Cigarette quality defect root cause analysis method and system driven by multi-source heterogeneous data

Abstract

The invention discloses a multi-source heterogeneous data driven cigarette quality defect root cause analysis method and system, and relates to the technical field of industrial intelligent diagnosis. The method comprises the following steps: performing feature extraction and space-time alignment processing on multi-source data of a cigarette production line to generate a unified feature vector set; performing graph embedding mapping and topological relation construction processing on the unified feature vector set to obtain an initial quality defect knowledge graph; based on the initial quality defect knowledge graph, performing causal path reasoning processing by taking the defect node as a starting point, and outputting a plurality of candidate root cause propagation paths; visually displaying the plurality of candidate root cause propagation paths, receiving a user operation instruction, and generating a path validity feedback instruction; and performing weight adjustment processing on the initial quality defect knowledge graph according to the path validity feedback instruction, and generating an optimized root cause analysis result. Through the method, the technical effects of improving root cause positioning precision, enhancing working condition adaptability and reducing analysis delay can be achieved.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent diagnostic technology, and more specifically to a method and system for root cause analysis of cigarette quality defects driven by multi-source heterogeneous data. Background Technology

[0002] With the deep application of intelligent manufacturing technology in the tobacco industry, root cause analysis of cigarette quality defects has become a core element in ensuring product quality and improving production efficiency. Accurately locating the source of quality defects plays a decisive role in reducing raw material waste, optimizing process parameters, and minimizing equipment downtime.

[0003] In traditional technologies, defect attribution methods based on statistical process control, while applied to quality defect analysis, cannot handle the multi-factor coupling effects of equipment, process, and environment, leading to the failure of composite defects. Static knowledge graph reasoning technology, while constructing causal relationship networks, suffers from insufficient root cause recall due to the solidification of graph topology when faced with drifting equipment sensor data or sudden changes in environmental temperature and humidity. Furthermore, path verification mechanisms relying on human experience lack the ability to align multimodal data in time and space, resulting in low efficiency of human-machine collaboration, and the feedback results cannot optimize the analysis model in real time.

[0004] Therefore, how to improve the accuracy of root cause localization, enhance adaptability to operating conditions, and reduce analysis delay are problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a multi-source heterogeneous data-driven root cause analysis method and system for cigarette quality defects that overcomes or at least partially solves the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for root cause analysis of cigarette quality defects driven by multi-source heterogeneous data, including: Feature extraction and spatiotemporal alignment are performed on the equipment sensor time-series data, process parameter configuration data, quality inspection structured data, environmental monitoring records, and operation log text data of the cigarette production line to generate a unified feature vector set; The unified feature vector set is subjected to graph embedding mapping and topological relationship construction to obtain an initial quality defect knowledge graph; Based on the initial quality defect knowledge graph, causal path reasoning is performed starting from the defect node, and multiple candidate root cause propagation paths are output. The system visualizes the multiple candidate root cause propagation paths, receives user operation commands, and generates path validity feedback commands. The initial quality defect knowledge graph is weighted according to the path validity feedback instruction to generate optimized root cause analysis results.

[0007] Furthermore, the step of performing graph embedding mapping and topological relationship construction on the unified feature vector set to obtain an initial quality defect knowledge graph specifically includes: Based on the preset ontology of the cigarette manufacturing field, entity types are defined, including equipment component entities, process node entities, defect type entities, environmental factor entities, and operation behavior entities; The unified feature vector set is subjected to graph neural network encoding to generate entity attribute vectors, wherein each entity attribute vector corresponds to a dynamic state representation of the entity type. Calculate the multidimensional statistical correlation between the entity types based on the entity attribute vectors, and generate initial relation edges; By integrating the association constraints defined in the cigarette manufacturing ontology with the initial relation edges, dynamic graph topology synthesis processing is performed to generate the initial quality defect knowledge graph.

[0008] Furthermore, the step of performing causal path reasoning based on the initial quality defect knowledge graph, starting from the defect node, and outputting multiple candidate root cause propagation paths specifically includes: Starting from the defect node in the initial quality defect knowledge graph, multi-order reverse graph propagation processing is performed along the causal relationship edges to generate an initial causal propagation path set. Based on the time-varying stability coefficient of the relation edges in the initial quality defect knowledge graph, the initial causal propagation path set is dynamically pruned to generate a simplified propagation path set. For each path in the simplified propagation path set, the confidence score is calculated by integrating the path length weight, relation edge density weight, and entity activity weight to obtain the overall confidence score of the path. An adaptive confidence threshold is generated based on the matching degree between the severity level of the current quality defect event and historical cases. The paths are sorted in descending order based on their comprehensive confidence scores, and paths with confidence scores higher than the adaptive confidence threshold are selected as candidate root cause propagation paths.

[0009] Furthermore, the step of visualizing the multiple candidate root cause propagation paths, receiving user operation instructions, and generating path validity feedback instructions specifically includes: The multiple candidate root cause propagation paths are subjected to topological hierarchical compression to generate a hierarchical path display tree, where the root node is the quality defect type and the branch nodes are the root cause entities. In response to the user's focus command on any node in the hierarchical path display tree, the multimodal raw data of the corresponding entity in the initial quality defect knowledge graph are dynamically retrieved. The system uses a preset instruction semantic template to perform structured parsing of the user's annotation behavior on the multimodal raw data, and generates intermediate instructions for path verification. The path validity feedback instruction is generated by fusing the intermediate path verification instruction with the topological context of the currently focused path.

[0010] Furthermore, the step of performing structured parsing of the user's annotation behavior on the multimodal raw data using a preset instruction semantic template to generate intermediate instructions for path verification specifically includes: The user's original operation command is matched with a preset command semantic template to generate intent classification labels. Based on the intent classification label, the spatiotemporal sensing waveforms and process parameter slices in the multimodal raw data are associated to generate data anchoring tuples; Logical consistency verification is performed on the multimodal evidence in the data anchoring tuple to generate conflict verification results; The intent classification label, the data anchoring tuple, and the conflict verification result are fused and structured to generate the path verification intermediate instruction.

[0011] Furthermore, the step of fusing the path verification intermediate instructions with the topological context of the currently focused path to generate the path validity feedback instructions specifically includes: Perform topological feature extraction on the graph structure of the current focused path to generate a topological feature vector; The path verification intermediate instructions and the topological feature vector are processed through a graph attention masking mechanism to perform semantic-topological collaborative reasoning to generate a collaborative reasoning result. The collaborative reasoning results are subjected to dynamic instruction casting processing using conflict resolution rules to generate structured instruction tuples. The structured instruction tuple is standardized and encapsulated to generate the path validity feedback instruction.

[0012] Furthermore, the step of adjusting the weights of the initial quality defect knowledge graph based on the path validity feedback instruction to generate optimized root cause analysis results specifically includes: Based on the confirmation instruction in the path validity feedback instruction, the relation edges in the initial quality defect knowledge graph are weighted and enhanced to generate a set of weighted enhanced relation edges. Based on the supplementary instructions in the path validity feedback instructions, graph insertion processing is performed on the newly added relation tuples in the supplementary instructions to generate a relation-enhanced knowledge graph. Incremental graph learning is performed on the weighted enhanced relation edge set and the relation-enhanced knowledge graph to generate an optimized knowledge graph. Root cause regression calculation is performed based on the optimized knowledge graph to generate the optimized root cause analysis results.

[0013] Secondly, the present invention provides a multi-source heterogeneous data-driven root cause analysis system for cigarette quality defects, applying the aforementioned multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects; the system includes: The feature alignment module is used to perform feature extraction and spatiotemporal alignment processing on the equipment sensor time-series data, process parameter configuration data, quality inspection structured data, environmental monitoring records and operation log text data of the cigarette production line, and generate a unified feature vector set. The graph construction module is used to perform graph embedding mapping and topological relationship construction processing on the unified feature vector set to obtain an initial quality defect knowledge graph. The path reasoning module is used to perform causal path reasoning processing based on the initial quality defect knowledge graph, starting from the defect node, and output multiple candidate root cause propagation paths. The interactive feedback module is used to visualize the multiple candidate root cause propagation paths, receive user operation instructions, and generate path validity feedback instructions. The optimization analysis module is used to adjust the weights of the initial quality defect knowledge graph according to the path validity feedback instruction, and generate optimized root cause analysis results.

[0014] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for multi-source heterogeneous data-driven root cause analysis of cigarette quality defects, which has the following beneficial effects: This invention extracts features and aligns them spatiotemporally to multi-source heterogeneous data from cigarette production lines, generating a unified feature vector set. This eliminates differences in format and spatiotemporal dimensions among different data, laying a solid data foundation for subsequent correlation of quality defects and potential influencing factors, thus improving the accuracy of root cause localization. It also reduces redundant operations in the data preprocessing stage, initially compressing analysis time and reducing latency. Based on the unified feature vector set, an initial quality defect knowledge graph is constructed using graph embedding technology combined with data correlation. This breaks free from the constraints of static topology. When equipment sensor data drifts or environmental conditions change abruptly, such as sudden changes in temperature and humidity, the graph can dynamically reflect the relationships between entities based on real-time data correlation, effectively alleviating the problem of insufficient root cause recall and enhancing adaptability to operating conditions.

[0017] This invention uses defect nodes as the starting point to perform causal path reasoning in the initial graph, outputting candidate root cause propagation paths. This focuses on the core causal chain between defects and root causes, filtering out irrelevant interference factors, making root cause localization more targeted and further improving accuracy. The candidate paths are visualized and user operation commands are received to generate feedback. This not only leverages user expertise to correct algorithmic reasoning biases and further optimize root cause localization accuracy, but also shortens the time required for manual verification and feedback through intuitive visual interaction, reducing time consumption in the analysis process.

[0018] This invention adjusts the weights of the initial map based on feedback instructions to generate optimized root cause analysis results. The map can be dynamically updated as actual verification progresses to adapt to changes in operating conditions, enhancing its adaptability. Furthermore, the incremental weight adjustment does not require map reconstruction, significantly reducing update time. Combined with optimization of preceding steps, this achieves a synergistic effect of improved root cause localization accuracy, enhanced operating condition adaptability, and reduced analysis latency. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects provided in this embodiment of the invention.

[0021] Figure 2 This is a flowchart illustrating the process of generating intermediate instructions for path verification by performing structured parsing of user annotation behavior on multimodal raw data using a preset instruction semantic template, as provided in this embodiment of the invention.

[0022] Figure 3This is a schematic diagram of the framework of a multi-source heterogeneous data-driven root cause analysis system for cigarette quality defects provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0024] First, the application scenarios of the embodiments of the present invention will be described. In the embodiments of the present invention, a multi-source heterogeneous data-driven method and system for root cause analysis of quality defects in cigarettes is provided, applicable to, but not limited to, the entire process of cigarette production. For example, in the stages of cigarette production such as tobacco processing, cigarette making, and packaging, when quality defects such as cigarette weight deviation, empty ends, and air leakage occur, this solution can be used to analyze information such as equipment operating data, process setting parameters, workshop temperature and humidity records, quality inspection results, and operation records to locate the root cause of the defect.

[0025] As an illustration, the multi-source heterogeneous data-driven root cause analysis method and system for cigarette quality defects provided in this embodiment of the invention can also be applied to other quality control scenarios related to cigarette production, such as tracing the source of quality anomalies in the raw material pretreatment stage and analyzing the root cause of quality changes in the finished product storage stage. This is only an example and does not limit the specific application scenarios.

[0026] like Figure 1 As shown, this embodiment of the invention provides a multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects, the method comprising: S101: Perform feature extraction and spatiotemporal alignment processing on the equipment sensor timing data, process parameter configuration data, quality inspection structured data, environmental monitoring records, and operation log text data of the cigarette production line to generate a unified feature vector set; For example, time-series feature extraction is performed on the equipment sensor time-series data of the cigarette production line to capture the changing trends and fluctuations of the data as the production process progresses; parameter feature extraction is performed on the process parameter configuration data to extract the setting attributes and adjustment features of the parameters; detection feature extraction is performed on the structured data of quality inspection to obtain the numerical and status features of the detection indicators; environmental feature extraction is performed on the environmental monitoring records to extract the changing and range features of the environmental parameters; and text feature extraction is performed on the text data of the operation log to mine the operational behavior features and descriptive features in the log.

[0027] Based on the above, using the production cycle time and key process time nodes of the cigarette production line as a benchmark, the extracted features are calibrated in terms of time dimension to ensure consistency in time stamps across different data sources. Spatial dimension associations are then established based on the corresponding production line equipment units, process steps, and inspection points to ensure spatial correspondence of features. Through this feature extraction and spatiotemporal alignment process, the features of various data types are integrated into vectors with a unified format and dimension, generating a unified feature vector set.

[0028] S102: Perform graph embedding mapping and topological relationship construction on the unified feature vector set to obtain the initial quality defect knowledge graph; For example, based on a unified feature vector set, the entity types corresponding to each vector in the unified feature vector set are analyzed, including equipment component entities, process node entities, defect type entities, environmental factor entities, and operational behavior entities. Semantic transformation of the unified feature vectors is performed using graph embedding technology, mapping the feature information to knowledge nodes carrying dynamic entity attributes. The data association features of the unified feature vector set are analyzed to determine various associations between entities corresponding to knowledge nodes, thereby constructing relational edges reflecting the strength and nature of these associations. All knowledge nodes and relational edges are integrated to form a structured network containing entities, attributes, and associations, generating an initial quality defect knowledge graph.

[0029] S103: Based on the initial quality defect knowledge graph, perform causal path reasoning processing starting from the defect node, and output multiple candidate root cause propagation paths; For example, based on an initial quality defect knowledge graph, the defect node corresponding to the current quality problem is located in the initial quality defect knowledge graph, determining the starting point for causal path reasoning. Starting from the defect node, tracing back along the causal relationship edges in the initial quality defect knowledge graph, layer by layer, upstream nodes that may affect the defect node are associated, generating multi-level causal propagation links. Combining the causal strength of the relationship edges in the initial quality defect knowledge graph with the rationality constraints of node associations, propagation links are screened, eliminating path fragments with logical contradictions or weak associations. Path features are extracted from the retained complete causal propagation links, and path effectiveness is evaluated. The causal propagation links that pass the evaluation are summarized, outputting multiple candidate root cause propagation paths.

[0030] S104: Visualize multiple candidate root cause propagation paths, receive user operation instructions, and generate path validity feedback instructions; For example, multiple candidate root cause propagation paths are structured and visualized to present the node sequences, node types, and inter-node relationship attributes of each path, generating an interactive path display interface. This interface receives user operation instructions for each candidate root cause propagation path, including confirmation of path validity, suggestions for correcting nodes or relationship edges within the path, and supplementary instructions for adding new relationships. The received user operation instructions are semantically parsed, extracting the path identifier, operation type, and associated node or relationship edge information. This information is then integrated with the structural features of the corresponding candidate root cause propagation path to form a structured instruction conforming to a preset format, generating a path validity feedback instruction.

[0031] S105: Adjust the weights of the initial quality defect knowledge graph based on the path validity feedback instruction to generate optimized root cause analysis results.

[0032] For example, the path validity feedback instruction is parsed to extract the weight adjustment requirements, association corrections, and supplementary information related to the initial quality defect knowledge graph. Based on this, the weights of corresponding relationship edges in the initial quality defect knowledge graph are adjusted, strengthening the weights of relationship edges on valid paths and reducing the weights of relationship edges on invalid paths, while simultaneously updating the node association status. The adjusted initial quality defect knowledge graph undergoes consistency verification to ensure it conforms to the principles of cigarette production quality analysis. Based on the verified adjusted initial quality defect knowledge graph, root cause analysis is performed again to generate optimized root cause analysis results.

[0033] The multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects provided in this invention extracts features and aligns them spatiotemporally from multi-source heterogeneous data from cigarette production lines, generating a unified feature vector set. This eliminates differences in format and spatiotemporal dimensions among different data, laying a solid data foundation for subsequent association of quality defects and potential influencing factors, thus improving the accuracy of root cause localization. It also reduces redundant operations in the data preprocessing stage, initially compressing analysis time to reduce latency. Based on the unified feature vector set, an initial quality defect knowledge graph is constructed using graph embedding technology combined with data correlation. This breaks free from the constraints of static topology. When equipment sensor data drifts or environmental conditions change abruptly, such as sudden changes in temperature and humidity, the graph can dynamically reflect the relationships between entities based on real-time data correlation, effectively alleviating the problem of insufficient root cause recall and enhancing adaptability to operating conditions.

[0034] Starting with defective nodes, causal path reasoning is performed in the initial graph, outputting candidate root cause propagation paths. This focuses on the core causal chain between defects and root causes, filtering out irrelevant interference factors, making root cause localization more targeted and further improving accuracy. Visualizing candidate paths and receiving user input to generate feedback not only leverages user expertise to correct algorithmic reasoning biases and further optimize root cause localization accuracy, but also shortens manual verification and feedback time through intuitive visual interaction, reducing time consumption in the analysis process.

[0035] The initial map is weighted according to the feedback instructions to generate optimized root cause results. The map can be dynamically updated as the actual verification progresses to adapt to changes in operating conditions, thereby enhancing its adaptability. Furthermore, the incremental weight adjustment does not require map reconstruction, which greatly reduces update time. Combined with the optimization of the preceding steps, this achieves a synergistic effect of improved root cause localization accuracy, enhanced adaptability to operating conditions, and reduced analysis latency.

[0036] Next, steps S102-S105 will be explained in detail.

[0037] In step S102 above, the unified feature vector set is processed by graph embedding mapping and topological relationship construction to obtain an initial quality defect knowledge graph; specifically including: (1) Based on the preset cigarette manufacturing domain ontology, define entity types, including equipment component entities, process node entities, defect type entities, environmental factor entities, and operation behavior entities.

[0038] For example, based on a pre-defined ontology of the cigarette manufacturing domain, the conceptual categories of cigarette production entities in the ontology are parsed and mapped into five entity types: equipment component entities, process node entities, defect type entities, environmental factor entities, and operational behavior entities. The attribute scope of each entity type is clarified. For instance, equipment component entities include cigarette making machine components, tobacco processing equipment components, and packaging equipment components; process node entities include tobacco processing process nodes, cigarette making process nodes, and packaging process nodes; defect type entities include empty cigarette ends, weight deviations, and packaging damage; environmental factor entities include workshop temperature and humidity, dust concentration, etc.; and operational behavior entities include equipment operation behaviors and process adjustment behaviors, thus completing the definition of the entity types.

[0039] Among them, the ontology of the cigarette manufacturing domain is a pre-constructed domain knowledge system. The attribute range of the entity type corresponds to the entity concepts of five dimensions in cigarette production: equipment, process, defect, environment, and operation.

[0040] (2) Perform graph neural network encoding on the unified feature vector set to generate entity attribute vectors, where each entity attribute vector corresponds to a dynamic state representation of an entity type.

[0041] For example, the vectors in the unified feature vector set are classified according to the categories of equipment component entities, process node entities, defect type entities, environmental factor entities, and operational behavior entities. The classified vectors are then input into a graph neural network model. Through the model's feature extraction and vector transformation functions, vector representations that reflect the real-time parameters, historical trends, and attribute association features of the entities are generated, i.e., entity attribute vectors.

[0042] Graph neural network models are used to uncover spatial relationships and feature patterns in vectors. The dynamic state representation of entity attribute vectors includes features such as entity parameter values, change patterns, and associated attributes.

[0043] (3) Calculate the multidimensional statistical correlation between entity types based on entity attribute vectors and generate initial relation edges.

[0044] For example, when calculating the multidimensional statistical correlation between entity types based on entity attribute vectors, multiple correlation indicators such as Pearson correlation coefficient, mutual information, and cosine similarity are selected. For attribute vectors of different entity types, such as equipment component entities and process node entities, equipment component entities and defect type entities, and process node entities and environmental factor entities, the selected correlation indicators are used to calculate the strength and type of association between entity types. Initial relation edges are generated based on the calculation results. Each initial relation edge contains the associated entity type, the association strength value, and the association type identifier.

[0045] The calculation of multidimensional statistical correlation covers various association analysis methods. The association types of the initial relation edges include causal association, influence association, and synergistic association, which are used to represent the interaction relationship between entity types.

[0046] (4) Integrate the association constraints and initial relation edges defined in the ontology of cigarette manufacturing, perform dynamic graph topology synthesis processing, and generate an initial quality defect knowledge graph.

[0047] For example, the process integrates the association constraints and initial relation edges defined in the cigarette manufacturing ontology. This includes parsing normal association rules between entity types in the cigarette manufacturing ontology, such as the necessary association constraints between equipment component entities and their corresponding process node entities, and the influence range constraints of environmental factor entities on process node entities. The aforementioned association constraints are compared with the initial relation edges. Initial relation edges that conform to the constraints are retained and their weights are adjusted, while initial relation edges that do not conform to the constraints are corrected or deleted. Based on the processed relation edges and entity types, a topology containing nodes and edges is constructed, and dynamic graph topology synthesis processing is performed to generate an initial quality defect knowledge graph.

[0048] In this context, the association constraints of the ontology in the cigarette manufacturing field are rules formulated based on industry knowledge and process standards. Dynamic graph topology synthesis is the process of organizing entities and relational edges into a structured network. The initial quality defect knowledge graph is a structured knowledge network containing entities, attributes, and relations.

[0049] In step S103 above, based on the initial quality defect knowledge graph, causal path reasoning is performed starting from the defect node, outputting multiple candidate root cause propagation paths; specifically including: (1) Starting from the defect node in the initial quality defect knowledge graph, perform multi-level reverse graph propagation along the causal relationship edge to generate the initial causal propagation path set.

[0050] For example, starting with the defect node in the initial quality defect knowledge graph, the reverse direction of causal propagation is determined. A first-order reverse traversal is performed along the causal relationship edges already constructed in the initial quality defect knowledge graph, identifying all upstream nodes directly causally related to the defect node. Using these first-order upstream nodes as new starting points, a second-order reverse traversal is performed, identifying nodes directly influencing the first-order upstream nodes, and so on, expanding to multiple orders of upstream nodes until the traversal depth reaches the preset maximum order. During each traversal, the traversal order of the nodes and the corresponding causal relationship edges are recorded, forming a complete propagation chain. All chains are then summarized to generate an initial set of causal propagation paths.

[0051] The preset maximum order is determined based on the common length of causal chains for cigarette production quality defects. Causal relationship edges are the edge elements in the initial quality defect knowledge graph representing causal connections between nodes. The initial set of causal propagation paths is a set containing all traversed generated links.

[0052] (2) Based on the time-varying stability coefficient of the relation edges in the initial quality defect knowledge graph, the initial causal propagation path set is dynamically pruned to generate a simplified propagation path set.

[0053] For example, based on the time-varying stability coefficient of each relation edge in the initial quality defect knowledge graph, each path in the initial causal propagation path set is examined one by one for all its relation edges, and relation edges with time-varying stability coefficients lower than a preset stability threshold are identified. Paths containing such relation edges are determined to be unstable paths and are deleted, retaining only paths where the time-varying stability coefficients of all relation edges meet the preset stability threshold requirement, thus generating a simplified propagation path set.

[0054] The time-varying stability coefficient reflects the temporal stability of causal relationships during cigarette production. The preset stability threshold is set based on historical stability data of the cigarette production process and expert experience. The simplified propagation path set is the set of paths after stability screening.

[0055] (3) Using the following formula, calculate the confidence score for each path in the simplified propagation path set by integrating the path length weight, relation edge density weight, and entity activity weight, and obtain the overall confidence score of the path:

[0056] in, This indicates the overall confidence score of the path. Indicates the path length; Represents the set of edges related to a path; Indicates the time-varying stability coefficient of the relation edge; Represents the edge variable of the relation; Indicates the number of nodes in the path; Indicates the defect detection timestamp; This represents the average last active time of all entities in the path; Indicates the preset length weight; Indicates the preset edge density weight; This indicates the preset entity activity weight.

[0057] For example, for each path in the simplified propagation path set, the path length weight, relation edge density weight, and entity activity weight are calculated separately. When calculating the path length weight, the score is obtained by taking the square root of the inverse of the path length, based on the path length. When calculating the relation edge density weight, the time-varying stability coefficients of each relation edge in the path relation edge set are summed, and this sum is divided by the number of path nodes to obtain the score corresponding to the relation edge density. When calculating the entity activity weight, the defect detection timestamp and the entity's last active time are obtained, and the time difference between the two is calculated. The inverse of this time difference is used to obtain the score corresponding to the entity activity. Then, according to the preset coefficient ratios of the length weight, the preset edge density weight, and the preset entity activity weight, the scores of the three are merged to obtain the comprehensive confidence score for each path, i.e., the path comprehensive confidence score.

[0058] The path relation edge set is the set of all relation edges in the path. The defect detection timestamp is the time stamp when a quality defect was detected. The entity last active time is the time when the entity node in the path last generated data. The path comprehensive confidence score is the comprehensive result after integrating the three weighted scores.

[0059] (4) Generate an adaptive confidence threshold based on the severity level of the current quality defect event and the matching degree of historical cases.

[0060] For example, the characteristics of current quality defect events are analyzed, and the severity level of the current quality defect events is classified according to the degree of impact of the defects on the quality of cigarette products and the production process. Simultaneously, key features of the current quality defect events are extracted and compared with features of historical quality defect cases to calculate the historical case matching degree. Combining the severity level with the historical case matching degree, a confidence threshold applicable to the current quality defect event is dynamically generated, i.e., an adaptive confidence threshold.

[0061] The severity level is determined by referencing industry-wide grading standards for cigarette quality defects. Historical case matching refers to the similarity between the current defect and historical defect cases. The adaptive confidence threshold is a judgment standard that is dynamically adjusted based on the current event.

[0062] (5) Sort the paths in descending order based on the comprehensive confidence score, and select the paths with confidence scores higher than the adaptive confidence threshold as candidate root cause propagation paths.

[0063] For example, the overall confidence score of each path in the simplified propagation path set is obtained, and the paths are sorted in descending order of score. The overall confidence score of each path is compared with the adaptive confidence threshold, and the paths with an overall confidence score higher than the adaptive confidence threshold are selected. These paths are then organized into a set, and multiple candidate root cause propagation paths are output.

[0064] The descending order sorting arranges the paths from highest to lowest overall confidence score. Multiple candidate root cause propagation paths form the final set of candidate paths selected for subsequent analysis.

[0065] In step S104 above, multiple candidate root cause propagation paths are visualized and user operation instructions are received, generating path validity feedback instructions; specifically including: (1) Perform topological hierarchical compression on multiple candidate root cause propagation paths to generate a hierarchical path display tree, where the root node is the quality defect type and the branch nodes are root cause entities.

[0066] For example, when performing topological hierarchical compression on multiple candidate root cause propagation paths, the process includes dividing the path into layers based on the propagation order of the causal relationship in each path, starting with the quality defect type, so that root cause entities at the same causal propagation stage are grouped into the same layer. Root cause entity nodes that appear repeatedly in different candidate root cause propagation paths are merged, and redundant nodes in the path display are removed. Based on the hierarchical results and deduplication, a tree structure is constructed with the quality defect type as the root node and root cause entities at each level as branch nodes, generating a hierarchical path display tree.

[0067] The basis for topological hierarchical compression processing includes the causal propagation order of the candidate root cause propagation path, and the node composition of the hierarchical path display tree includes the quality defect type as the root node and the root cause entity as the branch node.

[0068] (2) In response to the user’s focus command on any node in the hierarchical path display tree, dynamically retrieve the multimodal original data of the corresponding entity in the initial quality defect knowledge graph.

[0069] For example, in response to a user's focus command on any node in the hierarchical path display tree, the system includes capturing the node identifier pointed to by the focus command and locating the corresponding root cause entity or quality defect type entity in the initial quality defect knowledge graph using the node identifier. Based on the located entity type, the system retrieves multimodal raw data associated with that entity from the initial quality defect knowledge graph, including equipment sensor time-series data, process parameter configuration data, structured quality inspection data, environmental monitoring records, and operation log text data. The retrieved multimodal raw data is presented in an interactive format for user viewing and analysis.

[0070] The scope of multimodal raw data includes equipment sensing time-series data, process parameter configuration data, quality inspection structured data, environmental monitoring records, and operation log text data associated with the entity corresponding to the focus node.

[0071] (3) The user’s annotation behavior on the multimodal raw data is structured and parsed through the preset instruction semantic template to generate intermediate instructions for path verification.

[0072] For example, the specific types of user annotation behaviors are identified, including confirming the validity of candidate root cause propagation paths, negating entity associations in the path, and supplementing newly added entity associations. The annotation behaviors are matched with preset instruction semantic templates to extract candidate root cause propagation path identifiers, target entity information, and annotation intents involved in the annotation behaviors. The extracted information is then structured and encapsulated according to a preset data format to form instruction content containing path identifiers, entity associations, and annotation intents, generating intermediate path verification instructions.

[0073] The preset instruction semantic template is used to match user annotation behavior and extract key information. The content of the intermediate instructions for path verification includes candidate root cause propagation path identifiers, target entity information and annotation intent.

[0074] like Figure 2 As shown, the user's annotation behavior on multimodal raw data is structurally parsed using a preset instruction semantic template to generate intermediate instructions for path verification, including: S201: Use a preset instruction semantic template to perform intent matching processing on the user's original operation instruction and generate intent classification labels.

[0075] For example, a preset instruction semantic template is retrieved, which contains preset intent categories and corresponding instruction element matching rules. The key elements of the user's original operation instruction are broken down, including the operation object, operation action, and associated data identifier. The broken-down key elements are compared one by one with the instruction element matching rules in the preset instruction semantic template to determine the preset intent category to which the user's original operation instruction belongs. Based on the matching results, a corresponding category identifier is assigned to the user's original operation instruction, generating an intent classification label.

[0076] The preset intent categories include path confirmation, path negation, data supplementation, and association correction. The intent classification labels are standardized labels used to identify user operation intents.

[0077] S202: Based on intent classification labels, associate spatiotemporal sensing waveforms and process parameter slices in multimodal raw data to generate data anchoring tuples.

[0078] For example, the associated requirements are determined based on intent classification labels, with different intent classification labels corresponding to different ranges of associated data. Target data is filtered from the multimodal raw data based on the associated requirements. Target data includes spatiotemporal sensing waveforms related to the intent and corresponding process parameter slices. The spatiotemporal sensing waveforms must match the time interval and equipment identifier corresponding to the labeled behavior, and the process parameter slices must match the process step and parameter type corresponding to the labeled behavior. Feature identifiers of the filtered spatiotemporal sensing waveforms and parameter identifiers of the process parameter slices are extracted. Feature identifiers include the waveform acquisition time and the equipment to which the waveform belongs; parameter identifiers include the parameter configuration time and the process step to which the parameter belongs. The spatiotemporal sensing waveforms, process parameter slices, and their corresponding feature and parameter identifiers are combined to form a structured data group containing data entities and data identifiers, generating a data anchoring tuple.

[0079] Among them, the data anchoring tuple is a structured data set containing spatiotemporal sensing waveforms, process parameter slices and corresponding data identifiers. The association requirements are determined by the category attributes of the intent classification label, such as the path confirmation label corresponding to the spatiotemporal sensing waveforms and process parameter slices related to the associated defects.

[0080] S203: Perform logical consistency verification on the multimodal evidence in the data anchor tuple and generate conflict verification results.

[0081] For example, multimodal evidence is extracted from the data anchoring tuples. This multimodal evidence includes equipment status information reflected by spatiotemporal sensing waveforms and parameter configuration information reflected by process parameter slices. Logical consistency verification dimensions are determined, including temporal consistency, causal correlation, and parameter range dimensions. The temporal consistency dimension verifies the temporal matching between the spatiotemporal sensing waveforms and process parameter slices; the causal correlation dimension verifies the causal logic of the equipment status information and parameter configuration information; and the parameter range dimension verifies whether the process parameter slices conform to the preset process standard range. The multimodal evidence is verified one by one according to the verification rules corresponding to each verification dimension, and the results of passing or failing the verification and the corresponding reasons are recorded. The verification results of each dimension are summarized to form a verification report containing the verification dimension, verification result, and conflict reason, generating conflict verification results.

[0082] The verification rules include time sequence matching rules, causal logic rules, and parameter standard rules. The conflict verification result is a comprehensive verification conclusion that reflects the logical rationality of multimodal evidence.

[0083] S204: Integrate intent classification labels, data anchoring tuples, and conflict verification results into a structured encapsulation to generate intermediate instructions for path verification.

[0084] For example, a structured encapsulation format is determined, which includes fixed fields: an intent label field, a data anchoring field, a validation result field, and an encapsulation time field. The intent category label is filled into the intent label field, the complete content of the data anchoring tuple is filled into the data anchoring field, the validation report of the conflict validation result is filled into the validation result field, and the time information of the current encapsulation operation is recorded and filled into the encapsulation time field. The content of each field is formatted to ensure that the data within each field has a consistent format and clear semantics, avoiding ambiguity. All field content is integrated according to the structured encapsulation format to form instruction text with machine-readable attributes, generating intermediate instructions for path validation.

[0085] Among them, the structured encapsulation format is a preset standardized data format used to uniformly integrate various types of information, and the path verification intermediate instruction is a structured instruction containing user operation intent, associated data, and verification results, used to generate subsequent path validity feedback instructions.

[0086] (4) Integrate the intermediate instructions for path verification with the topological context of the current focused path to generate a path validity feedback instruction.

[0087] For example, the topological context information of the current focused path is extracted, including the node hierarchy distribution of the current focused path in the hierarchical path display tree, the relationship edge attributes between nodes, and the association logic between this path and other candidate root cause propagation paths. The annotation intent and entity association relationships in the path verification intermediate instructions are correlated and verified with the topological context information to ensure the consistency between the annotation intent and the topological logic. Based on the verification results, the path verification intermediate instructions are supplemented and improved to form a structured instruction that meets the adjustment requirements of the initial quality defect knowledge graph, and a path validity feedback instruction is generated.

[0088] The topological context of the current focused path includes the node hierarchy distribution of the current focused path, the relational edge attributes between nodes, and the association logic with other candidate root cause propagation paths.

[0089] In one embodiment, the path validation intermediate instructions are fused with the topological context of the currently focused path to generate a path validity feedback instruction, including: (1) Extract topological features from the graph structure of the current focused path to generate a topological feature vector.

[0090] For example, the graph structure of the current focused path is decomposed into its constituent elements, including the node types, the attributes of the edges between nodes, the hierarchical distribution of the path, and the connection density of nodes within the path. Feature quantization is performed on each constituent element, mapping node types to preset category codes, converting edge attributes into standardized numerical features, and statistically analyzing the hierarchical distribution and connection density into structural feature indicators. The quantized node features, edge features, and structural features are then integrated according to a preset feature dimension order. Vector encoding techniques are used to transform the integrated features into a fixed-dimensional vector form, generating a topological feature vector.

[0091] The node types in the graph structure of the current focused path include quality defect nodes, equipment component nodes, process parameter nodes, environmental factor nodes, and operational behavior nodes. The attributes of the relationship edges include causal association strength and association type, specifically direct causal association and indirect influence association. The fixed dimensions of the topological feature vector include three main categories: node feature dimension, edge feature dimension, and structural feature dimension.

[0092] (2) The intermediate instructions for path verification and the topological feature vector are processed by semantic-topological collaborative reasoning through graph attention masking mechanism to generate collaborative reasoning results.

[0093] For example, the semantic information of intermediate instructions in the path verification is parsed, and the intent classification labels, entity identifiers corresponding to data anchor tuples, and logical consistency conclusions in the conflict verification results are extracted. The extracted semantic information is transformed into semantic feature vectors, which contain intent encoding, entity identifier encoding, and logical conclusion encoding. The semantic feature vectors and topological feature vectors are input into a graph attention masking mechanism model. This model assigns attention weights to different feature dimensions in the two types of vectors based on the correlation between semantic and topological information, emphasizing features related to root cause analysis and weakening interference from irrelevant features. Based on the weighted feature vectors, collaborative reasoning is performed to determine the matching degree between semantic intent and topological structure, derive association conclusions that conform to the logic of cigarette quality defect analysis, and generate collaborative reasoning results.

[0094] The semantic information of the intermediate instructions for path verification includes intent classification labels such as path confirmation labels, path negation labels, supplementary association labels, and corrective association labels. The core function of the graph attention masking mechanism is to associate semantic and topological information through attention weight allocation. The collaborative reasoning results include semantic-topological matching conclusions, key association feature identifiers, and reasoning credibility scores.

[0095] (3) Apply conflict resolution rules to the collaborative reasoning results to perform dynamic instruction casting processing and generate structured instruction tuples.

[0096] For example, the collaborative reasoning results are traversed to identify conflicting information, specifically inconsistencies between semantic intent and topological logic, and conflicts between conclusions from different reasoning dimensions. Pre-defined conflict resolution rules are retrieved; these rules are formulated in conjunction with cigarette production process standards and historical root cause analysis experience. The conflicting information is then reconciled according to the conflict resolution rules, correcting contradictory reasoning conclusions to ensure that the processed conclusions conform to the actual logic of cigarette production quality control. The reconciled reasoning conclusions are then organized according to preset fields of structured instruction tuples to generate structured instruction tuples.

[0097] The conflict information in the collaborative reasoning results includes semantic-topological logic conflicts and reasoning credibility conflicts. The pre-defined conflict resolution rules specifically prioritize user-annotated intent over machine reasoning conclusions, cigarette production process specifications over general topological association logic, and high-credibility reasoning conclusions over low-credibility conclusions. The pre-defined fields of the structured instruction tuples include a reasoning conclusion field, a conflict resolution result field, and a key entity association field.

[0098] (4) Standardize and encapsulate the structured instruction tuples to generate path validity feedback instructions.

[0099] For example, a standardized encapsulation format for path validity feedback instructions is determined, which includes fixed fields. The complete content of a structured instruction tuple is filled into the structured instruction tuple field, and a unique instruction identifier is generated and filled into the instruction unique identifier field. The instruction unique identifier is generated based on a combination of a timestamp and a random code. The current operation time is recorded and filled into the instruction generation time field, and the root cause analysis scenario corresponding to the instruction is marked and filled into the instruction applicable scenario field. The format of each field is validated to ensure that the field information is free of syntactic errors, semantically clear, and conforms to machine-readable specifications. After successful validation, all field contents are integrated according to the standardized encapsulation format to form a unified instruction text, generating the path validity feedback instruction.

[0100] The standardized encapsulation format of path validity feedback instructions must conform to industrial data interaction specifications. Fixed fields include a unique instruction identifier field, a structured instruction tuple field, an instruction generation time field, and an instruction applicable scenario field. Applicable scenarios include defect analysis in the coiling process, defect analysis in the packaging process, and defect analysis in the yarn-making process. The purpose of path validity feedback instructions is for subsequent weight adjustment processing of the initial quality defect knowledge graph.

[0101] In step S105 above, the initial quality defect knowledge graph is weighted according to the path validity feedback instruction to generate optimized root cause analysis results; specifically including: (1) Based on the confirmation instruction in the path validity feedback instruction, the relation edges in the initial quality defect knowledge graph are weighted and enhanced to generate a set of weighted relation edges.

[0102] For example, based on the confirmation instruction in the path validity feedback instruction, the content of the confirmation instruction is parsed to extract the target path identifier and the corresponding list of relational edges contained therein. The associated path structure is located in the initial quality defect knowledge graph using the target path identifier. The specific relational edges in the initial quality defect knowledge graph that require weight adjustment are matched according to the relational edge list. The current base weight values ​​of these specific relational edges in the initial quality defect knowledge graph are retrieved, and the base weight values ​​are adjusted according to a preset enhancement coefficient rule. The preset enhancement coefficient rule divides different enhancement coefficient levels based on the path validity level, and multiplicative operations are performed to increase the base weight values. If the enhanced weight value exceeds the preset weight upper limit, the preset weight upper limit value is used as the final adjustment result for these relational edges. All relational edges that have undergone weight enhancement are collected, and the weight values ​​before and after adjustment and the corresponding target path identifier of each relational edge are recorded to generate a set of weight-enhanced relational edges.

[0103] The confirmation instruction includes the target path identifier and a list of corresponding relation edges. The preset enhancement coefficient rules are based on the path validity level. The set of weight-enhanced relation edges includes information such as the relation edges that have undergone weight enhancement, the weight values ​​of each relation edge before and after adjustment, and the target path identifier corresponding to each relation edge.

[0104] (2) Based on the supplementary instructions in the path validity feedback instructions, the newly added relation tuples in the supplementary instructions are processed by graph insertion to generate a relation-enhanced knowledge graph.

[0105] For example, based on the supplementary instructions in the path validity feedback instructions, the structure of the supplementary instructions is deconstructed, and the newly added relation tuples are extracted. Each newly added relation tuple contains the source entity ID, target entity ID, relation type, and relation attribute description. The newly added relation tuples undergo validity verification, checking whether the source entity ID and target entity ID exist in the entity database of the initial quality defect knowledge graph, whether the relation type conforms to the preset cigarette manufacturing domain relation category, and whether the relation attribute description is complete. If the verification passes, the newly added relation tuple is inserted into the topology of the initial quality defect knowledge graph, establishing a new relation edge between the source and target entities, and supplementing the associated relation attribute information. The entity association index of the initial quality defect knowledge graph is updated, generating a relation-enhanced knowledge graph.

[0106] The newly added relation tuples include the source entity ID, target entity ID, relation type, and relation attribute description. The validity verification is based on the entity database of the initial quality defect knowledge graph and the preset relation categories in the cigarette manufacturing field. The relation-expanded knowledge graph is the knowledge graph after inserting valid new relation tuples and updating the entity association index.

[0107] (3) Perform incremental graph learning on the weighted enhanced relation edge set and relation expanded knowledge graph to generate an optimized knowledge graph.

[0108] For example, each relation edge in the weighted augmented relation edge set is associated with the corresponding path in the relation-enhanced knowledge graph to determine the local region of the graph that needs to be updated. This local region includes the path containing the relation edges in the weighted augmented relation edge set, the path containing the newly added relation edges in the relation-enhanced knowledge graph, and the associated nodes surrounding these paths. An appropriate incremental graph learning algorithm is selected, including an incremental graph convolutional neural network.

[0109] Based on a relation-enhanced knowledge graph framework, this approach takes the weight changes of weight-enhanced relation edges and the feature information of newly added relation edges as input. It updates the embedding features of nodes within local regions of the graph, while retaining the original embedding features of nodes not involved in these local regions. Simultaneously, it optimizes the association strength calculation model for relation edges in the graph, ensuring that the adjusted weights and new relations are integrated into the overall graph's reasoning logic. After completing the node embedding feature update and association strength calculation model optimization, a global consistency check is performed on the graph to ensure that there are no logical conflicts in the associations between nodes and relation edges, generating the optimized knowledge graph.

[0110] The local region of the knowledge graph includes the path containing weighted augmented relationship edges, the path containing newly added relationship edges, and surrounding associated nodes. The adapted incremental graph learning algorithm includes an incremental graph convolutional neural network. The optimized knowledge graph completes local node feature updates, model optimization, and global verification.

[0111] (4) Root cause backtracking calculation is performed based on the optimized knowledge graph to generate optimized root cause analysis results.

[0112] For example, the starting point for backtracking is determined as the defect node corresponding to the current quality defect event. A backtracking depth is set based on the common length of the causal chain in cigarette production to avoid over-traversal. Traversal is performed in the reverse direction of the causal relationship edges in the optimized knowledge graph, from the defect node to the upstream influencing node. During the traversal, relationship edges with higher weight values ​​are selected first. Candidate root cause entities and the weight values ​​of the corresponding relationship edges on each traversal path are recorded. Candidate root cause entities include equipment component entities and process node entities.

[0113] The importance of candidate root cause entities is assessed, with indicators including the total weight of the entity's relational edges and the frequency of the entity's appearance in multiple backtracking paths. Based on the assessment results, candidate root cause entities are ranked, and the entities with the highest importance ranking are selected as core root causes. The core root cause entities, their corresponding propagation paths, and the weights of the relational edges within those paths are integrated to form a structured analysis report, generating optimized root cause analysis results.

[0114] The backtracking depth is set based on the common length of the causal chain in cigarette production. Importance assessment indicators include the total weight of the edges representing entity relationships and the frequency of an entity's appearance in multiple backtracking paths. The optimized root cause analysis result is a structured analysis report containing core root cause entities, corresponding propagation paths, and weight information.

[0115] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0116] In one embodiment, the present invention also provides a multi-source heterogeneous data-driven root cause analysis system 300 for cigarette quality defects, such as... Figure 3 As shown, the system 300 includes: The feature alignment module 301 is used to perform feature extraction and spatiotemporal alignment processing on the equipment sensing time-series data, process parameter configuration data, quality inspection structured data, environmental monitoring records and operation log text data of the cigarette production line, and generate a unified feature vector set. The graph construction module 302 is used to perform graph embedding mapping and topological relationship construction on the unified feature vector set to obtain an initial quality defect knowledge graph. The path reasoning module 303 is used to perform causal path reasoning based on the initial quality defect knowledge graph, starting from the defect node, and output multiple candidate root cause propagation paths. The interactive feedback module 304 is used to visualize multiple candidate root cause propagation paths, receive user operation instructions, and generate path validity feedback instructions. The optimization analysis module 305 is used to adjust the weights of the initial quality defect knowledge graph based on the path validity feedback instructions, and generate optimized root cause analysis results.

[0117] Specifically, the feature alignment module 301 processes the time-series data of equipment sensors, process parameter configuration data, structured quality inspection data, environmental monitoring records, and operation log text data of the cigarette production line. This includes extracting the core features of various data types, using the production cycle time of the cigarette production line as the time reference and equipment units and process links as the spatial reference, and calibrating the spatiotemporal dimensions of the data. The integrated and aligned features are then transformed into vectors with a unified format and dimensions, generating a unified feature vector set.

[0118] The graph construction module 302 is based on a unified feature vector set. This includes defining entity types according to a pre-defined cigarette manufacturing domain ontology, and mapping the unified feature vector set to knowledge nodes carrying entity attributes using graph embedding technology. It calculates the data relationships between nodes to generate initial relational edges, and integrates domain ontology relational constraints with these initial relational edges to construct a topological structure, generating an initial quality defect knowledge graph.

[0119] When processing the initial quality defect knowledge graph, the path reasoning module 303 generates an initial causal propagation path set by backpropagating multiple times along the causal relationship edges from the defect node. It then prunes the path according to the time-varying stability coefficient of the relationship edges to generate a simplified path set. The module calculates the comprehensive confidence score of each path and generates an adaptive threshold by combining event severity and historical cases. Paths exceeding the threshold are selected by sorting the scores, and multiple candidate root cause propagation paths are output.

[0120] When processing multiple candidate root cause propagation paths, the interactive feedback module 304 includes topological hierarchical compression to generate a hierarchical path display tree with the quality defect type as the root node. It responds to user focus commands by retrieving the corresponding entity's multimodal raw data, parsing user annotation behavior to generate intermediate path verification commands, and fusing the intermediate commands with the path topology context to generate path validity feedback commands.

[0121] When the optimization analysis module 305 processes the path validity feedback instructions, it enhances the weights of corresponding relation edges according to the confirmation instructions to generate a set of weighted relation edges, and inserts legally added relation tuples according to the supplementary instructions to generate a relation-expanded knowledge graph. Incremental graph learning is then performed on both to generate an optimized knowledge graph. Based on this graph, the root causes are traced back and the ranking is evaluated to generate optimized root cause analysis results.

[0122] The map construction module 302 is also used for: Based on the preset ontology of the cigarette manufacturing field, entity types are defined, including equipment component entities, process node entities, defect type entities, environmental factor entities, and operational behavior entities. A graph neural network encoding process is applied to a unified feature vector set to generate entity attribute vectors, where each entity attribute vector corresponds to a dynamic state representation of an entity type. Calculate the multidimensional statistical correlation between entity types based on entity attribute vectors, and generate initial relation edges; By integrating the association constraints and initial relation edges defined in the ontology of cigarette manufacturing, dynamic graph topology synthesis is performed to generate an initial quality defect knowledge graph.

[0123] Path reasoning module 303 is also used for: Starting from the defect node in the initial quality defect knowledge graph, multi-order back graph propagation processing is performed along the causal relationship edges to generate an initial causal propagation path set. The initial causal propagation path set is dynamically pruned based on the time-varying stability coefficient of the relation edges in the initial quality defect knowledge graph to generate a simplified propagation path set. Using the following formula, for each path in the simplified propagation path set, the confidence score is calculated by integrating the path length weight, relation edge density weight, and entity activity weight, resulting in a comprehensive path confidence score:

[0124] in, This indicates the overall confidence score of the path. Indicates the path length; Represents the set of edges related to a path; Indicates the time-varying stability coefficient of the relation edge; Represents the edge variable of the relation; Indicates the number of nodes in the path; Indicates the defect detection timestamp; Indicates the last time an entity was active; Indicates the preset length weight; Indicates the preset edge density weight; This indicates the preset entity activity weight.

[0125] An adaptive confidence threshold is generated based on the matching degree between the severity level of the current quality defect event and historical cases. The paths are sorted in descending order based on their overall confidence score, and paths with confidence scores higher than the adaptive confidence threshold are selected to output multiple candidate root cause propagation paths.

[0126] Interactive feedback module 304 is also used for: Multiple candidate root cause propagation paths are subjected to topological hierarchical compression to generate a hierarchical path display tree, where the root node is the quality defect type and the branch nodes are the root cause entities. In response to the user's focus command on any node in the hierarchical path display tree, the system dynamically retrieves the multimodal raw data of the corresponding entity in the initial quality defect knowledge graph. The system uses a pre-defined instruction semantic template to perform structured parsing of user annotation behavior on multimodal raw data and generates intermediate instructions for path verification. The intermediate instructions for path verification are merged with the topological context of the currently focused path to generate a path validity feedback instruction.

[0127] Interactive feedback module 304 is also used for: The user's original operation command is matched with a preset command semantic template to generate intent classification labels. Based on intent classification labels, spatiotemporal sensing waveforms and process parameter slices in multimodal raw data are associated to generate data anchoring tuples; Perform logical consistency checks on the multimodal evidence in the data anchoring tuples and generate conflict check results; The intent classification labels, data anchor tuples, and conflict verification results are integrated and encapsulated in a structured manner to generate intermediate instructions for path verification.

[0128] Interactive feedback module 304 is also used for: Perform topological feature extraction on the graph structure of the current focused path to generate a topological feature vector; The intermediate instructions for path verification and the topological feature vector are processed through a graph attention masking mechanism to perform semantic-topological collaborative reasoning to generate collaborative reasoning results. The results of collaborative reasoning are processed by dynamic instruction casting using conflict resolution rules to generate structured instruction tuples. The structured instruction tuples are standardized and encapsulated to generate path validity feedback instructions.

[0129] The optimization analysis module 305 is also used for: Based on the confirmation instruction in the path validity feedback instruction, the relation edges in the initial quality defect knowledge graph are weighted and enhanced to generate a set of weighted enhanced relation edges. Based on the supplementary instructions in the path validity feedback instructions, graph insertion processing is performed on the newly added relation tuples in the supplementary instructions to generate a relation-enhanced knowledge graph. Incremental graph learning is applied to the weighted enhanced relation edge set and relation-enhanced knowledge graph to generate an optimized knowledge graph. Root cause regression calculations are performed based on the optimized knowledge graph to generate optimized root cause analysis results.

[0130] In one embodiment, multiple types of production data from the cigarette manufacturing process are collected. A unified feature vector set is generated by calibrating the spatiotemporal dimensions using a feature alignment module. Entity types are defined using the cigarette manufacturing domain ontology. Attributed knowledge nodes are generated through graph embedding technology. Initial relational edges are generated by calculating node relationships, and domain constraints are integrated to construct an initial quality defect knowledge graph. This graph is used as a reinforcement learning environment, with a graph optimization and root cause reasoning module as the agent. Core actions and reward indicators are defined, and the module is trained using a reinforcement learning algorithm adapted to high-dimensional spaces. When quality defects exceed a reasonable range, the trained module optimizes the graph, traverses backward along the optimized graph, and evaluates candidate root cause entities to locate the core root cause. Based on the root cause, the defect rate is adjusted and monitored. Simultaneously, the interaction experience is stored in an experience replay pool to optimize the training strategy.

[0131] The production data includes various types such as equipment sensor data, process parameter configuration data, quality inspection data, environmental monitoring records, and operation log text data. Entity types include equipment component entities, process node entities, defect type entities, environmental factor entities, and operational behavior entities. Core actions include graph optimization actions and root cause localization actions. Graph optimization actions involve adjusting relation edge weights and adding new entity associations, while root cause localization actions involve selecting backtracking paths and filtering core root cause entities. Reward metrics include root cause localization accuracy, defect rate improvement effect, and graph inference confidence improvement. Reinforcement learning algorithms adapted to high-dimensional spaces include deep Q-network algorithms. An experience replay pool is used to store experience data such as states, actions, rewards, and new states.

[0132] In one embodiment, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0133] In one embodiment, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0134] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0135] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0136] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects, characterized in that, include: Feature extraction and spatiotemporal alignment are performed on the equipment sensor time-series data, process parameter configuration data, quality inspection structured data, environmental monitoring records, and operation log text data of the cigarette production line to generate a unified feature vector set; The unified feature vector set is subjected to graph embedding mapping and topological relationship construction to obtain an initial quality defect knowledge graph; Based on the initial quality defect knowledge graph, causal path reasoning is performed starting from the defect node, and multiple candidate root cause propagation paths are output. The system visualizes the multiple candidate root cause propagation paths, receives user operation commands, and generates path validity feedback commands. The initial quality defect knowledge graph is weighted according to the path validity feedback instruction to generate optimized root cause analysis results.

2. The multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects as described in claim 1, characterized in that, The process of performing graph embedding mapping and topological relationship construction on the unified feature vector set to obtain an initial quality defect knowledge graph specifically includes: Based on the preset ontology of the cigarette manufacturing field, entity types are defined, including equipment component entities, process node entities, defect type entities, environmental factor entities, and operation behavior entities; The unified feature vector set is subjected to graph neural network encoding to generate entity attribute vectors, wherein each entity attribute vector corresponds to a dynamic state representation of the entity type. Calculate the multidimensional statistical correlation between the entity types based on the entity attribute vectors, and generate initial relation edges; By integrating the association constraints defined in the cigarette manufacturing ontology with the initial relation edges, dynamic graph topology synthesis processing is performed to generate the initial quality defect knowledge graph.

3. The method for multi-source heterogeneous data-driven root cause analysis of cigarette quality defects as described in claim 1, characterized in that, The process of performing causal path reasoning based on the initial quality defect knowledge graph, starting from defect nodes, and outputting multiple candidate root cause propagation paths specifically includes: Starting from the defect node in the initial quality defect knowledge graph, multi-order reverse graph propagation processing is performed along the causal relationship edges to generate an initial causal propagation path set. Based on the time-varying stability coefficient of the relation edges in the initial quality defect knowledge graph, the initial causal propagation path set is dynamically pruned to generate a simplified propagation path set. For each path in the simplified propagation path set, the confidence score is calculated by integrating the path length weight, relation edge density weight, and entity activity weight to obtain the overall confidence score of the path. An adaptive confidence threshold is generated based on the matching degree between the severity level of the current quality defect event and historical cases. The paths are sorted in descending order based on their comprehensive confidence scores, and paths with confidence scores higher than the adaptive confidence threshold are selected as candidate root cause propagation paths.

4. The multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects as described in claim 1, characterized in that, The process of visualizing the multiple candidate root cause propagation paths, receiving user operation commands, and generating path validity feedback commands specifically includes: The multiple candidate root cause propagation paths are subjected to topological hierarchical compression to generate a hierarchical path display tree, where the root node is the quality defect type and the branch nodes are the root cause entities. In response to the user's focus command on any node in the hierarchical path display tree, the multimodal raw data of the corresponding entity in the initial quality defect knowledge graph are dynamically retrieved. The system uses a preset instruction semantic template to perform structured parsing of the user's annotation behavior on the multimodal raw data, and generates intermediate instructions for path verification. The path validity feedback instruction is generated by fusing the intermediate path verification instruction with the topological context of the currently focused path.

5. The multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects as described in claim 4, characterized in that, The step of performing structured parsing of user annotation behavior on the multimodal raw data using a preset instruction semantic template to generate intermediate instructions for path verification specifically includes: The user's original operation command is matched with a preset command semantic template to generate intent classification labels. Based on the intent classification label, the spatiotemporal sensing waveforms and process parameter slices in the multimodal raw data are associated to generate data anchoring tuples; Logical consistency verification is performed on the multimodal evidence in the data anchoring tuple to generate conflict verification results; The intent classification label, the data anchoring tuple, and the conflict verification result are fused and structured to generate the path verification intermediate instruction.

6. The multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects as described in claim 4, characterized in that, The process of fusing the path verification intermediate instructions with the topological context of the currently focused path to generate the path validity feedback instructions specifically includes: Perform topological feature extraction on the graph structure of the current focused path to generate a topological feature vector; The path verification intermediate instructions and the topological feature vector are processed through a graph attention masking mechanism to perform semantic-topological collaborative reasoning to generate a collaborative reasoning result. The collaborative reasoning results are subjected to dynamic instruction casting processing using conflict resolution rules to generate structured instruction tuples. The structured instruction tuple is standardized and encapsulated to generate the path validity feedback instruction.

7. The method for multi-source heterogeneous data-driven root cause analysis of cigarette quality defects as described in claim 4, characterized in that, The step of adjusting the weights of the initial quality defect knowledge graph based on the path validity feedback instruction to generate optimized root cause analysis results specifically includes: Based on the confirmation instruction in the path validity feedback instruction, the relation edges in the initial quality defect knowledge graph are weighted and enhanced to generate a set of weighted enhanced relation edges. Based on the supplementary instructions in the path validity feedback instructions, graph insertion processing is performed on the newly added relation tuples in the supplementary instructions to generate a relation-enhanced knowledge graph. Incremental graph learning is performed on the weighted enhanced relation edge set and the relation-enhanced knowledge graph to generate an optimized knowledge graph. Root cause regression calculation is performed based on the optimized knowledge graph to generate the optimized root cause analysis results.

8. A multi-source heterogeneous data-driven root cause analysis system for cigarette quality defects, characterized in that, The system employs a multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects as described in any one of claims 1 to 7; the system comprises: The feature alignment module is used to perform feature extraction and spatiotemporal alignment processing on the equipment sensor time-series data, process parameter configuration data, quality inspection structured data, environmental monitoring records and operation log text data of the cigarette production line, and generate a unified feature vector set. The graph construction module is used to perform graph embedding mapping and topological relationship construction processing on the unified feature vector set to obtain an initial quality defect knowledge graph. The path reasoning module is used to perform causal path reasoning processing based on the initial quality defect knowledge graph, starting from the defect node, and output multiple candidate root cause propagation paths. The interactive feedback module is used to visualize the multiple candidate root cause propagation paths, receive user operation instructions, and generate path validity feedback instructions. The optimization analysis module is used to adjust the weights of the initial quality defect knowledge graph according to the path validity feedback instruction, and generate optimized root cause analysis results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-source heterogeneous data-driven root cause analysis method for cigarette quality defects as described in any one of claims 1 to 7.

Images