Industrial alarm causal network construction method based on deep structure equation model
The industrial alarm causal network construction method based on deep structural equation modeling, utilizing graph neural networks and variational autoencoder frameworks, combined with system topology and acyclic constraints, solves the simplification and redundancy problems of existing causal discovery methods, and achieves efficient and accurate causal relationship analysis of discrete alarm data, enabling the identification of the root causes of industrial alarms.
Patent Information
- Application Number
- CN202510866096.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-11-18
AI Technical Summary
Existing causal discovery methods based on alarm data oversimplify the model when analyzing discrete alarm data, failing to capture the characteristics of discrete alarm data. Furthermore, traditional methods ignore multivariate interactions and system topology, leading to redundancy and spurious causal relationships, making it difficult to accurately identify the root cause of alarms.
An industrial alarm causal network construction method based on deep structural equation modeling is adopted. The model is trained by graph neural network and variational autoencoder framework to learn the causal relationship between alarm variables. System topology and acyclic constraints are introduced to optimize the causal relationship graph and eliminate redundancy and spurious connections.
It achieves efficient causal relationship discovery, accurately identifies causal relationships among multiple variables and latent variables, eliminates redundancy and spurious connections, and helps to quickly locate the root cause of alarms.
Smart Images
Figure CN120975136A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of industrial process fault monitoring and alarm monitoring, and more specifically, relates to a method for constructing an industrial alarm causal network based on a deep structural equation model. Background Technology
[0002] Industrial alarm systems serve as a critical safety guarantee in complex manufacturing facilities, enabling operators to promptly identify abnormal process conditions and take corrective measures. However, modern alarm systems face significant challenges, including interference alarms, alarm proliferation, and difficulties in root cause analysis. To address the root cause analysis of alarms, a common approach is to discover causal relationships between alarm variables. Existing alarm causal discovery methods are primarily based on two types of data: continuous process data and discrete alarm data. Process data-based causal discovery methods analyze causal relationships between variables by directly utilizing sensor measurement data, while alarm data-based methods offer a more computationally efficient alternative.
[0003] However, existing causal discovery methods based on alarm data primarily apply time-series causal discovery techniques designed for continuous process data to discrete alarm data. This may oversimplify the model and fail to capture the characteristics of discrete alarm data. Furthermore, traditional causal discovery methods, such as Granger causality and transitivity, mainly analyze pairwise causal relationships, neglecting multivariate interactions and latent variables. In addition, existing methods often identify redundant or even spurious causal relationships, making the causal graph complex and difficult to identify the root cause of the alarm.
[0004] Therefore, it is particularly important to design a causal discovery method suitable for discrete alarm data, while taking into account multivariate interactions and system topology, to achieve accurate identification of the root cause of alarms. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a method for constructing industrial alarm causal networks based on deep structural equation modeling. This method directly utilizes discrete multi-valued alarm sequences to discover causal relationships, analyzes multivariate causal relationships and latent variables, and accurately identifies the root cause of alarms through acyclic and system topology constraints.
[0006] To achieve the above objectives, in a first aspect, this application provides a method for constructing causal networks for industrial alarms based on deep structural equation modeling, for use in industrial alarm systems, comprising the following steps: S10, Extract multi-value alarm sequence data within a certain period of time. X , X The rows in the table represent the number of samples, and each column represents an alarm variable, with each alarm variable having multiple discrete state values. S20, for multi-value alarm sequence data X One-hot encoding is performed to map each discrete state value to a continuous space; S30 trains a pre-established structural equation model based on a graph neural network using a variational autoencoder framework, and learns the adjacency matrix in the model to represent the causal relationships between alarm variables during the training process. A The model architecture includes: an encoder, used to process the encoded data. X Mapping to latent space Z Decoder, used to extract from latent space Z Reconstruct alarm sequence X ; S40 introduces system topological constraints and acyclic constraints, and solves the constrained optimization problem using the augmented Lagrangian method to obtain the optimal adjacency matrix. A Among them, system topology constraints ensure that the obtained causal relationship conforms to the actual physical connection by utilizing the topological connection between the various parts of the system, and acyclic constraints ensure that the obtained causal relationship graph is a directed acyclic graph by controlling the parameters of constraint strength. S50, for the optimal adjacency matrix A Threshold processing is performed to obtain the causal relationship between alarm variables.
[0007] The beneficial effects of the industrial alarm causal network construction method based on deep structural equation model provided in this application are: (1) It directly uses discrete multi-valued alarm sequences to discover causal relationships, which is computationally efficient and the multi-valued alarm sequence data can directly reflect the alarm status at different times; (2) It analyzes multivariate causal relationships and latent variables at the same time, which is more comprehensive than the traditional pairwise causal analysis method; (3) It eliminates redundancy and false connections through acyclic and system topology constraints, thereby improving the accuracy of causal inference; (4) It can accurately identify the root cause of the alarm, helping operators to quickly locate the problem and take corrective measures.
[0008] As a further preferred embodiment, in step S10, the multi-value alarm sequence data X The discrete state values include high-high alarm, high alarm, normal, low alarm, and low-low alarm.
[0009] As a further preferred embodiment, in step S30, both the encoder and the decoder contain GRU networks to capture nonlinear and temporal features; The encoder is:
[0010] The decoder is:
[0011] In the formula, and This represents an identity mapping. , This corresponds to the GRU network in the decoder and encoder; I Represents the identity matrix.
[0012] As a further preferred embodiment, the GRU network in the encoder and decoder and The calculation formula is:
[0013]
[0014]
[0015]
[0016]
[0017] In the formula, This represents the parameter matrix of the output layer; Indicates the bias of the output layer; Indicates a hidden state; Indicates the candidate hidden layer state; Indicates an update to the door; Indicates that the door is being reset; ( ) represents the parameter matrix input to the gate. ( ) represents the parameter matrix from the previous hidden state to the gate; ( () indicates the bias under different gate mechanisms; The first part of the data represents the first part of the data. t One sample.
[0018] As a further preferred embodiment, in step S30, the encoder is used to convert the variational posterior... Parameterized to have a mean and standard deviation The factor Gaussian distribution; the decoder is used to... The parameterized distribution is a factor classification distribution, and its probability matrix is calculated using the softmax function.
[0019] As a further preferred embodiment, in step S40, the system topology constraints are as follows:
[0020] Acyclic constraints are:
[0021] In the formula, A binary matrix representing the physical connections of the system. Representing variables to variable There is an actual physical connection; m Indicates the number of alarm variables; These are parameters that control the strength of the constraint. It is a function that ensures no cycles; This indicates that updates will be made gradually during training. As a further preferred embodiment, in step S40, when the augmented Lagrangian method solves the constrained optimization problem, the optimization objective is a Lagrangian function, which is achieved by alternately minimizing the adjacency matrix. A and all neural network parameters in the encoder and decoder The Lagrange function is calculated, and the penalty coefficients and Lagrange multipliers are updated for optimization.
[0022] Secondly, this application provides an application of the industrial alarm causal network construction method based on deep structural equation model as described in any of the above claims, which can be applied to the root cause analysis of alarms in industrial processes, alarm proliferation management, and alarm optimization.
[0023] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the industrial alarm causal network construction method based on deep structural equation model as described in any of the preceding claims.
[0024] Fourthly, this application provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the steps of the industrial alarm causal network construction method based on a deep structural equation model as described in any of the preceding claims.
[0025] It is understandable that the beneficial effects of the second, third and fourth aspects mentioned above can be found in the relevant descriptions in the first aspect above, and will not be repeated here. Attached Figure Description
[0026] Figure 1 This is a flowchart of the industrial alarm causal network construction method based on deep structural equation model provided in this application; Figure 2 This is an overall framework diagram of the industrial alarm causal network construction method based on deep structural equation model provided in the embodiments of this application; Figure 3 This is a diagram of the deep structural equation model architecture provided in the embodiments of this application; Figure 4 This is a flowchart of the Tennessee-Eastman process provided in the embodiments of this application; Figure 5 This is a multi-value alarm sequence diagram in the Tennessee-Eastman process provided in the embodiments of this application; Figure 6 This is a causal relationship diagram between alarm variables in the Tennessee-Eastman process provided in the embodiments of this application. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0028] The purpose of this application is to provide a method for constructing causal networks for industrial alarms based on deep structural equation modeling. This method utilizes the deep learning method Graph Neural Network (GNN) to parameterize a novel structural equation model (SEM), namely Deep SEM (DSEM). This method directly utilizes discrete multi-valued alarm sequences, enabling simultaneous analysis of multivariate causal relationships and latent variables, while considering system topology and acyclic constraints, thereby more accurately identifying the root cause of alarms.
[0029] like Figure 1 As shown, this application provides a method for constructing causal networks for industrial alarms based on deep structural equation modeling. This method can be applied to root cause analysis of alarms in industrial processes, alarm proliferation management, and alarm optimization. It includes the following steps: Step S10: Extract multi-value alarm sequence data over a period of time. X The rows represent the number of samples, and each column represents an alarm variable. Each alarm variable has multiple discrete state values, including high-high alarm (HH), high alarm (HI), normal (NOR), low alarm (LO), and low-low alarm (LL).
[0030] Step S20: Process the multi-value alarm sequence data. X One-hot encoding is performed to map each discrete state value to a continuous space. The calculation formula is as follows:
[0031] In the formula, It is one-hot encoded data. i The first number of data points represents the first number of data points. i One sample, It is a one-hot encoding function. m Indicates the number of alarm variables.
[0032] Step S30: Train the pre-established structural equation model based on graph neural network using a variational autoencoder framework, and learn the adjacency matrix in the model to represent the causal relationship between alarm variables during the training process. A .
[0033] In this application, step S30 may specifically be: First, an encoder-decoder architecture is constructed based on graph neural networks and structural equation models.
[0034] The encoder will encode the processed data. X Mapping to latent space Z The formula is:
[0035] Decoder is used to extract from latent space Z Reconstruct alarm sequence X The formula is:
[0036] In the formula, I Represents the identity matrix; and For identity mapping, and For a Gated Recurrent Unit (GRU) network, Z represents the latent variable.
[0037] Secondly, the model parameters, including the causal adjacency matrix, are learned through the Variational Autoencoders (VAE) framework. A And the relevant parameters of the GRU network. a) Inferring the model, using variational posterior... Parameterized to have a mean and standard deviation The factor Gaussian distribution: b) Generate a model, Parameterized as a factor classification distribution, its probability matrix is : c) Maximizing the lower bound of evidence (ELBO): The KL divergence term is The reconstructed precision term is .
[0038] Step S40: Introduce system topological constraints and acyclic constraints, and solve the constrained optimization problem using the augmented Lagrangian method to obtain the optimal adjacency matrix. A .
[0039] In step S40, system topology constraints ensure that the resulting causal relationships conform to the actual physical connections by utilizing the topological connections between the various parts of the system. Acyclic constraints ensure that the resulting causal relationship graph is a directed acyclic graph (i.e., it does not contain any loops) by controlling the parameters of the constraint strength.
[0040] Specifically, the system topology constraints are:
[0041] In the formula, This indicates that updates will be made gradually during training. A binary matrix representing the physical connections of the system. Representing variables to variable There is an actual physical connection.
[0042] Acyclic constraints are:
[0043] In the formula, It is a function that ensures no cycles; It is a parameter that controls the strength of the constraint.
[0044] Solving constrained optimization problems using the augmented Lagrange method:
[0045]
[0046]
[0047] The optimization objective is the Lagrange function. .
[0048] Step S50: Perform threshold processing on the optimal adjacency matrix A to obtain the causal relationship between alarm variables.
[0049] The beneficial effects of the industrial alarm causal network construction method based on deep structural equation model provided in this application are: (1) It directly uses discrete multi-valued alarm sequences to discover causal relationships, which is computationally efficient and the multi-valued alarm sequence data can directly reflect the alarm status at different times; (2) It analyzes multivariate causal relationships and latent variables at the same time, which is more comprehensive than the traditional pairwise causal analysis method; (3) It eliminates redundancy and false connections through acyclic and system topology constraints, thereby improving the accuracy of causal inference; (4) It can accurately identify the root cause of the alarm, helping operators to quickly locate the problem and take corrective measures.
[0050] The present application will now be described in detail with reference to specific embodiments.
[0051] This embodiment provides a method for constructing an industrial alarm causal network based on a deep structural equation model, such as... Figure 2 As shown, the main steps include: Step 1: DSEM Architecture: Based on the principles of GNN and SEM, a new causal discovery architecture, DSEM, is defined, which uses an encoder-decoder structure to parameterize the causal relationships between alarm variables.
[0052] In this embodiment, structural equation modeling is a multivariate statistical analysis method that reveals causal relationships between variables by representing the relationship between multiple variables as a functional relationship between one or more latent variables and observed variables. It can be expressed as:
[0053] in, Indicates having The weighted adjacency matrix of a directed acyclic graph (DAG) with n nodes. express A joint distribution sample of variables. Represents latent variables. The matrix represents the nodes arranged in topological order. A strictly upper triangular generative process (generating observed variables from latent variables) can be represented as:
[0054] However, a significant drawback of SEM is its applicability only to linear processes, limiting its suitability for nonlinear industrial processes. Therefore, this embodiment proposes an improved SEM architecture based on GNN, namely the DSEM architecture. This architecture leverages the powerful nonlinear learning capabilities of GNN to extend SEM, thereby making it suitable for nonlinear industrial processes. This architecture directly integrates multi-value alarm sequences... As input, and reconstructed through an encoding-decoding mechanism. The adjacency matrix between each node AIt is a learnable parameter that represents the causal relationship between alarms. For example... Figure 3 As shown, the architecture includes the following components: 1) Encoder (Inference Model): This encodes the alarm sequence. Mapping to latent space ; 2) Decoder (Generative Model): From the latent space Reconstruct alarm sequence ; 3) Adjacency matrix : Indicates the causal relationship between alarm variables.
[0055] Specifically, based on the analysis of SEM, the equation It can be written as This is an abstraction of parametric graph neural networks in deep learning, which incorporates the features of input nodes. Convert to high-level representation Based on this framework, a deep SEM architecture, namely the DSEM architecture, is proposed:
[0056] Among them, parameterized functions and As a transformation operator, in and A nonlinear mapping is introduced. To extract the temporal dynamics from the multi-value alarm sequence, this embodiment employs an asymmetric configuration, wherein... Defined as an identity mapping, Defined as GRU.
[0057] Based on the decoder, the corresponding encoder is:
[0058] in, and As respectively and The inverse function, i.e. It's a GRU. It is an identity mapping function.
[0059] Both the encoder and decoder incorporate GRUs to capture nonlinear and temporal features. For the function in the encoder... , update door Reset door and candidate hidden layer states Given by the following formula:
[0060]
[0061]
[0062] in, ( ) represents the parameter matrix input to the gate. ( ) represents the parameter matrix from the previous hidden state to the gate; ( () indicates the bias under different gate mechanisms; It is the tanh function. It is the Hadamard (element-by-element) product; The first number of data points represents the first number of data points. t One sample. Final hidden state. Given by the following formula:
[0063] Then, based on the hidden state Output Given by the following formula:
[0064] in, It is the parameter matrix of the output layer. This refers to the output layer bias. Additionally, the functions in the decoder... It is also a GRU, and the computation process is similar to... similar.
[0065] Step 2: Parameter learning through variational inference: The model parameters are learned using a variational autoencoder framework improved for discrete alarm data.
[0066] In this embodiment, given that multi-value alarm sequences are discrete data, a variational autoencoder method is employed to learn the causal relationships between alarms. First, in the encoding stage, one-hot encoding is used to map discrete values to a continuous space. Then, in the decoding stage, a softmax function is integrated into the output layer to accommodate the discrete nature of the alarm data. For discrete multi-value alarm sequences, a one-hot encoding function is used. O() Map each element as:
[0067] To specify the distribution in the VAE, this embodiment uses the standard matrix normal distribution. As a prior distribution, the inference model will use the variational posterior. Parameterized to have a mean and standard deviation The factor Gaussian distribution is derived by the encoder:
[0068] For generative models, in order to handle the discrete nature of variables, let It becomes a factor classification distribution, and its probability matrix is Each row is a probability vector for the corresponding categorical variable. Then, a softmax function is added. It can be calculated from the decoder:
[0069] In variational reasoning, the lower bound of evidence (ELBO) is designed as an optimization objective, and its calculation formula is as follows:
[0070] in It can be calculated using the following formula:
[0071] The first term in the above formula represents the Kullback-Leibler (KL) divergence between the approximate posterior probability and the prior probability, ensuring that the latent representation conforms to the prior distribution. The second term measures the reconstruction accuracy, ensuring that the model can regenerate the alarm sequence. Given a sample The encoder maps it to a latent distribution. The decoder, on the other hand, attempts to use conditional density. Refactoring .
[0072] Based on the analysis of the above formula, the KL divergence term in ELBO is in closed-form:
[0073] The reconstruction accuracy term can be approximated using Monte Carlo methods:
[0074] in, This is the decoder output, in Monte Carlo sample form. As input, .
[0075] Step 3: Constraint-based alarm causality inference: The causal graph is simplified and spurious connections are eliminated by introducing acyclic and system topology constraints.
[0076] In this embodiment, when analyzing the causal relationships of multiple alarm variables simultaneously, excessive redundancy or even spurious causal connections often occur, making it difficult to accurately identify the root cause of the alarm. This embodiment introduces acyclic constraints and topological constraints to infer causal relationships, thereby eliminating redundancy and spurious connections.
[0077] For acyclic constraints: The causal relationships in the SEM model are inherently acyclic, so constraints are needed to ensure that the generated graph is a DAG. This embodiment uses differentiable constraints to ensure that the generated graph is a DAG, i.e., it satisfies the following conditions:
[0078] in, This parameter controls the strength of the acyclic constraint and needs to be set according to the complexity of the graph structure to ensure stable convergence of the algorithm. This constraint is equivalent to forcing the graph to contain no cycles.
[0079] For system topology constraints: Since purely data-driven causal relationship discovery methods often ignore system topology, leading to misleading causal relationships, this embodiment introduces topological constraints to maintain consistency with the system structure.
[0080] set up Let be a binary matrix representing the physical connections of the system, where Representing variables to variable There is a physical connection. This indicates that no connection exists. Adjacency matrix under topological constraints. Defined as:
[0081] This operation effectively masks physically impossible connections during the optimization process. Topological constraints, derived from process flow diagrams, engineering designs, or expertise in industrial systems, provide a mechanism for incorporating domain expertise into the causal relationship discovery process.
[0082] Constraint-based model training: Based on the above learning objectives and constraints, the learning problem of the method in this embodiment can be summarized as follows:
[0083]
[0084]
[0085] in, This represents all the neural network parameters in the encoder and decoder. This constrained optimization problem is solved using the augmented Lagrangian method, where the Lagrangian function is defined as:
[0086] in, Represents the Lagrange multipliers. This represents the penalty coefficient. The optimization process involves alternately minimizing the value of... and of and update and conduct.
[0087] The following section illustrates the application of the industrial alarm causal network construction method provided in this application to the Tennessee-Eastman process.
[0088] The Tennessee-Eastman process (TEP) is a classic chemical process simulation model widely used for the validation of process monitoring technologies. For example... Figure 4 As shown, the main units of the TEP include a reactor, a product condenser, a gas-liquid separator, a product stripping column, and a recycle compressor. The process involves gaseous reactants A, C, D, and E, with final products G and H. The simulator provides data for 41 measured variables and 12 manipulated variables. These 41 measured variables can be further divided into 22 continuous measured variables (including temperature, pressure, level, and flow rate) and 19 compositional analytical variables obtained through an analyzer.
[0089] In this embodiment, the alarm propagation mechanism of TEP under IDV(1) fault conditions was investigated. When IDV(1) occurs, the A / C feed ratio changes, while the composition of B remains stable. This change propagates through the process via a multidimensional physicochemical mechanism: the closed-loop controller initially responds to the change by increasing the flow rate of A (F101). This flow rate change directly leads to a stoichiometric imbalance in the A and C feed flow rates (F104), while simultaneously triggering a compensating response from the A feed flow valve (XMV3). The deviation in the A / C ratio significantly alters the reaction kinetics within the reaction zone, causing fluctuations in reactor pressure (P201). This pressure change affects the reactor level (L210) through thermodynamic and hydrodynamic mechanisms and is transmitted to the separator pressure (P401). The pressure disturbance continues to propagate through the two-phase material flow to the stripping column pressure (P501), subsequently causing a deviation in the stripping column temperature (T510), which prompts the stripping column steam valve (XMV9) to intervene.
[0090] Meanwhile, changes in reactant ratios directly affect the relative distributions of component A (C601) and component C (C603) concentrations in stream 9 via stoichiometry. These changes in reaction conditions ultimately affect the quality of downstream products, particularly the concentration of component E (C705) in stream 11, through a complex chemical equilibrium shift mechanism.
[0091] The alarm variables used in this embodiment and their descriptions are shown in Table 1: Table 1. Alarm variables and their descriptions in the Tennessee-Eastman process.
[0092] Configure alarms in the model to generate multi-valued alarm sequences, with each alarm variable having five different alarm states: HH, HI, NOR, LO, LL, as shown below. Figure 5 As shown. Based on the multi-value alarm sequences generated under the IDV(1) condition, the DSEM method proposed in this application is used to discover causal relationships.
[0093] First, the multi-valued alarm sequences of each variable in the TEP are converted into one-hot encoded feature matrices to capture the dynamic characteristics of alarm state transitions. Then, a DSEM model is trained using a VAE framework, where the encoder uses a GRU to extract temporal features from the alarm sequences, and the decoder reconstructs the alarm state distribution and adjacency matrix. As a learnable parameter representing the causal relationship between alarm variables.
[0094] Figure 6 The alarm cause-effect graph constructed by the method of this application under IDV(1) faults is shown, clearly demonstrating the causal connections between alarm variables. Further analysis shows that this cause-effect graph is highly consistent with the theoretical analysis of alarm propagation paths under IDV(1). For example, the connections from F101 to F104 and from F104 to P201 conform to the mechanism that the proportion of reactants affects reaction kinetics in the physical process. The graph also accurately captures the propagation chain under pressure anomalies, i.e., P201 affects P401, and P401 affects P501. Finally, through analysis, F101 is identified as the root cause alarm without preceding nodes, which is consistent with the theoretical analysis.
[0095] This result not only verifies the effectiveness of the proposed method in alarm causation detection, but also provides a reliable data-driven approach for root cause analysis in industrial processes, helping operators quickly locate the root cause alarm and implement targeted measures.
[0096] Furthermore, to further verify the superiority of the proposed method, it was compared with traditional SEM methods, transfer entropy (TE), and deep learning-based Neural Granger Causality (NGC). The accuracy of the causal discovery results of different methods is presented in Table 2 after comparing them with the actual results. The proposed method exhibits the highest accuracy in causal discovery, demonstrating its superiority.
[0097] Table 2: Accuracy Comparison of Different Causal Discovery Methods
[0098] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for constructing causal networks for industrial alarms based on deep structural equation modeling, used in industrial alarm systems, characterized in that, Includes the following steps: S10, Extract multi-value alarm sequence data within a certain period of time. X , X The rows in the table represent the number of samples, and each column represents an alarm variable, with each alarm variable having multiple discrete state values. S20, for multi-value alarm sequence data X One-hot encoding is performed to map each discrete state value to a continuous space; S30 trains a pre-established structural equation model based on a graph neural network using a variational autoencoder framework, and learns the adjacency matrix in the model to represent the causal relationships between alarm variables during the training process. A The model architecture includes: an encoder, used to process the encoded data. X Mapping to latent space Z Decoder, used to extract from latent space Z Reconstruct alarm sequence X ; S40 introduces system topological constraints and acyclic constraints, and solves the constrained optimization problem using the augmented Lagrangian method to obtain the optimal adjacency matrix. A Among them, system topology constraints ensure that the obtained causal relationship conforms to the actual physical connection by utilizing the topological connection between the various parts of the system, and acyclic constraints ensure that the obtained causal relationship graph is a directed acyclic graph by controlling the parameters of constraint strength. S50, for the optimal adjacency matrix A Threshold processing is performed to obtain the causal relationship between alarm variables.
2. The method for constructing an industrial alarm causal network based on a deep structural equation model as described in claim 1, characterized in that, In step S10, the multi-value alarm sequence data X The discrete state values include high-high alarm, high alarm, normal, low alarm, and low-low alarm.
3. The method for constructing an industrial alarm causal network based on a deep structural equation model as described in claim 1, characterized in that, In step S30, both the encoder and decoder contain GRU networks to capture nonlinear and temporal features; The encoder is: The decoder is: In the formula, and This represents an identity mapping. , This corresponds to the GRU network in the decoder and encoder; I Represents the identity matrix.
4. The method for constructing an industrial alarm causal network based on a deep structural equation model as described in claim 3, characterized in that, GRU network in encoder and decoder and The calculation formula is: In the formula, This represents the parameter matrix of the output layer; Indicates the bias of the output layer; Indicates a hidden state; Indicates the candidate hidden layer state; Indicates an update to the door; Indicates that the door is being reset; ( ) represents the parameter matrix input to the gate. ( ) represents the parameter matrix from the previous hidden state to the gate; ( () indicates the bias under different gate mechanisms; The first part of the data represents the first part of the data. t One sample.
5. The method for constructing an industrial alarm causal network based on a deep structural equation model as described in claim 1, characterized in that, In step S30, the encoder is used to convert the variational posterior... Parameterized to have a mean and standard deviation The factor Gaussian distribution; the decoder is used to... The parameterized distribution is a factor classification distribution, and its probability matrix is calculated using the softmax function.
6. The method for constructing an industrial alarm causal network based on a deep structural equation model as described in claim 1, characterized in that, In step S40, the system topology constraints are: Acyclic constraints are: In the formula, A binary matrix representing the physical connections of the system. Representing variables to variable There is an actual physical connection; m Indicates the number of alarm variables; These are parameters that control the strength of the constraint. It is a function that ensures no cycles. This indicates that updates will be made gradually during training.
7. The method for constructing an industrial alarm causal network based on a deep structural equation model as described in claim 1, characterized in that, In step S40, when the augmented Lagrangian method solves the constrained optimization problem, the optimization objective is the Lagrangian function, which is achieved by alternately minimizing the adjacency matrix. A and all neural network parameters in the encoder and decoder The Lagrange function is calculated, and the penalty coefficients and Lagrange multipliers are updated for optimization.
8. An application of the industrial alarm causal network construction method based on deep structural equation model as described in any one of claims 1 to 7, characterized in that, Applications include root cause analysis of alarms in industrial processes, alarm proliferation management, and alarm optimization.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the industrial alarm causal network construction method based on deep structural equation model as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by a processor, the computer instructions implement the steps of the industrial alarm causal network construction method based on a deep structural equation model as described in any one of claims 1 to 7.