A safety intelligent detection method for a pasta food product

Abstract

The present application relates to the technical field of food safety production intelligent monitoring, in particular to a kind of noodle food production safety intelligent detection method, comprising: integration real-time and historical production data, extract the correlation between key quality control points, build production process state diagram model;Introduce environment and real-time data of equipment as dynamic influence factor, update model parameters to form dynamic risk assessment network;Based on the network simulation potential propagation path of abnormal safety index, and utilize real-time sensor readings to verify and screen path;Identify high-risk transmission chain, quantify the influence weight of different risk sources, generate structured safety risk traceability report.The method builds a networked analysis model that can dynamically respond to changes in the production environment, enabling real-time tracking and precise positioning of risk transmission paths between production stages, effectively improving the accuracy and timeliness of risk traceability.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for food safety production, and in particular to an intelligent detection method for the safety of pasta products. Background Technology

[0002] In the industrial production of pasta products, safety monitoring and risk traceability of the production process are crucial to ensuring product quality. Existing safety detection technologies primarily rely on deploying sensors at key quality control points for independent threshold alarms, or on post-production data statistical analysis and random sampling. These methods treat each quality control link as an isolated unit, only able to identify anomalies in local parameters, failing to reveal the interrelationships and transmission mechanisms of risks between different links. This makes it difficult to pinpoint the root cause of problems, resulting in delayed risk warnings.

[0003] Current risk assessment models are mostly trained on historical data, with relatively fixed parameters and correlation rules. Dynamic factors in the pasta production environment, such as temperature and humidity fluctuations and equipment operating status, affect the stability and risk probability of each production stage in real time. Static models struggle to incorporate these real-time external influencing factors, failing to accurately assess the dynamic evolution of risks under current production conditions. This leads to discrepancies between risk assessment results and actual situations, resulting in insufficient accuracy and timeliness of early warnings. Therefore, an intelligent detection method is needed that can characterize the inherent correlations within production stages and dynamically respond to environmental changes to achieve accurate tracing and early warning of safety risks. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an intelligent detection method for the safety of pasta food production.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent detection method for the safety of pasta food production, comprising: The real-time production process data and historical batch sample data of the pasta food to be tested are obtained. The real-time production process data and historical batch sample data are standardized, cleaned and aligned to generate a production data sequence to be analyzed. Based on a multi-source information fusion framework, the temporal dependencies and correlations between key quality control points are extracted from the production data sequence to be analyzed, and a production process state diagram model is constructed. Environmental monitoring data and equipment operation logs are used as external influencing factors and introduced into the production process state diagram model to update the state transition probabilities between nodes and generate a dynamic production risk assessment network. Based on the dynamic production risk assessment network, the potential propagation paths of safety indicator anomalies under different risk source assumptions are simulated, and the authenticity of the paths is verified by combining real-time sensor readings on the production line. Based on the results of the path authenticity verification, high-risk transmission chains are identified, and the weighted impact of each risk source on the current key security indicators is calculated. By integrating the weighted impact calculation results with the risk transmission chain, a structured safety risk tracing report is generated, and the safety risk tracing report is pushed to the production quality control system.

[0006] As a further aspect of the present invention, the real-time production process data and historical batch sample data are standardized, cleaned, and aligned to generate a production data sequence to be analyzed, specifically including: For the collected real-time production process data, the fields are mapped and the units are unified according to the preset production process code and parameter name dictionary to generate a standardized set of real-time production parameters. Key events were extracted and timestamps were restored from historical batch sample data to identify and mark raw material information, process parameter records and test result items related to safety testing. The standardized set of real-time production parameters is aligned with the labeled historical batch sample data along the timeline, and logically related information units are merged according to the production batch and process sequence to construct a multi-dimensional data sequence of the production process. For outliers and missing items in the multi-dimensional data sequence of the production process, an interpolation and correction strategy based on process context is used to repair the data and ensure the continuity of the sequence. The repaired multi-dimensional data sequence of the production process is normalized by feature scale to eliminate the difference in magnitude between different parameters, and the final output is the production data sequence to be analyzed.

[0007] As a further aspect of the present invention, the step of extracting the temporal dependencies and correlations between key quality control points from the production data sequence to be analyzed based on a multi-source information fusion framework, and constructing a production process state diagram model, specifically includes: On the production data sequence to be analyzed, a constraint relationship discovery algorithm is applied to test the statistical independence of the parameter variables corresponding to each key quality control point under different sets of conditions; Record all associations that pass the significance test, and based on this, construct an initial undirected graph with quality control points as nodes and associations as edges; For each edge in the initial undirected graph, apply a production process timing constraint, that is, the observation time of the preceding process node must be earlier than the observation time of the following process node, and delete the edges that violate the process timing. For the graph structure after time-series constraint filtering, the causal direction inference mechanism is applied to determine the causal orientation of each edge, forming a directed acyclic production process state graph; The strength of the association between nodes in the directed acyclic production process state graph is quantified and assigned using conditional probability or mutual information, and used as the weight of the edge, thus completing the construction of the production process state graph model.

[0008] As a further aspect of the present invention, the step of introducing environmental monitoring data and equipment operation logs as external influencing factors into the production process state diagram model to update the state transition probabilities between nodes specifically includes: Temperature and humidity data, as well as air cleanliness data, are collected in the production workshop as environmental monitoring data, and the operating status logs and maintenance records of key production equipment are simultaneously acquired as equipment operation logs. Environmental monitoring data and equipment operation logs are mapped to new impact nodes in the production process state diagram, and potential connection relationships are established between the new impact nodes and existing quality control point nodes in the model. On the extended state diagram structure that includes newly added impact nodes, the intervention effect analysis method is used to assess the change in the transition probability of the existing quality control point node state when specific changes occur in environmental conditions or equipment state. Based on the changes in transition probabilities obtained from the analysis, adjust the weight coefficients of relevant edges in the production process state diagram model; After adjusting the weighting coefficients, the extended state diagram structure is simplified, edges with weighting coefficients below the set threshold are removed, and a dynamic production risk assessment network that integrates environmental and equipment impact information is generated.

[0009] As a further aspect of the present invention, based on the dynamic production risk assessment network, potential propagation paths of safety indicator anomalies under different risk source assumptions are simulated, and the authenticity of the paths is verified by combining real-time sensor readings on the production line, specifically including: In the dynamic production risk assessment network, starting from several typical risk source nodes, forward diffusion simulation is carried out along the state transition edge to deduce the sequence of downstream safety indicator abnormal nodes that are triggered, forming multiple potential risk transmission paths. For each potential risk transmission path, a verification node is set up that is directly or indirectly observed by real-time sensors. The verification node is a key quality control point or intermediate state node on the path. Continuously collect real-time sensor readings on the production line and extract process parameters or quality index values ​​corresponding to the set verification nodes. The process parameters or quality index values ​​obtained from the analysis are compared with the preset safety thresholds or normal fluctuation ranges of the verification nodes. If the comparison results exceed the allowable deviation, the potential path is determined to be supported by authenticity. Record the authenticity verification results of each potential risk transmission path. For paths that are not supported, mark them as low-risk paths and reduce their ranking priority in subsequent risk assessments.

[0010] As a further aspect of the present invention, the step of identifying high-risk transmission chains based on the path authenticity verification results and calculating the weighted impact of each risk source on the current key security indicators specifically includes: All risk transmission paths that have passed authenticity verification are summarized, categorized and integrated according to the risk source node, forming a set of transmission chains starting from the risk source; For each transmission chain under each risk source, calculate its path length, the average anomaly probability of nodes on the path, and the compliance score of the authenticity verification stage, and calculate the comprehensive risk confidence of each chain accordingly. The chains are sorted according to the comprehensive risk confidence level, and chains with a confidence level higher than the preset risk threshold are identified as high-risk transmission chains. For each high-risk transmission chain, extract the risk source node from the dynamic production risk assessment network and the risk contribution value of each indicator node in the current key safety indicator set; The extracted risk contribution values ​​are normalized, the relative impact weight of each risk source on each security indicator is calculated, and then the total weight impact of each risk source on the entire current set of key security indicators is obtained through weighted summation.

[0011] As a further aspect of the present invention, the integration of weighted influence calculation results with the risk transmission chain to generate a structured security risk tracing report specifically includes: The total weight of each risk source is arranged in descending order. Risk sources whose cumulative weight influence exceeds a certain proportion are selected as primary risk sources, and the rest are secondary risk sources. For each major and minor risk source, select the most representative chain from its corresponding high-risk transmission chain and describe its risk transmission logic and key nodes in structured language; The weighted impact of risk sources on the ranking results, risk source classification information, and corresponding risk transmission logic descriptions are then structured and filled in according to a predefined risk report template. The report includes key anomaly time points and related process parameters extracted from the production data sequence to be analyzed, as supporting information for risk tracing. The final result is a structured security risk tracing report that includes risk source ranking, transmission logic analysis, and source tracing evidence.

[0012] As a further aspect of the present invention, the application of the causal direction inference mechanism to determine the causal orientation of each edge and form a directed acyclic production process state diagram specifically includes: For each edge in the undirected graph after time-series constraint filtering, check the order in which the two nodes it connects first exhibit parameter abnormalities or state deviations in the production data sequence to be analyzed. Set the node that exhibits the abnormality first in time as the cause and the node that exhibits the abnormality later as the effect. If the chronological order cannot be clearly determined, then a conditional independence test is introduced for the two nodes under a broader set of conditional variables, and inference is made by utilizing the asymmetric nature of causal direction in conditional independence. If the direction still cannot be determined, query the inherent influence relationship of the parameters of the two connected quality control points in the pasta production process knowledge base, and specify the causal direction based on the process knowledge; After all edges are assigned a direction, the entire graph structure does not form a directed cycle. If a cycle path appears, the edge with the lowest weight in the cycle path is cut to break the cycle. After determining the direction, the directed graph is verified for consistency with the process logic to ensure that it does not conflict with the standard process flow and quality control principles for the production of pasta products.

[0013] As a further aspect of the present invention, the continuous acquisition of real-time sensor readings on the production line, and the parsing of process parameters or quality index values ​​corresponding to the set verification nodes, specifically includes: Based on the quality control point or intermediate state corresponding to the verification node, determine the type of sensor to be collected, including temperature sensor, humidity sensor, weight sensor, and image acquisition device. Through the industrial data acquisition interface, raw sensor readings on the production line are continuously acquired at a set sampling frequency; The acquired raw sensor readings are preprocessed, including outlier removal, smoothing filtering, and data alignment, to obtain a reliable sensor data sequence. From the preprocessed sensor data sequence, based on the characteristics of the verification node, feature parameters that can characterize its state are extracted, including the heating rate and steady-state value extracted from the temperature sensor data, and color and morphological features extracted from the image data. The extracted process parameters or quality index values ​​are standardized and converted to ensure consistency with the preset safety thresholds of the verification nodes.

[0014] As a further aspect of the present invention, the normalization of the extracted risk contribution values ​​and the calculation of the relative impact weight of each risk source on each safety indicator specifically include: For each high-risk transmission chain that is selected, obtain the original risk contribution value from its risk source node to each indicator node in the current set of key security indicators; Calculate the sum of the original risk contribution values ​​from all risk source nodes to the same security indicator node, and divide the original risk contribution values ​​from the risk source node to the security indicator node by the sum to obtain the normalized relative influence weight of the risk source on the security indicator. Iterate through each indicator node in the current set of key security indicators, repeat the calculation process, and obtain a set of relative influence weights of the risk source node on all security indicator nodes. The relative influence weights of a risk source node on all security indicator nodes are weighted and summed, with the weights being the importance coefficients of the security indicator in the current detection. Finally, the total weight influence of the risk source on the entire current set of key security indicators is obtained. The total weighted impact of all risk sources is normalized again to ensure that the sum is 100%, thus completing the weighted impact calculation.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By extracting the temporal dependencies and correlations between key quality control points from multi-source time-series data, a production process state diagram model is constructed. This model represents independent quality control points as nodes in the graph and their interrelationships as edges, thus forming a structured network. When conducting risk analysis based on this network, the upstream and downstream links that may be associated with an abnormal state of a node can be identified according to the connection relationships of the edges, transforming discrete alarm signals into directional risk transmission hypothesis chains. This allows the diagnosis of production anomalies to move beyond the point of occurrence itself, enabling the tracing of their possible sources and diffusion directions within the production process network, achieving a shift from outcome localization to process tracing.

[0016] Real-time environmental monitoring data and equipment operation logs are used as external influencing factors, dynamically applied to the production process state diagram model to update the state transition probabilities between nodes. This operation transforms the model from a fixed structure into a dynamically evolving risk assessment network. The continuous input of external factors causes the risk propagation probability distribution within the network to change in real time, ensuring that the assessment results for the same risk assumption reflect the specific environment and equipment status of the current production line. The assessment results no longer rely solely on static historical patterns but incorporate the specific conditions of the real-time production site, directly linking risk warnings to the actual vulnerability of the current production state. The credibility and priority of risk transmission paths can be dynamically adjusted and filtered according to the actual situation on site. Attached Figure Description

[0017] Figure 1This is a flowchart of the intelligent detection method for the safety of pasta food production according to the present invention; Figure 2 A flowchart for constructing a production process state diagram model; Figure 3 The graph shows the effect of temperature fluctuations in the proofing room on the proofing expansion rate of the dough. Figure 4 Radar chart for assessing the multi-dimensional risk characteristics of different risk sources; Figure 5 Line chart showing the weighted impact of risk sources. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 The system acquires real-time production process data and historical batch sample data of the pasta product to be tested. It then performs standardized cleaning and alignment operations on the real-time production process data and historical batch sample data to generate a production data sequence for analysis. Based on a multi-source information fusion framework, it extracts the temporal dependencies and correlations between key quality control points from the production data sequence to construct a production process state diagram model. Environmental monitoring data and equipment operation logs are introduced as external influencing factors into the production process state diagram model to update the state transition probabilities between nodes, generating a dynamic production risk assessment network. Based on the dynamic production risk assessment network, it simulates potential propagation paths of safety indicator anomalies under different risk source assumptions and verifies the authenticity of these paths by combining real-time sensor readings on the production line. Based on the path authenticity verification results, it identifies high-risk transmission chains and calculates the weighted impact of each risk source on the current key safety indicators. It integrates the weighted impact calculation results with the risk transmission chains to generate a structured safety risk tracing report, which is then pushed to the production quality control system.

[0021] In one embodiment of the present invention, taking the production of a certain batch of steamed buns as an example, the real-time production process data originates from multiple sensors and programmable logic controllers on the production line. The data includes multiple processes such as flour input weight, kneading water temperature, proofing room temperature, and steamer pressure. The field names of these data vary in different devices, and the units also differ. In specific implementation, based on a preset production process code and parameter name dictionary, the "inlet water temperature" field is mapped to a unified "kneading water temperature," all weight units are converted to "kilograms," and the temperature unit is unified to "degrees Celsius," thereby generating a standardized set of real-time production parameters. It can be understood that historical batch sample data is usually stored in a database, containing relevant records of past successful and abnormal batches. In one embodiment, key events are extracted and timestamps are repaired from the historical batch sample data to identify and mark raw material information, process parameter records, and test result items related to safety testing, such as "high-activity dry yeast input," "dough clumping detection pH value exceeding the standard," and "finished steamed bun center temperature not meeting the standard." Optionally, timestamp repair addresses situations where device clocks are out of sync or records are missing. It calculates and calibrates based on the standard time consumption between processes to ensure that the timestamp of each event is accurate.

[0022] In practical implementation, the standardized set of real-time production parameters is aligned with the labeled historical batch sample data along the timeline, and logically related information units are merged according to production batch and process sequence. This step constructs a multi-dimensional data sequence of the production process. An information unit may contain parameters such as environmental humidity, dough expansion volume, and elapsed time at a certain point in time for the "secondary proofing" process. These parameters come from the real-time data stream and records of corresponding time points in the historical samples. In one embodiment, outliers and missing items in the multi-dimensional data sequence of the production process are repaired using an interpolation and correction strategy based on process context relevance. Specifically, when measuring process context relevance, the spatiotemporal correlation of multiple parameters belonging to the same production batch and the same or adjacent processes in the production data sequence to be analyzed can be utilized. One implementation method is to calculate the covariance matrix of other process parameters closely related to the outlier or missing parameter in the process flow within the same batch of time-series data. Based on this covariance matrix, a multivariate Gaussian distribution model or a multivariate regression model can be used to estimate the reasonable values ​​of the outlier or missing parameters using the normal values ​​of the relevant parameters. Another model is a deep learning approach based on autoencoders. This method trains the model using multi-process sequence data from historical normal batches, learning the potential dependency patterns between parameters of each process. When anomalies or missing data appear in the sequence, the sequence segment containing the anomaly is input into the trained autoencoder, and its reconstructed output serves as a repair value. This method fully utilizes the complete process context information in the sequence for estimation, rather than relying solely on the historical mean of a single parameter. For example, when the instantaneous value of the "pressing roller gap" parameter far exceeds the mechanical limit, it is identified as an outlier, and the normal value from a nearby time point is used as a substitute, based on the normal pressure parameters of adjacent processes and the equipment operating status. For continuous missing data in the "steam pressure inside the steamer" due to a brief sensor malfunction, interpolation is performed based on the fixed steaming pressure curve required for that batch of steamed buns, combined with normal pressure data from other parallel production lines at the same time point, thus ensuring the continuity of the sequence.

[0023] In practical implementation, the repaired multi-dimensional data sequence of the production process is normalized by feature scale to eliminate the magnitude differences between different parameters. For example, the numerical range of flour input weight is on the order of hundreds of kilograms, while the numerical range of water temperature for kneading is on the order of tens of degrees Celsius; direct comparison is meaningless. An optional implementation method is to use a max-min normalization method to scale each parameter to the [0,1] interval. The normalization process is performed using the following formula:

[0024] Where: X represents the original value of a certain parameter. and These represent the minimum and maximum values ​​of the parameter in the multi-dimensional data sequence of the production process, respectively. This represents the normalized numerical value. In this way, parameters with different dimensions and ranges, such as feed weight, water temperature, pressure, and time, are transformed to the same scale, and the final output is a production data sequence to be analyzed, which can then be used to build a model.

[0025] See Figure 2 In one embodiment of the present invention, a constraint relationship discovery algorithm is applied to the production data sequence to be analyzed to test the statistical independence of the parameter variables corresponding to each key quality control point under different condition sets. It can be understood that key quality control points include multiple parameter variables such as flour protein content, final kneading temperature, dough proofing expansion rate, and specific volume of steamed buns. In one embodiment, the constraint relationship discovery algorithm uses conditional independence testing. For example, given the "amount of water added during kneading" and "kneading time," it tests whether the "final kneading temperature" and "dough proofing expansion rate" are independent. If the p-value of the test is less than a preset significance level, the independence hypothesis is rejected, and it is considered that there is a correlation between the two. All correlations that pass the significance test are recorded, and based on this, an initial undirected relation graph is constructed with quality control points as nodes and correlations as edges. For example, if the test results show that "flour protein content" and "dough proofing expansion rate" are correlated, then an undirected edge is established between the nodes representing these two parameters in the initial undirected relation graph.

[0026] For each edge in the initial undirected graph, a production process timing constraint is applied: the observation time of the preceding process node must be earlier than the observation time of the following process node. Edges that violate the timing constraint are deleted. In one embodiment, the initial undirected graph may contain an edge connecting the "specific volume of steamed buns after steaming" node and the "final temperature of dough kneading" node. However, according to the standard process flow, the dough kneading process precedes the steaming process. Therefore, the observation time of the "final temperature of dough kneading" must be earlier than the observation time of the "specific volume of steamed buns after steaming." It can be understood that if the data shows that the abnormal time point of the "specific volume of steamed buns after steaming" parameter is earlier than the abnormal time point of the "final temperature of dough kneading" parameter, this edge violates the timing constraint and needs to be deleted. After timing constraint filtering, the remaining edges connect node pairs that satisfy the basic logic that the time of the preceding process node is earlier than the time of the following process node.

[0027] For the graph structure filtered by time constraints, a causal direction inference mechanism is applied to determine the causal orientation of each edge. For each edge in the undirected graph after time constraint filtering, the order in which the two connected nodes first exhibit parameter anomalies or state deviations in the production data sequence to be analyzed is checked. The node that becomes abnormal earlier in time is designated as the cause, and the node that becomes abnormal later is designated as the effect. For example, for the edge connecting "dough proofing expansion rate" and "specific volume of steamed buns after steaming", in the historical abnormal batch data, if the time point when "dough proofing expansion rate" falls below the threshold is always earlier than the time point when "specific volume of steamed buns after steaming" falls below the threshold, then the directed edge pointing the "dough proofing expansion rate" node to the "specific volume of steamed buns after steaming" node is set. If the time order cannot be clearly determined, a conditional independence test of the two nodes under a broader set of conditional variables is introduced, and inference is made using the asymmetric characteristics of causal direction in conditional independence. Optionally, one inference method is to infer that if "flour protein content" and "dough proofing expansion rate" are independent given a "dough kneading end temperature," but not independent given a "dough proofing expansion rate," then it is likely that "flour protein content" is the cause of "dough kneading end temperature," and "dough kneading end temperature" is the cause of "dough proofing expansion rate." If the direction still cannot be determined, the inherent influence relationship between the two connected quality control point parameters is queried from the pasta production process knowledge base, and the causal direction is specified based on the process knowledge. For example, if the process knowledge clearly indicates that "relative humidity in the proofing room" directly affects "dough surface hardness," then the edge direction is specified as pointing from the "relative humidity in the proofing room" node to the "dough surface hardness" node. The pasta production process knowledge base is constructed by using key quality control point parameters of pasta production as nodes and the inherent influence relationship between them as directed edges to build a structured relationship network. Its data sources include: industry standards and specifications, enterprise standard operating procedures, and empirical knowledge verified by food process experts. The knowledge base update mechanism adopts a combination of periodic review and event triggering. For example, when production equipment or raw material formulas change, or when new correlation patterns are discovered based on real-time production data, quality engineers and domain experts jointly evaluate and revise the knowledge base.

[0028] In practice, the causal direction inference mechanism includes the following steps to determine the direction: For an undirected graph that has undergone time-series constraint filtering, an edge connects two nodes. and Query the timestamp of the first occurrence of an anomaly in the parameter of each node in the production data sequence to be analyzed. and .like Significantly earlier than Then set the direction of the edge to ;like Significantly earlier than Then the direction is set to Significance threshold It can be set according to the minimum time interval of the process (e.g.) ), that is, when Only then is the chronological order considered clear. Conditional independence asymmetry determination: if the chronological order cannot be clearly determined (i.e.... If the condition is true, proceed to this step. Perform a conditional independence test over a broader set of conditional variables. The hypothesis test finds that there exists a condition set S such that... and Independent when given S, but independent when given They are not independent; at the same time, there is no set of conditions. , making and In the given Independent at times, but given The causal relationship is not independent. This asymmetry can be used as a basis for inferring the direction of causality, tending to determine the direction as... This step can be achieved using the direction rules in classic causal discovery algorithms such as the PC algorithm or FCI algorithm. If the direction cannot be determined in the above two steps, then the knowledge base of pasta production technology is queried. The knowledge base predefines the inherent process influence relationships between parameters, such as "the relative humidity in the proofing room directly affects the hardness of the dough surface". Based on this process knowledge, undirected edges are directly determined as directed edges.

[0029] After ensuring all edges are assigned a direction, the entire graph structure does not form directed cycles. If a cycle path exists, the edge with the lowest weight in the cycle path is cut to break the cycle. Optionally, edge weights can be estimated in advance using mutual information. The directed graph with determined directions undergoes process logic consistency verification to ensure it does not conflict with standard processes and quality control principles for pasta production. For example, logic that violates physical timing, such as "steaming pressure" affecting "water addition for dough mixing," should not exist. After completing direction determination and cycle breaking, a directed acyclic production process state graph is formed. The correlation strength between nodes in the directed acyclic production process state graph is quantified and assigned using conditional probability or mutual information, and used as edge weights. One quantification method is to calculate the conditional probability that the result node will experience an anomaly under specific conditions where the cause node is in a particular state.

[0030] In one embodiment of the present invention, temperature and humidity data and air cleanliness data in the production workshop are collected as environmental monitoring data, and the operation status logs and maintenance records of key production equipment are simultaneously acquired as equipment operation logs. It can be understood that the environmental monitoring data includes temperature and humidity sensor readings in the proofing room, as well as records of airborne particulate matter concentration in the dough kneading area. The equipment operation logs may contain information such as motor current fluctuations in the dough mixer, bearing temperature of the mixing paddle, and tripping records of the steamer's safety valve. In a specific implementation, the environmental monitoring data and equipment operation logs are mapped to new influencing nodes in the production process state diagram. For example, based on the production process state diagram model, new nodes are added: "Proofing Room Temperature Fluctuation," "Dough Kneading Area Air Cleanliness," and "Dough Mixer Bearing Overheating." Potential connections are established between the new influencing nodes and existing quality control point nodes in the model. These connections are based on process knowledge or historical correlation analysis; for example, undirected edges are established between the "Proofing Room Temperature Fluctuation" node and the "Dough Proofing Expansion Rate" node, and between the "Dough Mixer Bearing Overheating" node and the "Dough Kneading End-Point Temperature" node.

[0031] On an expanded state graph structure incorporating newly added influencing nodes, intervention effect analysis is used to assess the change in transition probability of existing quality control point node states when specific changes occur in environmental conditions or equipment states. In some embodiments, intervention effect analysis is achieved by comparing the conditional probability difference of the "dough proofing expansion rate" node occurring abnormally when the "proofing room temperature fluctuation" node is in a normal state versus an abnormal state. Assuming the probability of the "dough proofing expansion rate" node being abnormal is 0.05 when the "proofing room temperature fluctuation" node is normal, and the probability of abnormality increases to 0.25 when the "proofing room temperature fluctuation" node is abnormal, then the change in transition probability is 0.20. Based on the change in transition probability obtained from the analysis, the weight coefficients of relevant edges in the production process state graph model are adjusted. The determination of the weight coefficients is mainly based on historical data training and incorporates intervention effect analysis of real-time external influencing factors. Specifically, the initial weight coefficients are derived from the production process state graph model constructed based on historical batch sample data and quantified using methods such as conditional probability or mutual information. After introducing real-time environmental and equipment data, the change in transition probability calculated using the intervention effect analysis method dynamically adjusts the initial weight coefficients. Therefore, the weighting coefficients are neither completely fixed nor purely dependent on expert experience, but rather based on historical data and dynamically updated according to real-time external influences. It can be understood that the edge weight connecting the "temperature fluctuation in the proofing room" node and the "dough proofing expansion rate" node will be strengthened due to changes in the transition probability, while the weight coefficients of some intrinsic nodes with no obvious statistical correlation to changes in external factors may be weakened. After adjusting the weighting coefficients, the extended state diagram structure is simplified, and edges with weighting coefficients below a set threshold are removed, generating a dynamic production risk assessment network that integrates environmental and equipment influence information. The specific value of the set threshold can be determined by analyzing the distribution of weighting coefficients of each edge in the dynamic production risk assessment network. For example, if the calculated weight of the edge connecting the "air cleanliness in the dough mixing area" node and the "total bacterial count in the finished steamed buns" node is below 0.01, it is removed from the network.

[0032] In practice, the intervention effect analysis method is used to quantify the changes in the state of quality control point nodes caused by external influencing factors. The specific implementation logic is as follows: A change in the state of a newly added node (denoted as E) mapped in environmental monitoring or equipment operation logs is defined as an "intervention". For example, an intervention... This indicates that the state of node E will be fixed as an abnormal state. Intervention This indicates that the system is now in a normal state. Calculating the intervention effect: On the extended state graph structure (i.e., the causal graph) containing the new node E, using causal inference rules such as the backdoor criterion or the frontdoor criterion, determine how the conditional probability of the target quality control point node Q changes with the intervention, given other observed variables Z. The intervention effect IE can be calculated by comparing the differences in conditional probabilities under different interventions:

[0033] in, Representatives intervened Given the observed Z, the conditional probability of Q is calculated. This probability can be estimated from historical batch sample data after adjusting for conditional independence, or calculated from the parameters (edge ​​weights) of the constructed production process state diagram model. The absolute value of the calculated intervention effect IE is the change in the transition probability of node Q caused by the state change of node E. .Right now Based on the calculated change in transition probability Adjust the weight coefficients of the edges connecting node E and node Q in the extended state graph. One adjustment method is:

[0034] in: Represents the learning rate parameter (e.g.) This is used to control the adjustment range of external influencing factors on the strength of historical correlations.

[0035] In a dynamic production risk assessment network, starting from several typical risk source nodes, a forward diffusion simulation is performed along the state transition edges to deduce the sequence of downstream safety indicator anomalies, forming multiple potential risk transmission paths. In some embodiments, typical risk source nodes can be "excessive moisture content in raw flour," "sudden drop in temperature in the proofing room," or "overheating of the dough mixer bearings." Starting with "sudden drop in temperature in the proofing room," the simulation may sequentially lead to "insufficient dough proofing expansion rate," "low specific volume of steamed buns," and "excessive hardness." For each potential risk transmission path, a verification node that can be directly or indirectly observed by real-time sensors is set. The verification node is usually a key quality control point or intermediate state node on the path. For example, for the above path, "dough proofing expansion rate" can be selected as the verification node. Real-time sensor readings on the production line are continuously collected, and the process parameters or quality indicator values ​​corresponding to the set verification nodes are extracted. Based on the quality control point or intermediate state corresponding to the verification node, the types of sensors to be collected are determined, including temperature sensors, humidity sensors, weight sensors, and image acquisition devices. For the verification node of "dough proofing expansion rate," the optional sensor combination is an image acquisition device installed in the proofing chamber to monitor dough volume changes. Raw sensor readings from the production line are continuously acquired via an industrial data acquisition interface at a set sampling frequency. The acquired raw sensor readings undergo signal preprocessing, including outlier removal, smoothing filtering, and data alignment, to obtain a reliable sensor data sequence. From the preprocessed sensor data sequence, feature parameters characterizing the verification node's state are extracted based on its characteristics. From the dough contour time-series image acquired by the image sensor, the rate of change of the dough's projected area can be extracted using image processing algorithms as a proxy feature for the expansion rate. The heating rate and steady-state value are extracted from the temperature sensor data. Color and morphological features are extracted from the image data. The extracted process parameters or quality index values ​​are standardized to ensure consistency with the verification node's preset safety thresholds. Standardization typically maps the extracted feature values ​​to the same dimensions and range as the production data sequence being analyzed. The parsed process parameters or quality index values ​​are then compared for compliance with the verification node's preset safety thresholds or normal fluctuation range. For example, the calculated "dough proofing expansion rate" for the current batch is 2.5, while its preset lower limit is 2.8. If the comparison result exceeds the allowable deviation, the potential path is deemed to be supported by the evidence. The verification result of each potential risk transmission path is recorded. For paths that are not supported, they are marked as low-risk paths, and their ranking priority is reduced in subsequent risk assessments.

[0036] See Figure 3This is a graph showing the impact of temperature fluctuations in the proofing room on the dough's proofing expansion rate. During the "proofing stage," when the temperature suddenly drops from the normal 38℃ to an abnormal 32℃, the dough's proofing expansion rate plummets from 2.8 to near zero, directly verifying that temperature is a key factor affecting proofing performance. This visualized causal chain provides direct data support for the subsequent construction of a dynamic production risk assessment network. The strong correlation between temperature and expansion rate is directly used to adjust the edge weights between the "proofing room temperature fluctuations" and "dough proofing expansion rate" nodes in the production process state diagram, strengthening this critical path. Simultaneously, it provides a reference for eliminating other weak edges with no clear correlation to external factors, helping to generate a more accurate and efficient dynamic production risk assessment network.

[0037] In one embodiment of the present invention, all risk transmission paths that pass the authenticity verification are summarized and categorized and integrated according to the risk source node to form a set of transmission chains starting from the risk source. For example, the verified paths may include: path A1 from the source of "excessive moisture in raw flour" to "excessive water activity in finished steamed buns"; path A2 from the same source to "excessive hardness in finished steamed buns"; and path B1 from the source of "sudden temperature drop in the proofing room" to "low specific volume in finished steamed buns". For each transmission chain under the risk source, its path length, the average anomaly probability of the nodes on the path, and the compliance score in the authenticity verification stage are calculated, and the comprehensive risk confidence of each chain is calculated accordingly. The path length refers to the number of edges traversed from the source node to the end safety indicator node. The average anomaly probability of the nodes on the path can be obtained from the historical statistics or status assessment of each node in the dynamic production risk assessment network. The compliance score in the authenticity verification stage can refer to the calculation result of the path authenticity support S. Understandably, one way to calculate the overall risk confidence level C is to assign different weights to the three factors mentioned above and then perform a weighted sum, for example:

[0038] Where: C represents the overall risk confidence level of the chain, L represents the path length, and the function f(L) is a function that maps the path length to the contribution value to the confidence level. S represents the average anomaly probability of nodes on the path, and S represents the support score for path authenticity verification. It is a pre-set weighting coefficient, and satisfies The chains are sorted based on their overall risk confidence levels. Chains with confidence levels higher than a preset risk threshold are identified as high-risk transmission chains. The preset risk threshold can be determined based on the statistical distribution of overall risk confidence levels corresponding to confirmed safety accident batches in historical batch sample data. Assuming the preset risk threshold is 0.7, the overall risk confidence level of path A1 is 0.85, the overall risk confidence level of path B1 is 0.75, and the overall risk confidence level of path A2 is 0.65, then paths A1 and B1 are identified as high-risk transmission chains.

[0039] For each high-risk transmission chain, the risk contribution value of each indicator node in the current key safety indicator set is extracted from the dynamic production risk assessment network, moving from the risk source node to the target indicator node. The risk contribution value can be quantified as an aggregation of the edge weights on all directed paths from the source node to the target indicator node, for example, by multiplying the edge weights on all paths and then summing them. The current key safety indicator set may include "water activity of finished steamed buns," "total bacterial count of finished steamed buns," "hardness of finished steamed buns," and "specific volume of finished steamed buns," etc. Refer to Table 1 for the extracted risk contribution values.

[0040] Table 1: Examples of Risk Contribution Values ​​of Risk Sources to Key Safety Indicators

[0041] In practical implementation, the extracted risk contribution values ​​are normalized, and the relative impact weight of each risk source on each safety indicator is calculated. For each selected high-risk transmission chain, the original risk contribution value from its risk source node to each indicator node in the current key safety indicator set is obtained. The sum of the original risk contribution values ​​from all risk source nodes to the same safety indicator node is calculated, and the sum is divided to obtain the normalized relative impact weight of the risk source on the safety indicator. The calculation process is repeated for each indicator node in the current key safety indicator set to obtain a set of relative impact weights of a risk source node on all safety indicator nodes. The relative impact weights of a risk source node on all safety indicator nodes are weighted and summed, with the weight being the importance coefficient of the safety indicator in the current detection, finally obtaining the total weight impact of the risk source on the entire current key safety indicator set. In some embodiments, the importance coefficient can be preset based on food safety standards or customer complaint rates; for example, the importance coefficient of "total bacterial count in finished steamed buns" is the highest. The calculation formula can be expressed as:

[0042] in: This represents the total weighted impact of the k-th risk source. This represents the relative weight of the k-th risk source on the j-th safety indicator. Let represent the importance coefficient of the j-th safety indicator, and m represent the total number of safety indicators. Assuming the importance coefficients are set as {water activity: 0.2, total bacterial count: 0.4, hardness: 0.2, specific volume: 0.2}, then the total weight of "excessive moisture content in raw flour" affects... The total weighted impact of all risk sources is then normalized again to ensure that the sum is 100%, thus completing the weighted impact calculation. For example, if the calculation shows that "the moisture content of the raw flour exceeds the standard", the weighted impact is calculated as follows: The value was 0.718, indicating a sudden drop in temperature in the proofing room. The value is 0.282, and the sum of the two is 1.0, so the final normalized total weights have an impact of 71.8% and 28.2%, respectively.

[0043] See Figure 4 This is a radar chart assessing the multi-dimensional risk characteristics of different risk sources. By comparing the scores of each risk source across different dimensions, production managers can clearly identify risks that require priority control. For example, "insufficient steaming time" and "excessive moisture content in raw flour" should be considered primary control priorities, while "high humidity in the cooling room" can be considered a secondary control target. The scores of each risk source in terms of "severity" and "scope of impact" directly correspond to the weight coefficients of each node in the dynamic production risk assessment network. The higher the score, the greater the influence of the risk source in the network, and the weight of its connected edges should be correspondingly increased. By analyzing the detection difficulty of each risk source, optimization suggestions can be provided for the monitoring frequency and data collection strategies of relevant nodes in the network, improving the efficiency and accuracy of the entire risk assessment network.

[0044] In one embodiment of the present invention, the total weight influence of each risk source is arranged in descending order. Risk sources whose cumulative weight influence exceeds a certain percentage are selected as primary risk sources, and the rest are designated as secondary risk sources. The certain percentage is set to a fixed value between 70% and 90%, for example, 80%. In operation, the total weight influence of the risk sources is sorted from high to low and accumulated. When the accumulated weight reaches or exceeds the set percentage for the first time, the accumulation stops, and all risk sources participating in the accumulation are identified as primary risk sources. For example, according to the calculation results of the embodiment, the total weight influence of "excessive moisture in raw flour" is 71.8%, and the total weight influence of "sudden drop in temperature in the proofing room" is 28.2%. If the cumulative percentage threshold is set to 80%, then the single weight influence of "excessive moisture in raw flour" has exceeded the threshold, or the sources with the highest cumulative weight are selected as primary risk sources. In some embodiments, the cumulative weight influence exceeding a certain percentage can be defined as: starting from the risk source with the highest weight, the accumulation continues until the total weight reaches a preset percentage, and these risk sources participating in the accumulation are identified as primary risk sources. It is understandable that in the example described, "excessive moisture content in the raw flour" is identified as the primary risk source, while "sudden drop in temperature in the proofing room" is identified as a secondary risk source.

[0045] For each primary and secondary risk source, the most representative chain is selected from its corresponding high-risk transmission chain, and its risk transmission logic and key nodes are described in structured language. In practical implementation, the most representative chain can be the one with the highest overall risk confidence, or the shortest and most direct chain. For the primary risk source "excessive moisture content in raw flour," the corresponding high-risk transmission chain might be: excessive moisture content in raw flour → dough that is too soft after kneading → insufficient support for the dough after shaping → dough collapse after proofing → smaller specific volume of the finished steamed bun. The structured description is as follows: "The risk source, 'excessive moisture content in the raw flour,' alters the flour's water absorption characteristics, causing the 'dough hardness after kneading' parameter to fall below the standard lower limit; the overly soft dough struggles to maintain its shape during the 'shaping' process, resulting in an abnormal 'dough height after shaping' parameter; insufficient support causes the dough to collapse during 'proofing,' manifesting as a 'volume expansion rate after proofing' failing to meet standards; ultimately, this leads to the critical safety indicator 'specific volume of the finished steamed bun' falling below specifications." The secondary risk source, "sudden temperature drop in the proofing room," is described using a similar format. Optionally, the description can explicitly identify the name of each key node in the chain and its abnormal state.

[0046] The risk source weighting ranking, risk source classification information, and corresponding risk transmission logic descriptions are structured and filled in according to a predefined risk report template. In practice, the predefined risk report template may include the following fixed sections: Report Summary, Risk Source Overview, In-depth Analysis of Major Risk Sources, Summary of Secondary Risk Sources, and Supporting Information for Risk Tracing. In the "Risk Source Overview" section, all identified risk sources, their total weighting percentage, and risk level classification are listed in tabular or list format. In the "In-depth Analysis of Major Risk Sources" section, a detailed logical description of the selected representative transmission chains is provided. Key abnormal time points and related process parameters extracted from the production data sequence to be analyzed are attached to the report as supporting information for risk tracing. For example, when describing the transmission chain of "excessive moisture in raw flour," the report includes the following supporting information: "For batch 2023-08-15-1015, the 'moisture content of raw flour' recorded at 08:15 was 15.8%; the 'dough hardness after kneading' recorded at 08:30 decreased to 65; and the 'volume expansion rate after proofing' recorded at 09:45 was 2.1." The correspondence between key abnormal time points and parameter values ​​provides data-level support for the risk assumptions.

[0047] The final output is a structured safety risk tracing report containing risk source ranking, transmission logic analysis, and source tracing evidence. In some embodiments, the report's structuring level can be defined using a preset XML or JSON schema to ensure information is organized in a unified format that is both machine-readable and human-readable. The report may incorporate the calculation of a comprehensive risk assessment index, which reflects the overall risk level faced by the current production batch. An optional formula for calculating the comprehensive risk assessment index R is:

[0048] Where: R represents the comprehensive risk assessment index, and n is the total number of identified primary and secondary risk sources. It is the normalized total weight of the k-th risk source. This is the average support score obtained during the authenticity verification phase for the high-risk transmission chain corresponding to the k-th risk source. It can be understood that this index integrates the weight of the risk source and the degree to which its transmission path is supported by real-time data. The structured security risk tracing report will be automatically pushed to the production quality control system in electronic document form, triggering corresponding early warning and handling procedures.

[0049] See Figure 5This is a line chart ranking the risk sources by their weight. This quantitative ranking provides a clear priority for production quality control. "Excessive moisture content in raw flour" and "sudden drop in temperature in the proofing room" are the two most critical risk sources requiring immediate attention and control, while the other three risks have zero weight in the current batch and can be considered secondary control targets. Based on their weight percentages, the quality control department can prioritize its limited resources on these two high-weight risk sources, achieving optimal resource allocation. For "excessive moisture content in raw flour," incoming raw material testing and batch management can be strengthened; for "sudden drop in temperature in the proofing room," the temperature and humidity monitoring system and early warning mechanism can be upgraded to achieve precise control.

[0050] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent detection of food safety in pasta production, characterized in that, Includes the following steps: The real-time production process data and historical batch sample data of the pasta food to be tested are obtained. The real-time production process data and historical batch sample data are standardized, cleaned and aligned to generate a production data sequence to be analyzed. Based on a multi-source information fusion framework, the temporal dependencies and correlations between key quality control points are extracted from the production data sequence to be analyzed, and a production process state diagram model is constructed. Environmental monitoring data and equipment operation logs are used as external influencing factors and introduced into the production process state diagram model to update the state transition probabilities between nodes and generate a dynamic production risk assessment network. Based on the dynamic production risk assessment network, the potential propagation paths of safety indicator anomalies under different risk source assumptions are simulated, and the authenticity of the paths is verified by combining real-time sensor readings on the production line. Based on the results of the path authenticity verification, high-risk transmission chains are identified, and the weighted impact of each risk source on the current key security indicators is calculated. By integrating the weighted impact calculation results with the risk transmission chain, a structured safety risk tracing report is generated, and the safety risk tracing report is pushed to the production quality control system.

2. The intelligent detection method for food safety in pasta production according to claim 1, characterized in that, The real-time production process data and historical batch sample data are standardized, cleaned, and aligned to generate a production data sequence to be analyzed, specifically including: For the collected real-time production process data, the fields are mapped and the units are unified according to the preset production process code and parameter name dictionary to generate a standardized set of real-time production parameters. Key events were extracted and timestamps were restored from historical batch sample data to identify and mark raw material information, process parameter records and test result items related to safety testing. The standardized set of real-time production parameters is aligned with the labeled historical batch sample data along the timeline, and logically related information units are merged according to the production batch and process sequence to construct a multi-dimensional data sequence of the production process. For outliers and missing items in the multi-dimensional data sequence of the production process, an interpolation and correction strategy based on process context is used to repair the data and ensure the continuity of the sequence. The repaired multi-dimensional data sequence of the production process is normalized by feature scale to eliminate the difference in magnitude between different parameters, and the final output is the production data sequence to be analyzed.

3. The intelligent detection method for food safety in pasta production according to claim 2, characterized in that, The multi-source information fusion framework extracts the temporal dependencies and correlations between key quality control points from the production data sequence to be analyzed, and constructs a production process state diagram model, specifically including: On the production data sequence to be analyzed, a constraint relationship discovery algorithm is applied to test the statistical independence of the parameter variables corresponding to each key quality control point under different sets of conditions; Record all associations that pass the significance test, and based on this, construct an initial undirected graph with quality control points as nodes and associations as edges; For each edge in the initial undirected graph, apply a production process timing constraint, that is, the observation time of the preceding process node must be earlier than the observation time of the following process node, and delete the edges that violate the process timing. For the graph structure after time-series constraint filtering, the causal direction inference mechanism is applied to determine the causal orientation of each edge, forming a directed acyclic production process state graph; The strength of the association between nodes in the directed acyclic production process state graph is quantified and assigned using conditional probability or mutual information, and used as the weight of the edge, thus completing the construction of the production process state graph model.

4. The intelligent detection method for food safety in pasta production according to claim 3, characterized in that, The step of incorporating environmental monitoring data and equipment operation logs as external influencing factors into the production process state diagram model to update the state transition probabilities between nodes specifically includes: Temperature and humidity data, as well as air cleanliness data, are collected in the production workshop as environmental monitoring data, and the operating status logs and maintenance records of key production equipment are simultaneously acquired as equipment operation logs. Environmental monitoring data and equipment operation logs are mapped to new impact nodes in the production process state diagram, and potential connection relationships are established between the new impact nodes and existing quality control point nodes in the model. On the extended state diagram structure that includes newly added impact nodes, the intervention effect analysis method is used to assess the change in the transition probability of the existing quality control point node state when specific changes occur in environmental conditions or equipment state. Based on the changes in transition probabilities obtained from the analysis, adjust the weight coefficients of relevant edges in the production process state diagram model; After adjusting the weighting coefficients, the extended state diagram structure is simplified, edges with weighting coefficients below the set threshold are removed, and a dynamic production risk assessment network that integrates environmental and equipment impact information is generated.

5. The intelligent detection method for food safety in pasta production according to claim 4, characterized in that, Based on the aforementioned dynamic production risk assessment network, potential propagation paths of safety indicator anomalies under different risk source assumptions are simulated, and the authenticity of these paths is verified by combining real-time sensor readings on the production line. Specifically, this includes: In the dynamic production risk assessment network, starting from several typical risk source nodes, forward diffusion simulation is carried out along the state transition edge to deduce the sequence of downstream safety indicator abnormal nodes that are triggered, forming multiple potential risk transmission paths. For each potential risk transmission path, a verification node is set up that is directly or indirectly observed by real-time sensors. The verification node is a key quality control point or intermediate state node on the path. Continuously collect real-time sensor readings on the production line and extract process parameters or quality index values ​​corresponding to the set verification nodes. The process parameters or quality index values ​​obtained from the analysis are compared with the preset safety thresholds or normal fluctuation ranges of the verification nodes. If the comparison results exceed the allowable deviation, the potential path is determined to be supported by authenticity. Record the authenticity verification results of each potential risk transmission path. For paths that are not supported, mark them as low-risk paths and reduce their ranking priority in subsequent risk assessments.

6. The intelligent detection method for food safety in pasta production according to claim 5, characterized in that, Based on the path authenticity verification results, high-risk transmission chains are identified, and the weighted impact of each risk source on the current key security indicators is calculated, specifically including: All risk transmission paths that have passed authenticity verification are summarized, categorized and integrated according to the risk source node, forming a set of transmission chains starting from the risk source; For each transmission chain under each risk source, calculate its path length, the average anomaly probability of nodes on the path, and the compliance score of the authenticity verification stage, and calculate the comprehensive risk confidence of each chain accordingly. The chains are sorted according to the comprehensive risk confidence level, and chains with a confidence level higher than the preset risk threshold are identified as high-risk transmission chains. For each high-risk transmission chain, extract the risk source node from the dynamic production risk assessment network and the risk contribution value of each indicator node in the current key safety indicator set; The extracted risk contribution values ​​are normalized, the relative impact weight of each risk source on each security indicator is calculated, and then the total weight impact of each risk source on the entire current set of key security indicators is obtained through weighted summation.

7. The intelligent detection method for food safety in pasta production according to claim 6, characterized in that, The integrated weighting affects the calculation results and the risk transmission chain, generating a structured security risk tracing report, specifically including: The total weight of each risk source is arranged in descending order. Risk sources whose cumulative weight influence exceeds a certain proportion are selected as primary risk sources, and the rest are secondary risk sources. For each major and minor risk source, select the most representative chain from its corresponding high-risk transmission chain and describe its risk transmission logic and key nodes in structured language; The weighted impact of risk sources on the ranking results, risk source classification information, and corresponding risk transmission logic descriptions are then structured and filled in according to a predefined risk report template. The report includes key anomaly time points and related process parameters extracted from the production data sequence to be analyzed, as supporting information for risk tracing. The final result is a structured security risk tracing report that includes risk source ranking, transmission logic analysis, and source tracing evidence.

8. The intelligent detection method for food safety in pasta production according to claim 7, characterized in that, The aforementioned causal direction inference mechanism determines the causal orientation of each edge, forming a directed acyclic production process state diagram, specifically including: For each edge in the undirected graph after time-series constraint filtering, check the order in which the two nodes it connects first exhibit parameter abnormalities or state deviations in the production data sequence to be analyzed. Set the node that exhibits the abnormality first in time as the cause and the node that exhibits the abnormality later as the effect. If the chronological order cannot be clearly determined, then a conditional independence test is introduced for the two nodes under a broader set of conditional variables, and inference is made by utilizing the asymmetric nature of causal direction in conditional independence. If the direction still cannot be determined, query the inherent influence relationship of the parameters of the two connected quality control points in the pasta production process knowledge base, and specify the causal direction based on the process knowledge; After all edges are assigned a direction, the entire graph structure does not form a directed cycle. If a cycle path appears, the edge with the lowest weight in the cycle path is cut to break the cycle. After determining the direction, the directed graph is verified for consistency with the process logic to ensure that it does not conflict with the standard process flow and quality control principles for the production of pasta products.

9. The intelligent detection method for food safety in pasta production according to claim 8, characterized in that, The continuous acquisition of real-time sensor readings on the production line, and the parsing of process parameters or quality index values ​​corresponding to the set verification nodes, specifically includes: Based on the quality control point or intermediate state corresponding to the verification node, determine the type of sensor to be collected, including temperature sensor, humidity sensor, weight sensor, and image acquisition device. Through the industrial data acquisition interface, raw sensor readings on the production line are continuously acquired at a set sampling frequency; The acquired raw sensor readings are preprocessed, including outlier removal, smoothing filtering, and data alignment, to obtain a reliable sensor data sequence. From the preprocessed sensor data sequence, based on the characteristics of the verification node, feature parameters that can characterize its state are extracted, including the heating rate and steady-state value extracted from the temperature sensor data, and color and morphological features extracted from the image data. The extracted process parameters or quality index values ​​are standardized and converted to ensure consistency with the preset safety thresholds of the verification nodes.

10. The intelligent detection method for food safety in pasta production according to claim 9, characterized in that, The process of normalizing the extracted risk contribution values ​​and calculating the relative impact weight of each risk source on each safety indicator specifically includes: For each high-risk transmission chain that is selected, obtain the original risk contribution value from its risk source node to each indicator node in the current set of key security indicators; Calculate the sum of the original risk contribution values ​​from all risk source nodes to the same security indicator node, and divide the original risk contribution values ​​from the risk source node to the security indicator node by the sum to obtain the normalized relative influence weight of the risk source on the security indicator. Iterate through each indicator node in the current set of key security indicators, repeat the calculation process, and obtain a set of relative influence weights of the risk source node on all security indicator nodes. The relative influence weights of a risk source node on all security indicator nodes are weighted and summed, with the weights being the importance coefficients of the security indicator in the current detection. Finally, the total weight influence of the risk source on the entire current set of key security indicators is obtained. The total weighted impact of all risk sources is normalized again to ensure that the sum is 100%, thus completing the weighted impact calculation.

Images