Causal tracing method and system for topological perception of cigarette quality anomaly and storage medium

Abstract

This invention relates to the field of causal attribution technology in industrial processes, specifically to a process topology-aware causal attribution method, system, and storage medium for cigarette quality anomalies during the cigarette-making process. The method includes: acquiring historical time-series data of the abnormal period in the cigarette-making process to obtain feature representations of each variable group; constructing an intra-group graph structure to obtain intra-group causal representations; designing a graph-level feature extractor to obtain graph-level representations of each variable group and constructing a process topology-aware mask constraint matrix; constructing an inter-group graph structure and, under the constraints of the process topology-aware mask constraint matrix, obtaining inter-group causal representations; constructing a grouped root cause analysis algorithm to build a global causal matrix and calculate the root cause contribution of each variable to obtain root cause variables; and identifying the anomaly propagation path through path weight analysis. This invention achieves root cause localization and propagation path identification of cigarette quality anomalies by prior construction of a mask constraint matrix, multi-level causal modeling, and combining it with a grouped root cause analysis algorithm.

Description

Technical Field

[0001] This invention relates to the field of industrial process causal tracing technology, specifically to a process topology-sensing causal tracing method, system, and storage medium for identifying abnormal cigarette quality during the cigarette rolling process. Background Technology

[0002] As industrial systems continue to develop towards integration and intelligence, modern industrial processes exhibit complex characteristics of multi-variable coupling and multi-process interaction. In cigarette manufacturing, draw resistance and ventilation, as key indicators of cigarette quality, are highly dependent on the operational status of the cigarette-making process. When process parameters fluctuate abnormally, the anomaly propagates and evolves along the information and material flows between different variables and processes, ultimately causing key indicators to exceed preset upper and lower limits, triggering cigarette rejection and affecting overall production output. Therefore, effectively identifying the propagation path of anomalies, accurately locating the root cause, and implementing targeted adjustments have become key technical issues in the quality control of the cigarette-making process. Currently, cigarette factories mainly rely on existing quality control systems to ensure the safety and stability of the production process. When cigarette quality is abnormal, the cigarette-making site typically relies on a comprehensive judgment based on operator experience, equipment alarm information, and manual inspection results, and intervenes through parameter adjustments or shutdown maintenance. While this approach possesses certain engineering practicality, it still suffers from significant shortcomings: on the one hand, root cause determination heavily relies on subjective experience, resulting in low diagnostic efficiency and consistency; on the other hand, it struggles to systematically characterize the propagation mechanism of anomalies among multiple variables and processes, and cannot identify potential root causes beyond the scope of expert experience, thus restricting production efficiency and quality control levels in the cigarette-making process to some extent. In recent years, with the development of big data technology, data-driven methods have been gradually applied to quality monitoring in the cigarette-making process. However, modeling of the cigarette-making process remains at the level of anomaly detection, only able to determine whether an anomaly has occurred; in terms of anomaly mechanism explanation, it relies on correlation analysis between variables, making it difficult to characterize direct and indirect causal relationships between variables, and unable to effectively reveal anomaly propagation paths and achieve root cause localization. Based on this, industrial process root cause diagnosis technology has begun to model the interaction relationships between variables from a causal inference perspective. Typical methods include Granger causality (GC), transfer entropy (TE), and dynamic Bayesian networks (DBN), among which neural GC methods are widely used for complex causal relationship modeling.

[0003] However, the aforementioned methods primarily rely on directed dependencies between variables in historical data for modeling, lacking effective constraints on prior knowledge such as the physical structure of industrial processes. In complex industrial scenarios with multiple variables and processes, a purely data-driven paradigm can easily introduce redundant or even erroneous causal connections, thus masking the true root cause. Specifically, in the cigarette rolling process, the rolling unit consists of three sub-units: shredding supply (VE), coil forming (SE), and cigarette forming (MAX), corresponding to three processes. Each sub-unit operates collaboratively according to a predetermined process sequence and is tightly coupled through feedback control loops such as weight and circumference. If the aforementioned process topology and prior constraints are ignored, existing root cause diagnosis methods are prone to learning redundant and spurious causal relationships that deviate from the actual mechanism or expert experience, reducing the accuracy and reliability of anomaly propagation path identification and root cause localization.

[0004] Therefore, in order to address the problem of the difficulty in accurately modeling causal relationships under the conditions of multi-process collaboration and the coexistence of multiple control loops in the cigarette rolling process, it is urgent to propose a multi-level causal modeling method with process topology awareness in order to achieve accurate causal tracing of cigarette quality anomalies. Summary of the Invention

[0005] The purpose of this invention is to provide a process topology-aware causal tracing method, system, and storage medium for identifying abnormal cigarette quality during the cigarette rolling process, in order to solve the problem of difficulty in accurately modeling causal relationships under conditions of multi-process collaboration and multiple control loops coexisting in the existing technology of cigarette rolling process.

[0006] To achieve the above objectives, embodiments of the present invention provide a process topology-aware causal tracing method for identifying cigarette quality anomalies during the cigarette rolling process, comprising: To obtain historical time-series data of abnormal periods in the cigarette rolling process, in order to obtain characteristic representations of each variable group; Based on the feature representations of each variable group, an intra-group graph structure is constructed, and a graph neural network is used for information propagation and aggregation to obtain intra-group causal representations. Design a graph-level feature extractor to obtain graph-level representations of each variable group, and construct a process topology-aware mask constraint matrix; Based on the graph-level representation of each variable group, an inter-group graph structure is constructed, and under the constraint of the process topology-aware mask constraint matrix, a graph neural network is used for information propagation and aggregation to obtain inter-group causal representations. A grouped root cause analysis algorithm is constructed, and a global causal matrix is ​​built based on the intra-group causal representation and the inter-group causal representation. The root cause contribution of each variable is calculated to obtain the root cause variable. Based on the root dependent variable, abnormal propagation paths are identified through path weight analysis.

[0007] Optionally, based on the feature representations of each variable group, an intra-group causal matrix is ​​constructed, and a graph neural network is used for information propagation and aggregation to obtain intra-group causal representations, including: The initial embedding features of each group of variables are used as the input node features of the corresponding graph network. A causal matrix is ​​constructed for each group of variables. Except for the last group, the causal matrix of each group is generated by training parameters through activation and normalization. The causal matrix of the last group is set as an identity matrix. Each group uses its causal matrix to perform multi-layer graph convolution operation on the node features within the group to obtain the intra-group causal representation of each group. Based on the intra-group causal representation described in each group, the predicted values ​​of the variables in that group at future times are obtained to obtain the prediction loss; The mean squared error between the predicted values ​​and the corresponding true values ​​of all groups is used as the prediction loss, and the sum of the L1 norms of all learnable causal matrices is used as the sparse loss. The prediction loss and the sparse loss are combined to jointly optimize all trainable parameters of the graph neural network, the multilayer perceptron, and the learnable causal matrices until convergence. Once the training converges, the learnable causal matrix is ​​the final learned within-group causal matrix.

[0008] Optionally, a graph-level feature extractor is designed to obtain graph-level representations of each variable group, and a process topology-aware mask constraint matrix is ​​constructed, including: The process topology-aware mask constraint matrix is ​​obtained according to formula (1). (1) in, This is the process topology-aware mask constraint matrix. This is a process sequence constraint matrix. Here is the feedback loop constraint matrix. This is the process distance attenuation constraint matrix. This indicates the Hadamard element-wise multiplication operation.

[0009] Optionally, the process sequence constraint matrix is ​​obtained according to formula (2): (2) in, This is a process sequence constraint matrix. Represents a vector index. Indicates the first The process index to which each variable belongs. Indicates the process index of the variable group. This represents the set of process pairs that have feedback control loops.

[0010] Optionally, the feedback loop constraint matrix is ​​obtained according to formula (3): (3) in, Here is the feedback loop constraint matrix. Represents a vector index. Indicates the first The process index to which each variable belongs. Indicates the process index of the variable group. This represents the set of process pairs that have feedback control loops. Indicates The first layer of graph convolutional information aggregation Individual variable characteristics, Indicates the first Graph-level embedding representation of groups of variables This represents the learnable scale parameter that controls the distribution of membership degrees.

[0011] Optionally, the process distance attenuation constraint matrix is ​​obtained according to formula (4): (4) in, This is the process distance attenuation constraint matrix. Represents a vector index. Indicates the first The process index to which each variable belongs. Indicates the process index of the variable group. This represents the set of process pairs that have feedback control loops. This is a hyperparameter used to control the degree to which causal strength decreases with distance.

[0012] Optionally, a grouped root cause analysis algorithm is constructed, which builds a global causal matrix based on the within-group causal representation and the between-group causal representation, and calculates the root cause contribution of each variable to obtain the root cause variables, including: Construct a global causal matrix according to formula (5). (5) in, This is the global causal matrix. Indicates the first The variable group and the first Block submatrices between variable groups , Indicates the total number of variable groups; The root cause contribution of each variable is obtained according to formula (6). (6) in, For the first The root cause contribution of each variable, For variables For variables Causal contribution weight, For the first The root cause contribution of each variable, The total number of variables; The root cause contribution is obtained according to formula (7). (7) in, Represents the root cause contribution vector; Choose the variable with the largest contribution as the root cause variable.

[0013] Optionally, identifying abnormal propagation paths through path weight analysis based on the root dependent variable includes: The global causal matrix is ​​sparsified to obtain a sparse global causal matrix; Using the root dependent variable and the downstream abnormal variable as the starting and ending points respectively, a set of candidate propagation paths is constructed based on the sparsed global causal matrix. Calculate the weights of all candidate propagation paths and select the path with the highest weight as the anomaly propagation path.

[0014] On the other hand, the present invention provides a process topology-aware causal tracing system for cigarette quality abnormalities during the cigarette rolling process, the system including a processor configured to perform any of the methods described above.

[0015] In another aspect, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement any of the methods described above.

[0016] The beneficial effects of this invention are: This invention addresses the challenge of accurately locating the root causes of quality anomalies during cigarette rolling under conditions of multi-process collaboration and multiple control loops. It proposes a process topology-aware causal tracing method for cigarette quality anomalies during the rolling process. First, variables are grouped according to their respective processes, and key quality indicators are separately categorized as downstream outcome variables, providing a unified representation basis for multi-level causal relationship modeling. By introducing triple prior constraints—process sequence, feedback control loops, and process distance attenuation—a process topology-aware mask constraint matrix is ​​constructed. This effectively embeds the physical structure information of the rolling process into the inter-group causal relationship learning process. Combined with intra-group causal graph modeling, unified modeling of multi-level causal relationships between intra-group variables and inter-group units is achieved. Compared to traditional purely data-driven causal modeling methods, this invention fully utilizes process topology priors to depict variable dependencies that better align with actual mechanisms and expert experience, thereby significantly improving the accuracy and reliability of causal graph structure learning. Furthermore, to accurately pinpoint the root cause, this invention proposes a grouped root cause analysis algorithm. By correcting the causal relationship between groups through the causal structure within groups, and combining Granger prediction contribution and causal graph, the algorithm can pinpoint the root cause of cigarette quality abnormalities and identify the abnormal propagation path, thus providing support for on-site production decisions.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart of a process topology-sensing causal tracing method for cigarette quality abnormalities in the cigarette rolling process according to an embodiment of the present invention. Figure 2 A flowchart illustrating a method for obtaining intra-group causal representations using graph neural networks according to an embodiment of the present invention; Figure 3 A flowchart of a method for identifying abnormal propagation paths based on root dependent variables through path weight analysis according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a hierarchical causal tracing framework for embedded process topology constraints according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a learned hierarchical causal structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram of an anomaly propagation path deduced according to an embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0020] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0021] like Figure 1 The diagram shows a flowchart of a process topology-aware causal attribution method for cigarette quality abnormalities during the cigarette-making process, according to an embodiment of the present invention. Figure 1 In this method, the steps may include: In step S10, historical time-series data of abnormal periods in the cigarette rolling process are obtained to obtain the characteristic representation of each variable group; In step S11, an intra-group graph structure is constructed based on the feature representations of each variable group, and a graph neural network is used for information propagation and aggregation to obtain intra-group causal representations; In step S12, a graph-level feature extractor is designed to obtain the graph-level representation of each variable group and to construct a process topology-aware mask constraint matrix; In step S13, an inter-group graph structure is constructed based on the graph-level representation of each variable group, and under the constraint of the process topology-aware mask constraint matrix, a graph neural network is used for information propagation and aggregation to obtain inter-group causal representations. In step S14, a grouped root cause analysis algorithm is constructed, a global causal matrix is ​​built based on the intra-group causal representation and the inter-group causal representation, and the root cause contribution of each variable is calculated to obtain the root cause variable; In step S15, abnormal propagation paths are identified through path weight analysis based on the root dependent variable.

[0022] In such Figure 1 In the process topology-aware causal tracing method for identifying cigarette quality anomalies during the cigarette rolling process, step S10 is used to acquire historical time-series data of the abnormal period in the cigarette rolling process to obtain the feature representations of each variable group. In this embodiment, the specific method for acquiring the feature representations of each variable group in step S10 can be of various forms known to those skilled in the art. In one example of the present invention, step S10 may include: Collect historical time-series data during abnormal periods in the splicing process. ,in and These represent the number of variables and the length of the time series, respectively. Next, the time series data is analyzed based on the process characteristics of the winding process. Grouping, resulting in Data subsets corresponding to each variable group ,in , Indicates the first The number of variables in each group, and satisfying Meanwhile, the quality indicators were grouped into the last group. Furthermore, Z-score standardization was performed on the variables in each group to eliminate dimensional differences between the variables.

[0023] Subsequently, the time-series data from each group are input into the parallel time-series coding module for time feature extraction. The parallel time-series coding module includes... Each variable is an independent one-dimensional convolutional neural network (CNN). To eliminate the interaction effects between variables, each variable... The features are independently fed into the corresponding CNN to extract their features, as expressed in the following expression: ; in, Represents the trainable parameters of a convolutional neural network; Indicates the first The first group One input variable, This represents the temporal characteristics of the convolution output. Represents convolution Feature dimensions after the layer.

[0024] Then the first The feature matrix extracted from each group is represented as follows: ; in, Indicates the first The initial node feature matrix of each group.

[0025] Step S11 is used to construct an intra-group graph structure based on the feature representations of each variable group, and to use a graph neural network for information propagation and aggregation to obtain intra-group causal representations. In this embodiment, the specific method for obtaining intra-group causal representations in step S11 can be of various forms known to those skilled in the art. In one example of the present invention, step S11 may include, for example... Figure 2 The steps shown are described in this. Figure 2 In this context, step S11 may include: In step S20, the initial embedding features of each group of variables are used as the input node features of the corresponding graph network. In step S21, a causal matrix is ​​constructed for each variable group. Except for the last group, the causal matrices of the remaining groups are generated by the activation and normalization of trainable parameters. The causal matrix of the last group is set as an identity matrix. In step S22, each group uses its causal matrix to perform multi-layer graph convolution operation on the node features within the group to obtain the intra-group causal representation of each group; In step S23, the predicted value of the variable in the group at a future time is obtained according to the intra-group causal representation of each group, so as to obtain the prediction loss; In step S24, the mean squared error between all group predicted values ​​and their corresponding true values ​​is used as the prediction loss, and the sum of the L1 norms of all learnable causal matrices is used as the sparse loss. The prediction loss and the sparse loss are used to jointly optimize all trainable parameters of the graph neural network, the multilayer perceptron, and the learnable causal matrix until convergence. After training convergence, the learnable causal matrix is ​​the finally learned intra-group causal matrix.

[0026] In such Figure 2 In the method shown, step S20 is used to use the initial embedding features of each group of variables as the input node features of the corresponding graph network. In this example, the initial embedding features of the graph network for each group of variables are obtained, denoted as: ,in, Indicates the first The initial embedding features of each group.

[0027] Step S21 is used to construct the causal matrix for each variable group, denoted as: ; in, Indicates the first Learnable causal matrix of 1 set of variables This represents the initial trainable parameters. Since there is usually no explicit causal dependency between quality metrics, the last set of corresponding causal matrices is set to non-learnable identity matrices to avoid introducing spurious causal connections.

[0028] Step S22 is used to perform multi-layer graph convolution operations on the node features within each group using its causal matrix, thereby obtaining the intra-group causal representation of each group. Specifically, in this example, the node embedding features of each group are input into their corresponding graph convolutional network (GCN) for information propagation, and the information aggregation process can be represented as follows: ; in, Represents the ReLU activation function. Indicates the first The first graph convolutional network of the _th Layer trainable parameters.

[0029] The node features aggregated from each variable group can then be represented as: ; in, Indicates the first The variable groups were processed The node features after layer information propagation and aggregation This indicates the hidden feature dimension.

[0030] Step S23 is used to obtain the predicted values ​​of the variables in each group at future times based on the intra-group causal representations, in order to obtain the prediction loss. Specifically, in this example, the node features output from the graph convolution of each group are input to... In a multilayer perceptron (MLP) with multiple layers, the values ​​of each set of variables are predicted at future times, thus completing the prediction task under the Granger causality paradigm, denoted as: ; in, This represents the trainable parameters of the multilayer perceptron. Indicates the first Group of variables The predicted value at time t, with the predictor variable having the following dimensions. .

[0031] Step S24 is used to jointly optimize all trainable parameters of the graph neural network, multilayer perceptron, and learnable causal matrix using the joint prediction loss and sparse loss until convergence. Specifically, in this example, the within-group prediction loss function can be expressed as: The mean squared error (MSE) between the predicted and actual values ​​of a group of variables is denoted as: ; in, The squared norm of a vector is the sum of the squares of its elements.

[0032] Furthermore, in the process of learning the within-group causal matrix, a method based on... The sparse constraint term of the norm, which induces sparsity in the within-group causal matrix, is denoted as: ; in, Indicates the first A causal matrix of variables, This represents the summation of the absolute values ​​of the elements of a matrix.

[0033] Step S12 is used to design a graph-level feature extractor to obtain the graph-level representation of each variable group and to construct a process topology-aware mask constraint matrix. In this example, the specific method for designing the graph-level feature extractor to obtain the graph-level representation of each variable group can be to use an extractor based on a cross-attention mechanism to adaptively weight and aggregate the node features within the group. Specifically, let the first step be... The learnable query vector for each variable group is The node representation after convolutional aggregation of this set of graphs Construct the query vector, key vector, and value vector respectively, denoted as: ; in, , All of these are trainable parameters.

[0034] Subsequently, the scaling dot product attention was used to calculate the first... Importance weights of each node within each variable group: ; in, Indicates the first The node attention coefficients of each variable group This indicates the dimension of the embedded features.

[0035] Next, the node value vectors are weighted and aggregated according to the attention coefficient, as expressed in the following expression: ; in, Indicates the first Graph-level embedding representation of groups of variables.

[0036] A process topology-aware mask constraint matrix is ​​constructed to guide subsequent hierarchical causal learning. In this example, the mask constraints include process sequence, feedback loops, and process distance attenuation constraints, which are used to characterize the modulation effect of different constraints on the causal connection strength. Its overall form is expressed as: (1) in, This is the process topology-aware mask constraint matrix. This is a process sequence constraint matrix. Here is the feedback loop constraint matrix. This is the process distance attenuation constraint matrix. Indicates the total number of variables. This represents the total number of variable groups (processes). This indicates the Hadamard element-wise multiplication operation.

[0037] Furthermore, the process sequence constraint matrix To constrain variables to establish causal connections only with their own process, upstream processes, and related processes with feedback paths, the element is defined as follows: (2) in, Indicates variable index, Indicates the first The process index to which each variable belongs. Indicates the index of the variable group (process). This represents the set of process pairs that have feedback control loops.

[0038] Furthermore, the feedback loop constraint matrix This method is used to characterize the cross-group association strength between variables and related processes when process feedback paths exist. Specifically, it introduces fuzzy membership degrees based on Gaussian kernel functions among process groups with feedback channels to model the soft membership relationships of variables across groups, thereby weighting and adjusting the causal relationships on the feedback path. Its elements are defined as follows: (3) in, Indicates The first layer of graph convolutional information aggregation Individual variable characteristics, Indicates the first Graph-level embedding representation of groups of variables The learnable scaling parameter represents the controllable membership distribution, and its value reflects the fuzzy membership range of the variable to the target process group: when When the value is large, the Gaussian kernel function distribution is more gradual, and different variables have a stronger cross-group correlation ability for this process group, with a wider membership relationship; when When the value is small, the Gaussian kernel function distribution is more concentrated, and only variables with similar characteristics to the process group have a high degree of membership, thus reflecting a stricter local membership relationship.

[0039] Furthermore, the process distance attenuation constraint matrix This method is used to characterize the impact of inter-process distance on the strength of potential causal relationships between variables. Specifically, it does not attenuate the causal relationships between variables within the same process or those with feedback pathways, but attenuates the causal strength of other cross-process connections exponentially. Its elements are defined as follows: (4) in, This represents a hyperparameter controlling the attenuation of causal strength with distance; its value reflects the attenuation sensitivity of cross-process effects. When the value is large, the exponential decay rate accelerates, and causal relationships between distant processes are rapidly suppressed, demonstrating stronger locality constraints; when When the value is smaller, the decay process is more gradual, and different processes can still maintain a certain degree of remote correlation, thus reflecting a more relaxed cross-process influence range.

[0040] Step S13 is used to construct an inter-group graph structure based on the graph-level representation of each variable group, and under the constraint of the process topology-aware mask constraint matrix, a graph neural network is used for information propagation and aggregation to obtain inter-group causal representations. Specifically, in this example, firstly, the graph-level embedding representation of each variable group is obtained by step S12. Construct the data features of the inter-group graph, denoted as: ; in, Indicates between groups The feature matrix of each unit.

[0041] Secondly, construct the inter-group causal adjacency matrix, denoted as: ; in, This represents the causal matrix that can be learned between groups. This represents the initial trainable parameters.

[0042] Next, the constructed process topology-aware mask constraint matrix is ​​obtained from step S12. This is integrated into the learning process of the inter-group causal adjacency matrix to guide the learning of causal relationships between variables and groups, denoted as: ; in, This represents the inter-group causality matrix with embedded mask constraints. This indicates a row normalization operation.

[0043] Subsequently, the prediction head is output through a single graph convolutional network and an MLP. The predicted values ​​of each variable at future times are obtained, thus completing the prediction task under the Granger causality paradigm, denoted as: ; in, Indicates all Variables in Predicted value at time, Represents the ReLU activation function. These are trainable parameters.

[0044] Furthermore, the between-group prediction loss function can be expressed as: The mean squared error (MSE) between the predicted and actual values ​​of a variable is denoted as: ; in, The squared norm of a vector is the sum of the squares of its elements.

[0045] Finally, based on steps S11 and S13, to jointly optimize the learning of intra-group and inter-group causal relationships and constrain the sparsity of the causal structure, the overall loss function is constructed as follows: ; in, Indicates the predicted loss within the group. Indicates the predicted loss between groups. This represents the sparse constraint loss within the group; This represents the sparse constraint coefficient.

[0046] Subsequently, with the overall loss function as the optimization objective, the model parameters are trained based on the training set, and the model parameters are iteratively updated using the backpropagation algorithm until the preset training epochs are reached. This allows us to obtain the optimal hierarchical causal model.

[0047] Step S14 is used to construct a grouped root cause analysis algorithm, which builds a global causal matrix based on within-group and between-group causal representations, and calculates the root cause contribution of each variable to obtain the root cause variable. Specifically, in this example, the within-group causal matrix is ​​obtained according to steps S11 and S13 respectively. ( ) and between-group causal matrix Next, the between-group causal matrix will be... Extend along the column direction to By considering the variable dimensions, we obtain the inter-group causal expansion matrix. It is represented as: ; in, This represents the matrix expansion function, which is based on the number of variables contained in each variable group. The first group of the inter-group causal matrix is ​​then sequentially... Column copying Next, the expanded matrix is ​​row normalized so that the sum of the elements in each row is 1.

[0048] Subsequently, based on the intragroup causal matrix To correct the inter-group causal extension matrix The diagonal block matrix yields the global causality matrix. It is represented as: ; in, This represents the diagonal block correction function, which performs element-wise weighted correction on the diagonal blocks of the inter-group causal expansion matrix based on the intra-group causal matrix of each variable group, while keeping the off-diagonal blocks unchanged.

[0049] Specifically, Represented as a block matrix divided by variable groups: ; in, Indicates the first The variable group and the first Block submatrices between variable groups .

[0050] Subsequently, based on the above-mentioned block structure, the intra-group causal matrix was used. ( For block matrices Correct the diagonal blocks to construct a global causal matrix. It is represented as: (5) Among them, matrix The non-negative sum of its rows is 1. Indicates the total number of variable groups. This represents the element-wise product of Hadamard.

[0051] Furthermore, based on the constructed global causal matrix Define the root cause contribution of each variable as: (6) in, For the first The root cause contribution of each variable, For variables For variables Causal contribution weight, For the first The root cause contribution of each variable, This represents the total number of variables.

[0052] The above relationship can be expressed as an equation, denoted as: ; in, Represents the root cause contribution vector, matrix Represents the global causal matrix The transpose of .

[0053] Under these conditions, according to the Perron–Frobenius theorem, the matrix There exists a non-negative eigenvector with a corresponding eigenvalue of 1. By solving for this eigenvector, the non-negative contribution scores of each variable as a root cause can be obtained. .

[0054] Finally, the root cause contributions of each variable are compared, and the variable with the largest contribution is selected. As the root dependent variable, its corresponding index is represented as: .

[0055] Step S15 is used to identify abnormal propagation paths based on root dependent variables through path weight analysis. In this embodiment, the specific method for identifying abnormal propagation paths based on root dependent variables through path weight analysis in step S15 can be of various forms known to those skilled in the art. In one example of the present invention, step S15 may include, for example... Figure 3 The steps shown are described. Figure 3 In this context, step S15 may include: In step S30, the global causal matrix is ​​sparsified to obtain a sparsified global causal matrix. In step S31, a set of candidate propagation paths is constructed based on a sparse global causal matrix, with the root dependent variable and the downstream abnormal variable as the starting and ending points, respectively. In step S32, the weights of all candidate propagation paths are calculated, and the path with the largest weight is selected as the abnormal propagation path.

[0056] In such Figure 3 In the method shown, step S30 is used to sparsify the global causal matrix to obtain a sparse global causal matrix. Specifically, in this example, to preserve the main information propagation relationships, the global causal matrix is ​​sparsified. A threshold pruning strategy is used for sparsification to obtain a sparse global causal matrix. Its elements are defined as: ; in, Represents the sparse global causal matrix The elements in Represents the original global causal matrix The elements in This represents the sparse pruning hyperparameter.

[0057] Step S31 is used to calculate the root dependent variable separately. and downstream abnormal variables As the starting and ending points, based on the sparsed global causal matrix Construct a set of candidate propagation paths Any candidate propagation path is represented as follows: ; in This represents a propagation path from the root dependent variable to downstream outliers. Indicates the path length.

[0058] Step S32 is used to calculate the weights of all candidate propagation paths and select the path with the highest weight as the anomaly propagation path. Specifically, in this example, for any candidate propagation path... Its propagation weight is defined as: ; in, Representing a path Edges between adjacent variables, This represents the causal propagation weight of the corresponding edge.

[0059] By comparing candidate path sets The propagation weights of each path are determined, and the path with the highest weight is selected as the anomaly propagation path, which can be expressed as: ; in, This represents the root dependent variable that was ultimately identified. Downstream abnormal variables The abnormal propagation path.

[0060] In one embodiment of the present invention, a typical case of suction resistance anomaly in the cigarette-making process of a cigarette factory is used as an example for illustration and validity verification. The cigarette-making process in this embodiment is a typical complex multi-process system, consisting of three sub-processes: shredding supply (VE), bundle forming (SE), and cigarette forming (MAX). This process is characterized by high variable dimensionality, strong nonlinearity, and highly coupled control loops. In this embodiment, the data sampling interval is 2 seconds, and 980 samples after the anomaly occurred are selected for modeling, containing a total of 24 process variables.

[0061] like Figure 4 The diagram shows a hierarchical causal tracing framework with embedded process topology constraints. This invention is a process topology-aware causal tracing method for cigarette quality anomalies during the cigarette rolling process. The specific modeling steps are as follows: Step 1: Collect historical time-series data from abnormal periods during the splicing process as the training set. Group the variables according to their type and the corresponding work unit, and extract the feature representations of each variable group using a parallel time-series coding module. The specific steps are as follows: Collect historical time-series data during abnormal periods in the splicing process. ,in and These represent the number of variables and the length of the time series, respectively. Next, the time series data is analyzed based on the process characteristics of the winding process. Grouping, resulting in Data subsets corresponding to each variable group ,in , Indicates the first The number of variables in each group, and satisfying Meanwhile, the quality indicators were grouped into the last group. Furthermore, Z-score standardization was performed on the variables in each group to eliminate dimensional differences between the variables.

[0062] In this embodiment, the total number of variables Set to 24, the total number of variable groups Set to 4, time series length setting The value was set to 1000, and training set samples were generated using a sliding window operation. Subsequently, based on the process characteristics of the cigarette rolling process, the variables were divided into 4 groups. The first three groups correspond to the three sub-processes of shredding supply (VE), rolling and forming (SE), and cigarette forming (MAX), respectively, while the last group corresponds to the cigarette quality index (Indicator). For detailed variable grouping information, please refer to Table 1.

[0063] Table 1. Results of process grouping for variables in cigarette rolling process.

[0064] As shown in Table 1, the number of variables involved in the VE process is... Set to 5, the number of variables included in the SE process. Setting it to 11, the MAX process contains the number of variables. Setting it to 6 increases the number of variables included in the Indicator indicator. Set to 2.

[0065] Subsequently, the time series data of each group were... The input is processed by a parallel time-series coding module for time feature extraction. The parallel time-series coding module includes... Each variable is an independent one-dimensional convolutional neural network (CNN). To eliminate the interaction effects between variables, each variable... The features are independently fed into the corresponding CNN to extract their features, as expressed in the following expression: ; in, Represents the trainable parameters of a convolutional neural network; Indicates the first The first group One input variable, This represents the temporal characteristics of the convolution output. Represents convolution The feature dimension after the layer. In this embodiment, the number of convolutional layers. Set to 2, feature dimension Set to 14.

[0066] Then the first The feature matrix extracted from each group is represented as follows: ; in, Indicates the first The initial node feature matrix of each group.

[0067] Step 2: Construct an intra-group graph structure based on the feature representations of each variable group. Utilize a graph neural network for information propagation and aggregation to learn the causal relationships between variables within the group, thus obtaining an intra-group causal representation. The steps are as follows: Based on step 1, the initial embedding features of the graph network for each variable group are obtained, denoted as: ; in, Indicates the first The initial embedding features of each group.

[0068] Subsequently, the causal matrix for each variable group is constructed, denoted as: ; in, Indicates the first Learnable causal matrix of 1 set of variables This represents the initial trainable parameters. Since there is usually no explicit causal dependency between quality metrics, the last set of corresponding causal matrices is set to non-learnable identity matrices to avoid introducing spurious causal connections.

[0069] Next, the embedded features of each group of nodes are input into their corresponding graph convolutional networks for information propagation. The information aggregation process can be represented as follows: ; in, Represents the ReLU activation function. Indicates the first The first graph convolutional network of the _th Layer trainable parameters.

[0070] The node features aggregated from each variable group can then be represented as: ; in, Indicates the first The variable groups were processed The node features after layer information propagation and aggregation This represents the hidden feature dimension. In this embodiment, the number of graph convolutional layers... Set to 2 to hide feature dimensions Set it to 32.

[0071] Subsequently, the node features output from the convolution of each set of images are input into... In a multilayer perceptron (MLP) with multiple layers, the values ​​of each set of variables are predicted at future times, thus completing the prediction task under the Granger causality paradigm, denoted as: ; in, This represents the trainable parameters of the multilayer perceptron. Indicates the first Group of variables The predicted value at time t, with the predictor variable having the following dimensions. In this embodiment, the number of layers in the multilayer perceptron... Set to 2, Setting it to 20 indicates the length of the constructed sample.

[0072] Furthermore, the within-group prediction loss function can be expressed as: The mean squared error (MSE) between the predicted and actual values ​​of a group of variables is denoted as: ; in, The squared norm of a vector is the sum of the squares of its elements.

[0073] Furthermore, in the process of learning the within-group causal matrix, a method based on... The sparse constraint term of the norm, which induces sparsity in the within-group causal matrix, is denoted as: ; in, Indicates the first A causal matrix of variables, This represents the summation of the absolute values ​​of the elements of a matrix.

[0074] Step 3: Design a graph-level feature extractor to obtain graph-level representations of each variable group, and construct a process topology-aware mask constraint matrix to guide the learning of hierarchical causal relationships. The specific steps are as follows: (1) Design a graph-level feature extractor to obtain graph-level representations of each variable group. The extractor is based on a cross-attention mechanism and performs adaptive weighted aggregation of node features within the group. The specific steps are as follows: Let the first The learnable query vector for each variable group is The node representation after convolutional aggregation of this set of graphs Construct the query vector, key vector, and value vector respectively, denoted as: ; in, , All of these are trainable parameters.

[0075] Subsequently, the scaling dot product attention was used to calculate the first... Importance weights of each node within each variable group: ; in, Indicates the first The node attention coefficients of each variable group This indicates the dimension of the embedded features.

[0076] Next, the node value vectors are weighted and aggregated according to the attention coefficient, as expressed in the following expression: ; in, Indicates the first A graph-level embedding representation of a set of variables. In this embodiment, the embedding feature dimension... Set it to 32.

[0077] (2) Construct a process topology-aware mask constraint matrix to guide subsequent hierarchical causal learning. The mask constraints include process sequence, feedback loop, and process distance attenuation constraints, which are used to characterize the modulation effect of different constraints on the causal connection strength. Its overall form is expressed as: ; in, Indicates the total number of variables. This represents the total number of variable groups (processes). This indicates the Hadamard element-wise multiplication operation.

[0078] Specifically, the process sequence constraint matrix To constrain variables to establish causal connections only with their own process, upstream processes, and related processes with feedback paths, the element is defined as follows: ; in, Indicates variable index, Indicates the first The process index to which each variable belongs. Indicates the index of the variable group (process). This represents the set of process pairs that have feedback control loops.

[0079] In this embodiment, , Based on the characteristics of the coiling process, This indicates the existence of a weight and circumference control loop that feeds back from process group 2 to process group 1.

[0080] Furthermore, the feedback loop constraint matrix This method is used to characterize the cross-group association strength between variables and related processes when process feedback paths exist. Specifically, it introduces fuzzy membership degrees based on Gaussian kernel functions among process groups with feedback channels to model the soft membership relationships of variables across groups, thereby weighting and adjusting the causal relationships on the feedback path. Its elements are defined as follows: ; in, Indicates The first layer of graph convolutional information aggregation Individual variable characteristics, Indicates the first Graph-level embedding representation of groups of variables The learnable scaling parameter represents the controllable membership distribution, and its value reflects the fuzzy membership range of the variable to the target process group: when When the value is large, the Gaussian kernel function distribution is more gradual, and different variables have a stronger cross-group correlation ability for this process group, with a wider membership relationship; when When the value is small, the Gaussian kernel function distribution is more concentrated, and only variables with similar characteristics to the process group have a high degree of membership, thus reflecting a stricter local membership relationship.

[0081] Furthermore, the process distance attenuation constraint matrix This method is used to characterize the impact of inter-process distance on the strength of potential causal relationships between variables. Specifically, it does not attenuate the causal relationships between variables within the same process or those with feedback pathways, but attenuates the causal strength of other cross-process connections exponentially. Its elements are defined as follows: ; in, This represents a hyperparameter controlling the attenuation of causal strength with distance; its value reflects the attenuation sensitivity of cross-process effects. When the value is large, the exponential decay rate accelerates, and causal relationships between distant processes are rapidly suppressed, demonstrating stronger locality constraints; when When the value is smaller, the decay process is smoother, and different processes can still maintain a certain degree of remote correlation, thus reflecting a more relaxed cross-process influence range. In this embodiment, the hyperparameter... Set it to 0.3.

[0082] Step 4: Construct a group-level graph structure based on the graph-level representation of each variable group. Under the constraints of the process topology-aware mask constraint matrix, use a graph neural network for information propagation and aggregation to learn the causal relationships between groups and obtain the causal representation between groups. The specific steps are as follows: First, the graph-level embedding representation of each variable group is obtained from step 3(1). Construct the data features of the inter-group graph, denoted as: ; in, Indicates between groups The feature matrix of each unit.

[0083] Secondly, construct the inter-group causal adjacency matrix, denoted as: ; in, This represents the causal matrix that can be learned between groups. This represents the initial trainable parameters.

[0084] Next, the constructed process topology-aware mask constraint matrix is ​​obtained from step 3(2). This is integrated into the learning process of the inter-group causal adjacency matrix to guide the learning of causal relationships between variables and groups, denoted as: ; in, This represents the inter-group causality matrix with embedded mask constraints. This indicates a row normalization operation.

[0085] Subsequently, the prediction head is output through a single graph convolutional network and an MLP. The predicted values ​​of each variable at future times are obtained, thus completing the prediction task under the Granger causality paradigm, denoted as: ; in, Indicates all Variables in Predicted value at time, Represents the ReLU activation function. These are trainable parameters.

[0086] Furthermore, the between-group prediction loss function can be expressed as: The mean squared error (MSE) between the predicted and actual values ​​of a variable is denoted as: ; in, The squared norm of a vector is the sum of the squares of its elements.

[0087] Finally, based on steps 2 and 4, to jointly optimize the learning of intra-group and inter-group causal relationships and constrain the sparsity of the causal structure, the overall loss function is constructed as follows: ; in, Indicates the predicted loss within the group. Indicates the predicted loss between groups. This represents the sparse constraint loss within the group; This represents the sparse constraint coefficient.

[0088] Subsequently, with the overall loss function as the optimization objective, the model parameters are trained based on the training set, and the model parameters are iteratively updated using the backpropagation algorithm until the preset training epochs are reached. This allows us to obtain the optimal hierarchical causal model.

[0089] In this embodiment, Set to 0.6, training rounds Set it to 80.

[0090] Step 5: Design a grouped root cause analysis algorithm. Construct a global causal matrix based on within-group and between-group causal representations, calculate the root cause contribution of each variable, and identify abnormal propagation paths through path weight analysis. The specific steps are as follows: (1) Identification of root dependent variable Based on steps 2 and 4, obtain the intra-group causal matrix respectively. ( ) and between-group causal matrix Next, the between-group causal matrix will be... Extend along the column direction to By considering the variable dimensions, we obtain the inter-group causal expansion matrix. It is represented as: ; in, This represents the matrix expansion function, which is based on the number of variables contained in each variable group. The first group of the inter-group causal matrix is ​​then sequentially... Column copying Next, the expanded matrix is ​​row normalized so that the sum of the elements in each row is 1.

[0091] Subsequently, based on the intragroup causal matrix To correct the inter-group causal extension matrix The diagonal block matrix yields the global causality matrix. It is represented as: ; in, This represents the diagonal block correction function, which performs element-wise weighted correction on the diagonal blocks of the inter-group causal expansion matrix based on the intra-group causal matrix of each variable group, while keeping the off-diagonal blocks unchanged.

[0092] Specifically, Represented as a block matrix divided by variable groups: ; in, Indicates the first The variable group and the first Block submatrices between variable groups .

[0093] Subsequently, based on the above-mentioned block structure, the intra-group causal matrix was used. ( For block matrices Correct the diagonal blocks to construct a global causal matrix. It is represented as: ; Among them, matrix The non-negative sum of its rows is 1. Indicates the total number of variable groups. This represents the element-wise product of Hadamard.

[0094] Furthermore, based on the constructed global causal matrix Define the root cause contribution of each variable as: ; in, Indicates the first The root cause contribution of each variable, Representing variables For variables The causal contribution weight.

[0095] The above relationship can be expressed as an equation, denoted as: ; in, Represents the root cause contribution vector, matrix Represents the global causal matrix The transpose of .

[0096] Under these conditions, according to the Perron–Frobenius theorem, the matrix There exists a non-negative eigenvector with a corresponding eigenvalue of 1. By solving for this eigenvector, the non-negative contribution scores of each variable as a root cause can be obtained. .

[0097] Finally, the root cause contributions of each variable are compared, and the variable with the largest contribution is selected. As the root dependent variable, its corresponding index is represented as: .

[0098] (2) Identification of abnormal propagation paths According to step 5(1), obtain the global causal matrix. and the identified root causes Used for identifying abnormal propagation paths. Next, to preserve the main information propagation relationships, the global causal matrix is... A threshold pruning strategy is used for sparsification to obtain a sparse global causal matrix. Its elements are defined as: ; in, Represents the sparse global causal matrix The elements in Represents the original global causal matrix The elements in This represents the sparse pruning hyperparameter.

[0099] Subsequently, the root dependent variable was used respectively. and downstream abnormal variables As the starting and ending points, based on the sparsed global causal matrix Construct a set of candidate propagation paths Any candidate propagation path is represented as follows: ; in, This represents a propagation path from the root dependent variable to downstream outliers. Indicates the path length.

[0100] In this embodiment, the sparse pruning hyperparameter Set to 0.2, path length Set to 4, downstream abnormal variables Set to suction resistance value ( ).

[0101] For any candidate propagation path Its propagation weight is defined as: ; in, Representing a path Edges between adjacent variables, This represents the causal propagation weight of the corresponding edge.

[0102] By comparing candidate path sets The propagation weights of each path are determined, and the path with the highest weight is selected as the anomaly propagation path, which can be expressed as: ; in, This represents the root dependent variable that was ultimately identified. Downstream abnormal variables The abnormal propagation path.

[0103] In this embodiment, the learned hierarchical causal structure is as follows: Figure 5 As shown. According to Figure 5 The two-level causal representation of within-group variables and between-group units, combined with the grouped root cause analysis algorithm, is used to infer the anomalous root cause of this case. It is the starting position of the suction ribbon ( "), its main abnormal propagation path like Figure 6 As shown.

[0104] Depend on Figure 6 It can be concluded that the abnormal cigarette draw resistance in this case can be explained as follows: in the VE unit of the cigarette rolling machine, the starting position of the draw tape ( As a root cause variable, when it deviates from the process setpoint, it will cause the actual position of the suction ribbon ( The tobacco bundle shifts, thus altering the working gap between itself and the leveling cutter disc. This abnormal gap causes unnatural cutting of the tobacco bundle as it passes through the cutter disc, resulting in fluctuations in the tobacco supply and affecting the actual weight of the tobacco stick. The change in tobacco filling amount is further transmitted axially and is reflected in the local segment quality as a value of 5 ( ). Deviations in the density of the cigarette pack can lead to uneven filling density, altering the characteristics of the airflow channels and ultimately affecting the draw resistance. Abnormal fluctuations. Expert verification of the process mechanism shows that the embodiments of this invention can accurately identify root cause variables and infer propagation paths consistent with actual mechanisms, demonstrating its good interpretability and root cause localization capabilities.

[0105] In this embodiment, to evaluate the consistency between the causal structure learned by different methods and the actual process mechanism, the proportion of violations of process topology constraints is used as an evaluation index, which is defined as follows: ; in, This represents the number of causal edges that violate process topology constraints. This represents the total number of causal edges learned by the model.

[0106] To further verify the superiority of the embodiments of the present invention, this embodiment compares it with three causal discovery methods: LG, cMLP, and MPGE, and sets up an ablation model (w / o topology) for removing process topology priors as a control group. All methods were tested using the same case data, and the experimental results are shown in Table 2.

[0107] Table 2 Comparison of results from different methods

[0108] In Table 2, For topological constraint violation of proportion; "√" indicates that the method cannot identify the root cause variable; "√" indicates that the propagation path can be identified. "Indicates that the propagation path cannot be identified."

[0109] As shown in Table 2, LG and cMLP have relatively low violation rates, mainly because sparse constraints cause the models to retain a small number of high-confidence causal edges, thereby reducing connections that do not conform to topological constraints. However, neither model can identify root causes or anomalous propagation paths. MPGE and ablation models, while enhancing structural expressiveness, introduce more potential causal relationships. In the absence of effective constraints, they are more prone to generating causal edges that violate process mechanisms, leading to a higher violation rate. Furthermore, their root cause identification results deviate from the true root causes. In contrast, the implementation method of this invention, after introducing process topological constraints, has a violation rate of 0, can accurately identify root causes and recover anomalous propagation paths, and demonstrates stronger causal discovery capabilities and interpretability compared to purely data-driven methods.

[0110] On the other hand, the present invention provides a process topology-aware causal tracing system for cigarette quality abnormalities during the cigarette rolling process. The system includes a processor configured to perform any of the methods described in the process topology-aware causal tracing method for cigarette quality abnormalities during the cigarette rolling process.

[0111] In another aspect, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement any of the methods described in the process topology-aware causal tracing method for abnormal cigarette quality during the cigarette rolling process.

[0112] The beneficial effects of this invention are: This invention addresses the challenge of accurately locating the root causes of quality anomalies during cigarette rolling under conditions of multi-process collaboration and multiple control loops. It proposes a process topology-aware causal tracing method for cigarette quality anomalies during the rolling process. First, variables are grouped according to their respective processes, and key quality indicators are separately categorized as downstream outcome variables, providing a unified representation basis for multi-level causal relationship modeling. By introducing triple prior constraints—process sequence, feedback control loops, and process distance attenuation—a process topology-aware mask constraint matrix is ​​constructed. This effectively embeds the physical structure information of the rolling process into the inter-group causal relationship learning process. Combined with intra-group causal graph modeling, unified modeling of multi-level causal relationships between intra-group variables and inter-group units is achieved. Compared to traditional purely data-driven causal modeling methods, this invention fully utilizes process topology priors to depict variable dependencies that better align with actual mechanisms and expert experience, thereby significantly improving the accuracy and reliability of causal graph structure learning. Furthermore, to accurately pinpoint the root cause, this invention proposes a grouped root cause analysis algorithm. By correcting the causal relationship between groups through the causal structure within groups, and combining Granger prediction contribution and causal graph, the algorithm can pinpoint the root cause of cigarette quality abnormalities and identify the abnormal propagation path, thus providing support for on-site production decisions.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0118] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0119] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0120] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0121] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A process topology-based causal tracing method for identifying cigarette quality anomalies during the cigarette rolling process, characterized in that, The method includes: To obtain historical time-series data of abnormal periods in the cigarette rolling process, in order to obtain characteristic representations of each variable group; Based on the feature representations of each variable group, an intra-group graph structure is constructed, and a graph neural network is used for information propagation and aggregation to obtain intra-group causal representations. Design a graph-level feature extractor to obtain graph-level representations of each variable group, and construct a process topology-aware mask constraint matrix; Based on the graph-level representation of each variable group, an inter-group graph structure is constructed, and under the constraint of the process topology-aware mask constraint matrix, a graph neural network is used for information propagation and aggregation to obtain inter-group causal representations. A grouped root cause analysis algorithm is constructed, and a global causal matrix is ​​built based on the intra-group causal representation and the inter-group causal representation. The root cause contribution of each variable is calculated to obtain the root cause variable. Based on the root dependent variable, abnormal propagation paths are identified through path weight analysis.

2. The method according to claim 1, characterized in that, Based on the feature representations of each variable group, an intra-group causal matrix is ​​constructed. A graph neural network is used for information propagation and aggregation to obtain the intra-group causal representation, including: The initial embedding features of each group of variables are used as the input node features of the corresponding graph network. A causal matrix is ​​constructed for each group of variables. Except for the last group, the causal matrix of each group is generated by training parameters through activation and normalization. The causal matrix of the last group is set as an identity matrix. Each group uses its causal matrix to perform multi-layer graph convolution operation on the node features within the group to obtain the intra-group causal representation of each group. Based on the intra-group causal representation described in each group, the predicted values ​​of the variables in that group at future times are obtained to obtain the prediction loss; The mean squared error between the predicted values ​​and the corresponding true values ​​of all groups is used as the prediction loss, and the sum of the L1 norms of all learnable causal matrices is used as the sparse loss. The prediction loss and the sparse loss are combined to jointly optimize all trainable parameters of the graph neural network, the multilayer perceptron, and the learnable causal matrices until convergence. Once the training converges, the learnable causal matrix is ​​the final learned within-group causal matrix.

3. The method according to claim 1, characterized in that, Design a graph-level feature extractor to obtain graph-level representations of each variable group, and construct a process topology-aware mask constraint matrix including: The process topology-aware mask constraint matrix is ​​obtained according to formula (1). ，（1） in, This is the process topology-aware mask constraint matrix. This is a process sequence constraint matrix. Here is the feedback loop constraint matrix. This is the process distance attenuation constraint matrix. This represents the Hadamard element-wise multiplication operation. Indicates the total number of variables. This indicates the total number of variable groups.

4. The method according to claim 3, characterized in that, The sequence constraint matrix of the processes is obtained according to formula (2): ，（2） in, This is a process sequence constraint matrix. Represents a vector index. Indicates the first The process index to which each variable belongs. Indicates the process index of the variable group. This represents the set of process pairs that have feedback control loops.

5. The method according to claim 3, characterized in that, The feedback loop constraint matrix is ​​obtained according to formula (3): ，（3） in, Here is the feedback loop constraint matrix. Represents a vector index. Indicates the first The process index to which each variable belongs. Indicates the process index of the variable group. This represents the set of process pairs that have feedback control loops. Indicates The first layer of graph convolutional information aggregation Individual variable characteristics, Indicates the first Graph-level embedding representation of groups of variables This represents the learnable scale parameter that controls the distribution of membership degrees.

6. The method according to claim 3, characterized in that, The process distance attenuation constraint matrix is ​​obtained according to formula (4): ，（4） in, This is the process distance attenuation constraint matrix. Represents a vector index. Indicates the first The process index to which each variable belongs. Indicates the process index of the variable group. This represents the set of process pairs that have feedback control loops. This is a hyperparameter used to control the degree to which causal strength decreases with distance.

7. The method according to claim 1, characterized in that, A grouped root cause analysis algorithm is constructed, which builds a global causal matrix based on the within-group causal representation and the between-group causal representation, and calculates the root cause contribution of each variable to obtain the root cause variables, including: Construct a global causal matrix according to formula (5). ，（5） in, This is the global causal matrix. Indicates the first The variable group and the first Block submatrices between variable groups , Indicates the total number of variable groups; The root cause contribution of each variable is obtained according to formula (6). ，（6） in, For the first The root cause contribution of each variable, For variables For variables Causal contribution weight, For the first The root cause contribution of each variable, The total number of variables; The root cause contribution is obtained according to formula (7). ，（7） in, Represents the root cause contribution vector; Choose the variable with the largest contribution as the root cause variable.

8. The method according to claim 1, characterized in that, Based on the root dependent variable, path weight analysis is used to identify abnormal propagation paths, including: The global causal matrix is ​​sparsified to obtain a sparse global causal matrix; Using the root dependent variable and the downstream abnormal variable as the starting and ending points respectively, a set of candidate propagation paths is constructed based on the sparsed global causal matrix. Calculate the weights of all candidate propagation paths and select the path with the highest weight as the anomaly propagation path.

9. A process topology-sensing causal tracing system for identifying abnormal cigarette quality during the cigarette rolling process, characterized in that, The system includes a processor configured to perform the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images