined, and the node centrality index in the risk transmission network is used to screen out the key regulatory nodes and their corresponding regulatory factor sets from each breeding batch.

[0037] Each breeding batch is used as a network node, and matrix elements in the risk factor correlation matrix that are greater than a preset correlation threshold are used as network edges to construct an initial risk transmission network for each breeding batch. The preset correlation threshold is set as follows: all off-diagonal elements in the risk factor correlation matrix are extracted, arranged in ascending order of value, and the element value at the 75th percentile after arrangement is taken as the preset correlation threshold. This ensures that approximately 25% of the off-diagonal elements in the risk factor correlation matrix meet the condition of being greater than the preset correlation threshold, thereby controlling the edge density of the initial risk transmission network. For any matrix element in the risk factor correlation matrix, if the value of the matrix element is greater than the preset correlation threshold, an undirected network edge is established between the breeding batch corresponding to the row index and the breeding batch corresponding to the column index of the matrix element; otherwise, no network edge is established.

[0038] The method for generating a directed risk transmission network is as follows: the two farming batches connected by any network edge in the initial risk transmission network are denoted as the source farming batch and the target farming batch, respectively. Factor dimensions with the same trend of change as the risk factor sub-vector of the source farming batch are extracted from the risk factor sub-vector of the target farming batch to obtain a set of consistent factor dimensions. Factor dimensions with the opposite trend of change as the risk factor sub-vector of the source farming batch and the risk factor sub-vector of the target farming batch are extracted from the risk factor sub-vector of the source farming batch to obtain a set of opposite factor dimensions.

[0039] The method for determining whether the trends are consistent or opposite is as follows: For each factor dimension in the risk factor sub-vector, obtain the median of the binning code values ​​for that factor dimension across all farming batches, and use this median as the global reference median for that factor dimension. For the source and target farming batches, obtain the source binning code values ​​and target binning code values ​​for that factor dimension, respectively. If both the source and target binning code values ​​are greater than the global reference median for that factor dimension, or both are less than the global reference median for that factor dimension, then the trends of the source and target farming batches are consistent in that factor dimension. If the source binning code value is greater than the global reference median for that factor dimension while the target binning code value is less than the global reference median for that factor dimension, or vice versa, then the trends of the source and target farming batches are opposite in that factor dimension.

[0040] The network edge is configured by comparing the number of dimensions in the consistent factor dimension set with the number of dimensions in the opposite factor dimension set. If the number of dimensions in the consistent factor dimension set is greater than that in the opposite factor dimension set, the direction attribute of the network edge is set to point from the source breeding batch to the target breeding batch; otherwise, the direction attribute is set to point from the target breeding batch to the source breeding batch. If the number of dimensions in the consistent factor dimension set is equal to that in the opposite factor dimension set, the network edge is configured as a bidirectional edge, possessing both the direction attribute from the source breeding batch to the target breeding batch and the direction attribute from the target breeding batch to the source breeding batch. After assigning direction attributes to all network edges, a directed risk propagation network is obtained.

[0041] Calculate the in-degree centrality and out-degree centrality of each node in the directed risk propagation network, and use the product of the in-degree and out-degree centrality as the bidirectional propagation strength of the node. The in-degree centrality is the total number of directed edges pointing to that node in the directed risk propagation network, and the out-degree centrality is the total number of directed edges pointing from that node in the directed risk propagation network. For any given node... bidirectional propagation intensity The calculation formula is:

[0042]

[0043] in, Represents network nodes bidirectional propagation intensity, Represents network nodes in-degree centrality Represents network nodes Out-degree and centrality. The value of is a non-negative integer, and it is obtained by statistically analyzing all network nodes in the directed risk transmission network. The number of directed edges ending at a destination is used to obtain the value; The value of is a non-negative integer, and it is obtained by statistically analyzing all network nodes in the directed risk transmission network. The number of directed edges originating from a given point is obtained.

[0044] All network nodes in the directed risk transmission network are sorted in descending order of bidirectional propagation strength. A predetermined percentage of these nodes are selected as key control nodes. The predetermined percentage is set to 20%, based on the fact that in a directed risk transmission network, the number of nodes with high bidirectional propagation strength typically does not exceed 20% of the total number of network nodes. Selecting the top 20% of nodes captures the set of nodes that play a major role in risk transmission. The top 20% of network nodes in the sorted list constitute the set of key control nodes.

[0045] For each key control node, all risk factors in the risk factor subvector are used to form the control factor set corresponding to that key control node. For each key control node in the set, the identifiers of all risk factor dimensions contained in the risk factor subvector corresponding to that key control node are obtained, and the identifiers of all risk factor dimensions are aggregated into a set, which is the control factor set corresponding to that key control node.

[0046] Example 3:

[0047] In specific implementation, please refer to Figure 4 Based on key regulatory nodes and their corresponding regulatory factor sets, a perturbation propagation simulation strategy is used to evolve and deduce the quality label vector, thereby obtaining the predicted risk label sequence for each aquaculture batch.

[0048] For each key control node, positive and negative perturbations of preset amplitudes are applied to each control factor in the control factor set corresponding to the key control node, generating multiple perturbation simulation scenarios for the key control node. The preset amplitude is set to 10% of the range of binned code values ​​for each control factor in its dimension. The rationale for setting the preset amplitude to 10% of the range of binned code values ​​for each control factor in its dimension is that 10% of the range of binned code values ​​can generate observable local fluctuations without disrupting the overall distribution structure of the risk factor subvectors. This avoids both excessively large perturbation amplitudes that distort the simulation results and excessively small perturbation amplitudes that cause the perturbation effect to be overwhelmed by noise. For a control factor set containing... The key regulatory nodes of each regulatory factor are generated by the combination of positive and negative perturbations. Each disturbance simulation scenario corresponds to a risk factor sub-vector at a key control node, in which only one control factor dimension is subjected to a positive or negative disturbance while the binning encoding values ​​of the other risk factor dimensions remain unchanged.

[0049] For each disturbance simulation scenario, the key control node corresponding to the disturbance simulation scenario is used as the propagation starting node. In the directed risk propagation network, all reachable network nodes that can be reached from the propagation starting node via multiple directed edges are determined. The determination method is as follows: starting from the propagation starting node, a breadth-first traversal is performed along the directions of the directed edges in the directed risk propagation network. All network nodes visited during the traversal are recorded as reachable network nodes. Traversal stops when a network node has no pointed-together directed edges or when all pointed-together network nodes have been visited.

[0050] Based on the direction of the directed edges, each path from the propagation starting node to any reachable network node is divided into multiple propagation levels, with each level corresponding to the order in which a directed edge is traversed. The division method is as follows: in a path from the propagation starting node to a reachable network node, the network node directly reached from the propagation starting node via a directed edge is marked as the first propagation level; the network node reached from the first propagation level via another directed edge is marked as the second propagation level, and so on. The path is divided into multiple levels based on the direction of the directed edges. The network node reached by the directed edge is labeled as the i-th Transmission levels A positive integer starting from 1, representing the order number of the directed edge in the path.

[0051] For all reachable network nodes in the first propagation level, the disturbance amplitude under the disturbance simulation scenario is directly multiplied by the propagation attenuation coefficient of the directed edge corresponding to the first propagation level to obtain the change in risk factor for each reachable network node in the first propagation level. The propagation attenuation coefficient represents the degree of intensity attenuation when the disturbance propagates along the directed edge. Each directed edge has an independent propagation attenuation coefficient, and the value of the propagation attenuation coefficient is a real number greater than 0 and less than 1. The propagation attenuation coefficient is determined by calculating the similarity of the risk factor sub-vectors between the two aquaculture batches connected by the directed edge. The similarity is calculated using the Pearson correlation coefficient. If the Pearson correlation coefficient is positive, the propagation attenuation coefficient is equal to the absolute value of the Pearson correlation coefficient; if the Pearson correlation coefficient is negative, the propagation attenuation coefficient is 0.1; if the Pearson correlation coefficient cannot be calculated, the propagation attenuation coefficient is 0.05.

[0052] For each propagation level greater than the first propagation level, the change in risk factor of reachable network nodes along the same propagation path in the previous propagation level is multiplied by the propagation attenuation coefficient of the directed edge corresponding to the current propagation level to obtain the change in risk factor of each reachable network node in the current propagation level. Assuming that under the same disturbance simulation scenario, along the same propagation path, the... The propagation level can reach network nodes. The change in risk factors is From reachable network nodes To the The propagation level can reach network nodes. The propagation attenuation coefficient of the directed edge is Then the network nodes can be reached. Changes in risk factors The calculation formula is:

[0053]

[0054] in, Indicates reachable network nodes The change in risk factors, Indicates reachable network nodes The change in risk factors, Indicates from reachable network nodes To reach network nodes The propagation attenuation coefficient of the directed edge, The value of is a real number greater than 0 and less than 1, and the reachable network nodes can be calculated. With reachable network nodes The Pearson correlation coefficient between the risk factor subvectors was determined.

[0055] The risk factor changes of each other network node are superimposed onto the original risk factor subvectors of those other network nodes to obtain the updated risk factor subvectors of the other network nodes under the perturbation simulation scenario. For any reachable network node in the directed risk propagation network other than the propagation initiation node, if the reachable network node is located on multiple propagation paths originating from the propagation initiation node, the risk factor changes from different propagation paths are superimposed onto the binning codes of the corresponding dimensions in the original risk factor subvectors of the reachable network node, and the superposition result for each dimension is the maximum value of all paths in that dimension. If the binning code value of a certain dimension after superposition exceeds the range of binning code values ​​for that dimension, the values ​​exceeding the range are truncated to the nearest boundary value.

[0056] According to a pre-defined risk mapping function, the updated risk factor sub-vectors of all network nodes in each perturbation simulation scenario are mapped to the corresponding simulated quality label vectors for that scenario. The pre-defined risk mapping function is an attention-based multilayer perceptron model. The input layer of this model receives an updated risk factor sub-vector, and the output layer outputs the predicted value of the corresponding quality label vector. The attention-based multilayer perceptron model consists of one input layer, one self-attention layer, two fully connected hidden layers, and one output layer. The number of neurons in the input layer is equal to the total dimension of the risk factor sub-vectors. The input vector is the updated risk factor sub-vector. The self-attention layer calculates the attention weights between each dimension of the input vector, multiplying these weights by the input vector to obtain a weighted representation vector. The output dimension of the self-attention layer is the same as the input dimension. The first fully connected hidden layer has 64 neurons, using a modified linear unit function (MRU) for activation; the second fully connected hidden layer has 32 neurons, also using an MRU. The output layer has the same number of neurons as the total dimensions of the quality label sub-vector, and uses a linear function for activation. The training process of the attention-based multilayer perceptron model is as follows: Risk factor sub-vectors from all historical aquaculture batches are used as training input samples, and the corresponding quality label sub-vectors from each batch are used as training output labels. Parameter updates are performed using a mean squared error loss function and an adaptive moment estimation optimizer. The training iterations are set to 500, the batch size to 16, and the initial learning rate to 0.001.

[0057] The simulated quality label vectors corresponding to multiple disturbance simulation scenarios are arranged in chronological order to obtain the predicted risk label sequence for each aquaculture batch. The chronological order is determined by the generation order of each disturbance simulation scenario, which is jointly determined by the order of the bidirectional propagation intensity of the key control nodes from largest to smallest and the preset arrangement order of the positive and negative disturbances applied by the control factors within each key control node. For the simulated quality label vectors generated by all disturbance simulation scenarios, the correspondence between the simulated quality label vectors and the disturbance simulation scenarios is maintained, and the simulated quality label vectors are stacked row by row according to the generation order to form the predicted risk label sequence.

[0058] Example 4:

[0059] In specific implementation, please refer to Figure 5 Based on the deviation distribution between the predicted risk label sequence and the product inspection results, the quality offset of each breeding batch is generated, and the risk level of each breeding batch is classified according to the quality offset.

[0060] The difference calculation is performed between each simulated quality label vector in the predicted risk label sequence and the product inspection result corresponding to the same breeding batch, yielding multiple predicted deviation values ​​for that breeding batch. Each simulated quality label vector in the predicted risk label sequence corresponds to a quality label prediction result under a perturbation simulation scenario, and the dimension of each simulated quality label vector is the same as the dimension of the product inspection result. The difference calculation method is as follows: For the same breeding batch, all simulated quality label vectors corresponding to that breeding batch are extracted from the predicted risk label sequence. For each simulated quality label vector, the predicted binning code value in each dimension of the simulated quality label vector is subtracted dimension by dimension from the actual binning code value in the corresponding dimension of the product inspection result, and the absolute value is taken to obtain the deviation component in each dimension. The sum of the deviation components in all dimensions is calculated as the predicted deviation value corresponding to that simulated quality label vector. Let the breeding batch... Total A simulated quality label vector, then the breeding batch The corresponding prediction deviation value sequence is denoted as ,in This equals the number of all perturbation simulation scenarios. Take the total number of breeding batches from 1.

[0061] Calculate the statistical characteristic values ​​of multiple prediction deviations and use these values ​​as the mass offset for the breeding batch. The statistical characteristic values ​​include the mean and standard deviation of the multiple prediction deviation values. The mean represents the average level of prediction deviation for the breeding batch across all perturbation simulation scenarios, and the standard deviation represents the fluctuation range of prediction deviation for the breeding batch across all perturbation simulation scenarios. Mass Offset The calculation formula is:

[0062]

[0063] in, Indicates the batch size of the breeding operation The mass offset; Indicates the batch size of the breeding operation The mean of multiple prediction bias values, by analyzing... The arithmetic mean is obtained; Indicates the batch size of the breeding operation The standard deviation of multiple prediction bias values, by analyzing the... The unbiased standard deviation is obtained by calculation; This represents the fluctuation penalty coefficient. The value is set to 1.5. The basis for setting it to 1.5 is that, in the statistical analysis of the deviation distribution, adding the fluctuation amplitude to the mean in the form of 1.5 times the standard deviation can make the quality offset cover about 93% of the one-sided confidence upper bound in a statistical sense. This makes the quality offset reasonably sensitive to deviation fluctuations under the condition that multiple predicted deviation values ​​follow an approximately normal distribution, while avoiding the risk level classification being too conservative due to excessive fluctuation penalties.

[0064] Obtain a preset set of risk level classification intervals. This set contains multiple consecutive numerical intervals, each corresponding to a risk level. The preset set of risk level classification intervals includes four risk levels: low risk, medium risk, high risk, and very high risk. The numerical interval corresponding to the low risk level is... The numerical range corresponding to the medium-risk level is The numerical range corresponding to the high-risk level is: The numerical range corresponding to the extremely high risk level is: Interval boundary values , , The determination method is as follows: Obtain samples of mass offsets from all historical breeding batches, arrange the mass offset samples in ascending order, and take the sample value located at the 25th percentile after sorting as... The sample value at the 50th percentile after sorting is taken as... The sample value at the 75th percentile after sorting is taken as... .

[0065] The quality offset of each breeding batch is matched with the numerical range in the risk level classification interval set to determine the risk level to which the breeding batch belongs. The matching method is as follows: for each breeding batch... mass offset Starting from the numerical range corresponding to the low-risk level, the judgment is made sequentially upwards. Less than Then the breeding batch The risk level is determined to be low risk; if Greater than or equal to and less than Then the breeding batch The risk level is determined to be medium risk; if Greater than or equal to and less than Then the breeding batch The risk level is determined to be high risk; if Greater than or equal to Then the breeding batch The risk level was determined to be extremely high. After all breeding batches completed the risk level matching, the risk level classification results for each breeding batch were obtained.

[0066] Example 5:

[0067] In specific implementation, please refer to Figure 6 The results of risk level classification and key control nodes are mapped to a preset management decision knowledge base, and a full-process quality and safety control plan for each breeding batch is output.

[0068] Based on the node identifiers of key control nodes in the directed risk transmission network, multiple candidate control measures associated with the node identifiers are retrieved from a pre-set management decision knowledge base. This pre-stored knowledge base contains a mapping relationship between node identifiers and control measures. The node identifier uses a unique batch number from the aquaculture batch. The pre-set management decision knowledge base is implemented using a relational database table structure, with each table containing a node identifier field and a control measure identifier field. There is a many-to-many mapping relationship between the node identifier and control measure identifier fields; one node identifier corresponds to multiple control measure records. Each control measure record includes a control measure identifier, control measure name, control measure parameter type, and a description of the control measure's applicable scope.

[0069] Retrieve the level identifier corresponding to each risk level in the risk level classification results, and the set of applicable level identifiers corresponding to each candidate control measure in the preset management decision knowledge base. The level identifiers are enumerated, with values ​​including low-risk, medium-risk, high-risk, and extremely high-risk level identifiers. Each candidate control measure is associated with a set of applicable level identifiers in the preset management decision knowledge base, which records several risk level identifiers that can be applied to the candidate control measure.

[0070] The matching level set of the candidate control measures is obtained by intersecting the set of applicable level identifiers of each candidate control measure with the level identifiers in the risk level classification results. For a given candidate control measure, let the set of applicable level identifiers of the candidate control measure be denoted as . The set consisting of all the level identifiers involved in the risk level classification results is Then match the set of levels The calculation formula is:

[0071]

[0072] in, This represents the set of applicable level identifiers for candidate regulatory measures. The elements in the text are one or more of the following: low-risk level identifier, medium-risk level identifier, high-risk level identifier, or extremely high-risk level identifier; This represents the set of risk level identifiers corresponding to the risk levels of all aquaculture batches in the risk level classification results. The elements in the text are one or more of the following: low-risk level identifier, medium-risk level identifier, high-risk level identifier, or extremely high-risk level identifier; This represents the set of matching levels for candidate regulatory measures. The elements in the middle belong to and .

[0073] The number of levels in the matching level set is counted, and the candidate control measures are sorted in descending order of the number of levels to obtain a candidate control measure sequence. For candidate control measures with an empty matching level set, the number of levels is zero, and they are placed at the end of the candidate control measure sequence.

[0074] The candidate control measures ranked at the top of the preset number in the candidate control measure sequence are identified as target control measures. The preset number is set to 5. The reason for setting the preset number to 5 is that, in management decision-making scenarios, the number of effective control measures that can be executed for a single breeding batch is constrained by human, material, and time resources. Setting an upper limit of 5 measures can achieve a balance between measure coverage and execution feasibility, ensuring multi-dimensional intervention in key risk factors while avoiding excessive measures that lead to fragmented execution.

[0075] For each key control node, the target control measures corresponding to the key control node are prioritized according to the ratio of the in-degree centrality to the out-degree centrality of the key control node, generating a node control sequence for the key control node. The ratio of in-degree centrality to out-degree centrality reflects the relative intensity of risk information received and emitted by the key control node in the directed risk transmission network. A larger ratio indicates that the key control node is more likely to be a convergence point of risk information, while a smaller ratio indicates that the key control node is more likely to be a diffusion point of risk information. The execution priority is ranked as follows: for target control measures with an in-degree centrality to out-degree centrality ratio greater than 1, execution priorities are assigned from high to low in descending order of the ratio; for target control measures with an in-degree centrality to out-degree centrality ratio less than or equal to 1, execution priorities are assigned from low to high in ascending order of the ratio.

[0076] Following the order of bidirectional propagation intensity from highest to lowest, the node control sequences corresponding to each key control node are concatenated and combined to generate a comprehensive quality and safety control plan. The concatenation and combination method is as follows: key control nodes are arranged sequentially according to their bidirectional propagation intensity from highest to lowest. For the first key control node in the arrangement, its corresponding node control sequence is placed at the beginning of the comprehensive quality and safety control plan. For subsequent key control nodes, their corresponding node control sequences are appended sequentially to the end of the plan, until all key control nodes have had their node control sequences appended. The comprehensive quality and safety control plan includes the batch number of the key control nodes, the risk level classification results, a list of target control measures, and the execution priority ranking of each target control measure.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for the whole-process management of aquaculture product quality and safety based on big data, characterized in that: include: We will acquire input data, environmental parameter sequences, and product inspection results for multiple aquaculture batches throughout the entire aquaculture chain, as a multidimensional initial dataset for quality and safety. Perform density distribution-based feature binning on the aforementioned quality and safety multidimensional initial dataset to construct quality label vectors and risk factor association matrices for each aquaculture batch; Based on the factor coupling paths in the risk factor association matrix, the risk transmission network of each breeding batch is determined, and the node centrality index in the risk transmission network is used to screen out key control nodes and their corresponding control factor sets from each breeding batch. Based on the key control nodes and their corresponding control factor sets, the quality label vector is evolved and deduced using a perturbation propagation simulation strategy to obtain the predicted risk label sequence for each aquaculture batch. Based on the deviation distribution between the predicted risk label sequence and the product inspection results, the quality offset of each breeding batch is generated, and the risk level of each breeding batch is classified according to the quality offset. The results of the risk level classification and the key control nodes are mapped to a preset management decision knowledge base, and the quality and safety full-process control plan for each breeding batch is output.

2. The method for full-process management of aquaculture product quality and safety based on big data according to claim 1, characterized in that, The initial multidimensional dataset for quality and safety is subjected to density distribution-based feature binning to construct the quality label vector and risk factor association matrix for each aquaculture batch, including: For each dimension of the initial multidimensional dataset for quality and safety, the numerical distribution interval corresponding to that dimension is calculated, and the numerical distribution interval is divided into multiple density intervals according to the equal density binning strategy, with each density interval containing the same number of data samples. Normalize the data samples within each density interval to obtain the binning coding sequence corresponding to that dimension of data; The binned coding sequences of all dimensions of data corresponding to the same breeding batch are concatenated in dimensional order to generate the initial feature coding vector corresponding to the breeding batch. According to the preset quality label lookup table, the quality label sub-vector corresponding to the breeding batch is extracted from the initial feature encoding vector, and the binning encoding sequence of the remaining dimensions is used to form the risk factor sub-vector corresponding to the breeding batch. The quality label subvectors corresponding to each breeding batch are combined in batch order to obtain the quality label vector of each breeding batch. Calculate the multidimensional spatial distance between risk factor subvectors of any two breeding batches, and construct the risk factor correlation matrix for each breeding batch based on the multidimensional spatial distance.

3. The method for full-process management of aquaculture product quality and safety based on big data according to claim 2, characterized in that, Based on the factor coupling paths in the risk factor association matrix, the risk transmission network for each aquaculture batch is determined. Then, using the node centrality index in the risk transmission network, key regulatory nodes and their corresponding regulatory factor sets are selected from each aquaculture batch, including: Using each of the breeding batches as a network node and the matrix elements in the risk factor association matrix that are greater than a preset association threshold as network edges, an initial risk transmission network corresponding to each breeding batch is constructed. For each network edge in the initial risk transmission network, a directional attribute is assigned to the network edge according to the coupling direction between the risk factor sub-vectors of the two breeding batches connected by the network edge, thereby generating a directed risk transmission network; Calculate the in-degree centrality and out-degree centrality of each network node in the directed risk propagation network, and use the product of the in-degree centrality and the out-degree centrality as the bidirectional propagation strength of the network node. All network nodes in the directed risk transmission network are sorted in descending order of bidirectional propagation intensity, and a number of network nodes that rank at the top of the preset proportion are selected as the key control nodes. All risk factors in the risk factor subvector corresponding to each key control node constitute the control factor set corresponding to that key control node.

4. The method for full-process management of aquaculture product quality and safety based on big data according to claim 3, characterized in that, Based on the key regulatory nodes and their corresponding regulatory factor sets, a perturbation propagation simulation strategy is used to evolve and deduce the quality label vector, resulting in the predicted risk label sequence for each aquaculture batch, including: For each of the key control nodes, positive and negative perturbations of preset amplitude are applied to each control factor in the control factor set corresponding to the key control node, thereby generating multiple perturbation simulation scenarios corresponding to the key control node. For each of the aforementioned disturbance simulation scenarios, based on the directed propagation path from the key control node to other network nodes in the directed risk transmission network, the change in risk factors of each other network node under the disturbance simulation scenario is calculated level by level. The risk factor change of each of the other network nodes is superimposed on the original risk factor subvector of that other network node to obtain the updated risk factor subvector of that other network node under the disturbance simulation scenario. According to the preset risk mapping function, the updated risk factor subvectors of all network nodes in each disturbance simulation scenario are mapped to the simulation quality label vector corresponding to that disturbance simulation scenario; Arrange the simulated quality label vectors corresponding to the multiple disturbance simulation scenarios in chronological order to obtain the predicted risk label sequence corresponding to each aquaculture batch.

5. The method for full-process management of aquaculture product quality and safety based on big data according to claim 4, characterized in that, Based on the deviation distribution between the predicted risk label sequence and the product inspection results, a quality offset is generated for each breeding batch, and risk level classification is performed on each breeding batch according to the quality offset, including: The difference between each simulated quality label vector in the predicted risk label sequence and the product inspection result corresponding to the same breeding batch is calculated to obtain multiple predicted deviation values ​​corresponding to the breeding batch. Calculate the statistical characteristic value of the plurality of predicted deviation values, and use the statistical characteristic value as the quality offset of the breeding batch. The statistical characteristic value includes the mean and standard deviation of the plurality of predicted deviation values. Obtain a preset set of risk level division intervals, wherein the set of risk level division intervals contains multiple consecutive numerical intervals, and each numerical interval corresponds to a risk level; The quality offset of each breeding batch is matched with the numerical range in the risk level classification interval set to determine the risk level to which the breeding batch belongs.

6. The method for full-process management of aquaculture product quality and safety based on big data according to claim 5, characterized in that, The output of the quality and safety control plan for each breeding batch includes: Based on the node identifier of the key control node in the directed risk transmission network, multiple candidate control measures associated with the node identifier are retrieved from the preset management decision knowledge base. The preset management decision knowledge base stores the mapping relationship between node identifiers and control measures in advance. Select target control measures that match the risk level classification results from the plurality of candidate control measures; For each key control node, the target control measures corresponding to the key control node are sorted by execution priority according to the ratio of the in-degree centrality to the out-degree centrality of the key control node, and a node control sequence corresponding to the key control node is generated. According to the bidirectional propagation intensity from large to small, the node control sequences corresponding to each key control node are connected in series to generate the quality and safety full-process control scheme.

7. The method for full-process management of aquaculture product quality and safety based on big data according to claim 2, characterized in that, The calculation of the multidimensional spatial distance between risk factor sub-vectors of any two aquaculture batches, and the construction of the risk factor correlation matrix for each aquaculture batch based on the multidimensional spatial distance, includes: Any two breeding batches are designated as the first breeding batch and the second breeding batch, and the first risk factor sub-vector corresponding to the first breeding batch and the second risk factor sub-vector corresponding to the second breeding batch are extracted. Calculate the absolute value of the difference between the element in each dimension of the first risk factor sub-vector and the element in the same dimension of the second risk factor sub-vector to obtain the difference in multiple dimensions; The weighted summation of the differences in the multiple dimensions yields the multidimensional spatial distance between the first and second breeding batches, wherein the weight of each dimension difference is positively correlated with the degree of data variation in that dimension. Based on the multidimensional spatial distance between any two breeding batches, a distance matrix with a dimension equal to the total number of breeding batches is constructed as the risk factor correlation matrix.

8. The method for full-process management of aquaculture product quality and safety based on big data according to claim 3, characterized in that, The generation of the directed risk transmission network includes: The two farming batches connected by any one of the network edges in the initial risk transmission network are respectively denoted as the source farming batch and the target farming batch; Extract the factor dimensions from the risk factor subvector of the source breeding batch that have the same trend of change as the risk factor subvector of the target breeding batch to obtain a set of consistent factor dimensions; Extracting the factor dimensions from the risk factor subvector of the source breeding batch that have the opposite trend to the risk factor subvector of the target breeding batch, and obtaining the set of opposite factor dimensions; Compare the number of dimensions in the consistent factor dimension set with the number of dimensions in the opposite factor dimension set. If the number of dimensions in the consistent factor dimension set is greater than the number of dimensions in the opposite factor dimension set, then set the direction attribute of the network edge to point from the source breeding batch to the target breeding batch; otherwise, set the direction attribute of the network edge to point from the target breeding batch to the source breeding batch.

9. The method for full-process management of aquaculture product quality and safety based on big data according to claim 4, characterized in that, The step-by-step calculation of the change in risk factors for each other network node in the disturbance simulation scenario includes: For each of the aforementioned disturbance simulation scenarios, the key control node corresponding to the disturbance simulation scenario is taken as the propagation starting node, and all reachable network nodes that can be reached from the propagation starting node through multiple directed edges are determined in the directed risk transmission network. According to the direction of the directed edge, each path from the propagation starting node to any reachable network node is divided into multiple propagation levels, and each propagation level corresponds to the order in which a directed edge is traversed. For all reachable network nodes in the first propagation level, the disturbance amplitude in the disturbance simulation scenario is directly multiplied by the propagation attenuation coefficient of the directed edge corresponding to the propagation level to obtain the change in risk factor for each reachable network node in the propagation level. For each propagation level greater than the first propagation level, the change in risk factor of reachable network nodes on the same propagation path under the previous propagation level is multiplied by the propagation attenuation coefficient of the directed edge corresponding to the current propagation level to obtain the change in risk factor of each reachable network node under the current propagation level.

10. The method for full-process management of aquaculture product quality and safety based on big data according to claim 6, characterized in that, Selecting target control measures from the multiple candidate control measures that match the risk level classification results includes: Obtain the level identifier corresponding to each risk level in the risk level classification results, and the set of applicable level identifiers corresponding to each candidate control measure in the preset management decision knowledge base; The set of applicable level identifiers for each candidate control measure is intersected with the level identifiers in the risk level classification result to obtain the set of matching levels for the candidate control measure. The number of levels in the matching level set is counted, and the multiple candidate control measures are sorted in descending order of the number of levels to obtain a candidate control measure sequence; The candidate control measures ranked at the top of the candidate control measure sequence by a predetermined number are determined as the target control measures.

Images