A method and system for diagnosing root causes of failures in industrial production

Abstract

The present application relates to the technical field of fault root cause diagnosis in the production of hydrocracking reaction process, and particularly relates to a fault root cause diagnosis method and system for industrial production, which comprises the following steps: obtaining historical production data of a hydrocracking reactor; modeling the historical production data using a probabilistic graphical model to obtain a fault probability propagation probability network of each predicted variable and control variable, and calculating the root cause variable of the actual predicted variable by optimizing the network path; the predicted variable comprises a target predicted bit number; inputting the bit number value of the fault into the probabilistic graphical model to obtain the fault probability of each control variable, and obtaining a fault solution by searching a production process knowledge graph.

Description

Technical Field

[0001] This invention relates to the technical field of root cause diagnosis of failures in hydrocracking production processes, and particularly to a method and system for root cause diagnosis of failures in industrial production. Background Technology

[0002] Hydrocracking is a petroleum refining process that involves the cracking of heavy oil under conditions of heating, high hydrogen pressure, and the presence of a catalyst, converting it into gases, gasoline, jet fuel, or diesel. The main characteristics of hydrocracking are high production flexibility and good product quality stability. However, the production conditions of hydrocracking are complex and demanding, exhibiting multiple processes and non-steady-state characteristics. Therefore, it is difficult to locate production faults in a timely manner, and relying solely on human experience for judgment is insufficient in terms of accuracy and timeliness.

[0003] Currently, there is no literature on root cause diagnosis of hydrocracking reactor failures. However, there are some studies on the production process knowledge base of hydrocracking units. Most of these studies involve upgrading and iterating the production knowledge base based on mechanisms, historical data, and specific enterprise production process rules, thereby ensuring the identification and handling of common production anomalies. The disadvantage of this approach is the slow iteration of the production knowledge base. Even with the reliance on the production knowledge base, the handling of production anomalies still heavily depends on human experience. If humans cannot accurately diagnose abnormal operating conditions, it will significantly increase production safety risks.

[0004] Currently, fault diagnosis for hydrocracking reactors largely employs two approaches. One focuses on optimizing production rules, using refined and customized rules to more accurately define and control production anomalies. The other is based on supervised machine learning models, quantifying production conditions at the data level, identifying these conditions through the model, and then processing them accordingly. In practice, these methods typically require experienced personnel to subjectively judge, label, and maintain production faults. However, neither approach considers the significant impact of differences in subjective experience among different personnel on the rules and labeled data.

[0005] Therefore, it is essential to combine production process knowledge bases and establish data-driven models to diagnose production failures in a timely and accurate manner, which can effectively reduce production risks and losses. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method and system for diagnosing the root causes of failures in industrial production, which solves the technical problem that it is difficult to locate production failures in a timely manner in the existing hydrocracking production process.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0010] In a first aspect, embodiments of the present invention provide a method for diagnosing the root causes of faults in industrial production, comprising the following steps:

[0011] Obtain historical production data from the hydrocracking reactor;

[0012] Historical production data is modeled using a probabilistic graphical model to obtain the failure probability propagation probability network for each predictor and control variable. By optimizing the network path, the root cause variable of the actual predictor variable failure is calculated. The predictor variable includes the target predictor number.

[0013] Input the faulty tag number into the probabilistic graphical model to obtain the fault probability of each control variable of the current fault, and obtain the fault solution by searching the production process knowledge graph.

[0014] The industrial production failure root cause diagnosis method proposed in this invention employs a production process knowledge graph and establishes a probabilistic graphical model to trace, reason about, and diagnose the root causes of production failures. By establishing a causal relationship model between production goals (predictive variables) and operational indicators (control variables), quantitative root cause results are provided to employees. This enables operators to trace and reason about industrial production failures more quickly and accurately, significantly reducing the safety risks for enterprises.

[0015] Optionally, after obtaining the failure probability of each control variable for the current fault, a fault solution is obtained by searching the production process knowledge graph, the fault tag value and the failure probability of the control variable are verified, and the verification data is updated to the historical production data periodically.

[0016] Optionally, historical production data includes: tag information data and mechanism data;

[0017] The tag information data includes: the tag number of the predictive indicator and the tag number of the operation control;

[0018] Mechanistic data include: inlet temperature of each bed in the reactor, outlet temperature of each bed in the reactor, temperature rise of each bed in the reactor, total temperature rise of the reactor, reactor pressure, inlet temperature of the heat exchanger, outlet temperature of the heat exchanger, as well as feed rate, fresh hydrogen rate and circulating hydrogen rate.

[0019] Optionally, a probabilistic graphical model can be used to model historical production data, including the following steps:

[0020] Based on production needs, determine the relevant predictive variables and the names of control variables that may cause anomalies in the target tag number, and obtain the training dataset; standardize the training dataset to obtain a new sample dataset;

[0021] Calculate the mutual information coefficient between each pair of variables in the training dataset to obtain the degree of correlation between variables; based on the calculated maximum mutual information coefficient, take each predictor variable and control variable as nodes, retain the edges between the two variables with the maximum mutual information coefficient value, and construct the maximum spanning tree model;

[0022] The number of edges connected to a node in the maximum spanning tree model is regarded as the number of the node’s maximum parent nodes, and the set of the maximum parent nodes of a node is determined based on the maximum spanning tree model.

[0023] Sort the nodes by importance;

[0024] The new sample dataset, the importance ranking of nodes, and the maximum number of parent nodes for each node are used as inputs to learn the parameters of the Bayesian network, and the K2 algorithm is used to solve for the optimal Bayesian network structure.

[0025] Optionally, after finding the optimal Bayesian network structure using the K2 algorithm, the Bayesian network structure can be selectively optimized and adjusted based on expert experience.

[0026] Alternatively, the formula for calculating mutual information is as follows:

[0027]

[0028] Where I(X,Y) is the mutual information coefficient between variables X and Y; P(X,Y) is the joint distribution function of variables X and Y; and P(X) and P(Y) are the independent probability distribution functions of variables X and Y, respectively.

[0029] Optionally, in learning the parameters of the Bayesian network, the maximum a posteriori probability method is used, with the following steps:

[0030] The prior probability of the failure phenomenon is obtained by statistical calculation of historical production data;

[0031] For the prior probability of the observed node, assume that the observed node p follows a Beta distribution with parameters α and β: Beta(p:α,β);

[0032] Calculate the posterior distribution function p:

[0033]

[0034] get:

[0035]

[0036] Where, pi q represents the number of times the observed node P in the dataset fails. i The conditional probability table for each node is obtained by representing the number of times no fault occurred, and a Bayesian network is constructed.

[0037] Optionally, the production process knowledge graph includes six types of nodes: production materials, production equipment, failure conditions, control variables, target variables, and solutions;

[0038] The production process knowledge graph includes multiple triples: production equipment - occurrence - failure condition; failure condition - solution - solution; control variable - cause - failure condition; production equipment - query - target variable; production equipment - processing - production materials and control variable - failure - solution.

[0039] In a second aspect, embodiments of the present invention provide a fault root cause diagnosis system for industrial production, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above methods.

[0040] (III) Beneficial Effects

[0041] The beneficial effects of this invention are as follows: The industrial production failure root cause diagnosis method and system of this invention, by combining production process knowledge graph and probabilistic graphical model to trace, reason, and diagnose the root causes of production failures, can establish a causal relationship model between production targets (predictive variables) and operational indicators (control variables) compared to the prior art. This enables operators to trace and reason about industrial production failures more quickly and accurately, greatly reducing the safety risks of enterprises. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating a preferred embodiment of the industrial production fault root cause diagnosis method of the present invention.

[0043] Figure 2 This is a flowchart illustrating the modeling process of the root cause analysis model according to a preferred embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the reasoning query process according to a preferred embodiment of the present invention. Detailed Implementation

[0045] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] This invention proposes a fault probability graphical model that establishes predictive and operational variables using different process data from historical production. It calculates the root causes of fault indicators and uses knowledge graphs for rapid reasoning to derive corresponding solutions. This approach improves both work efficiency and significantly enhances production process safety, enabling preventative measures to address potential problems.

[0047] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0048] Example 1

[0049] Figure 1 This is a flowchart illustrating a method for diagnosing the root cause of industrial production failures according to Embodiment 1 of the present invention. The method includes the following steps:

[0050] Obtain historical production data from the hydrocracking reactor. During implementation, this historical production data includes tag information data and mechanistic data. Tag information data includes predictive index tag numbers and operational control tag numbers. Mechanistic data includes: inlet temperature of each reactor bed, outlet temperature of each reactor bed, temperature rise of each reactor bed, total reactor temperature rise, reactor pressure, heat exchanger inlet temperature, heat exchanger outlet temperature, feed rate, fresh hydrogen rate, and recycle hydrogen rate. Mechanistic data is simulated data generated from mechanistic modeling.

[0051] Then, the historical production data is processed, including the removal of DCS outlier data, filling in null values, and standardization.

[0052] Historical production data is modeled using a probabilistic graphical model to obtain the failure probability propagation probability network for each predictor and control variable. By optimizing the network path, the root cause variable of the actual predictor variable failure is calculated. The predictor variable includes the target predictor number.

[0053] Input the tag number value of the current fault into the probabilistic graphical model to obtain the fault probability of each control variable of the current fault, and obtain the fault solution by searching the production process knowledge graph.

[0054] Through the above steps, the industrial production failure root cause diagnosis method proposed in this embodiment of the invention employs a production process knowledge graph and establishes a probabilistic graphical model to trace, reason about, and diagnose the root causes of production failures. By establishing a causal relationship model between production goals (predictive variables) and operational indicators (control variables), quantitative root cause results are provided to employees. This enables operators to trace and reason about industrial production failures more quickly and accurately, significantly reducing the safety risks for enterprises.

[0055] During implementation, after obtaining the failure probabilities of each control variable for the current fault, a fault solution is obtained by searching the production process knowledge graph. The fault tag values ​​and failure probabilities of the control variables are then verified, and the verification data is updated to historical production data periodically. The period can be one week, one month, etc. When the model is applied online, the production data of the current period is updated and iterated based on the original model after each production cycle, making the model better.

[0056] In the above steps, the data for each tag is collected by the DCS system or a real-time database according to the process requirements. There are two types of fault data: one is fault data marked by the DCS (data not collected by the DCS), which requires processing during the preprocessing stage; the other is actual fault data, which will be used as training data input into the model. Additionally, missing values ​​in the data are filled using the mean.

[0057] Bayesian networks (BNs) are probabilistic networks based on graph theory and probability theory. They qualitatively represent the causal relationships between variables within a system using directed acyclic graphs (DAGs), and precisely represent the strength of these relationships using probability values. This effectively addresses the challenges of expressing fault knowledge and the inaccuracy of inference results in complex systems. Learning a BN involves two parts: BN structure learning and BN parameter learning. BN structure learning algorithms fall into two main categories: constraint-based learning algorithms and scoring function-based learning algorithms. Constraint-based network learning methods focus on each pair of nodes as the research object, searching for the directed partition set between these two nodes among all remaining nodes.

[0058] In this embodiment, the BN structure learning adopts a fusion method. First, a constraint-based method is used to constrain the conditional independence of the learning to reduce the parsing difficulty of the search space or to obtain an undirected network graph. Then, the scoring function in the scoring search algorithm is combined with heuristic search to find the optimal BN structure.

[0059] See Figure 2 In this embodiment, a probabilistic graphical model is used to model historical production data. This involves using a Bayesian network to fit and model the fault propagation probability of each predictor and control variable. The steps include:

[0060] Based on production requirements, determine the relevant predictive variables and the names of control variables that may cause anomalies in the target designation, and obtain a training dataset. To eliminate heterogeneity among different variables, the training dataset is standardized to obtain a new sample dataset;

[0061] Calculate the mutual information coefficient (MIC) between each pair of variables in the training dataset to obtain the degree of association between the variables. Normalize the maximum mutual information by dividing it by log(min(X,Y)). Select the maximum mutual information value as the MIC value. The formula for calculating mutual information is as follows:

[0062]

[0063] Where I(X,Y) is the mutual information coefficient between variables X and Y; P(X,Y) is the joint distribution function of variables X and Y; and P(X) and P(Y) are the independent probability distribution functions of variables X and Y, respectively.

[0064] Based on the calculated maximum mutual information coefficient, each predictor variable and control variable is used as a node, and the edges between the two variables with the maximum mutual information coefficient are retained to construct a maximum spanning tree model.

[0065] The number of edges connected to a node in the maximum spanning tree model is regarded as the number of the node’s maximum parent nodes, and the set of the maximum parent nodes of a node is determined based on the maximum spanning tree model.

[0066] Nodes are ranked by importance. In practice, this can be done by performing a hierarchical analysis (AHP) on the original sample data to obtain the node importance ranking. The AHP is a decision analysis method that combines qualitative and quantitative approaches to solve complex multi-objective problems. This embodiment only uses process experience to subjectively rank the importance of variables qualitatively.

[0067] The new sample dataset, the importance ranking of nodes, and the maximum number of parent nodes for each node are used as inputs to learn the parameters of the Bayesian network, and the K2 algorithm is used to solve for the optimal Bayesian network structure.

[0068] In practice, after using the K2 algorithm to find the optimal Bayesian network structure, the Bayesian network structure is selectively optimized and adjusted based on expert experience.

[0069] In this embodiment, the K2 algorithm steps are as follows:

[0070]

[0071] Bayesian network parameter learning is the process of learning from sample data based on a determined Bayesian network structure. The process of creating a conditional probability table for the network can be divided into two types, depending on the sufficiency of the sample data: Maximum Likelihood Estimation (MLE) and Maximum A posteriori Estimation (MAP).

[0072] In this embodiment, the maximum posterior probability method is used in Bayesian network parameter learning, and the steps are as follows:

[0073] The prior probability of the failure phenomenon is obtained by statistical calculation of historical production data;

[0074] For the prior probability of the observed node, assume that the observed node p follows a Beta distribution with parameters α and β: Beta(p:α,β);

[0075] Calculate the posterior distribution function p:

[0076]

[0077] get:

[0078]

[0079] Where, p i q represents the number of times the observed node P in the dataset fails. i The conditional probability table for each node is obtained by representing the number of times no fault occurred, and a Bayesian network is constructed. Using the constructed BN network, the probability of a fault occurring under different parameters is calculated, and the value with the highest probability is the root cause of the fault.

[0080] Since the root cause of a failure needs to be reasoned and queried in a knowledge graph to obtain the corresponding solution, the knowledge graph design needs to meet a certain level of industry completeness. Therefore, the production process knowledge graph in this embodiment includes six types of nodes: production materials, production equipment, failure conditions, control variables, target variables, and solutions. The production process knowledge graph also includes multiple triple representations: production equipment - occurrence - failure condition; failure condition - solution - solution; control variable - causing - failure condition; production equipment - query - target variable; production equipment - processing - production materials; and control variable - failure - solution. When the control variable is abnormal, the solution to the corresponding root cause is obtained from the knowledge graph according to the query path "failure condition" - caused - "control variable" - failure - "solution", such as... Figure 3 As shown, it can achieve a high accuracy rate in fault diagnosis. In implementation, the trained fault diagnosis model can be deployed on the information system and combined with the production knowledge graph to realize online root cause diagnosis of production faults.

[0081] Based on the same inventive concept, embodiments of the present invention also provide a fault root cause diagnosis system for industrial production, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of any of the above embodiments.

[0082] Existing research on root cause analysis of production failures mainly focuses on optimizing empirical rules in production. However, this merely refines current empirical rules and cannot uncover unknown, potential knowledge and experience. Process engineers also need significant business experience to ensure a certain level of reliability, which poses a challenge to new employees. Furthermore, companies need to invest resources in employee training, incurring additional costs. This invention, from the perspective of probabilistic graphical models, quantitatively models the causal relationship between target and control tags in production. It provides a feasible method for root cause tracing and reasoning from a data perspective, offering assistance to personnel and significantly saving manpower and time costs. This can effectively reduce production risks and minimize corporate losses.

[0083] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0084] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0085] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," or "on top" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," or "beneath" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0086] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0087] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for diagnosing the root causes of failures in industrial production, characterized in that, Includes the following steps: Historical production data of the hydrocracking reactor is obtained; the historical production data includes: tag information data and mechanism data; the mechanism data is simulation data generated by mechanism modeling; the tag information data is collected by the DCS system or real-time database, and the tag information data includes: predictive index tag numbers and operation control tag numbers; Historical production data is modeled using a probabilistic graphical model to obtain the failure probability propagation probability network for each predictor and control variable. By optimizing the network path, the root cause variable of the actual predictor variable failure is calculated; the predictor variable includes the target predictor number. The process of modeling historical production data using a probabilistic graphical model includes the following steps: Based on production needs, determine the relevant predictive variables and the names of control variables that may cause anomalies in the target tag number, and obtain the training dataset; standardize the training dataset to obtain a new sample dataset; Calculate the mutual information coefficient between each pair of variables in the training dataset to obtain the degree of correlation between variables; based on the calculated maximum mutual information coefficient, take each predictor variable and control variable as nodes, retain the edges between the two variables with the maximum mutual information coefficient value, and construct the maximum spanning tree model; The number of edges connected to a node in the maximum spanning tree model is regarded as the number of the node’s maximum parent nodes, and the set of the maximum parent nodes of a node is determined based on the maximum spanning tree model. Based on experience with hydrocracking production processes, nodes are ranked by importance. The new sample dataset, the importance ranking of nodes, and the maximum number of parent nodes for each node are used as inputs to learn the parameters of the Bayesian network, and the K2 algorithm is used to solve for the optimal Bayesian network structure. Input the faulty tag number into the probability graph model to obtain the fault probability of each control variable of the current fault, and obtain the fault solution by searching the production process knowledge graph. The production process knowledge graph includes six types of nodes: production materials, production equipment, fault conditions, control variables, target variables, and solutions. The production process knowledge graph includes multiple triples: production equipment - occurrence - failure condition; failure condition - solution - solution; control variable - cause - failure condition; production equipment - query - target variable; production equipment - processing - production materials; and control variable - failure - solution.

2. The method for diagnosing root causes of industrial production failures as described in claim 1, characterized in that, After obtaining the failure probability of each control variable for the current fault, a fault solution is obtained by searching the production process knowledge graph. The fault tag value and the failure probability of the control variable are verified, and the verification data is updated to the historical production data periodically.

3. The method for diagnosing root causes of failures in industrial production as described in any one of claims 1 to 2, characterized in that, The mechanistic data include: inlet temperature of each bed in the reactor, outlet temperature of each bed in the reactor, temperature rise of each bed in the reactor, total temperature rise of the reactor, reactor pressure, inlet temperature of the heat exchanger, outlet temperature of the heat exchanger, as well as feed rate, fresh hydrogen rate and circulating hydrogen rate.

4. The method for diagnosing root causes of industrial production failures as described in claim 3, characterized in that, After finding the optimal Bayesian network structure using the K2 algorithm, the Bayesian network structure is selectively optimized and adjusted based on expert experience.

5. The method for diagnosing root causes of industrial production failures as described in claim 3, characterized in that, The formula for calculating the mutual information is as follows: ； Where I(X,Y) is the mutual information coefficient between variables X and Y; Let X be the joint distribution function of variables X and Y. and Let X and Y be the independent probability distribution functions of variables X and Y, respectively.

6. The method for diagnosing root causes of industrial production failures as described in claim 3, characterized in that, In the Bayesian network parameter learning, the maximum a posteriori probability method is used, and the steps are as follows: The prior probability of the failure phenomenon is obtained by statistical calculation of historical production data; For the prior probability of the observed node, assume that the observed node p satisfies the parameter as follows: Beta distribution: ; Calculate the posterior distribution function : ； get: in, This represents the number of times observation node P in the dataset has failed. The conditional probability table for each node is obtained by representing the number of times no fault occurred, and a Bayesian network is constructed.

7. A root cause diagnosis system for industrial production, 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 computer program, it implements the steps of the method described in any one of claims 1 to 6.

Images