A method for anomaly root cause localization using dynamic Bayesian networks incorporating causal memory matrices

By integrating causal memory matrices with dynamic Bayesian networks and utilizing material flow constraints and online update mechanisms, the problems of computational complexity and changing operating conditions in dynamic Bayesian networks under high-dimensional variables are solved, enabling accurate positioning and sorting in industrial processes.

CN123116049APending Publication Date: 2026-11-17NANJING RICHISLAND INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611174200.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-04
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for identifying the root causes of anomalies based on dynamic Bayesian networks are computationally complex and prone to getting trapped in local optima under high-dimensional variables. They lack process knowledge constraints, are difficult to adapt to changes in operating conditions, and lack effective means for root cause quantification and ranking.

Method used

By integrating causal memory matrices and dynamic Bayesian networks, and through DBN structure learning under material flow constraints, a structure and strength memory matrix under normal operating conditions is constructed. An exponentially weighted moving average algorithm is used for online updating, and the structure and strength offsets under abnormal conditions are calculated. Root cause sorting is performed by combining process level weights.

Benefits of technology

It enables accurate and reliable root cause localization in complex industrial processes, eliminates physically impossible dependencies, adapts to operating condition drift, provides a clear root cause investigation sequence, and improves the accuracy and interpretability of root cause localization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116049A_ABST
    Figure CN123116049A_ABST
Patent Text Reader

Abstract

This invention discloses a method for locating the root causes of anomalies using a dynamic Bayesian network that integrates causal memory matrices. The method constrains the direction of dynamic Bayesian network structure learning based on material flow direction; it divides time-series data into blocks using a sliding window, constructing a first-order vector autoregressive model with L1 regularization within each window to learn the causal structure between variables; it constructs a structural memory matrix and an intensity memory matrix under normal operating conditions as benchmarks, and updates them online using an exponentially weighted moving average algorithm; when an anomaly occurs, it calculates the offset between the current causal structure and the memory matrix in both structural and intensity dimensions, generating a comprehensive offset matrix, and then calculates the root cause score for each variable, achieving quantitative ranking and location of root cause variables. This invention, by integrating process knowledge constraints and memory matrix offset analysis, effectively eliminates spurious associations, providing accurate and interpretable location results for diagnosing the root causes of anomalies in industrial processes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of industrial process anomaly tracing and fault diagnosis technology, specifically a method for locating the root cause of anomalies by integrating causal memory matrices and dynamic Bayesian networks and using causal structure shift analysis. Background Technology

[0002] Industrial process variables are intricately coupled through material and energy flows. When an anomaly occurs in a localized process, the disturbance propagates downstream along the inter-variable relationships, causing a series of variables to deviate from their normal operating states. Because the influence between variables is both directional and temporal, relying solely on single-variable threshold alarms is insufficient to accurately identify the propagation path and trace the source from large-scale fluctuations. Therefore, a modeling method capable of characterizing the temporal dependencies between variables is needed to support the localization of anomaly root causes.

[0003] Dynamic Bayesian Networks (DBNs), as an extension of Bayesian Networks in the time-series domain, can characterize the temporal dependencies between variables by unfolding the network structure along the time axis. They have gained widespread attention in time-series data modeling and industrial process anomaly diagnosis. Compared to static models, dynamic Bayesian Networks can capture the lag and direction of influence between variables, providing a more suitable theoretical framework for tracing the propagation path of anomalies in processes.

[0004] However, existing root cause localization methods based on dynamic Bayesian networks (DBNs) still have the following shortcomings. First, the search space for DBN structure learning grows exponentially with the dimensionality of variables. Purely data-driven methods are computationally complex and prone to getting trapped in local optima under high-dimensional variables. Second, statistical correlations are easily misjudged as true associations, leading to learned dependencies that contradict the physical constraints of material flow. Existing methods lack effective means to integrate such process knowledge into the structure learning process to eliminate physically impossible dependencies. Furthermore, industrial processes are affected by factors such as equipment wear and environmental changes, resulting in slow drift in normal operating conditions. Existing methods typically use fixed benchmarks for anomaly detection, making it difficult to adapt to dynamic changes in operating conditions. At the root cause localization level, existing methods mostly rely on contribution maps or single statistical indicators, lacking effective means to quantify and rank root causes from the perspective of changes in dependency patterns between variables.

[0005] To address the aforementioned problems, this invention proposes a dynamic Bayesian network anomaly root cause localization method that integrates process knowledge constraints and causal memory matrices. This method first eliminates physically impossible dependencies by constraining the direction of DBN structure learning based on material flow direction; then, it constructs a memory matrix to encode the structure and strength under normal operating conditions as a benchmark and updates it online; finally, it achieves the quantitative ranking of anomaly root cause variables by calculating the offset of structure and strength relative to the benchmark during anomalies. Summary of the Invention

[0006] To address the aforementioned problems, this invention discloses a dynamic Bayesian network anomaly root cause localization method that integrates causal memory matrices to perform root cause localization analysis on industrial process anomalies.

[0007] The present invention comprises the following steps:

[0008] A dynamic Bayesian network anomaly root cause localization model integrating causal memory matrices was established, and this model was used to locate the root causes of anomalies in industrial processes.

[0009] The establishment of a causal memory dynamic Bayesian network model and the analysis of abnormal root causes are as follows:

[0010] (1) Selecting those related to industrial process anomaly analysis process variables Collect time-series data of each variable under normal operating conditions;

[0011] (2) Based on the material flow direction as described above Process variables are used to classify process levels. Based on the irreversibility of material flow, the causal direction is constrained:

[0012] ① Time Prohibition ;

[0013] ② Autoregression with the same variable is allowed;

[0014] ③ If a closed-loop control loop exists between variables, the corresponding feedback edge is allowed;

[0015] (3) Set the width of the sliding window, and process the time-series running data in blocks according to the time sequence, with the data in each window as a basic analysis unit;

[0016] (4) Under the constraints of the process level, based on the time series data within the sliding window, the transfer network of the dynamic Bayesian network is represented as a first-order vector autoregressive form:

[0017]

[0018] in, It is a variable exist The value at time, It is a variable exist The value at time, For the noise term, the regression coefficients are solved using least squares estimation with L1 regularization:

[0019]

[0020] in, It is the number of samples within the time window. The observation vector of the target variable within the time window. For a first-order lag observation matrix, For to be estimated Causality strength coefficient It is a variable The column vector consisting of all the corresponding regression coefficients The regularization coefficient is used.

[0021] (5) Solve and then let ,express Causal strength; setting structural stability threshold ,definition express Does a directed edge exist? If ,but ,otherwise ;

[0022] (6) Based on the causal graph model learned by multiple sliding windows under normal operating conditions, construct the structural memory matrix under normal operating conditions. and intensity memory matrix It is updated online using an exponentially weighted moving average algorithm:

[0023]

[0024]

[0025] For the One window, and These are the updated memory values, and This is the memory value from the previous window. and The instantaneous value learned by the current window. For memory retention coefficient, It is a logical OR operator. If either the previous memory value or the current window estimate is 1 (i.e., the edge exists), the updated memory value is 1; otherwise, it is 0.

[0026] (7) When an anomaly is detected, steps (4) and (5) are re-executed with the data of the current sliding window to obtain the current causal graph model. The offsets between the current causal graph model and the normal baseline in the structural and intensity dimensions are calculated:

[0027]

[0028]

[0029]

[0030] in, These are directed edges in the event of an anomaly. It is the causal strength during anomalies. To prevent small constants from being divided by zero, Finally, a comprehensive offset matrix is ​​generated. ;

[0031] (8) Based on the aforementioned composite offset matrix Calculate the total out-degree offset for each variable node. Total in-degree offset :

[0032]

[0033] in, For prior weights based on process level, As the in-degree penalty factor, all variables are calculated according to... Sort them from highest to lowest, and the one with the highest score is determined as the root cause of the problem.

[0034] Beneficial effects:

[0035] This invention discloses a dynamic Bayesian network anomaly root cause localization method that integrates causal memory matrices. Addressing the problems of traditional DBN structure learning, which easily introduces dependencies that contradict physical reality and lacks a normal operating condition benchmark, this method deeply integrates process mechanism knowledge with dynamic Bayesian network structure learning. By using material flow constraints to eliminate physically impossible dependencies, it avoids false associations introduced by statistical correlation misjudgments in data-driven methods. A structure memory matrix and a strength memory matrix are constructed to encode the graph pattern under normal operating conditions as a benchmark, and an exponentially weighted moving average algorithm is used for online updates, effectively tracking slow drift in operating conditions and avoiding false alarms caused by changes in operating conditions over long-term operation of a fixed benchmark. By calculating the two-dimensional offset of structure and strength relative to the benchmark during anomalies, and combining the total out-degree offset, total in-degree offset, and prior weights of the process level to calculate the root cause score, it achieves quantitative ranking and accurate localization of root cause variables, overcoming the limitations of traditional methods that rely on single statistical indicators and are difficult to rank root causes from the perspective of dependency pattern changes. This invention can provide accurate, reliable, and interpretable localization results for the diagnosis of anomalies in complex industrial processes. Attached Figure Description

[0036] Figure 1 This is a flowchart of a dynamic Bayesian network anomaly root cause localization method that integrates causal memory matrices.

[0037] Figure 2This is a schematic diagram of the comprehensive offset matrix during an abnormal period in the ironmaking process of a blast furnace in an embodiment.

[0038] Figure 3 This is the result of the root cause score ranking of the variables for abnormal sulfur content in a blast furnace ironmaking process in the example.

[0039] Specific implementation process

[0040] The following, with reference to the accompanying drawings and specific examples, illustrates the implementation effect of this method in the steel blast furnace production process through a concrete operational procedure. This embodiment is implemented based on the technical solution of this invention, but the scope of protection of this invention is not limited to the following embodiments.

[0041] This invention uses the blast furnace ironmaking process of a steel enterprise as an application case. This blast furnace plays a crucial role in reducing iron ore to qualified liquid pig iron, and its smelting process involves complex physicochemical reactions and multivariate dynamic coupling. Affected by factors such as operational adjustments and changes in equipment status, the blast furnace frequently experiences persistently excessive sulfur content in pig iron during actual production, which not only seriously affects product quality but also brings a chain of adverse effects to subsequent steelmaking processes. To verify the effectiveness of the method proposed in this invention, an abnormal event of excessive sulfur content in pig iron occurring during a certain period of the blast furnace was selected as the analysis object, and the specific application of this invention in the diagnosis of the root cause of the abnormality was described in detail. This study uses 30 days of continuous production data from March to April 2023 for this blast furnace. Data from the normal operating period was used to construct a causal memory matrix, and data from the period of excessive sulfur content was used to verify the effectiveness of anomaly detection and root cause localization.

[0042] The implementation process is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0043] (1) Based on the blast furnace structure, the location of the sensors was analyzed, and a total of 19 relevant control and measurement variables were selected. Among them, the output variables include: pig iron sulfur content, pig iron silicon content, pig iron manganese content, pig iron phosphorus content, and pig iron titanium content; the observation and control variables include: blower air pressure, blower air volume, hot blast stove hot blast temperature, hot blast stove hot blast pressure, furnace top pressure, furnace top temperature, furnace throat temperature, furnace body temperature, furnace bottom temperature, total pressure difference, oxygen enrichment, pulverized coal injection rate, operating fuel ratio, and pulverized coal injection ratio. The abbreviations and names of the variables are shown in Table 1;

[0044] Table 1 Variable Names and Abbreviations

[0045]

[0046] (2) Based on the process mechanism and material flow direction, the variables of the blast furnace system are divided into process levels. AVB, FAP, THF, PHF, OG, CI, OFR, and CIR are classified as feed layer variables; FToP, FToT, FThT, FT, FBT, and FDP are classified as reactor layer variables; and C(Si), C(Mn), C(S), C(P), and C(Ti) are classified as downstream variable layers. Based on the principle of irreversibility of material flow, causal direction constraints are defined: downstream variables cannot be the cause of changes in upstream variables, and causal edges between variables at the same level and autoregression of the same variable are allowed.

[0047] (3) The sliding window technique is used to divide the time series running data into blocks. The sliding window width and sliding step size are set to divide the continuous production data into several continuous analysis windows. The data in each window is used as a basic analysis unit for dynamic Bayesian network modeling.

[0048] (4) Within each sliding window, perform smoothing filtering on each variable to eliminate the influence of sampling error;

[0049] (5) Under the constraints of the aforementioned process level, a first-order vector autoregressive model with L1 regularization is constructed. With the sulfur content C(S) of pig iron as the target variable, the model takes the following form:

[0050]

[0051] in For to be estimated The causal strength coefficient;

[0052] (6) Estimate the regression coefficient vector corresponding to the objective variable C(S) by solving the least squares optimization problem with L1 penalty. :

[0053]

[0054] in The observation vector of the target variable within the window. For a first-order lag observation matrix, Let be the regularization coefficient. After solving, let . Set structural stability threshold Define directed edges express Does a directed edge exist? If ,but ,otherwise ;

[0055] (6) Construct a structural memory matrix based on the causal graph model learned by multiple sliding windows under normal operating conditions. and intensity memory matrix And it is updated online using an exponentially weighted moving average algorithm. For the first... One window, and These are the updated memory values, and This is the memory value from the previous window. and The update rule for the instantaneous value learned by the current window is as follows:

[0056]

[0057]

[0058] in The memory retention factor;

[0059] (7) When an abnormality in the sulfur content C(S) of pig iron is detected, steps (4) to (5) are re-executed using the data of the current sliding window to obtain the current causal graph model. The offsets between the current causal graph model and the normal baseline in the structural and strength dimensions are calculated:

[0060]

[0061]

[0062]

[0063] in, These are directed edges in the event of an anomaly. It is the causal strength during anomalies. To prevent small constants from being divided by zero, Generate the composite offset matrix . Figure 2 A schematic diagram of the comprehensive offset matrix during abnormal periods is given, where the gray level of each grid point represents the magnitude of the comprehensive offset corresponding to that directed edge. From Figure 2 It can be seen that the directed edge shift of the source variables representing furnace bottom temperature and furnace body temperature toward pig iron sulfur content is significantly higher than that of other variable combinations. The causal relationship between the two and sulfur content deviates sharply from the normal operating conditions, which is the key abnormal transmission path that induced the sulfur content to exceed the standard this time.

[0064] (8) Based on the comprehensive offset matrix Calculate the total out-degree offset for each variable node. Total in-degree offset Combining process level prior weights in-degree penalty factor Calculate the root cause scores for each variable:

[0065]

[0066] All variables Sort from highest to lowest. Figure 3 The ranking of root cause scores for each variable is presented. From Figure 3 It can be seen that the furnace bottom temperature (FBT) was identified as the most important root cause variable with the highest score, while the furnace body temperature (FT) ranked second. The blast system variables also entered the top ten, indicating that changes in operating parameters such as air pressure and air volume indirectly affect sulfur content by influencing the thermal state inside the furnace.

[0067] To verify the effectiveness of the method of the present invention, the method of the present invention was compared with the static Bayesian network and the contribution graph method, and the results are shown in Table 2.

[0068] Table 2 Comparison of root cause localization results using different methods

[0069]

[0070] As shown in Table 2, the method of this invention identifies the true root cause, FBT, in the top-1. The static Bayesian network ranks FToP first and FBT second in the top-3. Although it includes the true root cause in the top-3, it fails to accurately identify its root cause status. This is because the static Bayesian network lacks the ability to track changes in the dependency patterns between variables over time, relying solely on the static graph structure for root cause inference, which easily misclassifies highly correlated non-root cause variables as root causes. The contribution graph method is limited by the magnitude of the variables and is more sensitive to variables with large changes, thus ranking FT first, masking the contribution of the true root cause variable. The method of this invention, by constructing a memory matrix under normal operating conditions as a benchmark and calculating the two-dimensional shift of structure and intensity during anomalies, can effectively filter out background fluctuations of non-root cause variables, thereby improving the accuracy of root cause localization.

[0071] The above results demonstrate that by constructing a normal operating condition memory matrix as a benchmark, this invention can effectively capture substantial changes in the dependency patterns between variables when anomalies occur. In practical applications, this method provides on-site operators with a clear order for root cause investigation. Operators can directly prioritize checking the top-ranked variables and corresponding measurement points based on the root cause score ranking results, reducing the time cost of blind investigation and providing a basis for decision-making for rapid production recovery.

Claims

1. A method for locating the root causes of anomalies in a dynamic Bayesian network that integrates a causal memory matrix, characterized in that... To address anomalies in industrial processes, a dynamic Bayesian network is constructed to describe the temporal dependencies between variables, and a causal memory matrix is ​​established as a baseline for normal operating conditions. The root cause of the anomaly is located through offset analysis between the current causal structure and the memory matrix, including the following steps: (1) Selecting those related to industrial process anomaly analysis process variables Collect time-series data of each variable under normal operating conditions; (2) Based on the material flow direction and energy transfer logic, for the... The process variables are used to divide the process hierarchy and constrain the causal direction of dynamic Bayesian network learning. (3) Set the sliding window width and process the time-series running data in blocks according to time order. The data in each window is a basic analysis unit for dynamic Bayesian network modeling. (4) Based on the time series data within the sliding window, learn the dynamic Bayesian network causal graph model under that window. The causal graph model contains a set of variable nodes. Directed edge set and causal strength matrix ; (5) The exponentially weighted moving average algorithm is used to analyze the set of directed edges in the causal graph model. and causal strength matrix Perform online updates to obtain the structural memory matrix under normal operating conditions. and intensity memory matrix ; (6) When an anomaly is detected, step (4) is re-executed with the data of the current sliding window to obtain the current causal graph model, and the offset matrix between the current causal graph model and the structure memory matrix and the strength memory matrix in the structure dimension and strength dimension is calculated. , The two offset matrices are weighted and fused to generate a comprehensive offset matrix. ; (7) Based on the aforementioned composite offset matrix Calculate each variable node Total out-degree offset Total in-degree offset Used to calculate the root cause score for each variable node. The variable with the highest score is identified as the root cause of the failure.

2. The method for locating the root cause of anomalies in a dynamic Bayesian network by fusing a causal memory matrix as described in claim 1, characterized in that, The process hierarchy division and causal constraint described in step (2) specifically include: (2-1) Based on the material flow direction and energy transfer logic, for the... Process variables are used to classify process levels. ; (2-2) Define constraints based on the principle of irreversibility of material flow: ① Time Prohibition That is, prohibited variables Pointer variable The directed edges; ② Autoregression with the same variable is allowed; ③ If a closed-loop control loop exists between variables, then the corresponding feedback edge is allowed.

3. The method for locating the root cause of anomalies in a dynamic Bayesian network by fusing a causal memory matrix as described in claim 1, characterized in that... The learning of the dynamic Bayesian network causal graph model under this window in step (4) specifically includes: (4-1) The transition network of the dynamic Bayesian network is represented as a first-order vector autoregressive form: in, For to be estimated Causality strength coefficient It is a variable exist The value at time, It is a variable exist The value at time, Noise term; (4-2) Based on the aforementioned process level constraints, for the prohibited directions Force zeroing; (4-3) Solve for the regression coefficients using least squares estimation with L1 regularization: in, It is the number of samples within the time window. The observation vector of the target variable within the time window. For a first-order lag observation matrix, It is a variable The column vector consisting of all the corresponding regression coefficients is the regularization coefficient.

4. The method for locating the root cause of anomalies in a dynamic Bayesian network by fusing a causal memory matrix as described in claim 3, characterized in that... Solve and let Causal strength Set structural stability threshold Define directed edges like ,otherwise .

5. The method for locating the root cause of anomalies in a dynamic Bayesian network by fusing a causal memory matrix as described in claim 1, characterized in that... The online update mentioned in step (5) specifically includes: For the Learned from a window and Online recursive update rules: For the One window, and These are the updated memory values, and This is the memory value from the previous window. and The instantaneous value learned by the current window. For memory retention coefficient, It is a logical OR operator. If either the previous memory value or the current window estimate is 1, the updated memory value is 1; otherwise, it is 0.

6. The method for locating the root cause of anomalies in a dynamic Bayesian network by fusing a causal memory matrix as described in claim 5, characterized in that... The calculation of the offset matrix in step (6) specifically includes: in, These are directed edges in the event of an anomaly. It is the causal strength during anomalies. Represents a directed edge The structural existence benchmark value under normal operating conditions This represents the baseline value of causal strength after EWMA smoothing under normal operating conditions for the corresponding edge. To prevent small constants from being divided by zero, .

7. The method for locating the root cause of anomalies in a dynamic Bayesian network by fusing a causal memory matrix as described in claim 1, characterized in that... The calculation of root cause scores in step (7) specifically includes: in, For nodes The degree of exit, For nodes in-degree, For prior weights based on process level, As the in-degree penalty factor; assign all variables to... Sort them from highest to lowest, and the one with the highest score is determined as the root cause of the problem.

8. A method for locating the root cause of anomalies in a dynamic Bayesian network by fusing a causal memory matrix, as described in any one of claims 1 to 7, characterized in that... Under normal operating conditions, the sliding window moves along the time axis with a fixed step size, and steps (4) to (7) are executed independently in each window to output the trajectory of the root cause score of each variable as it evolves over time.