A method and system for automatically supervising the quality of soft candy production

Abstract

The application discloses a gummy production quality automatic supervision method and system, and belongs to the technical field of industrial data processing. The method comprises the following steps: according to process procedures, equipment logic and quality specifications, mode definition and initial instantiation of a gummy production process knowledge graph are completed, a current gummy production batch is supervised, a unique batch node is created in the gummy production process knowledge graph to identify a whole link of production, multi-modal production data of the current gummy production batch is collected, is mapped to a corresponding process node of the gummy production process knowledge graph and attribute labels are added, a target node set with abnormal attribute labels is screened out from the gummy production process knowledge graph, an abnormal path is constructed according to gummy production upstream and downstream process logic, a graph neural network model is introduced, and the effectiveness of the abnormal path is determined through risk assessment, so that the overall effectiveness of gummy production quality supervision in abnormal screening, risk grading and resource investment is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology, and in particular to an automated method and system for monitoring the quality of gummy candy production. Background Technology

[0002] The automated monitoring of gummy candy production quality is primarily used for the continuous sensing, comprehensive analysis, and coordinated intervention of key process parameters, key equipment status, and key quality indicators throughout the entire gummy candy production process. This ensures that each batch of gummy candies remains within controlled limits in terms of shape, size, and safety. Through this monitoring method and system, quality control that previously relied on manual experience and random sampling can be upgraded to data-driven, fully online automated monitoring. Furthermore, it can promptly identify and provide adjustment suggestions when quality risks are still showing abnormal trends, preventing problems from being discovered only after they have scaled up on the production line, thus avoiding the scrapping of entire batches or extensive rework.

[0003] The fundamental reason for the need for automated monitoring of the gummy candy production process is that gummy candy products are highly sensitive to process conditions: minute deviations in raw material ratios, slight drifts in the sugar melting temperature curve, slight decreases in mixing uniformity, minor changes in pouring viscosity and mold temperature, and slight loss of control over drying humidity can all be amplified in subsequent stages, ultimately resulting in quality defects such as a hard and sticky texture, uneven internal structure, surface collapse, and clumping of the sugar frosting. If relying solely on post-production sampling or manual inspection, by the time problems are discovered, a large number of defective products have often accumulated, leading to waste of raw materials and energy, disruption of production rhythm, increased cost of quality correction, and even damage to brand reputation. Therefore, it is necessary to build an automated quality monitoring mechanism that spans the entire process chain, shifting from post-production problem detection to proactive risk control during the process.

[0004] In terms of the objects of supervision, the system targets the entire process of equipment and technology in the gummy candy production line, from upstream ingredient preparation to final packaging. This includes, but is not limited to: metering and batching systems for raw materials such as raw sugar and gelatin, sugar dissolving kettles and gelatin dissolving tanks, mixing and homogenizing devices, insulation and conveying pipelines, online filtration and degassing devices, casting machines and mold components, cooling / drying channels, automatic demolding mechanisms, coating devices for powder / sugar / oil, online weighing modules, visual inspection modules, and downstream automatic packaging units. The system utilizes various types of sensors, including those for temperature, time, sugar content, viscosity, humidity, pressure, weight, and images, deployed on the aforementioned equipment and pipelines to form a multi-dimensional data acquisition network covering raw materials, processes, and finished products, providing a continuous and granular data foundation for quality supervision.

[0005] Under the existing quality monitoring model for gummies production, typical processes have already introduced a certain degree of automated testing, and some production lines have attempted to use knowledge graphs to provide structured descriptions of process steps and quality elements. Taking a common practice as an example: In the production process, the quantitative addition of each raw material is first completed by an automatic batching system to ensure that the ratio meets the process specifications. Then, in the sugar and gum stages, sensors such as temperature, time, and viscosity are used to monitor the dissolution and heating status, and the corresponding process nodes and key parameter nodes can be linked in a knowledge graph. After mixing, the syrup or gum solution is transported to the casting machine. After casting and molding, the soft candy enters the cooling and drying zone. Temperature and humidity control is used to ensure the curing effect. Some systems will also mark the cooling and drying stations and their quality characteristics in the knowledge graph. After cooling, the soft candy is automatically demolded and enters the preliminary quality inspection station, where visual recognition and weight sensors are used to check the appearance, size, and weight of each piece. If the process formula requires it, it will also undergo coating with powder, sugar, or oil. Finally, it enters the automatic packaging line, where online weighing, counting, and final inspection cameras are used to check the packaging quantity, label, and sealing status. Throughout the process, the production line typically records some process data and correlates the processes, equipment, parameters, and quality results to a certain extent in a knowledge graph. When a parameter exceeds the limit, an alarm is triggered so that operators can intervene.

[0006] However, the existing technology has the following technical problems: Although some production lines have used knowledge graphs to structure the relationships between process steps, equipment, and quality results, existing monitoring and early warning mechanisms still primarily rely on fixed threshold triggers for single parameters and single workstations in actual operation. They can only determine whether a specific measuring point exceeds limits, but cannot determine whether such an anomaly poses a substantial risk to the final product quality from the perspective of the entire process chain. On the one hand, slight deviations in local parameters such as sugar melting temperature, gelling viscosity, mixing concentration, pouring conditions, and drying environment can be compensated for by subsequent workstations through process adjustments, thus minimizing the actual impact of these anomalies on the final product quality. While the current system may have no substantial impact on the final product quality, it still triggers warnings in all cases, resulting in a large number of invalid warnings that do not match the actual quality risks. This leads to a low effective hit rate and high intervention costs. On the other hand, when multiple workstations exhibit small deviations within their respective thresholds, which amplify along the process chain and ultimately cause serious defects such as product collapse, bubbles, or uneven internal structure, the existing system cannot establish an effective judgment logic for deviation accumulation and quality defects in the knowledge graph because each individual parameter does not exceed the limit. It also struggles to mark such potential risks as valid anomalies, thus completely lacking early warning and intervention at critical evolution stages. Furthermore, the existing knowledge graphs mostly remain at the level of static relationship display and simple querying, lacking an anomaly effectiveness evaluation mechanism based on process evolution and causal chains. This means that even if the system records anomaly nodes, it cannot understand whether the anomaly truly needs to be addressed or which type of anomaly should be prioritized. This severely restricts the overall effectiveness of gummy candy production quality supervision in anomaly screening, risk classification, and resource allocation. Summary of the Invention

[0007] To address the technical problems existing in current technologies that limit the overall effectiveness of gummy candy production quality supervision in anomaly screening, risk classification, and resource allocation, this invention provides an automated method and system for gummy candy production quality supervision. The technical solution is as follows: On the one hand, an automated monitoring method for the quality of gummy candy production is provided, the method comprising: S1. Based on the process specifications, equipment logic, and quality standards, define the pattern and initial instantiation of the gummy candy production process knowledge graph. Monitor the current gummy candy production batch, create a unique batch node in the gummy candy production process knowledge graph to identify the entire production chain, collect multimodal production data of the current gummy candy production batch, map it to the corresponding process node in the gummy candy production process knowledge graph, and add attribute tags. S2. Filter the target node set with abnormal attribute tags from the gummy candy production process knowledge graph. Based on the upstream and downstream process logic of gummy candy production, construct abnormal paths, introduce a graph neural network model, and determine the validity of abnormal paths through risk assessment. S3. Input the batch process subgraph corresponding to the valid abnormal path into the graph neural network. Through attention weight allocation and reverse attribution analysis, locate various abnormal nodes corresponding to the valid abnormal path, associate the control parameters corresponding to various abnormal nodes, and automatically generate a gummy candy production quality abnormality cause analysis report and control parameter optimization instructions.

[0008] On the other hand, an automated quality monitoring system for gummy candy production is provided. This system applies methods for automated quality monitoring in gummy candy production and includes: a graph data mapping module, an anomaly path construction module, and an anomaly path attribution module. The graph data mapping module, based on process specifications, equipment logic, and quality standards, defines and initially instantiates the gummy candy production process knowledge graph, monitors the current gummy candy production batch, creates a unique batch node in the gummy candy production process knowledge graph to identify the entire production chain, collects multimodal production data of the current gummy candy production batch, and maps it to the corresponding process node in the gummy candy production process knowledge graph. The system includes: an attribute tagging module; an anomaly path construction module, which selects target node sets with anomaly attribute tags from the gummy candy production process knowledge graph, constructs anomaly paths based on the upstream and downstream process logic of gummy candy production, introduces a graph neural network model, and determines the validity of anomaly paths through risk assessment; and an anomaly path attribution module, which inputs the batch process subgraphs corresponding to valid anomaly paths into the graph neural network, locates various anomaly nodes corresponding to valid anomaly paths through attention weight allocation and reverse attribution analysis, associates the control parameters corresponding to various anomaly nodes, and automatically generates a gummy candy production quality anomaly cause analysis report and control parameter optimization instructions.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention constructs a knowledge graph of the gummy candy production process covering ingredients, cooking, gel forming, cooling, and packaging. Based on real-time collection of batch process parameters and quality indicators, it marks slight deviations and anomaly candidates at each process node. Based on the process topology, it automatically forms a set of batch-level anomaly paths. Under the joint constraints of node importance weight and edge association strength weight, it quantifies the risk and determines the effectiveness of each anomaly path, screening out the effective anomaly paths that truly cause batch quality control failures. Furthermore, it achieves matching and reverse inference between reported abnormal quality problems and historical effective anomaly paths. The located target anomaly path is used to adaptively update the closed-loop optimization effect of the node-level risk judgment strategy and the path-level risk aggregation strategy. This allows the regulatory strategy to continuously evolve with production fluctuations and quality problem forms, effectively solving the problems of existing technologies that rely only on single-point threshold alarms, have difficulty identifying effective anomaly paths at the process chain scale, rely on manual experience for quality problem tracing and lack strategy self-learning ability, resulting in high false alarm rates, serious missed detections, and insufficient guidance for rectification.

[0010] 2. This invention constructs an abnormal path validity verification mechanism based on the knowledge graph of gummy candy production process and batch process data. When a slight deviation or abnormal candidate marker is detected in a batch node, the node sequence of that batch in the process topology is extracted as an abnormal path. By combining verification, the correspondence between abnormal paths and actual quality problems is backtracked, compared, and statistically evaluated. This dynamically distinguishes between valid abnormal paths that causally contribute to final inspection defects and invalid abnormal paths caused only by occasional fluctuations. On the one hand, it automatically corrects the node-level risk judgment threshold and path risk aggregation weight, suppressing false alarms caused by single-point noise amplification and reducing unnecessary downtime for investigation and rework. On the other hand, it gradually amplifies the influence of key nodes and critical paths highly correlated with batch defects in the whole-process supervision strategy, enabling the system to prioritize process links that are prone to inducing batch defects, triggering early warnings and process interventions in advance. This transforms the discovery of gummy candy production quality problems from passive sampling inspections after the fact to proactive prevention based on the validity verification of abnormal paths, ensuring product quality stability while taking into account the comprehensive optimization of production rhythm and resource input.

[0011] 3. This invention utilizes a knowledge graph of the gummy candy production process for reverse backtracking and path analysis of abnormal quality issues. On one hand, it enables the system to accurately map a single abnormal quality issue to a specific node link in the process topology. Through reverse path search from downstream to upstream, it filters out effective backtracking abnormal paths with a real causal relationship to the issue, thus avoiding the generalization of a single-point deviation or local fluctuation into a high-risk state for the entire process. This significantly reduces false alarm rates and the downtime, rework, and waste of testing resources caused by excessive intervention. On the other hand, by structurally analyzing the inducing and transmission factors of each effective backtracking abnormal path, and dynamically updating the node-level risk assessment strategy and the path risk aggregation risk assessment strategy accordingly, the risk model can adaptively tighten around the path that actually causes the quality problem. High-risk nodes and highly sensitive paths are subject to higher weight for continuous monitoring and early warning, while nodes and paths that have not formed actual quality defects in multiple backtracking verifications are gradually relaxed or downweighted. This achieves a refined and differentiated evolution of the risk assessment strategy from static experience configuration to data-driven approaches. This mechanism not only improves the accuracy and robustness of abnormal path identification and strengthens the ability to trace the source of quality problems and define responsibilities, but also achieves macro-level matching and optimization between production process supervision resources and quality risk distribution. This enables gummy candy manufacturers to effectively control supervision costs and adjust the pace of intervention while ensuring stable and qualified products, ultimately improving the overall quality management level and operational efficiency of the production system. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart of an automated quality monitoring method for gummy candy production is provided as an embodiment of this application; Figure 2 A schematic diagram of an automated quality monitoring system for gummy candy production provided in this application embodiment; Figure 3 A schematic diagram of the knowledge graph of the gummy candy production process provided in the embodiments of this application; Figure 4 Example diagram of multimodal production data provided in the embodiments of this application; Figure 5 This is a schematic diagram of the training loss fluctuation of the graph neural network model provided in the embodiments of this application. Detailed Implementation

[0014] The following provides explanations for some of the terms used in this application. It should be noted that these explanations are for the convenience of those skilled in the art and do not constitute a limitation on the scope of protection claimed in this application.

[0015] The embodiments of this application involve at least one, including one or more; where "multiple" means two or more. Furthermore, it should be understood that in the description of this specification, terms such as "first," "second," and "third" are used only for descriptive purposes and should not be construed as indicating relative importance or order. For example, "first device" and "second device" do not represent the degree of importance of the two or their order, but are merely for descriptive distinction. In the embodiments of this application, "and / or" merely describes an association relationship, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0016] References to "one embodiment," "in some examples," or "some embodiments" as described in the embodiments of this application mean that one or more embodiments of this specification include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in some examples," "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0017] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0018] Example 1 provides an automated method for monitoring the quality of gummy candy production, such as... Figure 1The diagram shown is a flowchart of an automated quality monitoring method for gummy candy production provided in this application embodiment. The method includes the following steps: Based on the company's existing gummy candy process specifications and the corresponding operation manual for the gel gummy candy casting production line, collect standard process steps, process parameters, and equipment configuration information for the entire gummy candy production process. The process specifications refer to operational documents that clearly define the control parameters such as temperature, sugar content, and time for processes such as raw material acceptance, ingredient mixing, gelling, sugar melting, mixing and blending, storage, vacuum concentration, casting, cooling and shaping, drying, coating, inner packaging, and outer packaging. For example, the sugar melting stage requires heating to 116.5 to 118°C and achieving a sugar content of 86% to 87%. The sugar content of the syrup is controlled before casting. The drying temperature is 55℃±2℃ and the moisture content is controlled at 10%±1%, with a content of 76% to 77%. Quality specifications correspond to the key control items marked in the process specifications and the internal control standards for intermediate / finished products, such as completing packaging within 24 hours after the weight loss during drying is qualified. Based on this, the equipment logic refers to identifying the process support links and logical constraints such as operating sequence, speed setting, and temperature control between equipment such as the sugar boiling machine, storage tank, mixer, casting machine, cooling conveyor, drying device, coating device, and electrical control system, according to the equipment composition of the soft candy casting production line (including sugar melting pot, gum melting pot, storage and insulation tank, mixer, casting machine, cooling conveyor, drying device, coating device, electrical control system, etc.) and their material flow, start-stop dependencies, and electrical interlocking relationships.

[0019] After completing the above basic analysis, the basic model of the knowledge graph of gummy candy production process is defined based on process specifications, equipment logic, and quality standards: sugar dissolving, gum dissolving, mixing and blending, vacuum concentration, casting, cooling and shaping, drying, coating, and inner packaging are abstracted as process nodes; sugar boiling pot, gum dissolving pot, storage tank, casting machine, cooling conveyor belt, drying box, and coating machine are abstracted as equipment nodes; intermediate product testing, loss on drying testing, metal detection, and finished product release are abstracted as quality testing nodes; and workshop temperature, humidity, vacuum degree, and steam pressure are abstracted as environmental nodes. At the same time, the directed sequence relationship between different nodes, the execution relationship of which equipment completes the process, the constraint relationship of which quality standards constrain the process, and the influence relationship between process parameter fluctuations and quality results are defined as relationship types in the knowledge graph. Subsequently, after the model is determined, historical production data and corresponding quality records are imported to initially instantiate the knowledge graph: using each historical production batch as a sample, the actual process path, equipment start-up and shutdown status, key parameter records (such as a batch of vacuum concentration vacuum degree 0.02MPa, casting temperature 92℃, drying time 28h) and intermediate / final inspection conclusions of each batch are mapped to specific node instances and relationship instances, forming a multi-dimensional relational subgraph of process node-equipment node-quality inspection node-environment node. Finally, based on historical statistical results and the experience of process experts, the nodes and edges in the knowledge graph are weighted. The importance weight of the nodes is used to quantify the influence of each process node, equipment node, quality inspection node, and environmental node on the risk of final inspection failure of the gummies. The weight values ​​are preferably in the range of [0, 1]. For example, the co-occurrence frequency of deviations from control parameters in each process and the batches that fail final inspection are statistically analyzed. Nodes with high-frequency associations, such as vacuum concentration, drying, and casting, are given higher weights, while auxiliary nodes with relatively small influences are given lower weights. The association strength weight is used to characterize the tightness of deviation transmission between upstream and downstream nodes and its linkage effect on quality results, so as to quantify the tightness of coupling between processes, between equipment and processes, and between processes and quality results. For example, it can be determined based on indicators such as the correlation coefficient of upstream and downstream node parameter fluctuations, the joint frequency of deviations occurring simultaneously in the same abnormal batch, and the proportion of samples sharing the same abnormal quality problem. This completes the pattern definition and initial instantiation of the knowledge graph of the gummies production process, providing a structured basis for subsequent abnormal path identification and abnormal validity verification.

[0020] Figure 3This is a schematic diagram of the knowledge graph of the gummy production process provided in this application embodiment. The diagram illustrates the relationships between nodes in the entire gummy production process, specifically including nodes such as the gelling station, casting station, and drying station, reflecting the core process flow logic of gummy production; key control parameters for each node (sugar solution viscosity, gelling temperature, etc.); and defects caused by abnormal parameters at each node (such as surface cracks due to increased sugar solution viscosity). This graph achieves a structured mapping between process parameter deviations and quality risks, and serves as the core input for subsequent graph neural network (GCN) prediction of risk levels and transmission paths. It should be explained that... Figure 3 This is just a small example of the knowledge graph of gummy candy production processes. The actual annotations of the graph can be formulated by relevant technical personnel according to actual production.

[0021] When the system monitors the entire production process of the current batch of pectin gummies (Formula A) on December 10, 2025, the early shift, Line 2, it first creates a unique batch node in the gummies production process knowledge graph. This batch node can generate a batch ID by combining information such as production date, production line number, formula code, and shift. This ID is used to identify the unique identity of the batch throughout the entire process, including raw sugar boiling, gumming, mixing, vacuum concentration, pouring, cooling and shaping, drying, coating, and packaging. This enables the subsequent aggregated management of all process behaviors of the batch. Multimodal production data refers to data from different sources and of different types collected simultaneously during the production process of this batch. For example, process parameters such as temperature, sugar content, vacuum degree, rotation speed, and valve opening from equipment such as sugar boiling pot, gelling pot, vacuum concentration tank, casting machine, and drying oven; and quality inspection data from conveyor belt weighing, online metal detectors, and online appearance cameras. These heterogeneous data are mapped to the corresponding process nodes in the gummy candy production process knowledge graph and attribute tags are added. That is, the system writes the data collected in the sugar boiling stage into the sugar boiling process node, the data in the casting stage into the casting process node, etc., according to the topological sequence of the process specification from sugar boiling to gelling to mixing to concentration to casting to cooling to drying to coating to packaging, and records the corresponding timestamp, equipment number, and deviation of key process parameters (such as the deviation of sugar boiling temperature from the sugar boiling reference temperature) in the node attributes.

[0022] Figure 4 The diagram illustrates multimodal production data provided in this application. The left side shows the original form of multimodal production data (i.e., multimodal heterogeneous data as described in the diagram), such as sugar solution viscosity and drying humidity at core work stations like gelling and drying. The right side shows the normalized values ​​on the horizontal axis and the number of samples on the vertical axis, demonstrating the distribution characteristics of the data after Z-Score normalization. This indicates that the data acquisition and preprocessing scheme of the invention can support the initialization of node features of dynamic process knowledge graphs and the input requirements of subsequent graph neural networks.

[0023] On the production line, each station (such as the casting station, cooling station, and drying station) is divided into several station time windows, for example, 30 seconds or one mold cycle time. Within each time window, the system collects multimodal production data for the current batch and uses denoising methods such as median filtering and moving average to eliminate instantaneous sensor jitter and occasional anomalies. Subsequently, for the batch node created in the knowledge graph for that batch, if it is detected that the batch has already been created in a previous process, the system will not create the node again. Instead, it will update the node information synchronously based on the real-time production progress, such as updating its current process stage, cumulative output, and whether drying has been completed. All denoised multimodal production data are written to the corresponding node attributes. Each process node and its subordinate workstation time windows together constitute a set of multi-dimensional process snapshots for that batch under that process, thus forming a structured form of single batch node association and workstation time windows carrying multi-dimensional process snapshots: the batch node is the root, which is associated with a process node link covering the entire process. Each process node is also connected to a series of time-ordered window sub-nodes. The window sub-nodes store multi-dimensional data such as temperature, sugar content, vacuum degree, image features and detection results within that time period.

[0024] Based on this, the system assigns attribute tags to each process node and its time window. For example, when all parameters of a node are within the process specification within the time window, the node is marked as a normal node. When a node has parameters deviating from the allowable fluctuation range within the time window, and the deviation is less than the maximum deviation stored in the database, it is marked as a slight deviation node. For instance, the system first configures the allowable fluctuation range specified in the process specification for each key process parameter. Simultaneously, based on historical qualified batch operation data, it maintains a maximum deviation for each parameter in the database, representing the maximum acceptable deviation that the parameter has exhibited without causing actual quality defects. For example, for the drying process, the process specification requires the drying temperature to be controlled at 55℃±2℃, i.e., the allowable fluctuation range is 53℃ to 57℃. However, historical data analysis shows that in some batches, the drying temperature once rose to 58℃ to 60℃, exceeding the allowable range of 53℃ to 57℃. However, after appropriate compensation measures such as extending the drying time, the final product was still deemed qualified. Therefore, the system records +3℃ relative to the upper limit of 57℃ as the maximum deviation of this parameter without causing defects. Based on this, when the current production batch is monitored to have a drying temperature of 58℃, on the one hand, this temperature exceeds the allowable fluctuation range of 53℃ to 57℃, which is a case of parameter deviation from the allowable fluctuation range; on the other hand, the deviation from the upper limit of 57℃ is only +1℃, which is less than the maximum deviation recorded in the database of +3℃. Therefore, the system marks this drying process node as a slightly deviated node, considering that its risk has increased but has not yet reached a serious abnormality level, and uses it as a slightly risky node in the subsequent abnormal path construction and validity verification. When a node has a parameter deviation from the allowable fluctuation range within the time window, and the degree of deviation is not less than the maximum deviation stored in the database, it is marked as an abnormal candidate node. This not only realizes the full-link structured expression with a unique batch node as the core, but also provides a data foundation and topology carrier for the subsequent construction of abnormal paths based on node attribute marking and its validity verification.

[0025] For the gummy candy production batch with batch number B20251201, the system first filters out process nodes marked with slight deviation or anomaly candidate nodes during the monitoring period from the gummy candy production process knowledge graph, forming a target node set. For example, in this batch, the vacuum concentration process is marked as an anomaly candidate node because the vacuum degree deviates from the allowable range and is greater than or equal to the historical maximum deviation; the drying process is marked as a slight deviation node because the drying temperature repeatedly exceeds the upper limit of the allowable range but is lower than the historical maximum deviation; and the metal detection process before final inspection is marked as an anomaly candidate node because the online metal detection rate shows a significant abnormal upward trend. Starting from these marked process nodes, the system identifies upstream and downstream process relationships between two nodes based on the process topology structure (e.g., vacuum concentration to casting to cooling to drying to powder coating to metal detection to packaging) that is fixed in the knowledge graph beforehand. Nodes with material flow or quality impact relationships are connected in series from upstream to downstream to construct abnormal paths: for example, starting from the vacuum concentration node, the casting, cooling and drying nodes are connected downstream along the topology to obtain an abnormal path P1 covering multiple stations; then, the powder coating and metal detection nodes are connected downstream from the drying node to obtain an abnormal path P2.

[0026] For some target nodes that lack effective upstream or downstream connections in the knowledge graph, such as an independent environmental monitoring node (e.g., high humidity in the workshop) that is only weakly associated with the current batch of anomalies, the system cannot continue to extend upstream and downstream when expanding the association. In this case, the anomaly path corresponding to the node naturally degenerates into a single-node anomaly path that only contains the node itself, which is used to record isolated but potentially meaningful anomaly signals. Meanwhile, during the time-axis-based association expansion process, the system also counts the marking status of each process node within a sliding time window divided by workstation. For example, if a drying process node is marked as a slightly deviated node or an anomaly candidate node in a consecutive limited number of workstation time windows (e.g., 3), and its downstream metal detection node appears as an anomaly candidate node multiple times in the same batch, then it is considered that there is a stronger anomaly transmission relationship between drying and metal detection. The system extracts the association strength weight increment from the database and adds it to the current association strength weight between the two, thereby increasing the association strength weight of the upstream and downstream edge in the knowledge graph. Through the above process, on the one hand, a multi-workstation anomaly path from vacuum concentration to final inspection is formed for subsequent anomaly path validity verification and quality problem tracing; on the other hand, by amplifying the weight of consecutive deviation nodes, the key processes with truly continuous impact are given higher priority in subsequent risk assessment and early warning strategies, realizing the refinement and dynamic adaptation of anomaly path construction.

[0027] After constructing the abnormal paths for batch B20251201 of gummy candy production, the system introduces a graph neural network model based on the knowledge graph of gummy candy production processes to conduct risk assessments on the effectiveness of each abnormal path. Effectiveness refers to the fact that the process corresponding to the abnormal path poses a substantial risk to the quality of the finished gummy candy. First, the system uses the process subgraph corresponding to the current batch as input, constructing the graph neural network topology by combining process nodes (such as vacuum concentration, casting, cooling, drying, coating, metal detection, etc.) and edges between nodes representing the process sequence, material transfer relationships, and risk correlation strength. For each node, a multimodal feature vector is extracted, including the degree of deviation of process parameters, the label type of slightly deviated nodes / abnormal candidate nodes, the corresponding online detection results, the workshop environment status, the frequency of historical non-conformities, and associated complaint tags, as the node input features of the graph neural network. When a single-node abnormal path exists, the graph neural network executes a node-level risk judgment strategy through a multi-layer message passing and aggregation mechanism, that is, comprehensively considering the node's own deviation... After analyzing the differential features and their information interaction with neighboring nodes, the system outputs a risk score for each node, which characterizes the direct contribution of the node to the quality issues of the current batch. For single-node abnormal paths containing only a single abnormal node, the system makes a judgment based solely on the node's risk score and a preset node risk threshold in the database. The node risk threshold refers to the minimum allowed value of the risk score. If the risk score is lower than the node risk threshold, the single-node abnormal path is considered an invalid path, and the multimodal features of the node and its adjacent edges are written as negative samples into the model training sample set to weaken the weight of similar noisy samples in subsequent iterations. If the risk score is not lower than or higher than the node risk threshold, the batch process subgraph containing the node is retained as a valid sample.

[0028] In this invention, the risk score acquisition process for each node can be reproduced according to fixed calculation steps: First, for the target node p to be evaluated, the system extracts its initial feature vector to characterize the process status of the node in the current batch, including the deviation of process parameters from the target value and allowable fluctuation range, whether it is marked as a slightly deviated node or an abnormal candidate node, the statistical characteristics of the corresponding online detection results, the environmental parameters of the workshop, and the co-occurrence frequency with historical non-conforming batches, etc. These values ​​are concatenated in a predetermined order to form the node initial vector, which serves as the node representation of the 0th layer of the graph neural network. Subsequently, in each layer of the graph neural network, based on the set of neighbors of node p, the system first calculates the attention weights of each neighbor node for node p according to the attention subnetwork, obtaining the weight coefficient of each neighbor. Then, these weights are used as coefficients to perform a weighted summation of the representation vectors of each neighbor node in the current layer, forming the neighbor information aggregation vector of node p. Then, the neighbor aggregation vector is concatenated with the representation vector of node p itself in the current layer in the feature dimension, and an affine transformation is performed through a set of linear transformation parameters and bias terms. A pre-selected nonlinear activation function is applied to obtain the representation vector of node p in the next layer. The aforementioned attention-weighted aggregation, affine transformation, and activation operations are repeated for a preset number of layers until the risk representation vector of node p in the last layer is obtained. Finally, the system connects a single-output risk judgment subnetwork to this risk representation vector. That is, a set of trainable weight vectors and biases are used to map the representation to a real number score, and then the real number is compressed to the interval of 0 to 1 by normalization using the Sigmoid function to obtain the risk score of node p.

[0029] For non-single-node abnormal paths covering multiple process nodes, path length, node criticality weights within the path, edge association strength weights, anomaly occurrence time order, and co-occurrence frequency are introduced as path-level features. These features are then used to output an overall path risk score through the readout layer of a graph neural network. When the overall risk score of an abnormal path falls below a preset path validity threshold in the database (the minimum allowed value for the overall risk score), the system uniformly marks the multimodal features of the nodes and their connecting edges on that path as negative samples and adds them to the training sample set to correct the model's sensitivity to false association paths. When the overall risk score exceeds the path validity threshold, the entire batch process subgraph corresponding to the valid abnormal path is input into the graph neural network for enhanced training, enabling the model to more accurately identify similar high-risk anomaly combinations in subsequent production batches. This risk assessment strategy, combining node-level and path-level approaches, achieves differentiated processing for single-node and multi-node abnormal paths. It allows the graph neural network to gradually strengthen its ability to identify truly valid abnormal paths through continuous iteration and suppresses the interference of invalid abnormal paths caused by occasional fluctuations or weak associations on quality warnings.

[0030] For example, the system obtains the final representation vector and corresponding risk score of each process node on the abnormal path, and combines it with the node importance weights pre-labeled in the knowledge graph and the node risk contribution obtained through online training. Simultaneously, it reads the association strength weights of each edge on the path in the knowledge graph, the timestamps of each node's abnormal occurrence, and the co-occurrence frequency of the path with the same quality problem in historical batches. Then, the system uses the path length (total number of nodes), the weighted average and maximum values ​​of node criticality weights, the average and maximum values ​​of edge association strength on the path, the statistical measures of the time interval between the abnormality and the downstream (such as maximum total lag and average lag), and the normalized co-occurrence frequency of the path in historical non-conforming batches as components, and concatenates them in a fixed order into a path-level feature vector. Next, this path-level feature vector is input into the readout layer of the graph neural network (e.g., a one- or two-layer fully connected network), and a real-valued score is obtained through linear transformation and nonlinear activation. This score is then normalized to the 0-1 range using the Sigmoid function, serving as the overall risk score for the non-single-node abnormal path. The closer the score is to 1, the more likely the path is to be the true cause of the quality problem. Based on this, the path validity is determined by comparing it with a preset path validity threshold.

[0031] In this invention, the input, output, and training process of the graph neural network model can be specifically described as follows: First, on the input side, the model uses batch process subgraphs extracted from the knowledge graph of gummy candy production process as samples. Each subgraph contains multiple nodes and edges. Nodes include at least process nodes (such as vacuum concentration, casting, cooling, drying, coating, metal detection, etc.), equipment nodes, quality inspection nodes, and environmental nodes. Edges are used to characterize the sequence of processes, material transfer relationships, and correlation strength weights. At the feature level, each node corresponds to a set of multimodal feature vectors, which may include: the deviation between the current value of the process parameter and the target value / allowable fluctuation range / historical maximum deviation, whether it is marked as a slightly deviated node or an abnormal candidate node, the statistical characteristics of the corresponding online detection results (such as moisture, weight, metal detection rate, etc.), the current workshop environmental status (temperature, humidity, etc.), and the co-occurrence frequency with historical non-conforming batches or complaint records, etc.; edge features may include upstream and downstream correlation strength weights, abnormal transmission co-occurrence times, abnormal propagation time lag, etc., used to characterize the tightness of risk transmission between nodes.

[0032] On the output side, the graph neural network model needs to support both types of risk assessment strategies mentioned above. Firstly, after multiple layers of message passing and feature aggregation, the model outputs a node-level risk score for each node in the graph. This score is used for node-level risk assessment strategies, specifically determining the direct risk contribution of a single process node to the quality issues of the current batch. This is particularly useful for handling single-node abnormal paths containing only one abnormal node. Secondly, the model aggregates node embedding vectors and path-level features (such as path length, key node weights, edge correlation strength distribution, and the temporal order of anomalies) on an abnormal path at the readout layer. This outputs a path-level overall risk score for the abnormal path, used for path risk aggregation and assessment strategies. This allows for an overall effectiveness assessment of non-single-node abnormal paths covering multiple workstations. Combining preset node risk thresholds and path effectiveness thresholds, the model ultimately provides a judgment result on whether a node is a high-risk node and whether the abnormal path is a valid abnormal path.

[0033] In terms of the training process, this invention adopts a combination of supervised learning and online self-iteration: In the initial stage, a training set is constructed using quality problem samples from existing production history that have been verified manually or by rules. Among them, the batch process subgraphs corresponding to abnormal paths that have been confirmed to have a real causal relationship through final inspection or complaint tracing are used as positive samples, while the subgraphs corresponding to abnormal paths that have been verified as false alarms or have no actual quality consequences are used as negative samples. Labels are assigned to the node-level and path-level outputs respectively (such as whether the node is a critical causal node and whether the path is a valid abnormal path). During subsequent online operation, when a certain abnormal path in the current batch is deemed "invalid" through quality issue backtracking and rule-based judgment, the multimodal features of the corresponding nodes and associated edges of that abnormal path are added to the negative sample training set. When an abnormal path is deemed valid through actual defect confirmation or complaint analysis, the batch process subgraph corresponding to the valid abnormal path is used as a new positive sample input to the graph neural network. The batch process subgraph corresponding to the valid abnormal path refers to a local graph structure with complete process and parameter semantics formed by aggregating the process nodes, equipment nodes, quality inspection nodes, and their direct upstream and downstream associated edges involved in the process link of a certain gummy candy production batch within the overall process subgraph. During training, the model uses the above subgraph as input and, through multiple rounds of forward and backward propagation, jointly optimizes the node-level risk judgment loss function and the path risk aggregation judgment loss function (e.g., weighted summation of node classification loss and path classification loss), and updates the parameters of the graph neural network using gradient descent. By simultaneously incorporating the node-level risk assessment strategy for single-node abnormal paths and the path risk aggregation risk assessment strategy for multi-node abnormal paths into the training closed loop, the model continuously strengthens its ability to distinguish real and valid abnormal paths during continuous iteration, gradually reducing the interference of invalid abnormal paths caused by occasional fluctuations or weak correlations on the overall quality warning.

[0034] Figure 5 This diagram illustrates the training loss fluctuation of the Graph Neural Network (GCN) model provided in this application embodiment. The horizontal axis of the diagram represents the training step size of the GCN model, indicating the number of training iterations of the graph neural network model; the vertical axis represents the training loss (gummy candy quality risk prediction), reflecting the degree of deviation between the model's prediction results and the actual situation. The diagram shows the training process of the graph neural network model on the dynamic process knowledge graph dataset: as the training step size increases, the model loss value continues to decrease and gradually stabilizes. This trend proves that the graph neural network model of this invention has a good convergence state, and also shows that the model can stably learn the rules in the process knowledge graph and has a reliable gummy candy quality risk prediction capability.

[0035] Collect data corresponding to all valid abnormal paths in the current gummy production batch, such as process paths that have actually caused gummies to stick together, not dry enough, or have excessive metal foreign objects. Export the batch process sub-graphs (including process nodes, equipment nodes, quality inspection nodes and their connecting edges) corresponding to these paths, remove samples with missing key parameters, incorrect timestamps or inconsistent labels, and correct or delete abnormal values ​​to ensure that the remaining data can truly reflect the abnormal transmission process. The total data volume is the sum of all data in each node of the valid abnormal paths obtained after cleaning.

[0036] All valid abnormal paths' corresponding data are archived according to defect type as follows: gummy candy adhesion path dataset, insufficient drying path dataset, and excessive metal foreign object path dataset. After compression and format standardization, the system reads the file sizes corresponding to these three datasets from the storage system: adhesion dataset occupies approximately s1=1.2GB, insufficient drying dataset occupies approximately s2=0.6GB, and excessive metal foreign object dataset occupies approximately s3=0.2GB, with a total data volume s_sum of 2GB. Based on the enterprise's quality supervision needs, the system sets regulatory importance coefficients for different defect types. For example, considering that excessive metal foreign object content involves food safety, it is given higher importance, with β3=1.5; adhesion is assigned β1=1.0; and insufficient drying is assigned β2=0.8. Based on this, the system constructs unnormalized path-level loss function weight vector components: path-level loss function weight w1=β1×(s1 / s_sum) for adhesion class, path-level loss function weight w2=β2×(s2 / s_sum) for insufficient drying class, and path-level loss function weight w3=β3×(s3 / s_sum) for metal foreign object exceeding the limit class.

[0037] Normalize w1, w2, and w3, update w1, w2, and w3, and the path-level loss function weight vector is (w1, w2, w3). Complete the path-level loss function weight vector and perform model validation until the model prediction error rate drops to within the preset maximum allowable error rate.

[0038] After initializing the graph neural network and determining that a certain batch (e.g., batch B20251201) has a valid abnormal path corresponding to the gummies sticking quality problem, the system feeds the batch process subgraph corresponding to the path into the graph neural network as input. Here, the batch process subgraph refers to a directed subgraph formed by extracting process nodes (e.g., vacuum concentration, casting, cooling, drying, powder coating), equipment nodes (e.g., drying oven No. 1, casting machine A line), quality inspection nodes (e.g., moisture detection after drying, final inspection appearance sticking judgment) and their connecting edges (representing the process sequence and risk correlation strength) related to the valid abnormal path from the gummies production process knowledge graph, with the batch as the unit. When performing forward inference on the subgraph, the graph neural network introduces an attention weight allocation mechanism. That is, when aggregating features between nodes, a weight coefficient is automatically learned for different neighboring nodes and connecting edges to characterize the influence of the neighbor on the risk representation of the current node. This makes the model focus more on process nodes and connecting edges that have a greater impact on the result when calculating quality risk. Subsequently, based on attention weights and gradient information, the system performs reverse attribution analysis on the network output. The predicted output that ultimately points to the target quality risk node of adhesion defect (i.e., the quality inspection node that represents the conclusion that the adhesion of this batch is unqualified) is decomposed in reverse to each process node and its adjacent edge on the path to obtain the risk contribution of each node, that is, the quantitative contribution value of the node deviation to the adhesion risk of this batch.

[0039] For example, the output of adhesion defects in a certain batch of graph neural networks is denoted as the target quality risk output y, and the feature representations of the v-th process node on the path before and after attention aggregation are denoted as h, respectively. v with h v ′.

[0040] First, during the backpropagation phase, the feature gradient of the node with respect to the target output is calculated, which can be defined as the gradient norm G of the node embedding vector with respect to y. v , Meanwhile, the attention weights α of each neighbor node for node v learned through the attention mechanism are... u→v Based on this, calculate the overall attention coefficient A of the node along the entire path. v : N(v) represents the set of neighbors of node v, and A v This is used to characterize the degree to which the neighbors and edges connected to node v are considered by the model during the risk information propagation process, and then the norm Δh of the feature difference before and after node aggregation is used. v , As a measure of the change in activation of the node in this forward inference, it reflects the strength of its participation in this adhesion risk representation. The risk contribution s of node v can be defined. v For s v =A v ×Gv ×Δh v .

[0041] The database is pre-configured with high contribution thresholds (i.e., the first contribution threshold) and low contribution thresholds (i.e., the second contribution threshold). In this example, if the risk contribution of the drying process node is greater than the high contribution threshold (e.g., >0.45), the system marks the drying process node as a dominant abnormal node, indicating that deviations in drying temperature, moisture control, etc., are the main causes of this adhesion problem. If the risk contribution of the cooling process node is between the high and low contribution thresholds (e.g., 0.25 to 0.45), it is marked as a secondary amplifying factor node, indicating that insufficient cooling amplifies the impact of upstream drying deviations on adhesion. If the risk contribution of the vacuum concentration process node is lower than the second contribution threshold (e.g., <0.25), it is marked as a weakly correlated influence node, indicating that its status has a certain background influence on this adhesion, but it is not a key target for rectification in this incident. After completing the above node role division, the system further associates the control parameters corresponding to various abnormal nodes in the knowledge graph. For example, the drying process is associated with parameters such as drying temperature setpoint, drying time, wind speed level, and online moisture threshold; the cooling process is associated with parameters such as cooling channel temperature, cooling time, and conveyor belt speed; and the vacuum concentration process is associated with parameters such as vacuum degree and sugar content at the concentration endpoint. Based on this, the system automatically generates a report on the causes of abnormalities in the production quality of gummy candies and corresponding control parameter optimization instructions.

[0042] During the generation phase, the system first determines whether the high-priority anomaly condition is met. The high-priority anomaly condition means that a certain node (such as the drying process node) is repeatedly marked as the dominant anomaly node in no less than a preset number of valid anomaly paths. For example, in the most recent statistical period, the drying node is the dominant anomaly node in no less than 10 valid anomaly paths. If the condition is met, the control parameters corresponding to the node are determined to play a dominant role in the quality risk of the current and subsequent batches. The system will uniformly mark the control parameters associated with the node as high-priority control parameters. For example, it will set stricter control windows for drying temperature and drying time, improve the sensitivity of online moisture alarms, and increase the frequency of sampling inspections in the drying section. The system will also generate corresponding control parameter optimization instructions, such as lowering the upper limit of drying temperature from 57℃ to 56℃ and increasing the lower limit of drying time from 26h to 28h. These will be highlighted or marked as high-priority anomalies in the anomaly cause analysis report to prompt process engineers to prioritize rectification. If the current statistical results do not meet the high-priority anomaly condition, that is, no node is stably the dominant anomaly node in a sufficient number of effective anomaly paths, the system will maintain the default priority of the control parameters corresponding to various anomaly nodes and only generate the gummy candy production quality anomaly cause analysis report and routine control parameter optimization suggestions according to the original regulatory strategy, such as using the existing thresholds and sampling inspection frequencies. This embodiment demonstrates that the present invention utilizes the attention and reverse attribution of graph neural networks at the front end to precisely identify the roles of various abnormal nodes, and at the back end, it drives the dynamic priority adjustment of control parameters through high-priority abnormal conditions, thus realizing an integrated closed loop from abnormal path identification and abnormal cause analysis to the generation of differentiated control instructions.

[0043] Suppose that the target gummy candy production batch B20251201 is identified by the online visual inspection system as having a severe adhesion problem during the final inspection stage, meaning that the proportion of individual gummy candies adhering to each other is significantly higher than the internal control limit during final inspection sampling. The system first locates the directly related nodes in the gummy candy production process knowledge graph that are directly connected to this quality problem: such as the final inspection adhesion defect detection node, the post-drying moisture detection node, and the pre-packaging appearance inspection node. These nodes are linked to the severe adhesion quality problem label through the quality inspection-to-quality result relationship in the graph, and are therefore considered as the related nodes of this abnormal quality problem in the knowledge graph. Subsequently, based on the pre-constructed gummy candy production process topology (e.g., a directed process link from vacuum concentration to casting to cooling to drying to coating to metal detection to packaging to final inspection), the system performs a reverse path search from downstream to upstream along the opposite direction of the directed edges, starting from downstream related nodes such as the final inspection adhesion defect detection node. That is, it backtracks step by step along the path from final inspection to packaging to drying to cooling to casting, connecting process nodes that are marked as slightly deviated nodes or abnormal candidate nodes in the current batch (e.g., nodes with low drying temperature, shortened drying time, and insufficient cooling time) to form several candidate backtracking paths. Combined with the path-level risk score output by the aforementioned graph neural network, when the overall risk score of a backtracking path exceeds the path validity threshold under the model's judgment, the path is identified as a valid backtracking abnormal path with a real causal relationship to the serious adhesion quality problem; for example, path P1: from the low drying temperature node to the slightly shortened drying time node to the pre-packaging appearance inspection node to the final inspection adhesion defect node is marked as a valid backtracking abnormal path.

[0044] After identifying all valid backtracking anomaly paths, the system analyzes the inducing and transmission factors for each path. Inducing factors refer to upstream process deviations identified as dominant anomaly nodes in the graph neural network's attention weights and reverse attribution analysis, such as a node where the drying temperature is consistently 2°C below the target value and greater than or equal to the historical maximum deviation. Transmission factors refer to midstream process nodes marked as secondary amplifying factors or weakly correlated influencing nodes, such as slightly shortened drying time across multiple time windows or decreased cooling efficiency due to high workshop humidity. These nodes may not be sufficient to cause adhesion on their own, but they significantly amplify the impact of upstream inducing factors. Based on these analysis results, inducing factor nodes and their subgraphs are used as positive samples for enhanced training, automatically increasing the feature weights of these nodes and the attention coefficients of their associated edges during training. Conversely, nodes that frequently appear in invalid paths but have low contribution and their connections are added to the training set as negative samples, automatically reducing the model's sensitivity to them during training. By employing the closed-loop iteration of statistically effective backtracking of abnormal paths—analyzing inducing / transmission factors—and incrementally updating the graph neural network model, the model can more accurately focus on key process combinations that truly have causal and amplifying effects in subsequent batches, thereby improving the accuracy and stability of abnormal path identification and quality risk prediction.

[0045] In Example 2, while keeping other conditions unchanged from Example 1, to further refine the identification of key nodes with dominant inducing and amplifying effects in the path, thereby facilitating the subsequent formulation of targeted process rectification measures and differentiated risk control strategies, the system further manages the abnormal paths that have been determined to be valid by length classification, dividing the abnormal paths into long abnormal paths and short abnormal paths: Specifically, the system counts the total number of nodes contained in each abnormal path. When the total number of nodes is greater than the preset total number of defined nodes in the database, which is obtained through statistical analysis based on the node scale of typical multi-process linkage abnormal paths in historical cases and is used to distinguish between single or a few process deviations and multi-process cascaded abnormalities, the path is marked as a long abnormal path; otherwise, it is marked as a short abnormal path.

[0046] Short anomaly paths typically involve only a limited number of critical processes, with relatively concentrated anomaly mechanisms and simple transmission links. Therefore, the effectiveness of the entire path can be assessed primarily through a path risk aggregation risk assessment strategy, focusing on the overall path risk score to avoid over-amplifying the interference of minor local fluctuations on the model's judgment at the node level. In contrast, long anomaly paths often span multiple processes and equipment nodes, exhibiting complex anomaly superposition and multi-stage transmission effects. Relying solely on the overall path score makes it difficult to distinguish the role and contribution of each node in the anomaly formation process. Therefore, for long anomaly paths, the system simultaneously executes node-level risk assessment strategies and path risk aggregation risk assessment strategies. When the node-level risk assessment strategy determines that a few nodes are invalid (i.e., the number of invalid nodes is less than the maximum allowed number of invalid nodes preset in the database), and the path risk aggregation risk assessment strategy determines that the overall path is valid, it indicates that the long anomaly path is valid. This ensures accurate identification while supporting the formulation of subsequent targeted process rectification and differentiated risk control strategies.

[0047] Example 3: Based on the unchanged conditions of Example 1 or Example 2, the structural form of the abnormal path is further expanded into a divergent path topology. Specifically, in Examples 1 and 2, the abnormal paths mainly exist as single process links, extending sequentially from the upstream abnormal node along the process topology to the downstream quality result node, with the path structure approximating a linear chain. However, in this example, when the system performs reverse or forward path searches in the gummy candy production process knowledge graph, it allows for divergent scenarios where a single upstream abnormal node simultaneously connects to multiple downstream process nodes or quality inspection nodes. For example, a drying temperature abnormal node may simultaneously point to multiple downstream nodes in the graph, such as a post-drying moisture detection node, a pre-coating appearance inspection node, and a pre-packaging buffer station node, thus forming a set of divergent abnormal paths with a tree-like or network structure.

[0048] When divergence occurs, the graph neural network executes a node-level risk assessment strategy to obtain the risk contribution of the root anomalous node and each branch node, determining whether the root anomalous node is the dominant anomalous node constituting the current quality problem. Simultaneously, it executes a path risk aggregation risk assessment strategy for each branch sub-path. If the root anomalous node is the dominant anomalous node, and the average path-level overall risk score of the branch sub-paths exceeds the maximum allowed value in the database, the system determines the anomalous path under this divergent structure as a valid anomalous path and retains the corresponding branch for subsequent analysis of inducing and propagating factors.

[0049] like Figure 2 The diagram shown is a schematic of an automated monitoring system for gummy candy production quality provided in this application embodiment, including: a map data mapping module, an anomaly path construction module, an anomaly path attribution module, and a database.

[0050] The database, developed by relevant staff, is used to store parameters involved in an automated quality monitoring system for gummy candy production.

[0051] The map data mapping module is connected to the abnormal path construction module; the abnormal path construction module is connected to the abnormal path attribution module; and the map data mapping module, the abnormal path construction module, and the abnormal path attribution module are all connected to the database.

[0052] The graph data mapping module is used to define the pattern and initial instantiation of the gummy candy production process knowledge graph based on the process specifications, equipment logic and quality standards, monitor the current gummy candy production batch, create a unique batch node in the gummy candy production process knowledge graph to identify the entire production chain, collect multimodal production data of the current gummy candy production batch, map it to the corresponding process node in the gummy candy production process knowledge graph and add attribute tags.

[0053] The abnormal path construction module is used to filter out the set of target nodes with abnormal attributes from the knowledge graph of gummy candy production process, construct abnormal paths based on the upstream and downstream process logic of gummy candy production, introduce a graph neural network model, and determine the effectiveness of abnormal paths through risk assessment. The anomaly path attribution module is used to input the batch process subgraphs corresponding to valid anomaly paths into a graph neural network. Through attention weight allocation and reverse attribution analysis, it locates various anomaly nodes corresponding to valid anomaly paths, associates the control parameters corresponding to various anomaly nodes, and automatically generates a gummy candy production quality anomaly cause analysis report and control parameter optimization instructions.

[0054] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope and intent of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A method for automated quality monitoring in gummy candy production, characterized in that, Includes the following steps: S1. Based on the process specifications, equipment logic and quality standards, complete the pattern definition and initial instantiation of the gummy candy production process knowledge graph, monitor the current gummy candy production batch, create a unique batch node in the gummy candy production process knowledge graph to identify the entire production chain, collect multimodal production data of the current gummy candy production batch, map it to the corresponding process node in the gummy candy production process knowledge graph and add attribute tags. S2. Select the set of target nodes with abnormal attributes from the knowledge graph of gummy candy production process, construct abnormal paths based on the upstream and downstream process logic of gummy candy production, introduce a graph neural network model, and determine the effectiveness of abnormal paths through risk assessment. S3. Input the batch process subgraph corresponding to the effective abnormal path into the graph neural network. Through attention weight allocation and reverse attribution analysis, locate the various abnormal nodes corresponding to the effective abnormal path, associate the control parameters corresponding to the various abnormal nodes, and automatically generate a gummy candy production quality abnormality cause analysis report and control parameter optimization instructions.

2. The automated quality monitoring method for gummy candy production as described in claim 1, characterized in that: The process of mapping the process nodes to the knowledge graph of gummy candy production process and adding attribute tags is as follows: Based on process specifications, equipment logic, and quality standards, a basic model of the knowledge graph for gummy candy production process is defined, including node types and relationship types. After the pattern is determined, the knowledge graph of gummy candy production process is initially instantiated based on historical production data or quality records; In the knowledge graph of gummy candy production process, the importance weight of each node and the association strength weight of each edge are marked respectively; Collect multimodal production data of the current gummy candy production batch within the workstation time window and perform noise reduction processing; For the current batch of gummies produced, a unique batch node is created in the gummies production process knowledge graph. If the batch node already exists, the node information is updated synchronously based on the real-time production progress. The denoised multimodal production data is written into the corresponding node attributes, and each node is attached with an attribute label, forming a structured form of single batch node association and workstation time window carrying multi-dimensional process snapshots. The attribute markers include normal node markers, slightly deviated node markers, or abnormal candidate node markers.

3. The automated quality monitoring method for gummy candy production as described in claim 1, characterized in that: The specific construction process for the abnormal path is as follows: The target node set labeled with slightly deviated node tags or abnormal candidate node tags is associated and constructed. Identify the upstream and downstream process relationships between each node, and connect the nodes with process relationships sequentially to form a path; During the association extension process, for target nodes that cannot continue to connect to upstream or downstream nodes in the process topology, the abnormal path naturally degenerates into a single-node abnormal path that only contains the node itself. For target nodes that can extend upstream or downstream along the process topology, an abnormal path covering multiple workstations is formed by continuously supplementing their upstream and downstream nodes. If a node has a slightly deviated node marker or an abnormal candidate node marker within a consecutive threshold number of workstation time windows, then the association strength weight between that node and the corresponding downstream node is increased.

4. The automated quality monitoring method for gummy candy production as described in claim 3, characterized in that: The abnormal path construction also includes: Count the total number of nodes in the abnormal path; Abnormal paths are classified based on the total number of nodes in the abnormal path; Based on the classification results, different types of abnormal paths are constructed.

5. The automated quality monitoring method for gummy candy production as described in claim 1, characterized in that: The validity of the abnormal path is determined through risk assessment, and the specific determination process is as follows: The graph neural network executes node-level risk assessment strategies and / or path risk aggregation risk assessment strategies to determine whether abnormal paths are valid. If the abnormal path is invalid, the multimodal features of the node corresponding to the abnormal path and the associated edge are written into the model training sample set as negative samples. If the abnormal path is valid, the batch process subgraph corresponding to the valid abnormal path is input into the graph neural network.

6. The automated quality monitoring method for gummy candy production as described in claim 5, characterized in that: The step of inputting the batch process subgraph corresponding to the effective abnormal path into the graph neural network also includes initializing the graph neural network. The specific initialization process is as follows: Collect data corresponding to valid abnormal paths and complete data cleaning; Calculate the total amount of data in the effective abnormal paths after cleaning; Based on the total data volume and quality supervision requirements, a path-level loss function weight vector is constructed and normalized; Based on the weight vector of the path-level loss function, the graph neural network is initialized and the model is validated until the model prediction error rate drops to within the preset error rate threshold.

7. The automated quality monitoring method for gummy candy production as described in claim 1, characterized in that: The specific process for locating various abnormal nodes corresponding to valid abnormal paths is as follows: Forward reasoning and backward attribution are performed on each node on the effective abnormal path by assigning attention weights. Calculate the risk contribution of each node and its adjacent edges to the predicted output of the target quality risk node; When the risk contribution of a node exceeds the first preset contribution threshold in the database, the node is marked as the dominant abnormal node. When the risk contribution of a node is between the first contribution threshold and the second contribution threshold, the node is marked as a secondary amplification factor node. When the risk contribution of a node is lower than the second contribution threshold, the node is marked as a weakly correlated influence node.

8. The automated quality monitoring method for gummy candy production as described in claim 1, characterized in that: The automated generation process for the gummy candy production quality anomaly cause analysis report and control parameter optimization instructions is as follows: If a high-priority abnormal condition exists, the control parameter corresponding to the node is determined to have a dominant role in the quality risk in the current batch. The control parameter associated with the node is marked as a high-priority control parameter, and a control parameter optimization instruction is generated. The high-priority abnormality is marked in the abnormality cause analysis report. If there are no high-priority abnormal conditions, the default priority of the control parameters corresponding to various abnormal nodes will be maintained, and only the original regulatory strategy will be used to generate the gummy candy production quality abnormality cause analysis report and control parameter optimization instructions. High-priority anomaly conditions: A node is marked as the dominant anomaly node in no fewer than a preset number of valid anomaly paths.

9. The automated quality monitoring method for gummy candy production as described in claim 1, characterized in that: The introduction of the graph neural network model also includes: when an abnormal quality problem is detected in the target gummy candy production batch, obtaining the associated nodes of the abnormal quality problem in the gummy candy production process knowledge graph; Perform a reverse path search along the process topology from downstream nodes to upstream nodes to determine at least one valid backtracking path corresponding to the abnormal quality problem; Each valid backtracking anomaly path is statistically identified, the inducing and transmission factors of each valid backtracking anomaly path are analyzed, and the graph neural network model is updated.

10. An automated quality monitoring system for gummy candy production, employing the automated quality monitoring method for gummy candy production as described in any one of claims 1-9, characterized in that, include: Graph data mapping module, abnormal path construction module, and abnormal path attribution module; The graph data mapping module is used to define the pattern and initial instantiation of the gummy candy production process knowledge graph according to the process specifications, equipment logic and quality specifications, monitor the current gummy candy production batch, create a unique batch node in the gummy candy production process knowledge graph to identify the entire production chain, collect multimodal production data of the current gummy candy production batch, map it to the corresponding process node in the gummy candy production process knowledge graph and add attribute tags. The abnormal path construction module is used to select a set of target nodes with abnormal attribute markers from the knowledge graph of gummy candy production process, construct abnormal paths based on the upstream and downstream process logic of gummy candy production, introduce a graph neural network model, and determine the effectiveness of abnormal paths through risk assessment. The abnormal path attribution module is used to input the batch process subgraph corresponding to the effective abnormal path into the graph neural network. Through attention weight allocation and reverse attribution analysis, it locates various abnormal nodes corresponding to the effective abnormal path, associates the control parameters corresponding to various abnormal nodes, and automatically generates a gummy candy production quality abnormality cause analysis report and control parameter optimization instructions.

Images