Food safety risk assessment method and system based on model analysis

Abstract

The invention discloses a food safety risk assessment method and system based on model analysis, and relates to the technical field of food safety production monitoring, a production line virtual model synchronized with a physical production line in real time is constructed based on the structural parameters of a target production line, and the deviation of production line data is corrected through a model fusion algorithm, so that the risk assessment accuracy is improved. And identifying abnormal physical parameter inflection points strongly related to microbial pollution. According to the method, a production line virtual model is constructed through digital twinning, a model fusion algorithm is introduced, time-domain and frequency-domain multi-dimensional comparison and correction are performed on real-time misalignment data and historical steady-state data, instantaneous interference and systematic drifting are effectively filtered out, the corrected production line data are output, and the real-time misalignment data and the historical steady-state data are compared and corrected. And on the basis, abnormal physical parameter inflection points strongly related to microbial pollution are accurately identified, so that weak abnormal signals which are difficult to perceive by a traditional method can be captured in the initial stage of continuous production, and the early pollution risk perception capability for pathogenic bacteria is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of food safety production monitoring technology, specifically to a food safety risk assessment method and system based on model analysis. Background Technology

[0002] With the intensification of global trade and the rapid development of the food industry, food safety issues have become increasingly complex, and the risk factors involved have become more diversified. These risks not only come from the natural components of food, but may also be affected by factors such as pollutants, food additives, pesticide residues, and microbial contamination during the production process. Therefore, it is particularly important to establish a scientific risk assessment system.

[0003] In continuous food production, traditional sensor data suffers from severe inaccuracies due to transient interference and drift, making it impossible to accurately detect the early contamination risk of heat-resistant pathogens such as Salmonella. Furthermore, it is difficult to accurately locate specific hidden growth sites of pathogens in a large production line. Even after identifying the specific source of contamination, it is inconvenient to predict the diffusion path of pathogens released from the source along the production line without stopping the machine or disassembling the equipment, in order to analyze the risk of cross-contamination of different batches of downstream end products. Therefore, a food safety risk assessment method and system based on model analysis is proposed to solve the above-mentioned problems. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a food safety risk assessment method based on model analysis, comprising the following steps:

[0005] Step 1: Based on the structural parameters of the target production line, a virtual production line model that is synchronized with the physical production line in real time is constructed using digital twin technology. The model fusion algorithm is then used to compare the misaligned sensor data with historical steady-state data to correct the deviation of the production line data and identify the inflection points of abnormal physical parameters that are strongly correlated with microbial contamination.

[0006] Step 2: Map the corrected production line data onto the production line topology map constructed based on graph neural network, analyze the abnormal patterns of the propagation of abnormal physical parameter inflection points in the production line topology map, identify hotspot paths that conform to the microbial growth pattern, and locate the source of contamination in the blind spots of routine cleaning.

[0007] Step 3: Based on the identified pollution source, construct a microbial migration and diffusion model that characterizes the dynamic conditions of media flow and material exchange, and simulate the migration path and spatiotemporal distribution of pathogens in the production line.

[0008] Step 4: Apply particle swarm optimization algorithm to adaptively calibrate the motion parameters in the microbial migration and diffusion model to predict the contamination dose and probability of pathogens on downstream batches of terminal products at different time points.

[0009] Step 5: Integrate data on pollution source location, migration path, and terminal pollution probability to generate early warning information from point-of-origin risk to flow-direction diffusion, and conduct real-time assessment and intervention guidance for cross-contamination risk without shutting down or dismantling the system.

[0010] Preferably, in step 1, the process of correcting deviations in production line data includes:

[0011] Collect the physical production line parameters of the target production line, including its three-dimensional geometric structure and equipment attribute parameters;

[0012] Based on the physical production line parameters, a virtual production line model that is synchronized with the physical production line in real time is constructed using digital twin technology, thereby establishing a precise virtual mapping foundation.

[0013] By using time series alignment, Fourier transform, and Kalman filtering estimation, the system compares inaccurate sensor data with historical steady-state data to correct skewed production line data, effectively improving data reliability and consistency, and outputting corrected production line data.

[0014] Preferably, in step 1, the process of identifying inflection points of abnormal physical parameters includes:

[0015] Based on the corrected production line data, physical parameters of the production line, including temperature, humidity and flow rate, are extracted, and the moving mean and standard deviation of each physical parameter are calculated to capture the dynamic fluctuation characteristics of the parameters.

[0016] By setting dynamic judgment boundaries based on steady-state thresholds, continuous deviations of production line physical parameters can be identified, effectively filtering out transient interference.

[0017] The moment when a continuously deviating event first exceeds the dynamic judgment boundary is marked as an inflection point of an abnormal physical parameter strongly correlated with microbial contamination, thereby accurately pinpointing the starting point of the risk.

[0018] Preferably, in step 2, the analysis process of abnormal patterns includes:

[0019] The equipment of the target production line is abstracted as topological nodes, and the material and medium flow paths are abstracted as connecting edges. A production line topology graph based on graph neural network is constructed to realize the digital mapping of the physical production line.

[0020] The corrected production line data and the inflection points of abnormal physical parameters are mapped to the corresponding topological nodes in the production line topology diagram to establish a precise correlation between data and space.

[0021] By using graph neural networks to extract the spatiotemporal correlation features of abnormal patterns between topological nodes, we can analyze the abnormal patterns of the propagation of inflection points of abnormal physical parameters in the production line topology and accurately capture the pollution diffusion path. The spatiotemporal correlation features of abnormal patterns refer to the set of features extracted after modeling the abnormal propagation behavior in the production line topology through graph neural networks, which reflect the propagation law of abnormalities in both spatial and temporal dimensions. These features include spatial propagation path features, temporal evolution sequence features, node influence strength and correlation weight features, spatiotemporal coupling features of abnormal patterns, and propagation features dependent on the topology.

[0022] Preferably, in step 2, the process of locating the source of pollution within the blind spot of conventional cleaning includes:

[0023] Based on the node status of the production line topology, hotspot scores representing the intensity of anomaly propagation are calculated to accurately quantify the risk level of each node.

[0024] Based on the laws of microbial growth, the upstream propagation path of nodes with high hotspot scores is traced back and marked as hotspot paths, effectively tracing the source direction of pollution spread;

[0025] The centrality metric is calculated for all nodes on the marked hotspot paths to assess the topological importance of the nodes in the entire anomaly propagation network. The calculated centrality metric is then cross-referenced with the spatial attribute data of the nodes. The nodes with the highest centrality metric value and located in the blind zone of the regular cleaning process are identified as the actual pollution sources, thus achieving accurate spatial location of hidden pollution sources.

[0026] Preferably, in step 3, the process of constructing the microbial migration and diffusion model includes:

[0027] Based on fluid dynamics equations and the location of pollution sources, a microbial migration and diffusion model is constructed to characterize the dynamic conditions of media flow and material exchange, thereby accurately quantifying the dynamic transport process of pathogens.

[0028] The source of pollution is defined as the release point of pathogenic bacteria particles with an initial concentration, providing accurate initial conditions for dose and probability prediction;

[0029] We established a migration equation for pathogenic bacteria particles under the influence of convection and diffusion in a dynamic field, thus fully characterizing the spatiotemporal evolution of pollutants in a complex flow field.

[0030] Preferably, in step 3, the process of simulating the migration path and spatiotemporal distribution of pathogens includes:

[0031] Discretization is used to solve the migration equation, and the position and concentration of pathogenic bacteria particles are calculated iteratively to achieve accurate simulation of dynamic evolution;

[0032] Record the trajectory of each pathogenic particle over a continuous time step to form the migration path of pathogens in the production line and clearly trace the spread of contamination.

[0033] By aggregating the spatiotemporal data of all pathogenic bacteria particles, a spatiotemporal distribution map of the pollution cloud in the production line is generated, intuitively presenting the overall risk situation.

[0034] Preferably, step 4 specifically includes:

[0035] The particle swarm initialization algorithm uses position vectors to represent motion parameters in the microbial migration and diffusion model, including convection rate and diffusion coefficient. The objective function is the good fit between simulated data and limited detection data, which effectively overcomes the bottleneck of model parameter uncertainty.

[0036] Iterative updates of particle positions and adaptive calibration of motion parameters, including convection rate and diffusion coefficient, significantly improve the model's prediction accuracy.

[0037] By using a calibrated microbial migration and diffusion model, the contamination dose and probability of pathogens on downstream batches of end products at different time points are predicted. Specifically, the contamination dose output by the integrated microbial migration and diffusion model is used to calculate the contamination probability of each batch of end products, thereby achieving accurate quantitative early warning of cross-contamination risk.

[0038] Preferably, step 5 specifically includes:

[0039] By integrating the pollution source locations and migration paths, a heat map of pollution diffusion is generated, transforming the abstract pollution risk into an intuitive spatial distribution image;

[0040] By associating the contamination probability of each batch of end products with product batch information, the risk level of cross-contamination can be assessed, thus realizing the transformation from numerical prediction to actionable risk classification.

[0041] Based on the risk assessment results, an optimal intervention guidance strategy is generated that includes specific intervention points and timings under the condition of no shutdown and no dismantling, effectively balancing the needs of production continuity and risk control.

[0042] Secondly, a model-based food safety risk assessment system is provided to implement the aforementioned model-based food safety risk assessment method, comprising the following modules:

[0043] The data perception and digital twin construction module is used to establish a virtual production line model that is synchronized with the physical production line in real time, and to complete the collection, correction and feature extraction of multi-source data to output corrected production line data and inflection point markers of abnormal physical parameters.

[0044] The pollution source topology localization and graph analysis module identifies pollution propagation paths based on graph neural networks, accurately locates pollution sources in blind spots of conventional cleaning, and outputs the coordinates of pollution sources, providing a spatial starting point for migration simulation;

[0045] The microbial migration and diffusion simulation module is used to simulate the migration path and spatiotemporal distribution of pathogenic bacteria from the source of pollution under the action of media flow, so as to output the migration path and spatiotemporal concentration field of pathogenic bacteria, providing a basis for dose calculation;

[0046] The parameter calibration and risk quantification module is used to calibrate model parameters using the particle swarm optimization algorithm, quantify the contamination dose and contamination probability of the end product, output the contamination dose and contamination probability of each batch of products, and realize risk quantification.

[0047] The intervention strategy generation module is used to integrate data from the entire chain to generate a visualized risk map and a non-stop intervention guidance strategy, thus achieving a closed loop from risk perception to control.

[0048] This invention provides a model-based method and system for food safety risk assessment. It offers the following advantages:

[0049] (I) This model-based food safety risk assessment method and system constructs a virtual production line model through digital twins and introduces a model fusion algorithm to compare and correct real-time inaccurate data with historical steady-state data in multiple dimensions of time and frequency domains. This effectively filters out transient interference and systematic drift, outputs reliable corrected production line data, and accurately identifies inflection points of abnormal physical parameters that are strongly correlated with microbial contamination. This enables the capture of weak abnormal signals that are difficult to detect by traditional methods in the early stages of continuous production, significantly improving the ability to perceive early contamination risks of pathogenic bacteria.

[0050] (II) This model-based food safety risk assessment method and system maps the corrected production line data onto a production line topology map constructed based on a graph neural network. It analyzes the dynamic spatiotemporal patterns of abnormal parameters propagating in the topology network, identifies hotspot paths that conform to the laws of microbial growth, not only revealing the propagation trajectory of anomalies, but also, based on the node hotspot score and centrality metric, backtracks and locks the nodes that have both topological importance and spatial concealment (i.e., are in the cleaning blind zone) as the actual source of pollution, thus achieving accurate spatial positioning of pollution sources in the traditional detection and cleaning blind zone. Attached Figure Description

[0051] Figure 1 This is a schematic diagram illustrating the workflow of a model-based food safety risk assessment method and system according to the present invention.

[0052] Figure 2This is a schematic diagram of the process flow of a food safety risk assessment method based on model analysis according to the present invention. Detailed Implementation

[0053] 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.

[0054] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a food safety risk assessment method based on model analysis, comprising the following steps:

[0055] Step 1: Based on the structural parameters of the target production line, a virtual production line model synchronized with the physical production line in real time is constructed using digital twin technology. The model fusion algorithm is used to compare the inaccurate sensor data with historical steady-state data to correct the deviation of the production line data. The inflection points of abnormal physical parameters strongly correlated with microbial contamination are identified. The physical production line parameters of the target production line are collected, including three-dimensional geometric structure and equipment attribute parameters. Based on the physical production line parameters, a virtual production line model synchronized with the physical production line in real time is constructed using digital twin technology, thereby establishing an accurate virtual mapping foundation. Through time series alignment, Fourier transform and Kalman filter estimation, the inaccurate sensor data is compared with historical steady-state data to correct the deviation of the production line data, effectively improving the reliability and consistency of the data, and outputting the corrected production line data.

[0056] The specific work involves: collecting physical production line parameters of the target production line, including using 3D laser scanning and engineering drawing analysis to obtain the spatial coordinates, dimensions, and topological connections of each component of the production line, such as pipes, tanks, valves, and conveying devices, forming a complete 3D geometric structure database. Simultaneously, it extracts equipment attribute parameters such as material properties, surface roughness, rated working pressure, and flow rate from equipment nameplates and control system interfaces. Based on these physical production line parameters, it uses digital twin technology to construct a virtual production line model in virtual space that has geometric similarity and dynamic consistency with the physical production line. Furthermore, it establishes a real-time data connection with the sensor network and programmable logic controller of the physical production line through industrial communication protocols, ensuring that the virtual model can synchronously map the instantaneous operating state of the physical production line. Based on the real-time synchronization of the virtual production line model, it initiates a model fusion algorithm for data correction (the model fusion algorithm relies on the construction of the virtual production line model and the accumulation of steady-state data; based on the 3D geometric structure and equipment attribute parameters of the target production line, it uses digital twin technology to establish a virtual model that is synchronized with the physical production line in real time; simultaneously, it extracts and establishes a steady-state database of the production line under normal operating conditions from historical operating data, serving as a benchmark for data correction). This process extracts out-of-range sensor data from the real-time data stream, which is drifted due to environmental interference or sensor aging, manifesting as abnormal fluctuations or deviations from the normal range. Simultaneously, it retrieves historical steady-state data accumulated from long-term, verified operation of the production line under normal conditions from the historical database, representing the baseline state of the system under undisturbed conditions. Then, it performs multi-dimensional comparisons between the out-of-range sensor data and historical steady-state data in the time and frequency domains. Using time series alignment technology, it compares the real-time out-of-range data with historical steady-state data in the time domain to identify the patterns and magnitudes of data deviations, and introduces frequency domain analysis using Fourier transform methods. The signal is decomposed to distinguish between normal production cycle fluctuations and abnormal interference frequency components. After data comparison, the Kalman filter estimation algorithm is used to dynamically correct the identified deviations. Based on the noise characteristics of real-time sensor data and the system dynamics model, the system state at the next moment is recursively predicted, and the optimal estimate is made in combination with the actual observation values. This effectively filters out instantaneous interference and compensates for systematic drift. The correction process takes into account the deviation at the current moment and integrates historical data trends to ensure the continuity and smoothness of the output data. In this way, a set of calibrated and reliable production line physical parameters are output, along with data quality identification and timestamps.

[0057] In addition, the process of identifying inflection points of abnormal physical parameters includes: extracting production line physical parameters, including temperature, humidity and flow rate, based on the corrected production line data, and calculating the moving average and standard deviation of each production line physical parameter to capture the dynamic fluctuation characteristics of the parameters, setting a dynamic judgment boundary based on the steady-state threshold, identifying continuous deviation events of production line physical parameters, effectively filtering out instantaneous interference, and marking the moment when the continuous deviation event first exceeds the dynamic judgment boundary as an inflection point of abnormal physical parameters strongly related to microbial contamination, thereby accurately locking the risk starting point;

[0058] The specific work involves: Based on the corrected production line data, discretizing production line physical parameters, including temperature, humidity, flow rate, and microbial growth, from the continuous data stream at a preset sampling frequency. A sliding time window algorithm is applied to each physical parameter sequence, with the window width dynamically set according to the production cycle. Within each window, the arithmetic mean of the data within that time period is calculated as the sliding mean, and its standard deviation is also calculated to characterize the central tendency and dispersion of the parameter over a short period, transforming the production line data into a set of statistical characteristic sequences that evolve with time and operating conditions. After obtaining the sliding mean and standard deviation statistical sequences of each production line physical parameter, based on historical data accumulated during the stable operation of the production line, statistical analysis is used to determine the upper and lower limits of the normal fluctuation range of each production line physical parameter under steady-state conditions. These upper and lower limits are defined as steady-state thresholds, and using these steady-state thresholds as a benchmark, a system is constructed that adapts to different production stages and load conditions. The adaptively adjusted dynamic judgment boundary continuously compares the real-time calculated moving average with the dynamic judgment boundary. When the moving average of any production line physical parameter is continuously outside the dynamic judgment boundary for multiple consecutive sampling periods, a continuous deviation event is determined to have occurred, and the start time, duration, and deviation magnitude of the continuous deviation event are recorded. For each identified continuous deviation event, an inflection point marking program is initiated. The inflection point marking program uses the start time of the continuous deviation event as a benchmark to trace back the original data sequence of high-frequency sampling, search for the specific data point where the production line physical parameter first breaks through the dynamic judgment boundary, and determines the timestamp corresponding to the data point, i.e., the exact moment when the production line physical parameter first enters the abnormal range from the normal fluctuation range, as the inflection point of the abnormal physical parameter. At the same time, the inflection point of the abnormal physical parameter is marked as a key event node with a strong correlation to the risk of microbial contamination, and its time, location, and parameter type information are encapsulated into a structured event object.

[0059] Step 2: Map the corrected production line data onto a production line topology graph constructed based on a graph neural network. Analyze the abnormal patterns of the propagation of inflection points of abnormal physical parameters in the production line topology graph, identify hotspot paths that conform to the laws of microbial growth, and locate the source of pollution in the blind spots of conventional cleaning. Abstract the equipment of the target production line as topology nodes and the material and media flow paths as connecting edges, construct a production line topology graph based on a graph neural network, realize the digital mapping of the physical production line, and map the corrected production line data and inflection points of abnormal physical parameters to the corresponding topology nodes in the production line topology graph, establishing a precise correlation between data and space. Use the graph neural network to extract the spatiotemporal correlation features of abnormal patterns between topology nodes to analyze the abnormal patterns of the propagation of inflection points of abnormal physical parameters in the production line topology graph, and accurately capture the pollution diffusion path. Among them, the spatiotemporal correlation features of abnormal patterns refer to the set of features extracted after modeling the abnormal propagation behavior in the production line topology structure through graph neural network, reflecting the propagation law of abnormality in both space and time dimensions. These features include spatial propagation path features, temporal evolution sequence features, node influence strength and correlation weight features, spatiotemporal coupling features of abnormal patterns, and topology features. The propagation characteristics of topology-dependent features include spatial propagation path features to describe the diffusion trajectory of abnormal signals on the physical structure of the production line, revealing the sequential path of the abnormality starting from the initial source node and affecting downstream nodes along the direction of material flow, energy flow, or information flow; temporal evolution sequence features to characterize the development law of the abnormality in time dynamics, recording the precise timestamp of activation of each node on the propagation path, the change curve of signal strength, and the time delay of transmission between different nodes; node influence strength and correlation weight features to quantify the interaction strength between nodes during the abnormality propagation process, calculating the correlation weight of any two connected nodes in the abnormal event through the message passing mechanism of graph neural networks, characterizing the influence of upstream nodes on downstream nodes; spatiotemporal coupling features of abnormality patterns to integrate the coupling relationship between spatial path and time sequence, describing whether the propagation of the abnormality on a specific path has temporal periodicity, continuity, or suddenness; and topology-dependent propagation features to reflect the dependence of abnormality propagation on the production line topology, including the influence of graph structure indices such as node degree, betweenness centrality, and path length on abnormality propagation.

[0060] The specific work involves: abstracting each physical piece of equipment in the target production line, including but not limited to mixing tanks, transfer pumps, heaters, buffer containers, and packaging machines, into discrete topological nodes in a graph structure based on their functional units in the process flow. Simultaneously, the material conveying pipelines, energy transfer paths, and media flow channels connecting each physical piece of equipment are abstracted into connecting edges. The direction of the edges is determined by the inherent order of the process flow, and the weight of the edges is assigned based on physical properties such as flow rate, distance, or resistance. By defining node sets and edge sets, their connection relationships are established, constructing a production line topology graph based on a graph neural network that represents the physical connections and functional logic of the production line. After completing the production line topology... After the topology diagram is constructed, the corrected production line data, including time series of production line physical parameters such as temperature, humidity, and flow rate, is assigned to the corresponding equipment nodes in the topology diagram based on the actual location of their physical measurement points. This serves as the dynamic attribute feature of the equipment nodes. Simultaneously, identified inflection points of abnormal physical parameters are marked as specific types of events and associated with the corresponding source nodes in the production line topology diagram based on their spatial location. This establishes a correspondence from continuous data flow to discrete graph nodes, anchoring data-driven observation events within the spatial structure of the production line. After the data and events are mapped to nodes, the constructed production line topology diagram serves as the underlying structure, with each node attached... The corrected production line data and inflection points of abnormal physical parameters are used as input features. Through the message passing mechanism of a graph neural network, the information diffusion process along the connection edges between nodes is simulated. The feature information of adjacent nodes is iteratively aggregated, and the spatiotemporal correlation features that can reflect the spatial propagation and temporal evolution of anomalies are automatically learned and extracted. This allows for the identification of anomaly patterns in the propagation of inflection points of abnormal physical parameters in the production line topology. The anomaly patterns include spatial propagation paths, temporal evolution sequences, and node influence strength and correlation weights. The spatial propagation path reveals the diffusion trajectory of anomalies on the physical structure of the production line, describing the spread of the abnormal signal from the initial source node along material flow, energy flow, or signal transmission. The direction of the information flow sequentially affects the serialized path of each downstream node, consisting of a series of topological nodes activated in chronological order, clarifying the potential spatial spread channels of pollution risk; the temporal evolution sequence characterizes the development law of the anomaly in time dynamics, recording the precise timestamp of each node being activated, the change curve of signal strength, and the time delay of transmission between different nodes along the propagation path of the anomaly signal; the node influence strength and correlation weight quantify the interaction strength between nodes during the anomaly propagation process. Through the message passing mechanism of the graph neural network, the correlation weight of any two connected nodes in the anomaly event is calculated and output, representing the magnitude of the influence of upstream nodes on downstream nodes;

[0061] In addition, the process of locating the source of pollution in the blind zone of routine cleaning includes: calculating the hot spot score representing the intensity of abnormal propagation based on the node status of the production line topology map, accurately quantifying the risk level of each node, tracing back the upstream propagation path of nodes with high hot spot scores according to the microbial growth law, marking them as hot spot paths, effectively tracing the source direction of pollution spread, calculating the centrality of nodes on all marked hot spot paths, and determining the node with the highest centrality value and located in the blind zone of routine cleaning as the actual source of pollution, thus achieving accurate spatial positioning of hidden pollution sources;

[0062] The specific work involves: Calculating a comprehensive hotspot score for each node in the production line topology based on anomaly patterns analyzed by a graph neural network. This score integrates information from multiple dimensions, including the number and severity of inflection points in the node's own anomalous physical parameters, the frequency and intensity of the node sending anomalous signals to neighboring nodes during graph neural network message passing, and the degree to which the node is identified as a key hub in the spatiotemporal correlation characteristics of anomalous propagation. By weighted fusion of these dimensions, a normalized hotspot score representing the node's contribution to the intensity of anomalous propagation is generated for each node. After obtaining the hotspot scores for all nodes, path analysis is performed based on microbial growth patterns. These patterns are then translated into constraints for graph search, including a tendency for propagation between nodes with suitable temperatures, slow flow rates, and a high biofilm formation rate. Automatic identification of hotspots is also performed. Nodes with scores exceeding a preset hotspot threshold are identified. Starting from these high-scoring nodes, a path backtracking process is performed in the production line topology map, following the reverse direction of material flow (i.e., upstream). All upstream node sequences connected to these high-scoring nodes are recorded. Connection paths with consecutive high hotspot scores that conform to microbial growth patterns are marked as potential hotspot paths. Centrality measures are calculated for all nodes on the marked hotspot paths to assess their topological importance in the entire abnormal propagation network. The calculated centrality measures are cross-referenced with the spatial attribute data of the nodes to select nodes that simultaneously possess the highest centrality measure and whose physical location is confirmed by the database to be outside the coverage area of ​​the regular cleaning procedure (the regular cleaning blind zone). Nodes that simultaneously meet both the conditions of topological importance and spatial concealment are identified as actual pollution sources.

[0063] The expression for calculating the hot topic score is as follows:

[0064] ;

[0065] ;

[0066] ;

[0067] In the formula: For nodes The normalized hotspot score is used to quantify the importance of each node in the anomaly propagation network; For nodes Its own abnormal strength, This represents the number of anomalous physical parameter inflection points occurring at this node. For the first The severity of an abnormal physical parameter inflection point, i.e., the product of the magnitude and duration of deviation from the steady-state threshold, is determined by... The weighting coefficient for the inflection point of abnormal physical parameters; For all nodes within the current calculation cycle The minimum and maximum values; For nodes The set of neighboring nodes in the graph; For the node to neighboring nodes The weight of the connecting edges; To ensure that within the observation time window, the anomalous signal originates from the node propagation to nodes The average strength; For nodes The number of neighboring nodes is used to calculate the average influence strength; For nodes The betweenness centrality of a network indicates that among all shortest paths in the network, the path passing through node is the most frequent. The proportion of paths; For nodes To the node The number of shortest paths, For the nodes passed through in the path Quantity; This represents the maximum value of the centrality of the number of intermediate nodes in the current graph. These are weighting coefficients, and they satisfy... This is used to adjust the relative contribution of three dimensions (self-anomaly, signal transmission, and topology hub) to the hotspot score. The higher the value, the more popular the node is in the anomaly propagation network, and the more likely it is to be a core node or key transit point for anomaly propagation.

[0068] The expression for calculating the centrality metric is as follows:

[0069] ;

[0070] In the formula: For nodes The centrality measure; For nodes The in-degree of a node is the number of edges pointing to that node. This represents the maximum out-degree of all nodes in the graph, where out-degree is the number of edges from a node to other nodes. For nodes In a graph, a high PageRank value for a node means that there are important nodes pointing to that node, or that the node is pointed to by a few extremely important nodes. The maximum PageRank value among all nodes in the graph; This is the harmonic coefficient, used to balance the weights between standard topological centrality and source properties (in-degree and reputation); The higher the value, the stronger the overall topological importance of the node in the entire anomaly propagation network;

[0071] Step 3: Based on the identified pollution source, construct a microbial migration and diffusion model that characterizes the dynamic conditions of media flow and material exchange, and simulate the migration path and spatiotemporal distribution of pathogens in the production line.

[0072] Step 4: Apply particle swarm optimization algorithm to adaptively calibrate the motion parameters in the microbial migration and diffusion model to predict the contamination dose and probability of pathogens on downstream batches of terminal products at different time points.

[0073] Step 5: Integrate data on pollution source location, migration path, and terminal pollution probability to generate early warning information from point-of-origin risk to flow-direction diffusion, and conduct real-time assessment and intervention guidance for cross-contamination risk without shutting down or dismantling the system.

[0074] Example 2, as Figure 1 , Figure 2 As shown, based on Example 1, the present invention provides a technical solution: In step 3, the process of constructing a microbial migration and diffusion model includes: constructing a microbial migration and diffusion model that characterizes the dynamic conditions of medium flow and material exchange based on fluid dynamics equations and the location of the pollution source, thereby accurately quantifying the dynamic transport process of pathogens, defining the pollution source as the release point of pathogenic bacteria particles with an initial concentration, providing accurate initial conditions for dose and probability prediction, establishing the migration equation of pathogenic bacteria particles under the action of convection and diffusion in the dynamic field, and completely depicting the spatiotemporal evolution law of pollutants in complex flow fields;

[0075] The specific work involves: constructing the basic framework of a microbial migration and diffusion model, using the Navier-Stokes equations from computational fluid dynamics as the core governing equations to describe the flow behavior of media (liquid, gas, or semi-fluid materials) in the production line. Simultaneously, combining the spatial coordinates of the pollution source, accurately located through graph neural network analysis, this pollution source is set as the initial and boundary conditions of the basic model framework. A microbial migration and diffusion model characterizing the dynamic environment formed by the media flow field and material exchange is constructed, and the simulation computational domain is defined, ensuring its geometric boundaries are consistent with the actual physical structure of the production line. Within the established basic model framework, the pollution source is defined, abstracting the pollution source in physical space as one or more discrete release points on the computational grid. Each release point is assigned an initial pathogen concentration value. The concentration value is set based on historical pollution data, biofilm characteristics, or conservative estimates. Each release point is used as a continuous pathogenic particle source term in the microbial migration and diffusion model to inject tracer particles representing pathogenic bacteria during the simulation. After defining the pollution source, a migration equation is established to describe the migration behavior of pathogenic particles in the system. This migration equation is a convection-diffusion equation, specifically in the form of a partial differential equation. The convection term in the migration equation is driven by the flow field velocity vector obtained from the Navier-Stokes equations, characterizing the advection transport process of pathogenic particles along with the overall flow of the medium. The diffusion term is described by Fick's diffusion law, characterizing molecular diffusion caused by concentration gradients and turbulent diffusion caused by turbulence and other effects. Thus, a complete mathematical relationship between the spatial position and concentration of pathogenic particles in the dynamic field and the evolution of concentration over time is established.

[0076] The computational expression for the simulated computational domain is as follows:

[0077] ;

[0078] ;

[0079] In the formula: Represents the entire three-dimensional volume of the simulation computational domain; Spatial coordinates within the computational domain; This is the signed distance function, used to define the geometry of the computational domain. The area defines the space where the fluid exists, which is consistent with the physical structure of the production line (such as the inner wall of the pipes and the volume of the tank); The boundary of the computational domain; This is the inflow boundary, which is the inflow velocity or pressure. The outflow boundary is the outflow pressure or outflow condition. For a solid boundary, a no-slip boundary condition is adopted; This serves as the boundary of the pollution source, corresponding to the location of the identified pollution source, where the release conditions for microorganisms are set.

[0080] The calculation expression for the migration equation is as follows:

[0081] ;

[0082] ;

[0083] In the formula: The mass concentration of pathogenic bacterial particles is a function of space and time. , For time; The velocity vector of the flow field is obtained by solving the Navier-Stokes equations. Let be the diffusion coefficient tensor; Source term, used to describe the origin of pollution. Released microorganisms, For the release rate, For the Dirac function, This represents the location vector of the pollution source, i.e., the fixed location where pathogenic particles are released. The source point location; This is a local time variation term, representing the rate of change of concentration at a certain point over time; This is the convection term, representing the transport of microorganisms due to the overall flow of the fluid; For diffusion, it represents the dispersal of microorganisms from high-concentration areas to low-concentration areas due to concentration gradients (molecular diffusion) and turbulent fluctuations (turbulent diffusion);

[0084] In addition, the process of simulating the migration path and spatiotemporal distribution of pathogens includes: discretizing and solving the migration equation, iteratively calculating the position and concentration of pathogen particles, realizing accurate simulation of dynamic evolution, recording the movement trajectory of each pathogen particle at a continuous time step, forming the migration path of pathogens in the production line, clearly tracing the pollution diffusion path, aggregating the spatiotemporal data of all pathogen particles, generating a spatiotemporal distribution map of the pollution cloud in the production line, and intuitively presenting the global risk situation.

[0085] The specific work involves: using the Lagrange particle tracking method to discretize the continuous partial differential form of the migration equations on a pre-divided computational grid; directly calculating the force and motion of each tracer particle using Newton's second law; initiating an iterative calculation process; and at each discrete time step, solving the discretized algebraic equations or updating the particle motion state based on the current flow field velocity, diffusion coefficient, and source term conditions to calculate the pathogen concentration field in each grid cell of the computational domain at that time step. After completing the calculation for each time step, the task of tracking and recording the particle motion trajectory is performed. For the particle-based Lagrange method, the three-dimensional spatial coordinate sequence of each particle at continuous time steps is directly recorded, creating an independent data structure for each pathogen particle released in the simulation and continuously recording data in chronological order. The spatial position, instantaneous velocity, and current concentration of particles at each computational time point transform the continuous migration motion into a discrete but complete time-space coordinate sequence, forming each independent migration path of pathogens in the production line. When the simulation reaches the preset termination time or all particles leave the computational domain, the spatiotemporal data of all pathogen particles are aggregated and analyzed. The path information and concentration information of all particles are mapped back to the three-dimensional computational grid system according to their spatial coordinates. Through kernel density estimation, the distribution of pathogen concentration experienced by each spatial region on the production line throughout the entire simulation time range is statistically determined, thereby generating a spatiotemporal distribution map of the pollution cloud with time as the fourth dimension, clearly showing the entire dynamic evolution process of the pollution cloud from source release, flow transport, diffusion and dilution to reaching downstream end products.

[0086] The formula for calculating the pathogen concentration field is as follows:

[0087] ;

[0088] ;

[0089] ;

[0090] In the formula: In spatial location (a three-dimensional coordinate) and time Calculated pathogen concentration field; In time The total number of pathogenic particles that still exist within the computational domain; For particle indices, from 1 to ; For particles The mass carried represents the number of pathogenic bacteria; Normalized distance; For particles In time The three-dimensional spatial coordinates at that time; For calculation points With particles Location The Euclidean distance between them; For kernel functions; The smooth length of the kernel function. A higher value results in a smoother concentration field, but spatial details are lost. The smaller the value, the richer the details, but the more noisy the scene becomes; This is a normalization factor used to ensure the accuracy of concentration dimensions;

[0091] Step 4 specifically includes: initializing the particle swarm of the particle swarm optimization algorithm, whose position vectors represent the motion parameters in the microbial migration and diffusion model, including convection rate and diffusion coefficient. The fit between simulated data and limited detection data is used as the objective function to effectively overcome the bottleneck of model parameter uncertainty. The particle positions are iteratively updated, and the motion parameters, including convection rate and diffusion coefficient, are adaptively calibrated to significantly improve the model prediction accuracy. The calibrated microbial migration and diffusion model is used to predict the contamination dose and probability of pathogens on downstream batches of terminal products at different time points. The contamination probability of each batch of terminal products is calculated by integrating the contamination dose output by the microbial migration and diffusion model, so as to achieve accurate quantitative early warning of cross-contamination risk.

[0092] The specific work involves: implementing the initial configuration of the particle swarm optimization algorithm, defining the parameter space to be optimized, and combining the motion parameters in the microbial migration and diffusion model—namely, the convection rate correction factor characterizing advection transport intensity and the diffusion coefficient characterizing random dispersion intensity—to form a position vector. This is then used to randomly generate an initial particle swarm containing a preset number of individuals. Each particle's position vector is randomly assigned within a preset parameter value range, representing a set of candidate motion parameter combinations. Simultaneously, the optimization objective is established as the fit between simulated data and actual microbial detection data collected from limited points in the production line. This objective function is defined as the negative of the error between the simulated concentration value and the measured concentration value. After initialization, an iterative optimization loop is entered. In each iteration, each particle adjusts its flight speed and direction according to its own historical best position and the global best position of the entire particle swarm, following a preset velocity update formula, thus updating its position vector in the parameter space. The process enables the particle swarm to adaptively explore and develop in the parameter space, gradually gathering towards the region that optimizes the objective function. Through multiple iterations, the convection rate correction factor and diffusion coefficient of the motion parameters are calibrated, so that the output of the microbial migration and diffusion model achieves the best match with the limited measured data. After obtaining the calibrated optimal motion parameters, they are substituted into the microbial migration and diffusion model for final simulation. The calibrated microbial migration and diffusion model is run to predict the spatiotemporal distribution of pathogens in the entire production line at different time points. For each downstream terminal product, the time series of pathogen concentration is extracted during the entire product exposure period. By numerically integrating the concentration-time curve, the cumulative contamination dose received by the batch of products is calculated. Based on historical statistical patterns, the Beta-Poisson model is used to map the calculated contamination dose into a specific contamination probability value, thus completing the quantitative assessment of the contamination risk of each batch of terminal products.

[0093] The formula for calculating the cumulative pollution dose is as follows:

[0094] ;

[0095] In the formula: Indicates batch number The cumulative contamination dose received by the end product; For batch The starting time when the product begins to pass through the risk point; For batch The product passed the risk point completion time completely; In time Located in batch The concentration of pathogens at specific spatial locations through which the product flows, this data comes from the output of a calibrated microbial migration and diffusion model; In time The volumetric flow rate of the product or medium flowing through this location;

[0096] The expression for calculating the probability of pollution is as follows:

[0097] ;

[0098] In the formula: For batch The probability that the end product is contaminated; This represents the median dose that caused infection in 50% of exposed individuals; Shape parameters used to describe the variability of a host’s susceptibility to pathogens;

[0099] Step 5 specifically includes: integrating the location of pollution sources and migration paths to generate a heat map of pollution diffusion, transforming abstract pollution risks into intuitive spatial distribution images, associating the pollution probability of each batch of end products with product batch information, assessing the risk level of cross-contamination, realizing the transformation from numerical prediction to actionable risk classification, and generating optimal intervention guidance strategies with specific intervention points and timings under the condition of no shutdown and no dismantling, effectively balancing production continuity and risk control needs;

[0100] The specific work involves: fusing the location coordinates of the identified pollution source with the migration paths of all pathogenic bacteria particles simulated through a microbial migration and diffusion model; based on the spatial distribution density and residence time of the particle paths, using a kernel density estimation algorithm to calculate and render a continuous pollution diffusion heat map representing the pollution state on a virtual production line model; statically displaying the potential distribution range of pollution; and dynamically replaying the entire process of the pollution cloud spreading downstream from the source along the migration path by sliding along the time axis; after generating the heat map, the calculated quantitative pollution probability value of each batch of end products is structurally associated with the product identification information of that batch, including batch number, production time window, equipment unit passed through, and planned inventory location; and mapping the pollution probability value to high-risk, medium-risk, and medium-risk levels according to a preset risk level classification matrix. Discrete risk level labels for risk level and low risk level are used, combined with the production sequence and spatial distribution of product batches, to assess the cross-contamination range that a single contamination event may cause across batches and regions. Based on the completed risk assessment results, intervention guidance strategies are generated. The formulation of intervention guidance strategies follows the constraints of not stopping the machine and not disassembling the equipment. The heat map of contamination diffusion, the determined risk level of each batch of products, and the real-time production status are comprehensively analyzed. Through a preset decision rule base, the optimal intervention guidance strategy is calculated. The intervention guidance strategy is output in the form of structured instructions, which clearly specifies the specific equipment locations that need to be cleaned online or disinfectant injected, the recommended type of intervention operation, and the best time to start the intervention measures. The aim is to cut off the contamination path and carry out in-situ control of high-risk and medium-risk areas in the context of continuous production.

[0101] In addition, the intervention strategy for high-risk levels is to immediately implement targeted cleaning and disinfection enhancement. Intervention points include the equipment node where the contamination source is located and its adjacent downstream key nodes (such as the mixing tank outlet and the transfer pump inlet). The timing of the intervention is the time window after the current batch of products has flowed through and before the next batch enters. The specific measures are as follows: activate the online cleaning system to carry out targeted high-temperature or chemical cleaning at the contamination source node; inject food-grade disinfectant at key nodes in the contamination path and continue until the downstream concentration drops significantly; adjust the flow rate of the medium in this area, temporarily increasing the flow rate to flush away residual contaminants; and monitor in real time whether the abnormal parameters of this path return to steady state after cleaning.

[0102] The intervention strategy for medium-risk levels involves localized enhanced cleaning and process fine-tuning. Intervention points are key nodes along the contamination path (buffer containers, elbows, valves, etc.). Intervention is conducted during production breaks or when equipment is idle to avoid direct conflict with product flow. Specific measures include: implementing localized cyclic cleaning of the path without interrupting the main process; setting up temporary filtration or UV sterilization devices downstream of the contamination path; fine-tuning temperature or humidity control parameters to inhibit microbial growth; and increasing the sampling and testing frequency of the path to evaluate the intervention effect in real time.

[0103] The intervention strategy for low-risk levels is preventative monitoring and parameter optimization. The intervention points are the pollution source and its adjacent nodes. The intervention timing is the maintenance window in the regular production cycle. The specific measures are: to implement preventative online cleaning of the pollution source nodes and shorten the cycle; to optimize the medium flow rate and temperature control in the area to avoid the formation of stagnation zones; to simulate the effects of different cleaning strategies in the digital twin model and select the optimal solution; and to record the abnormal patterns of the path for subsequent risk prediction and model calibration.

[0104] For the risk of cross-contamination across batches, the intervention strategy is batch isolation and path blocking. Intervention points are key interfaces such as the intersection of contaminated and clean paths and the packaging machine inlet. Intervention timing is before and after the flow of high-risk batches. Specific measures are as follows: set up physical or fluid isolation zones downstream of the contaminated path to prevent cross-contamination; implement separate path processing or temporary storage for high-risk batches of products, and release them after evaluation; mark the batch as "to be observed" in the system and track its subsequent flow; initiate reverse flushing to clean the intersection nodes and prevent residual contamination.

[0105] For contamination sources within cleaning blind spots, the intervention strategy involves targeted intervention in the blind spots and model-guided cleaning. Intervention points are the cleaning blind spot nodes identified by the system (dead corners, bends, near non-contact sensors, etc.). Intervention is conducted during periods of low production load or planned maintenance windows. Specific measures include: using the cleaning agent type and concentration recommended by the model for targeted injection; enhancing the cleaning effect of the blind spots by adjusting flow field parameters (such as pulsed flow); verifying the cleaning coverage in the digital twin model to ensure no omissions; and regularly updating the cleaning blind spot database to optimize the cleaning strategy.

[0106] Example 3, as Figure 1 As shown, based on Examples 1-2, the present invention also provides a model-based food safety risk assessment system for implementing the above-mentioned model-based food safety risk assessment method, comprising the following modules:

[0107] The data perception and digital twin construction module is used to establish a virtual production line model that is synchronized with the physical production line in real time, and to complete the acquisition, correction and feature extraction of multi-source data to output corrected production line data and inflection point markers of abnormal physical parameters. Specifically, through 3D scanning and equipment parameter integration, a digital twin model with consistent geometry and dynamics is constructed. Using model fusion algorithms, real-time sensor data is compared with historical steady-state data to correct inaccurate data caused by interference and drift. Based on sliding window statistics and dynamic thresholds, inflection points of abnormal physical parameters are identified as early signals of microbial contamination.

[0108] The pollution source topology localization and graph analysis module identifies pollution propagation paths based on graph neural networks, accurately locates pollution sources in conventional cleaning blind spots, and outputs pollution source coordinates to provide a spatial starting point for migration simulation. Specifically, production line equipment is abstracted as nodes, and material / media flow paths are abstracted as edges to construct a production line topology graph. The corrected production line data and inflection points of abnormal physical parameters are mapped to the corresponding nodes in the graph. The spatiotemporal features of abnormal propagation are extracted using GNN, the hot spot scores and centrality measures of nodes are calculated, high hot spot paths are traced back, and combined with cleaning blind spot information, the actual location of the pollution source is determined.

[0109] The microbial migration and diffusion simulation module is used to simulate the migration path and spatiotemporal distribution of pathogenic bacteria from the source of pollution under the action of media flow, so as to output the migration path and spatiotemporal concentration field of pathogenic bacteria, providing a basis for dose calculation. Specifically, a fluid dynamics model is constructed based on the Navier-Stokes equations, the pollution source is defined as the particle release point, the convection-diffusion equation is used to describe the movement of microorganisms in the flow field, the Lagrange particle tracking method is used to iteratively solve the particle trajectory, generate the spatiotemporal distribution map of concentration, and the particle data is reconstructed into a continuous concentration field through kernel density estimation.

[0110] The parameter calibration and risk quantification module is used to calibrate model parameters using particle swarm optimization algorithm, quantify the contamination dose and contamination probability of terminal products, output the contamination dose and contamination probability of each batch of products, and realize risk quantification. Specifically, the particle swarm optimization algorithm is used to adaptively calibrate the convection rate and diffusion coefficient based on measured data. Based on the calibrated model, the concentration time series at the downstream terminal location is extracted, the cumulative contamination dose is calculated by integration, and the dose is mapped to the contamination probability based on the Beta-Poisson model.

[0111] The intervention strategy generation module integrates data from the entire chain to generate a visualized risk map and intervention guidance strategies that do not require shutdown. It integrates pollution source, migration path and probability data to generate a pollution diffusion heat map, associates pollution probability with product batch information, classifies risk levels, and outputs intervention guidance strategies under the conditions of no shutdown and no disassembly based on real-time status and rule base, thus realizing a closed loop from risk perception to control.

[0112] 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0113] 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 food safety risk assessment method based on model analysis, characterized in that, Includes the following steps: Step 1: Based on the structural parameters of the target production line, construct a virtual production line model that is synchronized with the physical production line in real time, and correct the deviation of the production line data through a model fusion algorithm to identify the inflection point of abnormal physical parameters that are strongly correlated with microbial contamination. Step 2: Map the corrected production line data onto the production line topology map constructed based on graph neural network, analyze the abnormal patterns of the propagation of abnormal physical parameters inflection points, identify hotspot paths that conform to the laws of microbial growth, and locate the source of contamination in the blind spots of routine cleaning. Step 3: Based on the identified pollution source, construct a microbial migration and diffusion model to simulate the migration path and spatiotemporal distribution of pathogens; Step 4: Apply particle swarm optimization algorithm to adaptively calibrate the microbial migration and diffusion model to predict the contamination dose and probability of pathogens on downstream batches of end products at different time points. Step 5: Integrate data on pollution source location, migration path, and terminal pollution probability to generate early warning information from site risk to flow and diffusion.

2. The food safety risk assessment method based on model analysis according to claim 1, characterized in that: In step 1, the process of correcting deviations in production line data includes: Collect the physical production line parameters of the target production line, including its three-dimensional geometric structure and equipment attribute parameters; Based on the physical production line parameters, a virtual production line model that is synchronized with the physical production line in real time is constructed using digital twin technology; By aligning time series data, performing Fourier transform and Kalman filtering estimation, the system compares inaccurate sensor data with historical steady-state data, corrects the skewed production line data, and outputs the corrected production line data.

3. The food safety risk assessment method based on model analysis according to claim 1, characterized in that: In step 1, the process of identifying inflection points of abnormal physical parameters includes: Based on the corrected production line data, the physical parameters of the production line, including temperature, humidity and flow rate, are extracted, and the moving mean and standard deviation of each physical parameter are calculated. Set dynamic judgment boundaries based on steady-state thresholds to identify continuous deviation events of production line physical parameters; The moment when the continuous deviation event first exceeds the dynamic judgment boundary is marked as the inflection point of the abnormal physical parameter strongly correlated with microbial contamination.

4. The food safety risk assessment method based on model analysis according to claim 1, characterized in that: In step 2, the analysis process of abnormal patterns includes: The equipment of the target production line is abstracted as topological nodes, and the material and medium flow paths are abstracted as connecting edges to construct a production line topology graph based on graph neural network. Map the corrected production line data and the inflection points of abnormal physical parameters to the corresponding topology nodes in the production line topology diagram; Graph neural networks are used to extract spatiotemporal correlation features of anomaly patterns between topological nodes to analyze the anomaly patterns of inflection points of abnormal physical parameters propagating in the production line topology. The spatiotemporal correlation features of anomaly patterns include spatial propagation path features, temporal evolution sequence features, node influence strength and correlation weight features, spatiotemporal coupling features of anomaly patterns, and propagation features dependent on topology.

5. The food safety risk assessment method based on model analysis according to claim 1, characterized in that: Step 2, the process of locating the source of pollution within the blind spot of conventional cleaning, includes: Based on the node status of the production line topology, calculate the hotspot score that characterizes the intensity of anomaly propagation. Based on the laws of microbial growth, the upstream propagation path of nodes with high hotspot scores is traced back and marked as hotspot paths; Centrality metrics are calculated for all nodes on the marked hotspot paths to assess their topological importance in the entire anomaly propagation network. The calculated centrality metrics are then cross-referenced with the spatial attribute data of the nodes, and the nodes with the highest centrality metrics that are located in the regular cleaning blind zone are identified as the actual sources of contamination.

6. The food safety risk assessment method based on model analysis according to claim 1, characterized in that: Step 3, the process of constructing the microbial migration and diffusion model, includes: Based on fluid dynamics equations and the location of pollution sources, a microbial migration and diffusion model is constructed to characterize the dynamic conditions of media flow and material exchange. The source of pollution is defined as the point where pathogenic bacteria particles with an initial concentration are released. Establish the migration equation of pathogenic bacterial particles under the influence of convection and diffusion in a dynamic field.

7. The food safety risk assessment method based on model analysis according to claim 6, characterized in that: Step 3, the process of simulating the migration path and spatiotemporal distribution of pathogens, includes: Discretize the migration equation and iteratively calculate the position and concentration of pathogenic bacteria particles; Record the trajectory of each pathogenic particle at a continuous time step to form the migration path of pathogens in the production line; Aggregate the spatiotemporal data of all pathogenic bacteria particles to generate a spatiotemporal distribution map of the pollution cloud in the production line.

8. The food safety risk assessment method based on model analysis according to claim 1, characterized in that: Step 4 specifically includes: The particle swarm for initializing the particle swarm optimization algorithm has a position vector that represents the motion parameters in the microbial migration and diffusion model, including convection rate and diffusion coefficient, with the good fit between simulated data and limited detection data as the objective function. Iterative updates of particle positions and adaptive calibration of motion parameters including convection rate and diffusion coefficient; Using a calibrated microbial migration and diffusion model, the contamination dose and probability of pathogens on downstream batches of end products at different time points are predicted. Specifically, the contamination dose output by the integrated microbial migration and diffusion model is used to calculate the contamination probability of each batch of end products.

9. The food safety risk assessment method based on model analysis according to claim 1, characterized in that: Step 5 specifically includes: By integrating the locations of the pollution sources and their migration paths, a thermal map of pollution diffusion is generated. By associating the contamination probability of each batch of end products with product batch information, the risk level of cross-contamination can be assessed. Based on the risk assessment results, an optimal intervention guidance strategy is generated, including specific intervention points and timings, without shutting down or dismantling the system.

10. A model-based food safety risk assessment system, used to implement the model-based food safety risk assessment method according to any one of claims 1-9, characterized in that, Includes the following modules: The data perception and digital twin construction module is used to establish a virtual production line model that is synchronized with the physical production line in real time, and to complete the collection, correction and feature extraction of multi-source data to output corrected production line data and inflection point markers of abnormal physical parameters. The pollution source topology localization and graph analysis module identifies pollution propagation paths based on graph neural networks, accurately locates pollution sources in blind spots of conventional cleaning, and outputs the coordinates of the pollution sources. The microbial migration and diffusion simulation module is used to simulate the migration path and spatiotemporal distribution of pathogenic bacteria from the source of pollution under the action of media flow, so as to output the migration path and spatiotemporal concentration field of pathogenic bacteria. The parameter calibration and risk quantification module is used to calibrate model parameters using the particle swarm optimization algorithm, quantify the contamination dose and contamination probability of the end product, and output the contamination dose and contamination probability of each batch of products. The intervention strategy generation module is used to integrate data from the entire chain to generate a visualized risk map and a non-stop intervention guidance strategy.

Images