Seafood frozen food safety detection method

Abstract

The invention relates to the technical field of food safety detection, and discloses a seafood frozen food safety detection method. According to the method, a tracing record covering the whole process of the supply chain is constructed, and key quality control nodes are extracted. And establishing a dynamic change model for each node, wherein the dynamic change model is used for simulating evolution paths of pollutant concentration, total number of microbial colonies and ice crystal growth size. A sensing device is deployed at a node to collect actual measurement data, a corresponding model is input, and safety state vectors of the current node and the predicted future node are obtained through calculation. And aggregating the security state vectors of all the nodes, generating a global security portrait for comprehensive evaluation, and executing product release, directional reinspection or tracing recall decisions according to the portrait. According to the method, the conversion of the safety state from static detection to dynamic prediction is realized, differentiated accurate control is supported through the global risk portrait, and the risk early warning capability and the decision-making efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of food safety testing technology, specifically a method for testing the safety of frozen seafood. Background Technology

[0002] Currently, the safety management of frozen seafood mainly relies on discrete sampling and static testing at key nodes in the supply chain. Laboratory analyses of indicators such as biotoxins and microorganisms are conducted according to standards during the fishing, processing, and storage stages, and the temperature and humidity of the cold chain are monitored and recorded. This method, by comparing the detected values ​​with preset thresholds, can only determine the compliance status at the moment of sampling and cannot reflect the continuous changes in safety indicators during the process.

[0003] Existing methods have limitations. Data from each stage is isolated, making it difficult to establish complete traceability. The lack of correlation between discrete detection points makes it impossible to assess the impact of preceding conditions on subsequent safety. Its fundamental flaw lies in the inability to model and predict dynamic processes such as pollutant migration and microbial proliferation, leading to delayed control and an inability to achieve risk early warning.

[0004] The decision-making mechanism is also imperfect. Current decisions on release or recall are mostly based on simple compliance judgments of single or multiple test results, lacking integrated analysis of multi-dimensional, cross-node risk information. Such rigid rules are unable to identify systemic risks formed by the accumulation of multiple minor anomalies, easily leading to inaccurate measures.

[0005] There is a need for a detection method that can integrate data from the entire chain, build dynamic models of key quality control nodes to simulate the evolution of security indicators, and generate a global risk profile based on the aggregation and analysis of security status of multiple nodes, so as to achieve the transformation from static judgment to dynamic prediction and from single-point compliance to global risk assessment. Summary of the Invention

[0006] The purpose of this invention is to provide a method for testing the safety of frozen seafood, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for detecting the safety of frozen seafood, the method comprising: Construct a traceability record for frozen seafood that covers the entire supply chain. The traceability record integrates information on place of origin, catch or farm batch, initial biotoxin content, processing plant environmental parameters, freezing process curve, cold chain transportation trajectory, and storage temperature and humidity history. Based on food safety regulations and industry standards, key quality control nodes are extracted from the traceability records, and a dynamic change model is established for each key quality control node. The dynamic change model is used to simulate the evolution path of pollutant concentration, total number of microbial colonies, and ice crystal growth size under specific conditions. Deploy sensing devices at key quality control nodes to collect physical samples and convert them into digital signals to form a node measured data stream; The measured data stream of the node is input into the dynamic change model of the corresponding key quality control node to calculate the safety status vector of the frozen seafood at the current node and the predicted future node. The security status vectors of all key quality control nodes are aggregated to generate a global security profile, which is used to guide whether to release, conduct targeted re-inspection, or initiate a traceability recall procedure.

[0008] Preferably, establishing a dynamic change model for each key quality control node includes: Identify the core and auxiliary variables that affect the food safety of the key quality control nodes. The core variables include time, temperature, and humidity, and the auxiliary variables include the air permeability of packaging materials and the risk of cross-contamination between adjacent products. Collect historical batches of data on the decay or growth of safety indicators under different combinations of the core and auxiliary variables, and construct a training sample set. Using the training sample set, train a time series prediction network that can receive the current variable value and output the predicted values ​​of security indicators at multiple future time points; The trained time-series prediction network, its corresponding variable input range, and constraints are collectively encapsulated into the dynamic change model.

[0009] Preferably, the calculation of the safety state vector of the frozen seafood at the current node and the predicted future nodes includes: The specific values ​​of the core variables and auxiliary variables at the current moment are parsed from the measured data stream of the nodes; The specific values ​​are input into the time-series prediction network to obtain a continuous safety indicator prediction curve from the current time to a preset future time. The safety indicator prediction curve is piecewise integrated to calculate the cumulative change in the safety indicator within each time period. The safety state vector is formed by combining the measured safety indicators at the current moment, the worst extreme value in the predicted curve of future safety indicators, and the time period number where the cumulative amount exceeds the threshold.

[0010] Preferably, generating a global security profile includes: Each key quality control node's safety state vector is assigned a node weight, which is dynamically set based on the degree of influence and irreversibility of the key quality control node on the final food safety. Following the supply chain process sequence, all weighted security state vectors are concatenated to form a multi-dimensional security evolution path; On the multidimensional security evolution path, mark all abnormal points where the measured or predicted values ​​of security indicators exceed the legal standards, and record the node location, the extent of exceeding the standard, and the predicted duration of the abnormal points; Based on the labeling results, a global security profile is generated, which includes a visual map of the security evolution path, a list of anomalies, and an overall risk score.

[0011] Preferably, the method further includes an online calibration step for the dynamically changing model: After a batch of frozen seafood is actually released or processed at the key quality control node, the batch of frozen seafood released is continuously tracked until the next key quality control node or end consumption stage, and its actual safety indicator change data is collected. The actual safety indicator change data are compared with the prediction sequence previously made by the dynamic change model to calculate the prediction deviation; If the prediction deviation continues to exceed the allowable range, model correction is initiated, and the internal parameters of the time series prediction network are fine-tuned using the latest collected actual data to update the dynamic change model.

[0012] Preferably, the startup model calibration includes: Extract the complete traceability records of the batch of frozen seafood that caused the prediction deviation to exceed the standard, as well as all its measured data at the relevant nodes; Analyze the patterns of the prediction deviations to distinguish whether the deviations originate from inherent model errors, unrecorded sudden interference events, or variations in the product's own characteristics. If the bias stems from inherent model error, then incremental training of the time series prediction network is performed using new data. If the deviation originates from an unrecorded sudden disturbance event, the characteristics of the sudden disturbance event are abstracted into a new auxiliary variable and added to the input of the dynamic change model. If the deviation originates from variations in the product's own characteristics, a sub-model is established under the label of the category or place of origin of the frozen seafood that caused the variation, and the sub-model is used preferentially in subsequent detection.

[0013] Preferably, the method further includes a decision feedback step based on a global security profile: Analyze the list of anomalies in the global security profile, and formulate tiered intervention strategies based on the node location, the extent of exceedance, and the predicted duration of the anomalies. The tiered intervention strategy includes: for a single node with a slight exceedance that does not spread, increasing the sampling frequency of subsequent batches of the single key quality control node; for multiple nodes with consecutive exceedances or a single node with a serious exceedance, triggering a traceability review of the upstream links of the single key quality control node; and for predictions that indicate the risk will spread rapidly to the terminal, immediately notifying the terminal warehouse to implement interception and isolation. The new detection data generated after implementing the tiered intervention strategy will be fed back into the system as a new node measured data stream to update the global security profile.

[0014] Preferably, the triggering of a traceability review of the upstream links of the single key quality control node includes: Starting from the critical quality control point where serious exceedances occurred, trace back along the supply chain; Retrieve complete traceability records and historical safety status vectors of all upstream links for batches that exceeded the standards at the critical quality control node where serious exceedances occurred; Correlation analysis was conducted between historical operating parameters of upstream processes and the current terminal exceedance indicators to pinpoint the upstream processes and operating parameter ranges most likely to cause the problem. Generate a traceability review report, which includes the hypothesis of the source of the problem, the relevant chain of evidence, and the recommended process modification parameters.

[0015] Preferably, the construction of a traceability record for frozen seafood covering the entire supply chain includes: Each batch of frozen seafood is assigned a unique identification code, and the unique identification code is automatically read by sensing devices at each link in the supply chain; When reading the unique identifier, the static attributes and dynamic process data of each supply chain link are automatically or manually entered. The static attributes include the processing plant number, equipment number, and operator number. The dynamic process data includes the start time, end time, process temperature curve, and concentration of disinfectant used. Data from each stage is linked and stored using the unique identifier as an index, forming a chain-like traceability record.

[0016] Preferably, the step of extracting key quality control nodes from the traceability records includes: Analyze historical food safety incidents and statistically analyze the frequency and severity of these incidents at each stage of the supply chain; By combining industry standards and expert knowledge, we selected the links that have a decisive impact on the safety of the final product and whose data are available as candidate nodes. Evaluate the technical feasibility and economic cost of deploying sensing devices at each candidate node; Based on the frequency, severity, decisive impact, and feasibility of occurrence, the final multiple key quality control nodes and their order in the supply chain are determined.

[0017] Compared with the prior art, the beneficial effects of the present invention are: By establishing dynamic change models for indicators such as contaminant concentration, total microbial colony count, and ice crystal growth size at each key quality control node, and inputting real-time collected node data streams into the corresponding models for calculation, continuous simulation and prediction of product safety status across the supply chain can be achieved. This allows safety assessments to move beyond isolated instantaneous measurements at isolated nodes, enabling the calculation of product compliance at the current node and theoretical safety values ​​at subsequent nodes based on current measured data and pre-defined evolution paths. Control actions can therefore be initiated earlier, before actual measured values ​​exceed thresholds, based on predicted trends, issuing warnings and launching interventions at a more advanced stage, shifting the management focus from reactive post-event handling to proactively blocking potential risk paths.

[0018] A global safety profile is generated by aggregating the safety status vectors of key quality control nodes. Each vector integrates the node's measured data, model fitting information, and predicted status. The global profile then uses algorithms to comprehensively analyze the interactions and cumulative effects of multiple nodes and various types of risk factors, forming a structured and quantitative assessment of the overall risk situation of the product. The decision-making mechanism based on this profile can implement differentiated handling: batches with low overall risk levels are automatically released; batches showing abnormalities or uncertainties in specific node vectors within the profile are triggered for targeted re-inspection; and only batches showing significant correlations between multiple node risks in the profile are subject to traceability and recall. This represents a fundamental shift from rigid judgments relying on fixed rules to flexible decision-making based on dynamic and comprehensive risk situations. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the working principle of the seafood frozen food safety testing method described in this invention. Figure 2 A flowchart for establishing a dynamic change model; Figure 3 A flowchart for calculating the safety state vector; Figure 4 A comparison chart of prediction deviations before and after calibration of the seafood frozen food detection model; Figure 5 A radar chart showing the safety risks at each quality control stage of frozen seafood products. Detailed Implementation

[0020] 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, and 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] Please see Figure 1This invention provides a method for testing the safety of frozen seafood. The method includes: constructing a traceability record for frozen seafood covering the entire supply chain, which integrates origin information, catch or aquaculture batches, initial biotoxin content, processing plant environmental parameters, freezing process curves, cold chain transportation trajectory, and storage temperature and humidity history; extracting key quality control nodes from the traceability record according to food safety regulations and industry standards, and establishing a dynamic change model for each key quality control node, which is used to simulate the evolution path of contaminant concentration, total microbial colony count, and ice crystal growth size under specific conditions; deploying sensing devices at the key quality control nodes to collect physical samples and convert them into digital signals to form a node measured data stream; inputting the node measured data stream into the dynamic change model of the corresponding key quality control node to calculate the safety status vector of the frozen seafood at the current node and predicted future nodes; aggregating the safety status vectors of all key quality control nodes to generate a global safety profile, which is used to guide whether to release, conduct targeted re-inspection, or initiate a traceability recall procedure.

[0022] In one embodiment of the present invention, see [reference] Figure 2 The process of establishing a dynamic change model for each key quality control node includes: identifying core and auxiliary variables affecting food safety at key quality control nodes; core variables including time, temperature, and humidity; and auxiliary variables including the permeability of packaging materials and the risk of cross-contamination between adjacent products; collecting historical batch data on the decay or increase of safety indicators under different combinations of core and auxiliary variables to construct a training sample set; using the training sample set to train a time-series prediction network that can receive current variable values ​​and output predicted values ​​of safety indicators for multiple future time points; and encapsulating the trained time-series prediction network, its corresponding variable input range, and constraints into a dynamic change model.

[0023] In practical implementation, after extracting key quality control nodes from traceability records based on food safety regulations and industry standards, the process of establishing a dynamic change model for each key quality control node needs to be systematically executed. Identifying the core and auxiliary variables affecting the food safety of key quality control nodes is a fundamental step. Core variables include time, temperature, and humidity, while auxiliary variables include the air permeability of packaging materials and the risk of cross-contamination between adjacent products. Taking the frozen storage stage as an example, the core variables are specifically the storage duration, the average temperature and temperature fluctuation range of the cold storage, and the relative humidity inside the storage. The auxiliary variables are specifically the oxygen permeability of the seafood product packaging film and other types of aquatic products stored in the same storage space. Data on the decay or increase of safety indicators under different combinations of core and auxiliary variables in historical batches are collected to construct a training sample set. This work involves extracting complete records of all batches in the frozen storage stage from the enterprise's quality management database over the past three years. The records include the values ​​of each temperature recording point, the total number of colonies in each microbial sampling, and the corresponding packaging specifications and information on adjacent products. These records are cleaned and aligned according to the correspondence between variable combinations and safety indicator outputs to form a structured sample set.

[0024] Using a training sample set, a time-series prediction network is trained that can receive current variable values ​​and output predicted values ​​of safety indicators at multiple future time points. In some embodiments, the time-series prediction network adopts a long short-term memory network structure, where its input layer nodes correspond to a set of normalized real-time values ​​of core and auxiliary variables, and its output layer nodes correspond to the logarithmic values ​​of the predicted total colony count at seven equally spaced future time points. The goal of the training process is to minimize the difference between the network's predicted sequence and the actual observed changes in safety indicators in the training sample set. A loss function formula is used to calculate this difference. Expressed as:

[0025] Where: M represents the total number of training batches, and T represents the number of prediction time points. This represents the actual safety indicator observation value of the i-th training sample at the t-th time point in the future. This represents the predicted safety index value of the i-th training sample at the t-th time point in the future by the time-series prediction network. Training iteratively adjusts the network's internal parameters using the backpropagation algorithm until the loss function converges to below a preset threshold. Essentially, the trained time-series prediction network, its corresponding variable input range, and constraints are encapsulated into a dynamically changing model. This encapsulation involves saving the trained network parameters, network structure definition, standardized parameters of the input variables, and the upper and lower limits of the effective values ​​for each input variable as an independent software module or configuration file that can be called by the detection system. This module, upon receiving real-time variable values ​​that meet the required format, can output a structured safety index prediction sequence. Optionally, the variable input range explicitly defines the applicable scenarios for the model. For example, the temperature input range is set between -25 degrees Celsius and -18 degrees Celsius; input values ​​exceeding this range will be rejected by the system or trigger a warning.

[0026] In one embodiment of the present invention, see [reference] Figure 3 The safety state vector of frozen seafood at the current node and the predicted future node is calculated as follows: the specific values ​​of the core variables and auxiliary variables at the current moment are parsed from the measured data stream of the node; the specific values ​​are input into the time series prediction network to obtain a continuous safety index prediction curve from the current moment to the preset future moment; the safety index prediction curve is integrated piecewise to calculate the cumulative amount of safety index change in each time period; the measured safety index at the current moment, the worst extreme value of the future safety index prediction curve, and the time period number where the cumulative amount exceeds the threshold are combined together to form the safety state vector. Generating a global safety profile includes: assigning node weights to the safety status vectors of each key quality control node, with the node weights dynamically set based on the degree and irreversibility of the key quality control node's impact on the final food safety; concatenating all weighted safety status vectors according to the supply chain process sequence to form a multi-dimensional safety evolution path; marking all anomalies where the measured or predicted values ​​of safety indicators exceed the legal standards on the multi-dimensional safety evolution path, and recording the node location, the extent of exceeding the standard, and the predicted duration of the anomalies; and generating a global safety profile based on the marking results, including a visual map of the safety evolution path, a list of anomalies, and an overall risk score.

[0027] In practical implementation, the process of calculating the safety state vector of frozen seafood at the current node and predicting future nodes is specifically executed. At the critical quality control node of cold chain transportation, deployed sensing devices collect data streams of temperature, humidity, GPS location, and vibration frequency inside the vehicle. From the actual measured data stream at the node, the specific values ​​of the core variables and auxiliary variables at the current moment are parsed out. Specific values ​​include, for example, a temperature of -20.3 degrees Celsius, relative humidity of 85%, continuous transportation time of 18 hours, and packaging box stacking density of 0.75. These specific values ​​are input into a time-series prediction network dedicated to the cold chain transportation link to obtain a continuous safety indicator prediction curve from the current moment to a preset future moment. The prediction curve is plotted on the horizontal axis with the predicted logarithm of the total number of microbial colonies on the vertical axis, showing the hourly prediction values ​​for the next 24 hours. The safety indicator prediction curve is piecewise integrated to calculate the cumulative amount of safety indicator changes within each time period. Specifically, the next 24 hours are divided into four 6-hour time periods, and the area of ​​the microbial predicted growth curve within each time period is calculated. This area value represents the cumulative intensity of microbial activity within that time period. The current measured safety index, the worst extreme value in the future safety index prediction curve, and the time period number where the cumulative amount exceeds the threshold are combined to form a safety state vector.

[0028] Generating a global safety profile includes subsequent aggregation and visualization steps. Each key quality control node's safety state vector is assigned a node weight. These weights are dynamically set based on the degree and irreversibility of the key quality control node's impact on final food safety. In some embodiments, a weight allocation model based on the analytic hierarchy process (AHP) is used. The initial biotoxin control weight in the processing stage is set to 0.35, the ice crystal morphology control weight in the freezing stage is set to 0.30, the microbial control weight in the cold chain transportation stage is set to 0.25, and the weight in the warehousing stage is set to 0.10. These weight values ​​are calculated by combining expert scores and historical accident statistics. Following the supply chain process sequence, all weighted safety state vectors are concatenated to form a multi-dimensional safety evolution path. This path is represented as a matrix in the data structure, where each row represents a weighted safety state vector of a key quality control node. The order of the rows strictly corresponds to the order of the supply chain stages. On the multidimensional safety evolution path, all anomalies where the measured or predicted values ​​of safety indicators exceed the legal standards are marked, and the node location, exceedance range, and predicted duration of the anomalies are recorded. For example, the path matrix is ​​marked with "Processing node: Histamine predicted value continuously exceeds the legal standard of 50 mg / kg for the next 8 hours, with a maximum exceedance range of 120%" and "Transportation node: The current measured value of total bacterial count exceeds the standard limit." Based on the marking results, a global safety profile is generated, including a visual map of the safety evolution path, a list of anomalies, and an overall risk score. The visual map transforms the multidimensional safety evolution path matrix into a two-dimensional display combining a line graph with marked points and a heatmap. The overall risk score R is calculated using an aggregation function.

[0029] Where: N is the total number of key quality control nodes, It is the dynamic weight of the k-th node. It is the measured or predicted value of the core security indicator of the k-th node. This is the legally mandated standard value for this safety indicator. This is the predicted total duration of the exceedance at this node. It is the total time span of the assessment. and These are adjustment coefficients for the magnitude and duration of exceeding the standard, respectively. Optionally, the list of anomalies is presented in tabular form, with columns including node name, exceeding indicator, current value, standard value, predicted worst value, predicted duration of exceeding the standard, and recommended measures. In some embodiments, the overall risk score R is mapped to three levels: "low risk," "medium risk," and "high risk," and serves as the primary criterion for determining whether to release the vehicle.

[0030] In one embodiment of the present invention, the method further includes an online calibration step for the dynamic change model: after a batch of frozen seafood is actually released or processed at a key quality control node, the currently released batch of frozen seafood is continuously tracked until the next key quality control node or end-consumer stage, and its actual safety index change data is collected; the actual safety index change data is compared with the prediction sequence previously made by the dynamic change model, and the prediction deviation is calculated; if the prediction deviation continues to exceed the allowable range, model calibration is initiated, and the internal parameters of the time series prediction network are fine-tuned using the latest collected actual data to update the dynamic change model. The model calibration process includes: extracting the complete traceability records of the batch of frozen seafood that caused the prediction deviation to exceed the limit, as well as all measured data at the relevant nodes; analyzing the pattern of prediction deviation, distinguishing whether the deviation originates from inherent model error, unrecorded sudden interference events, or product-specific variation; if the deviation originates from inherent model error, incremental training of the time-series prediction network is performed using new data; if the deviation originates from unrecorded sudden interference events, the characteristics of the sudden interference events are abstracted into new auxiliary variables and expanded into the input of the dynamic change model; if the deviation originates from product-specific variation, a sub-model is established under the label of the category or place of origin of the frozen seafood that caused the variation, and the sub-model is preferentially called in subsequent detection.

[0031] In its implementation, the method also includes an online calibration step for the dynamic change model. Taking a batch of frozen shrimp products coded "SF20250315A" as an example, at key quality control points in the processing stage, the dynamic change model outputs a predicted microbial sequence for the next 24 hours based on the received initial total bacterial count, washing water temperature, and time parameters. After the actual release or processing of a batch of frozen seafood, the system activates continuous tracking for that batch. It continuously tracks the currently released "SF20250315A" batch of frozen shrimp products until the next key quality control point or end-consumer stage, collecting data on actual safety indicator changes. Specifically, through RFID tags attached to the product packaging and fixed readers along the route, it automatically correlates the measured total bacterial count values ​​obtained through random sampling during subsequent cold chain transportation and storage, forming an actual safety indicator change curve from the moment of release in the processing stage. The actual changes in safety indicators are compared with the predicted sequences previously made by the dynamic change model to calculate the prediction deviation. The deviation calculation is performed on the predicted values ​​and measured values ​​at multiple equally spaced time points, providing an indicator for quantifying the overall prediction deviation. The formula is expressed as:

[0032] Where: P represents the number of batches participating in the deviation assessment, and Q represents the number of time points for comparison for each batch. This represents the predicted safety index value for the j-th batch at time s, provided by the dynamic change model. This represents the actual measured value of the safety index for the j-th batch at time s. If the prediction deviation... If the data continues to exceed the allowable range, model calibration is initiated. The internal parameters of the time series prediction network are fine-tuned using the latest collected data from the "SF20250315A" batch and other batches from the same period. The dynamically changing model is updated. The fine-tuning process uses the gradient descent method, but a small learning rate is set to prevent significant overwriting of existing knowledge.

[0033] Initiating model calibration involves a series of diagnostic and corrective operations. This includes extracting the complete traceability records of the frozen seafood batch that caused the excessive prediction deviation, as well as all measured data at relevant nodes. For example, retrieving all environmental parameters, operation logs, and test reports for batch "SF20250315A" from catch to processing. The pattern of prediction deviation is analyzed to distinguish whether the deviation originates from inherent model errors, unrecorded sudden interference events, or variations in product characteristics. In some embodiments, cluster analysis is used to classify the morphology of the deviation sequences, and cross-validation is performed using text logs from the traceability records. If the deviation originates from inherent model errors, incremental training of the time-series prediction network is performed using new data. Incremental training is conducted based on the original network weights, and the training dataset includes batch data with significant deviations collected within the past month. If the deviation originates from an unrecorded sudden disturbance event, the characteristics of the sudden disturbance event are abstracted into a new auxiliary variable and added to the input of the dynamic change model. For example, after a prediction deviation event caused by a temporary refrigeration failure of a transport vehicle, the system adds a binary auxiliary variable, "refrigeration equipment abnormality flag," to the input of the dynamic change model in the transportation process. This auxiliary variable is then manually or automatically labeled by the driver or sensors through the system interface during subsequent data collection. If the deviation originates from variations in the product's own characteristics, a sub-model is established under the label of the category or origin of the frozen seafood that caused the variation. For instance, when it is found that the microbial growth rate of scallops from a specific origin is significantly different from the prediction of the conventional model due to seasonal physiological changes, the system will create a dedicated time-series prediction network sub-model for "scallop-origin X-autumn." This sub-model is trained using historical data from the autumn batches of that origin. Optionally, the sub-model is prioritized in subsequent detection. When the system identifies a matching sub-model between the category and origin label in the product traceability record, it will automatically call the corresponding sub-model for safety status prediction, rather than the general model. In some embodiments, the criterion for judging the inherent error of the model is deviation. It shows a slow increasing trend over time and cannot be correlated with any specific incident record.

[0034] In one embodiment of the present invention, the method further includes a decision feedback step based on a global security profile: parsing the list of anomalies in the global security profile, and formulating a graded intervention strategy based on the node location, exceedance magnitude, and predicted duration of the anomalies; the graded intervention strategy includes: increasing the sampling frequency of subsequent batches of a single key quality control node for a single node with only slight exceedance that does not spread; triggering a traceability review of the upstream links of a single key quality control node for multiple nodes with continuous exceedance or a single node with severe exceedance; and immediately notifying the terminal warehouse to implement interception and isolation for a prediction that indicates the risk will spread rapidly to the terminal; and feeding back the new detection data generated after executing the graded intervention strategy as a new node measured data stream into the system to update the global security profile. Triggering a traceability review of the upstream links of a single critical quality control node includes: starting from the critical quality control node where serious exceedances occurred, tracing back along the supply chain; retrieving complete traceability records and historical safety status vectors of all upstream links for batches that exceeded the standards at the critical quality control node where serious exceedances occurred; constructing a correlation analysis between the historical operating parameters of upstream links and the current terminal exceedance index to locate the upstream links and operating parameter ranges most likely to cause the problem; and generating a traceability review report, which includes the hypothesis of the problem's source location, the associated evidence chain, and recommended process correction parameters.

[0035] In its implementation, the method also includes a decision feedback step based on a global safety profile. This involves analyzing the list of anomalies in the global safety profile and formulating tiered intervention strategies based on the location of the anomalies, the extent of exceedance, and the predicted duration. For example, taking a batch of frozen cod products with three anomalies marked in the global safety profile as an example, the anomaly list provides a specific quantitative description of each issue. The tiered intervention strategies include: for minor exceedances at a single node that do not spread, increasing the sampling frequency of subsequent batches at that key quality control node. For instance, if the list shows that the temperature only briefly exceeds the limit by 0.5 degrees Celsius during storage and the prediction model shows that microbial growth has not accelerated, the system automatically increases the sampling frequency of the next three batches of similar products from that warehouse from 10% to 30%. For multiple nodes continuously exceeding the limit or a single node severely exceeding the limit, triggering a traceability review of the upstream links of that key quality control node. For instance, if the list shows that the initial total bacterial count exceeds the limit during processing and the predicted total bacterial count during transportation will continue to severely exceed the limit, the system generates an instruction to trigger a traceability review of the upstream links of the processing stage. For forecasts indicating a rapid spread of risk to end-users, the system immediately notifies the end-user warehouse to implement interception and isolation. For example, if the manifest shows that the predicted histamine content of a batch of products will exceed safety standards within the next two hours before arriving at the end-user warehouse, the system automatically sends an interception instruction containing the product batch code to the warehouse's management system. New testing data generated after implementing the tiered intervention strategy is fed back into the system as new node-specific data streams to update the overall safety profile. For instance, after increasing the sampling frequency in the warehousing process, newly acquired microbial testing data is collected and input in real time, driving the generation of an updated safety evolution path for that batch of products, including the latest measured data points.

[0036] Triggering a traceability review of the upstream links of a single key quality control node involves a series of reverse analysis and location operations. Starting from the key quality control node where serious exceedances occur, the process traces back along the supply chain, for example, from the terminal distribution warehouse node where the total bacterial count was found to be seriously exceeded, and then back through cold chain transportation, frozen processing, and finally to the fishing or aquaculture batches. The system retrieves complete traceability records and historical safety status vectors of all upstream links for the batches that exceeded the limits at the key quality control node where serious exceedances occurred. The system indexes and aggregates the raw data of all upstream links through the product's unique identifier. A correlation analysis is constructed between the historical operating parameters of upstream links and the current terminal exceedance index to locate the upstream links and operating parameter ranges most likely to cause the problem. In some embodiments, a combination of Pearson correlation coefficient and regression analysis is used to calculate the statistical correlation strength between variables in each upstream link and the terminal total bacterial count. A traceability review report is generated, including the problem source location hypothesis, the chain of related evidence, and suggested process correction parameters. The report is output as a structured document, where the source location hypothesis is derived from the correlation analysis results, and the chain of related evidence is presented in the form of a timeline and data comparison table. See Table 1.

[0037] Table 1: List of Anomalies and Related Upstream Parameters

[0038] It is understandable that a specific quantitative method of correlation analysis is to calculate a priority intervention index. This is used to prioritize the review of multiple suspicious upstream links, and its formula is expressed as:

[0039] in: V represents the absolute value of the correlation coefficient between an upstream operational parameter and the terminal exceeding the standard indicator, and V represents the actual value of the operational parameter of that upstream link in this batch exceeding the standard. This represents the standard process setting value for this operating parameter. This indicates the interval between the occurrence of the upstream operation and the terminal detection time. It is a time decay coefficient. The higher the value of the upstream link, the higher the priority of its source identification hypothesis in the traceability review report. Optionally, the chain of evidence may include direct references to original data snapshots in the traceability records, such as screenshots of log records showing that the chiller fan speed sensor reading was below the standard value for one consecutive hour at a specific time point in the processing stage. In some embodiments, the suggested process correction parameters are given based on the data distribution range of historical successful batches, for example, revising the suggested parameter for "effective chlorine concentration in cleaning water" from "≥1.5ppm" to "≥2.2ppm".

[0040] See Figure 4 This is a comparison chart of prediction deviations before and after calibration of a seafood frozen food detection model. It's a visualization of the "Online Calibration of Dynamically Changing Models" phase, demonstrating the improvement in prediction accuracy achieved through model calibration. Before calibration, the prediction deviations for all batches were significantly higher than the acceptable threshold (0.85-1.05), indicating insufficient model accuracy. After calibration, the prediction deviations for all batches decreased to below the acceptable threshold (0.15-0.25), showing a significant improvement in accuracy. Furthermore, the deviation fluctuations for each batch were smaller after calibration, indicating more stable model prediction results. This visually demonstrates the effectiveness of model calibration, ensuring the accuracy of subsequent safety indicator predictions. A high-precision prediction model can more accurately identify food safety risks, avoiding misjudgments or omissions. Improved model accuracy can reduce unnecessary re-inspection costs and increase testing efficiency.

[0041] In one embodiment of the present invention, constructing a traceability record for frozen seafood covering the entire supply chain includes: assigning a unique identifier to each batch of frozen seafood, and automatically reading the unique identifier through a sensing device at each link in the supply chain; automatically or manually recording static attributes and dynamic process data for each link in the supply chain when reading the unique identifier, including processing plant number, equipment number, and operator number, and dynamic process data including start time, end time, process temperature curve, and concentration of disinfectant used; associating and storing the data of each link using the unique identifier as an index to form a chain-like traceability record. Extracting key quality control nodes from the traceability record includes: analyzing historical food safety incidents and statistically analyzing the frequency and severity of incidents at each link in the supply chain; combining industry standards and expert knowledge to select links that have a decisive impact on the safety of the final product and whose data is available as candidate nodes; evaluating the technical feasibility and economic cost of deploying sensing devices at each candidate node; and determining the final multiple key quality control nodes and their order in the supply chain by comprehensively considering the frequency, severity, decisive impact, and feasibility.

[0042] In practice, building a traceability record for frozen seafood that covers the entire supply chain requires specific technical solutions and data standards. Each batch of frozen seafood is assigned a unique identification code, and the unique identification code is automatically read by sensing devices at each link in the supply chain. For example, a batch of tuna fillets is assigned a barcode that conforms to the GS1-128 standard on the fishing vessel. The barcode encodes the country of origin code, the registration number of the farm or fishing vessel, the production date, and the batch number. When reading the unique identifier, static attributes and dynamic process data for each link in the supply chain are automatically or manually entered. Static attributes include the processing plant number, equipment number, and operator number. Dynamic process data includes start time, end time, process temperature curve, and disinfectant concentration used. When a batch of tuna fillets enters the processing plant, workers scan the barcode using handheld terminals in the unloading area. The system automatically records the current time as the "entry start time." Simultaneously, workers manually select the "cleaning line 3 equipment" number and "operator employee number 1024" for this operation through the terminal interface. In the subsequent freezing process, temperature sensors automatically record and upload the complete temperature-time curve of the tuna fillet's core temperature dropping from 15 degrees Celsius to -18 degrees Celsius to the system. This curve is associated with the product's unique identifier. Data from each link is linked and stored using the unique identifier as an index, forming a chain-like traceability record. At the database level, this manifests as a master tracking table with the unique identifier as the primary key, and multiple detailed tables with the unique identifier combined with the link sequence number as the joint primary key, which are linked and queried through database relationships.

[0043] Extracting key quality control nodes from traceability records involves a systematic screening and decision-making process. This includes analyzing historical food safety incidents and statistically analyzing the frequency and severity of these incidents at each stage of the supply chain. This work is accomplished by querying the company's quality incident database from the past five years. The root causes of each incident are categorized and mapped to specific supply chain stages. For example, statistics show that 65% of "microbial contamination" incidents are associated with the "pre-processing" stage, and 20% with the "cold chain transportation" stage. Each incident at each stage is assigned a severity score based on its consequences. Combining industry standards and expert knowledge, stages with decisive impact on the safety of the final product and for which data is readily available are selected as candidate nodes. In some embodiments, food safety experts, production engineers, and quality managers conduct Delphi method discussions to score the entire supply chain list. Stages with scores above a threshold and whose key control parameters can be reliably acquired through existing or deployable sensors are selected, forming a candidate node list. The technical feasibility and economic cost of deploying sensing devices at each candidate node are assessed. The technical feasibility assessment includes examining the sensor's operational stability in humid, low-temperature industrial environments, data interface compatibility, and the impact of installation on existing production processes. The economic cost assessment includes sensor procurement costs, installation and commissioning costs, long-term maintenance costs, and data traffic fees. Considering the frequency, severity, decisive impact, and feasibility, the final key quality control nodes and their order in the supply chain are determined. This comprehensive decision-making process can be understood as employing a quantitative scoring model, a method for calculating node importance scores. The formula is expressed as:

[0044] Where: F represents the normalized frequency of historical events, S represents the normalized average event severity score, and I represents the normalized expert decisive influence score. , , These are the weighting coefficients assigned to the three dimensions respectively. This is the normalized technical and economic comprehensive cost assessment value. The higher the cost, the larger the denominator of this item, which affects the overall score. The smaller the bonus, the better. Calculate the bonus for all candidate nodes. After the values ​​are calculated, the nodes with the highest scores are selected and arranged in their natural flow order, thus determining the sequence of key quality control nodes. Optionally, in the technical feasibility assessment of deploying sensing devices, for the "deep-sea fishing" stage, although it has a decisive influence on the initial biotoxin content, due to the harsh marine operating environment and unstable power and network supply, the assessment considers the feasibility of deploying chemical sensors in real time to be low. Therefore, this stage may not be identified as a key quality control node requiring the deployment of real-time sensing devices, and instead, batch-based laboratory test reports will be used as traceability records.

[0045] See Figure 5 This is a radar chart of safety risks at various quality control points in frozen seafood products. It's a visualization result from the "Global Safety Profile" phase, used to compare the risk distribution characteristics of different points from multiple dimensions. Through the polygonal outline of the radar chart, it visually presents the differences in the distribution of four risk indicators—toxins, microorganisms, temperature, and ice crystals—at different quality control points, avoiding the one-sidedness of single-indicator analysis and achieving a holistic understanding of risks across all dimensions. It quickly identifies the core risks at processing points that require priority intervention; clarifies the risk focus of different points to support targeted prevention and control; and visually presents the risk profile of each point to assist in formulating a full-chain safety management strategy.

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the safety of frozen seafood, characterized in that, include: Construct a traceability record for frozen seafood that covers the entire supply chain. The traceability record integrates information on place of origin, catch or farm batch, initial biotoxin content, processing plant environmental parameters, freezing process curve, cold chain transportation trajectory, and storage temperature and humidity history. Based on food safety regulations and industry standards, key quality control nodes are extracted from the traceability records, and a dynamic change model is established for each key quality control node. The dynamic change model is used to simulate the evolution path of pollutant concentration, total number of microbial colonies, and ice crystal growth size under specific conditions. Deploy sensing devices at key quality control nodes to collect physical samples and convert them into digital signals to form a node measured data stream; The measured data stream of the node is input into the dynamic change model of the corresponding key quality control node to calculate the safety status vector of the frozen seafood at the current node and the predicted future node. The security status vectors of all key quality control nodes are aggregated to generate a global security profile, which is used to guide whether to release, conduct targeted re-inspection, or initiate a traceability recall procedure.

2. The method for detecting the safety of frozen seafood according to claim 1, characterized in that, The establishment of a dynamic change model for each key quality control node includes: Identify the core and auxiliary variables that affect the food safety of the key quality control nodes. The core variables include time, temperature, and humidity, and the auxiliary variables include the air permeability of packaging materials and the risk of cross-contamination between adjacent products. Collect historical batches of data on the decay or growth of safety indicators under different combinations of the core and auxiliary variables, and construct a training sample set. Using the training sample set, train a time series prediction network that can receive the current variable value and output the predicted values ​​of security indicators at multiple future time points; The trained time-series prediction network, its corresponding variable input range, and constraints are collectively encapsulated into the dynamic change model.

3. The method for detecting the safety of frozen seafood according to claim 2, characterized in that, The calculated safety state vectors of the frozen seafood at the current node and predicted future nodes include: The specific values ​​of the core variables and auxiliary variables at the current moment are parsed from the measured data stream of the nodes; The specific values ​​are input into the time-series prediction network to obtain a continuous safety indicator prediction curve from the current time to a preset future time. The safety indicator prediction curve is piecewise integrated to calculate the cumulative change in the safety indicator within each time period. The safety state vector is formed by combining the measured safety indicators at the current moment, the worst extreme value in the predicted curve of future safety indicators, and the time period number where the cumulative amount exceeds the threshold.

4. The method for detecting the safety of frozen seafood according to claim 3, characterized in that, The generation of the global security profile includes: Each key quality control node's safety state vector is assigned a node weight, which is dynamically set based on the degree of influence and irreversibility of the key quality control node on the final food safety. Following the supply chain process sequence, all weighted security state vectors are concatenated to form a multi-dimensional security evolution path; On the multidimensional security evolution path, mark all abnormal points where the measured or predicted values ​​of security indicators exceed the legal standards, and record the node location, the extent of exceeding the standard, and the predicted duration of the abnormal points; Based on the labeling results, a global security profile is generated, which includes a visual map of the security evolution path, a list of anomalies, and an overall risk score.

5. The method for detecting the safety of frozen seafood according to claim 1, characterized in that, The method also includes an online calibration step for the dynamically changing model: After a batch of frozen seafood is actually released or processed at the key quality control node, the batch of frozen seafood released is continuously tracked until the next key quality control node or end consumption stage, and its actual safety indicator change data is collected. The actual safety indicator change data are compared with the prediction sequence previously made by the dynamic change model to calculate the prediction deviation; If the prediction deviation continues to exceed the allowable range, model correction is initiated, and the internal parameters of the time series prediction network are fine-tuned using the latest collected actual data to update the dynamic change model.

6. The method for detecting the safety of frozen seafood according to claim 5, characterized in that, The startup model calibration includes: Extract the complete traceability records of the batch of frozen seafood that caused the prediction deviation to exceed the standard, as well as all its measured data at the relevant nodes; Analyze the patterns of the prediction deviations to distinguish whether the deviations originate from inherent model errors, unrecorded sudden interference events, or variations in the product's own characteristics. If the bias stems from inherent model error, then incremental training of the time series prediction network is performed using new data. If the deviation originates from an unrecorded sudden disturbance event, the characteristics of the sudden disturbance event are abstracted into a new auxiliary variable and added to the input of the dynamic change model. If the deviation originates from variations in the product's own characteristics, a sub-model is established under the label of the category or place of origin of the frozen seafood that caused the variation, and the sub-model is used preferentially in subsequent detection.

7. The method for detecting the safety of frozen seafood according to claim 4, characterized in that, The method also includes a decision feedback step based on a global security profile: Analyze the list of anomalies in the global security profile, and formulate tiered intervention strategies based on the node location, the extent of exceedance, and the predicted duration of the anomalies. The tiered intervention strategy includes: for a single node with a slight exceedance that does not spread, increasing the sampling frequency of subsequent batches of the single key quality control node; for multiple nodes with consecutive exceedances or a single node with a serious exceedance, triggering a traceability review of the upstream links of the single key quality control node; and for predictions that indicate the risk will spread rapidly to the terminal, immediately notifying the terminal warehouse to implement interception and isolation. The new detection data generated after implementing the tiered intervention strategy will be fed back into the system as a new node measured data stream to update the global security profile.

8. The method for detecting the safety of frozen seafood according to claim 7, characterized in that, The triggering of a retrospective review of the upstream links of the single key quality control node includes: Starting from the critical quality control point where serious exceedances occurred, trace back along the supply chain; Retrieve complete traceability records and historical safety status vectors of all upstream links for batches that exceeded the standards at the critical quality control node where serious exceedances occurred; Correlation analysis was conducted between historical operating parameters of upstream processes and the current terminal exceedance indicators to pinpoint the upstream processes and operating parameter ranges most likely to cause the problem. Generate a traceability review report, which includes the hypothesis of the source of the problem, the relevant chain of evidence, and the recommended process modification parameters.

9. The method for detecting the safety of frozen seafood according to claim 1, characterized in that, The establishment of a traceability record system for frozen seafood covering the entire supply chain includes: Each batch of frozen seafood is assigned a unique identification code, and the unique identification code is automatically read by sensing devices at each link in the supply chain; When reading the unique identifier, the static attributes and dynamic process data of each supply chain link are automatically or manually entered. The static attributes include the processing plant number, equipment number, and operator number. The dynamic process data includes the start time, end time, process temperature curve, and concentration of disinfectant used. Data from each stage is linked and stored using the unique identifier as an index, forming a chain-like traceability record.

10. A method for detecting the safety of frozen seafood according to claim 9, characterized in that, The extraction of key quality control nodes from the traceability records includes: Analyze historical food safety incidents and statistically analyze the frequency and severity of these incidents at each stage of the supply chain; By combining industry standards and expert knowledge, we selected the links that have a decisive impact on the safety of the final product and whose data are available as candidate nodes. Evaluate the technical feasibility and economic cost of deploying sensing devices at each candidate node; Based on the frequency, severity, decisive impact, and feasibility of occurrence, the final multiple key quality control nodes and their order in the supply chain are determined.

Images