Control loop dynamic reconfiguration method and system for anomaly event stream causal inference

Abstract

The application relates to the technical field of industrial automation and safety control, and particularly discloses a control loop dynamic reconstruction method and system for abnormal event stream causal inference. The application collects multi-source abnormal events in real time, identifies a driving propagation path through lightweight causal discovery, and calculates comprehensive causal strength according to path length, confidence and time coupling. A four-layer nested judgment mechanism is adopted to ensure that reconstruction is triggered only for a continuously high-threat path. The system combines reconstruction strategies based on an atomic action library, realizes smooth migration of new and old control structures through an S-shaped function after pre-performance verification by digital twinning, and continuously monitors and supports abnormal freezing and rollback in the process. The application realizes a closed-loop autonomy from abnormal perception, causal diagnosis to safety reconstruction, and significantly improves system resilience.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation and safety control, and in particular to a control loop dynamic reconstruction method and system for abnormal event stream causal inference. BACKGROUND

[0002] As the core of key infrastructure, the stable and safe operation of industrial control system is of great importance. During the operation, the system will continuously generate a large number of heterogeneous abnormal event streams due to factors such as equipment aging, complex working conditions, network attacks or human errors. The traditional mainstream abnormal processing method mainly relies on alarm systems based on fixed thresholds and diagnosis rules based on expert experience. The traditional method has significant defects: first, its essence is a passive response mode from perception to alarm, and the diagnosis and decision-making lag behind the occurrence of abnormalities, which is difficult to meet the real-time requirements of high dynamic industrial scenarios; second, the rule system is rigid and inflexible, and when facing unknown, coupled or evolving abnormal patterns, the diagnosis accuracy will drop sharply, and false positives and false negatives will occur frequently; more importantly, even if the abnormality is identified, the existing technology generally lacks the ability to automatically and safely convert the diagnosis conclusion into control system structure adjustment or parameter optimization, and it is heavily dependent on manual intervention, which is inefficient and inconsistent.

[0003] In recent years, in order to improve the fault root location capability, data-driven causal inference technology has been introduced into the industrial field. However, the existing solutions based on causal graphs and Granger causality mainly focus on offline and post-mortem analysis of historical data, which cannot perform lightweight and fast online reasoning on real-time event streams, and thus cannot provide immediate decision support for dynamic control. More importantly, even if the potential causal path can be identified online, how to evaluate the threat of the causal path and design a safe, reliable and smooth control loop reconstruction strategy remains an unsolved problem. Blind or aggressive online reconstruction is likely to cause system oscillation, secondary failure or even safety accidents due to model mismatch, constraint overrun or switching impact. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a control loop dynamic reconstruction method and system for abnormal event stream causal inference. The method can process multi-source abnormal event streams in real time, identify driving causal paths online, and realize safe, dynamic and adaptive reconstruction of industrial control loops through multi-layer safety judgment, virtual pre-play, constraint migration and adaptive learning mechanism, thereby improving the resilience and autonomous recovery capability of the system.

[0005] The technical solution of the present application is as follows:

[0006] According to an aspect of the present application, a control loop dynamic reconstruction method for abnormal event stream causal inference is provided, comprising:

[0007] Real-time acquisition of multi-source abnormal event stream generated in the running process of an industrial control system, online lightweight causal inference on the event sequence of the multi-source abnormal event stream, and identification of driving causal paths with controller input or output variables as endpoints;

[0008] For the identified driving causal paths, a four-layer nested judgment mechanism is started, and only when the four-layer judgment is established, the reconstruction preparation state is activated;

[0009] Among them, the first layer judgment mechanism is to screen the key causal paths associated with the endpoint controlled variables;

[0010] The second layer judgment mechanism is to calculate the comprehensive causal strength of each path;

[0011] The third layer judgment mechanism is to judge whether there is a path strength exceeding a preset safety threshold;

[0012] The fourth layer judgment mechanism is to verify whether the path exceeding the threshold value appears twice or more in a continuous number of cycles;

[0013] In the reconstruction preparation state, according to the structure, variable and time sequence characteristics of the current driving causal path, a plurality of basic reconstruction actions are matched and combined from the pre-set atomic action library to form a reconstruction strategy package; pre-rehearsal verification is carried out in a virtual environment, and only when the verification is successful, execution is approved;

[0014] Safety boundary constraints are set for each action in the reconstruction strategy package, and the strategy package is deployed to the real control loop under the constraints; the new and old control structures are migrated in multiple cycles using an S-shaped function; if a new high-intensity (exceeding the safety threshold) driving causal path is detected during the migration process, the transition migration is immediately frozen, and the four-layer nested judgment mechanism is restarted for reevaluation;

[0015] After the reconstruction is deployed, the abnormal event stream is monitored within a predetermined time scale; if the abnormal event stream disappears, the verified reconstruction strategy package is marked as a steady state, and after being labeled by a scene, it is included in the atomic action library; if the abnormality still exists or a new abnormal pattern appears, the strategy rollback is triggered, and the causal inference and evaluation cycle is restarted to form an adaptive reconstruction mechanism.

[0016] As a further option of the method of the present application, the online lightweight causal inference on the event sequence of the multi-source abnormal event stream comprises:

[0017] Data acquisition agents are deployed at key positions of the system to collect controller process variables, set values, control outputs, network messages and system log data;

[0018] Based on preset rules or models, the collected data is anomaly detected, and a uniformly formatted anomaly event triplet is generated; the anomaly event triplet includes a timestamp, variable identifier, and confidence level;

[0019] The abnormal event stream is pushed to the central processing engine. Within the sliding time window, the continuous time axis is discretized into equally spaced time slots, and the confidence scores of multiple events for each variable within the time slots are aggregated to form a multivariate time series.

[0020] The multivariate time series was analyzed using a causal discovery algorithm to generate a directed causal graph, and all candidate sets of driving causal paths ending with key control variables were extracted from it.

[0021] As a further option of the method of the present invention, the first-layer judgment mechanism specifically includes:

[0022] Define and maintain a list of key controlled variables based on process safety information, key product quality parameters, and equipment protection interlock lists;

[0023] For each path in the set of driving causal paths, obtain the endpoint variable;

[0024] If the endpoint variable is a controller output variable, then query the control loop configuration library to determine whether the final execution object manipulated by the controller output variable is used to call a variable in the list of key controlled variables;

[0025] If the endpoint variable is a process variable, trace the control loop to which the endpoint variable belongs and determine whether the controller output of the control loop is ultimately used to adjust the variables in the list of key controlled variables.

[0026] The first-level judgment mechanism filters the paths into a subset of key causal paths.

[0027] As a further option of the method of the present invention, in the second-layer judgment mechanism, the comprehensive causal strength of each path is calculated. The calculation formula is: ;in: It is the path length factor. , It is the path length; It is the average confidence factor. , It is the initial average confidence level of the path events. It is the time coupling tightness factor. , It is a path The maximum time interval between adjacent events, The attenuation coefficient is... These are the weight coefficients of the corresponding factors, and satisfy the following conditions: .

[0028] As a further choice of the method of the application, the third layer judgment comprises:

[0029] Setting the intensity safety threshold by analyzing the intensity distribution of the causal paths under normal working conditions in the long-term historical operation data of the system .

[0030] Traversing the subset of key causal paths, all paths that satisfy are selected to form a high-intensity path subset.

[0031] If the high-intensity path subset is empty, it indicates that there is no path intensity exceeding the safety line at present, and the system determines that immediate reconstruction is not needed, and the process continues to monitor. If the high-intensity path subset is not empty, the fourth layer judgment is entered.

[0032] As a further choice of the method of the application, the fourth layer judgment mechanism comprises:

[0033] The system maintains a ring buffer of the high-intensity path subset identified in a number of control cycles;

[0034] For each path in the high-intensity path subset of the current cycle, matching is performed in the historical buffer, and the matching conditions include the same variable sequence and the intensity thereof exceeding the intensity safety threshold in each historical cycle;

[0035] The number of times of appearance of the path in a number of consecutive cycles is counted;

[0036] If there is at least one path satisfying the number of times of appearance being greater than or equal to 2, the fourth layer judgment is established, the reconstruction preparation state is activated, and the path set satisfying the condition is recorded as a trigger path set.

[0037] As a further choice of the method of the application, in the reconstruction preparation state, a reconstruction strategy package is formed by matching and combining a plurality of basic reconstruction actions from a preset atomic action library according to the structure, variable and timing characteristics of the current driving causal path, comprising:

[0038] Deep feature extraction is performed on the paths in the trigger path set, including identifying the start and end variable types, analyzing the event timestamp sequence to determine the abnormal pattern, and analyzing the variable belonging to the physical device and control loop in combination with the factory information model;

[0039] Based on the extracted features, the abnormality is induced to at least one reconstruction dimension of signal source correction, control law adjustment, execution path switching, and logic association optimization;

[0040] From a pre-built library of atomic actions, matching atomic actions are retrieved based on reconstruction dimension labels, semantic similarity, or correlation with historical success cases; the atomic actions are independently executable and reversible control system modification instructions.

[0041] Based on path strength, scope of influence, and dependencies between actions, multiple atomic actions are combined into an ordered refactoring strategy package, and static conflict and dependency analysis is performed on the refactoring strategy package.

[0042] As a further option of the method of the present invention, the step of performing pre-performance verification in a virtual environment and approving execution only when the verification is successful includes:

[0043] Call upon a digital twin model that matches the actual control system;

[0044] Load the current real system state snapshot into the virtual environment as the initial condition;

[0045] The actions in the reconstruction strategy package are executed sequentially in the virtual environment, and the simulation runs for a predetermined duration at accelerated or real-time speed.

[0046] During and after the simulation, monitor whether the original abnormal event flow disappears, whether the dynamic response of the key controlled variables meets the standard, whether all process variables are within the safe operating range, and whether new high-intensity abnormal event flows or causal paths are generated.

[0047] The refactoring strategy package is approved to proceed to the actual deployment phase only if all predefined success criteria are met.

[0048] As a further option of the method of the present invention, a safety boundary constraint is set for each action in the reconstructed strategy package, and the strategy package is deployed to the real control loop under the constraint; the old and new control structures are gradually transitioned over multiple cycles using a S-shaped function, including:

[0049] Define safety boundary constraints for each atomic action in the refactoring strategy package, including upper and lower limits for parameter adjustment, preconditions for mode switching, execution timeout, number of retries for failure, and rollback trigger conditions, to form a refactoring strategy package with safety constraints;

[0050] Determine the total number of migration cycles, and calculate the mixing coefficient for each migration cycle based on the sigmoid function. : ,in, For shape parameters; This represents the total number of migration cycles. This is the current migration cycle;

[0051] Calculate the actual effective control action in each migration cycle. and will When applied to the actual control loop, when hour, , , migration is completed, and the system completely operates under the new structure; wherein , wherein is an output of the old control structure, is an output of the new control structure;

[0052] During the migration process, parallel execution of lightweight abnormal event stream monitoring and causal strength calculation is performed, if the integrated causal strength of the new path exceeds the safety threshold, the migration is immediately frozen, the current mixed coefficient and control action are maintained, and the four-layer nested judgment mechanism is returned to re-evaluate.

[0053] As a further option of the method of the application, after the reconstruction deployment, the abnormal event stream is monitored within a predetermined time scale; if the abnormal event stream disappears, the verified reconstruction strategy package is marked as a steady state, and after being tagged by a scene, it is included in the atomic action library; if the abnormality still exists or a new abnormal pattern appears, the strategy rollback is triggered, and the causal inference and evaluation cycle is restarted, including:

[0054] After the migration is successfully completed, a predetermined observation period is entered, and it is continuously monitored whether the original abnormal event stream disappears, whether a new high-intensity path with the same or similar pattern as the original trigger path set appears, and whether the key controlled variable operation index is restored;

[0055] If all the monitoring conditions are met during the observation period, it is determined that the reconstruction is successful, the topological features of the reconstructed trigger path set, the variable process semantic label, and the system steady state working point are encoded as a scene feature vector, and the successful reconstruction strategy package with safety constraints is associated with the scene feature vector and stored in the reconstruction case knowledge base;

[0056] If the original abnormality appears during the observation period, the periodic appearance or a completely new high-intensity abnormal pattern appears, it is determined that the reconstruction fails, and an automatic rollback program is started to restore the control structure to the state before the reconstruction by a smooth transition curve;

[0057] After the rollback is completed, the failure log is recorded, a new round of diagnosis and reconstruction cycle is started from the real-time acquisition and causal inference step, and the failed strategy combination is avoided in the subsequent action matching, realizing the continuous adaptive learning of the system.

[0058] Another aspect of the application provides an abnormal event stream causal inference control loop dynamic reconstruction system, the system comprising:

[0059] A multi-source event acquisition and causal inference module is used for real-time acquisition of multi-source abnormal event streams generated during the operation of an industrial control system, online lightweight causal inference is performed on the event sequence within a recent time window, and driving causal paths with controller input or output variables as endpoints are identified;

[0060] A multi-layer path evaluation and decision module is configured to start a four-layer judgment mechanism for the identified driving causal path, and only when the four-layer judgment is all established, the reconstruction preparation state is activated.

[0061] A reconstruction strategy generation and verification module is configured to match and combine multiple basic reconstruction actions from a preset atomic action library to form a reconstruction strategy package according to the structure, variables and timing characteristics of the current driving causal path in the reconstruction preparation state, and to perform a pre-verification in a virtual environment, and only when the verification is successful, the execution is approved.

[0062] A safe migration deployment module is configured to set a safety boundary constraint for each action in the reconstruction strategy package, and to deploy the strategy package to a real control loop under the constraint; the new and old control structures are controlled to migrate in a multi-periodic S-shaped function; if a new high-intensity driving causal path is detected during the migration process, the transition migration is immediately frozen, and the multi-layer path evaluation and decision module is re-evaluated.

[0063] A monitoring feedback and knowledge learning module is configured to monitor abnormal event streams in a predetermined time scale after the reconstruction deployment, and if the abnormal event streams disappear, the verified reconstruction strategy package is marked as a steady state, and is labeled by a scene and then included in the atomic action library; if the abnormality still exists or a new abnormal pattern appears, the strategy rollback is triggered, the multi-source event collection and causal inference module is restarted, and an adaptive reconstruction cycle is formed.

[0064] The beneficial effects of the present application are:

[0065] The method significantly improves the accuracy and timeliness of abnormal diagnosis by introducing real-time online causal inference and multi-layer intensity judgment mechanism. The traditional method relies on fixed rules and is difficult to deal with unknown abnormal patterns, with a false positive rate of more than 30%, and a response delay of several minutes to several hours. The method uses a sliding time window and online lightweight causal discovery to complete event stream analysis within seconds, and combines a four-layer judgment mechanism to effectively filter incidental noise, reduce the false positive rate to less than 5%, and ensure that only the continuous and high-threat abnormal path triggers reconstruction, greatly improving the diagnosis accuracy and system response speed.

[0066] In terms of control loop reconstruction, the method realizes safe and undisturbed online structure adjustment through digital twin pre-verification and S-shaped smooth migration mechanism. Traditional reconstruction often relies on manual trial and error, which is prone to secondary failures and has a high risk of system oscillation. The pre-verification success rate in the virtual environment can reach more than 90%, and the S-shaped transition function is used to realize the smooth switching of new and old control structures in multiple cycles, reducing the control output change rate by more than 70% and avoiding step disturbance. In addition, during the migration process, real-time monitoring of new abnormal paths is performed, and the migration is immediately frozen upon discovery, thereby controlling the system risk during the reconstruction process within an acceptable range and significantly improving the resilience and autonomous recovery capability of industrial control systems. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 Flowchart of the control loop dynamic reconstruction method for abnormal event stream causal inference;

[0068] Figure 2 Flowchart of the control loop dynamic reconstruction method for abnormal event stream causal inference S100;

[0069] Figure 3 Flowchart of the control loop dynamic reconstruction method for abnormal event stream causal inference S200;

[0070] Figure 4 Flowchart of the control loop dynamic reconstruction method for abnormal event stream causal inference S300;

[0071] Figure 5 Flowchart of the control loop dynamic reconstruction method for abnormal event stream causal inference S400;

[0072] Figure 6 Flowchart of the control loop dynamic reconstruction method for abnormal event stream causal inference S500. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0074] Industrial control systems, as the core of modern production operations, integrate numerous heterogeneous subsystems and high-frequency interactive control loops, such as process controllers, safety instrumented systems, and equipment management systems. These control loops generate a large number of status and alarm events during operation, and complex causal relationships may exist between these events. Especially when anomalies occur, disturbances can propagate along the control chain, leading to cascading effects. Traditional anomaly diagnosis and handling often rely on expert experience or fixed rules, resulting in delayed responses, poor flexibility, and difficulty in dynamically adjusting the control structure to suppress anomaly propagation during system operation. Existing methods lack systematic, online inference of causal relationships in anomaly event flows and intelligent reconfiguration mechanisms based on causal strength, leading to low operational efficiency and insufficient system resilience.

[0075] The theoretical foundation of this invention is built upon three pillars: the theory of temporal causal discovery, the theory of comprehensive evaluation of path strength, and the theory of safe migration in dynamic systems. By constructing an event flow causal network, implementing multi-layer strength judgment, and designing a controlled, progressive reconfiguration migration, adaptive recovery and continuous learning of industrial control loop anomalies are ultimately achieved.

[0076] The definitions of the core variables and the derivation of the formulas are as follows:

[0077] 1. Exceptional events and event flow:

[0078] Defined at time The abnormal event that occurred was a triple. .in, The system variable identifier associated with the abnormal event. It is a confidence level calculated based on predefined rules, model residuals, or statistical process control, representing the probability that an anomalous event is a true anomaly. (The last part, "in length...", appears to be incomplete and requires further context.) Within the time window, all events are sorted by timestamp to form an abnormal event stream. .

[0079] 2. Cause-effect graphs and cause-effect paths:

[0080] Through the By performing causal discovery, a directed graph is obtained. Vertex set Corresponding to all the variables that have appeared edge set Directed edges in Representing variables For variables There is a statistically significant causal effect. A causal relationship exists from the source variable. To target variable causal path Represented as a vertex sequence Path length The number of edges on a path, i.e. .

[0081] 3. Path integrated causality strength computation model:

[0082] To quantify the importance of different causal paths in driving abnormal propagation, the integrated causality strength of a path is defined as . The integrated causality strength is a weighted integration of path length, event confidence, and time coupling: ; where: is the path length factor. The shorter the path, the more direct the abnormal propagation, and the larger the factor value, the stronger the influence. is the arithmetic mean of all event confidences on the path, representing the overall confidence of abnormal signals on the path. is the time coupling tightness factor. is the maximum time interval between adjacent events on the path. is the decay coefficient. This factor indicates that the more concentrated the time of event occurrence on the path, the tighter the causal correlation in time, and the larger the factor value. are the weight coefficients of the corresponding factors, and satisfy .

[0083] The value range of the integrated causality strength is , and the larger the value, the stronger the driving of the causal path and the more attention it needs.

[0084] 4. Path significance decision function:

[0085] Whether a path is significant enough to trigger reconstruction is determined by whether its strength continuously exceeds the threshold and stably appears. The significance decision function of a path is defined as ; where: is the preset safety strength threshold. is an indicator function, taking the value 1 when the condition is true, and 0 otherwise. represents the set of high-strength paths identified in the th control period, i.e., the paths of . represents the current control period sequence number.

[0086] This function indicates that only when the current strength of the path exceeds the threshold, and in the last three consecutive control periods, at least two periods are identified as high-strength paths, the path is judged to be significant . ​

[0087] 5. Control loop migration S-shaped transition function: to ensure the control loop from the original structure to the new structure Smooth transition, avoid step disturbance, use S-shaped (Sigmoid) function to control the migration process. The hybrid coefficient of the th migration period in the total period is defined as: ; Where, is the shape parameter, which controls the steepness of the S-shaped function. Then the actual effective control effect of the migration period is: ; When , , ; When , , . The whole migration process is smooth and continuous.

[0088] The above theoretical framework provides a solid mathematical foundation for the present application, ensuring the accuracy and reliability of the whole process from abnormal event perception, causal strength quantification to safe reconstruction migration.

[0089] The specific embodiments of the present application will be described in detail below.

[0090] Example 1:

[0091] Please refer to Figure 1 , which shows the overall flowchart of an abnormal event stream causal inference control loop dynamic reconstruction method provided by an embodiment of the present application, which comprises:

[0092] S100: Real-time acquisition of multi-source abnormal event streams generated in the running process of the industrial control system, online lightweight causal inference on the event sequence in the recent time window, and identification of driving causal paths with controller input or output variables as the terminal.

[0093] S200: For the identified driving causal path, start the four-layer nested judgment mechanism, and only when all conditions are met, activate the reconstruction preparation state.

[0094] S300: In the reconstruction preparation state, according to the characteristics of the current driving causal path, match and combine multiple basic reconstruction actions from the pre-set atomic action library to form a reconstruction strategy package; pre-verification in a virtual environment, and only when the verification is successful, approve execution.

[0095] ​​S400: Set safety boundary constraints for each action in the reconfigured policy package, under which the policy package is deployed to the real control loop; the new and old control structures are gradually migrated in a multi-period according to a S-shaped function; during the migration process, continuous monitoring is performed, and if a new high-intensity driving causal path is detected, the migration is immediately frozen and the S200 is returned to re-evaluate.

[0096] S500: After reconfiguration and deployment, monitor abnormal event flow in a predetermined time scale; according to the monitoring result, the successful reconfigured policy package is marked as a steady state and is included in the knowledge base, or the policy rollback is triggered and the diagnosis and reconfiguration cycle is restarted.

[0097] The specific scheme is as follows:

[0098] In a control loop dynamic reconfiguration method of abnormal event flow causal inference, S100 realizes the conversion from raw data to preliminary causal diagnosis result through the deployment of data acquisition and event generation agents on each node of the system and the central flow processing engine.

[0099] Please refer to Figure 2 , which shows a flowchart of the S100 step in an exemplary control loop dynamic reconfiguration method of abnormal event flow causal inference of the present application, the content of which includes:

[0100] S110: In the present application, software or hardware acquisition agents are deployed at key positions of the industrial control network. The key positions include but are not limited to: operator stations and controllers of distributed control systems, CPU modules of programmable logic controllers, intelligent interfaces of key sensors and actuators, and network switches.

[0101] In a possible implementation, the collected content includes:

[0102] For the controller, the process variable PV, the set value SP, the control output OP, the working mode and the internal state word are collected.

[0103] For network traffic, the message of a specific protocol and its response time, error code are collected.

[0104] For system logs, the event logs of operating systems and application programs are collected.

[0105] According to the collected content, the abnormal event flow is judged according to the designed rules. In a possible implementation, the rule types include:

[0106] When or , it is determined that there is an abnormal event flow, and the confidence degree may increase with the degree of overrunning.

[0107] When , it is determined that there is an abnormal event flow.

[0108] When the device state jumps from running to failure, or the communication quality jumps from good to bad, it is determined that an abnormal event flow occurs.

[0109] In a possible implementation, all generated events are formatted into a unified structure based on the collected content and the abnormal event flow determination rule , wherein, is the time of the occurred abnormal event flow, is the system variable identification associated with the abnormal event flow, is the confidence of the abnormal event flow calculated based on a predefined rule, a model residual or a statistical process control.

[0110] As a possible option, the events are pushed to the central event flow processing engine in real time through a lightweight message middleware. The time of each node is synchronized through the NTP protocol to ensure global timing consistency.

[0111] In the present application, the central event flow processing engine maintains a fixed-length sliding time window. When a new event arrives, it is added to the window, and the event that is the oldest in the window is removed. The central event flow processing engine executes a causal discovery algorithm on the event sequence in the window.

[0112] In a possible implementation, the execution of online causal discovery includes:

[0113] In order to adapt to the causal discovery algorithm, the continuous time axis is discretized into equally spaced time slots. For each variable , in each time slot, if multiple events occur, their confidences are aggregated as the activity of the variable in this time slot. Finally, a set of multivariate time series with a length of is obtained.

[0114] The multivariate time series are analyzed by using a causal discovery algorithm. In a preferred embodiment, a PC algorithm based on conditional independence test is used. The PC algorithm sequentially tests the conditional independence of each pair of variables given a subset of other variables . If the independence hypothesis is rejected, an edge or is added to the causal graph . In another preferred embodiment, an algorithm based on transfer entropy is used, which calculates the contribution of the historical information of the variable to the future of the predicted variable . The contribution that exceeds the threshold is considered to have a causal effect.​​

[0115] From the generated causal graph , extract all paths that end with a key control variable . A key control variable refers to a controller input or output that directly participates in the control law calculation. Starting from each key control variable , traverse backward until either the starting variable is found or the maximum search depth is reached . All found paths constitute the initial set of driving causal path candidates .

[0116] S130: In the present application, for each path in the set of driving causal path candidates , compute the underlying attributes for subsequent strength evaluation based on the original information contained in the events.

[0117] In one possible implementation, the path attribute computation includes: recording the path sequence , computing the path length , extracting the confidence and timestamp of each event on the path, computing the initial average confidence of the path events , computing the maximum time interval between adjacent events on the path , where , , are the events corresponding to the adjacent variable pairs on the path.

[0118] Encapsulate the above information as structured data and output as the set of driving causal paths , which is the input of S200.

[0119] In one method of dynamic reconstruction of control loop in abnormal event stream causal inference, S200 receives as input and filters out the significant abnormal propagation paths that are truly threatening and need to be reconstructed by sequentially performing four layers of logical rigorous judgments.

[0120] Please refer to Figure 3 , which shows the flowchart of S200 in one exemplary method of dynamic reconstruction of control loop in abnormal event stream causal inference, the content of which includes:

[0121] S210: In the present application, the first layer of judgment is to filter the key paths that end with the core controlled object. Specifically, the first layer of judgment is to filter out the paths that involve control loops but do not have a direct impact on the core process safety, quality or equipment safety, thereby focusing on the most critical risks.

[0122] In one possible implementation, the execution of the first-level judgment includes:

[0123] Based on process safety information, key product quality parameters, and equipment protection interlock lists, define and maintain a list of key controlled variables. .

[0124] for Each path in Obtain its endpoint variable .

[0125] if If the controller output variable OP is itself, then the control loop configuration library is queried directly to determine whether the final execution object manipulated by the controller output variable OP is used for regulation. A variable in [the system]. If so, then... This is the critical path.

[0126] if If it is a process variable PV, then trace the control loop to which the process variable PV belongs and determine whether the controller output variable OP of the control loop is ultimately used for regulation. The variables in. If so, then This is the critical path.

[0127] Based on the above analysis, from Filter out all endpoint associations The paths constitute a subset of key causal paths. Paths that fail the screening will be recorded but will not trigger subsequent refactoring processes.

[0128] S220: In this invention, the second-level judgment is to calculate the comprehensive causal strength of the critical path. Specifically, the second-level judgment... Each path in Calculate the path using the aforementioned formula Comprehensive causal strength This transforms qualitative judgments of causal relationships into comparable quantitative indicators.

[0129] In one possible implementation, the execution of the second-level judgment includes:

[0130] for Read the path length calculated in S130. Initial average confidence level of path events and path Maximum time interval between adjacent events .

[0131] Calculate each intensity factor, specifically including:

[0132] Path length factor: ;

[0133] Average confidence factor, used directly ;

[0134] Time coupling factor: Attenuation coefficient Set according to the typical system response time;

[0135] According to preset weights Calculate the overall causal strength .

[0136] for Each path in Add the calculated value.

[0137] S230: In this invention, the third-level judgment is a preliminary strength screening based on a safety threshold. Specifically, the third-level judgment sets a strength safety threshold. ,Will Paths with moderate to high intensity were selected as potential high-risk paths.

[0138] In one possible implementation, the execution of the third-level judgment includes:

[0139] A strength safety threshold is set by analyzing the intensity distribution of causal paths under normal operating conditions in the system's long-term historical operating data. .

[0140] Traversal Select all that satisfy The paths constitute a subset of high-intensity paths. .

[0141] if If the set is empty, it indicates that no path strength exceeds the safety threshold, the system determines that immediate reconstruction is unnecessary, and the process returns to S100 to continue monitoring. If If it is not empty, proceed to the fourth level of judgment.

[0142] S240: In this invention, the fourth layer of judgment is a recurrence stability verification. Specifically, the fourth layer aims to avoid overreacting to occasional, transient abnormal fluctuations, requiring high-intensity paths to appear stably over multiple consecutive control cycles to confirm their persistence and authenticity.

[0143] In one possible implementation, the execution of the fourth-level judgment includes:

[0144] The system maintains a circular buffer to record the most recent... identified in a control cycle , denoted as , where is the current cycle.

[0145] For the current cycle , each path in is matched in the history buffer . The match requires not only the same variable sequence, but also the same strength in the history cycle. .

[0146] The statistics is the number of times appears in the last consecutive cycles.

[0147] If there exists at least one path whose , the fourth layer of judgment is passed. At this point, all four layers of nested judgments are passed, the system officially activates the reconstruction preparation state, and the path set that meets the repetitive condition is denoted as the trigger path set . The system will carry and all related path information into S300.

[0148] In a control loop dynamic reconstruction method for abnormal event flow causal inference, S300, in the reconstruction preparation state, intelligently combines the pre-defined correction operations based on the abnormal propagation pattern disclosed in to form a feasible reconstruction scheme and perform safety verification in a virtual environment.

[0149] Please refer to Figure 4 , which shows the flow chart of S300 in an exemplary control loop dynamic reconstruction method for abnormal event flow causal inference, the content of which includes:

[0150] S310: In the present application, multi-dimensional feature extraction is performed on the paths in to provide a basis for subsequent matching and reconstruction actions.

[0151] In one possible implementation, the multi-dimensional feature extraction of the paths in includes:

[0152] Identify the type of the starting point , the intermediate transfer node, and the type of the end point . This helps to determine whether the abnormality is from the measurement link, the control calculation link, or the execution link.

[0153] Analyze the timestamp sequence of events on the path to determine whether it is a sudden abnormality or a slowly varying drift.

[0154] Combine with the plant information model, resolve the physical device, control loop and process unit that the variable in the path belongs to.

[0155] Based on the analysis results, the problem is summarized into one or more reconstruction dimensions, such as: signal source correction, control law adjustment, execution path switching and logic association optimization.

[0156] S320: In the present application, the system maintains a preset atomic action library Each atomic action is an atomic, independently executable and reversible control system modification instruction, such as: a1: Set_PID_Mode('TIC101', 'MANUAL'), a2: Switch_Input('FIC202.PV', 'FT202B'), a3: Adjust_Param('PIC301.Kp', +0.15).

[0157] In one possible implementation, the intelligent matching and combination of reconstruction actions from the atomic action library is as follows:

[0158] According to the reconstruction dimensions summarized in S310, candidate actions with matching tags are retrieved in . Retrieval can be based on keywords, semantic similarity or association degree of historical successful cases.

[0159] According to the strength, influence range and dependency relationship between actions of the path in , one or more selected atomic actions are combined into an ordered reconstruction strategy package . The strategy reconstruction strategy package contains the execution object, parameter, expected effect of each action and the timing or logic constraints between actions.

[0160] Perform static analysis on to check if there are resource conflicts or missing logic dependencies between actions.

[0161] S330: In the present application, before executing real system modification, must be simulated in a high-fidelity virtual environment , evaluate its effect and safety.

[0162] In one possible implementation, the execution of virtual preview verification includes:

[0163] Call the digital twin model matching the current real control system.

[0164] Load the current system state snapshot to the virtual environment as the initial condition. Execute the actions in in turn. Run in the virtual environment for a long enough time at accelerated or real-time speed.

[0165] During and after the simulation run, monitor whether the original abnormal event stream disappears, the dynamic response of the key controlled variables , whether all process variables are always in the safe operating interval, whether new, high-intensity abnormal event streams or causal paths are generated.

[0166] Only when the evaluation results meet all the predefined success criteria, the system approves to enter the real deployment phase. Otherwise, the system will reject the combined reconstruction strategy package and generate a pre-run failure report, optionally returning S320 to regenerate the strategy or triggering manual intervention.

[0167] In a control loop dynamic reconstruction method of abnormal event stream causal inference, S400 is responsible for safely and smoothly deploying the verified reconstruction strategy package to the real control system, and implementing strict monitoring during the migration process to ensure that the process is controlled.

[0168] Please refer to Figure 5 , which shows the flowchart of S400 step in an exemplary control loop dynamic reconstruction method of abnormal event stream causal inference of the present application, the content of which includes:

[0169] S410: In the present application, for each atomic action in , a clear operation boundary is defined as a hard constraint for its execution in the real system.

[0170] In one possible implementation, the setting of the safety boundary constraint includes:

[0171] For parameter adjustment type actions, set the upper and lower limits of the allowed adjustment.

[0172] For mode switching type actions, set the prerequisite state allowed for switching.

[0173] Set the maximum waiting time for each action execution, the number of failure retries. Predefine the action level rollback trigger condition.

[0174] The above constraints are attached to , forming a reconstruction strategy package with safety constraints .

[0175] S420: In the present application, to avoid the impact of control output jump on the production process, the switching between new and old control structures is controlled by a Sigmoid function and completed smoothly within multiple cycles.

[0176] In one possible implementation, the execution of the migration between the new and old control structures includes:

[0177] Determine total number of migration cycles . Depending on the dominant time constant of the process. Set the S-shaped function shape parameter , take to make the migration smooth.

[0178] Migration control initialization phase, let the migration cycle counter , the current control structure is the old structure .

[0179] For each migration cycle from 1 to :

[0180] a. Calculate the current mixing coefficient .

[0181] b. Get the new control structure expected after all actions in take effect. .

[0182] c. Calculate the actual control action that should be applied in the current cycle .

[0183] d. Apply to the real control loop.

[0184] e. Wait for a migration cycle duration.

[0185] f. Perform migration cycle according to the execution procedure of a-e of the migration cycle .

[0186] When , , , the migration is complete, and the system is fully running under the new structure.

[0187] S430: In the present invention, throughout the migration, the system performs parallel lightweight abnormal event stream monitoring and causal strength calculation, and judges in real time whether the migration introduces new risks.

[0188] In one possible implementation, while the migration loop is running, a separate monitoring thread executes the S100 and S230 logic, i.e. collects events, makes quick causal inference, and calculates the strength of the new-born path at a frequency comparable to the control cycle .

[0189] In one possible implementation, the freezing condition is: whether there is a new-born path , the new-born path comprehensive causal strength Exceeding a safety threshold .

[0190] Once the freeze condition is triggered, the system immediately freezes the migration process. In one possible implementation, the freeze operation includes: stopping the increment of the migration cycle counter ; keeping the current mixing coefficients unchanged, thereby stabilizing the control action at the current intermediate state ; and suspending the further execution of the actions that have not yet taken full effect in the system.

[0191] After the freeze, the system immediately jumps to S200 with all the current state information, and restarts the four-layer nested judgment process. This time, the judgment will consider both the original triggering path and the newly born path.

[0192] In one method of dynamically reconstructing a control loop of an abnormal event flow causal inference, S500 enters the effect confirmation and experience learning phase after the reconstruction migration is completed, forming a complete link of diagnosis, reconstruction, and learning.

[0193] Please refer to Figure 6 , which shows a flowchart of the S500 step in one exemplary method of dynamically reconstructing a control loop of an abnormal event flow causal inference, the content of which includes:

[0194] S510: In the present application, after the S-shaped function migration is successfully completed, i.e., the freeze is not triggered and S200 is reached smoothly, the system does not immediately end the task, but enters a predetermined observation period . The length of the observation period should be sufficient to cover the main dynamics and possible disturbance periods of the process.

[0195] In one possible implementation, the effect determination logic is as follows:

[0196] If all the following conditions are met during the entire predetermined observation period , it is determined that this reconstruction is successful:

[0197] a. The original triggering abnormal event flow completely disappears.

[0198] b. There is no new high-intensity ( ) causal path of the same or similar pattern as the triggering path set .

[0199] 3. The operating index of the key controlled variable returns to an acceptable level.

[0200] If any of the following occurs during the entire predetermined observation period , it is determined that this reconstruction fails:​

[0201] a. The original trigger abnormal event stream still persists or occurs periodically.

[0202] b. A brand new, high-intensity (e.g., high-frequency, high-impact) abnormal pattern emerges.

[0203] S520: In the present application, for the reconfiguration case with successful determination, its strategy package and scenario information are converted into reusable knowledge and stored in the system knowledge base.

[0204] Specifically, the complete context of this reconfiguration is encoded to generate a scenario feature vector. The scenario feature vector includes: the topological features of the trigger path set , the process semantic labels of key variables, the steady-state operating point of the system before reconfiguration, etc. The reconfiguration strategy package with safety constraints that is verified to be successful is associated with the scenario feature vector and is labeled with success, time stamp, and effect score. Then this scenario-strategy pair is stored as a new record in the reconfiguration case library, which is an extension of the atomic action library.

[0205] S530: In the present application, for the reconfiguration case with failed determination, the system automatically performs a strategy rollback and learning cycle, and optimizes subsequent decisions using failure information.

[0206] Specifically, the system starts an automatic rollback program. The rollback process uses a smoother transition curve than the forward migration, which restores the control structure from the current state to the original state before reconfiguration. After rollback, the system records the complete log of this failure in detail, and continues to collect and analyze the abnormal event stream from the S100 step. At this time, due to the rollback, the system state is basically restored, but the root cause of the abnormality may still not be eliminated, so the new round of diagnosis will be based on updated information. At the same time, in the future S320 step, the matching algorithm refers to this failure case to avoid recommending the strategy combination that has been proven to be ineffective or harmful, thereby realizing the continuous adaptive learning and optimization of the system.

[0207] Embodiment Two:

[0208] The present application was deployed and verified in the main steam temperature control system of a large thermal power plant for nine months. The system includes four boilers, and the main steam temperature control of each boiler involves a cascade PID loop, multiple desuperheating water regulating valves, and feedforward compensation logic. The monitoring range covers more than 500 related variables. During implementation, the following specific configurations were used:

[0209] ​​A data collection agent is deployed on each boiler's DCS engineer station to read about 3000 data points through an OPC DA interface in real time with a sampling period of 1 second. The event generation rule base contains more than 80 rules for cases such as large temperature deviation, frequent valve action, decoupling of cascade loops, actuator failure, etc.

[0210] The edge computing server adopts a time window W = 90 seconds and uses a lightweight causal discovery algorithm based on VAR-LiNGAM, which performs inference every 30 seconds. The integrated causal strength weight is set to , and the safety threshold .

[0211] A special atomic action library for temperature control loops is preconfigured, containing 45 actions in 12 categories, such as switching the main PID to the backup parameter set, adjusting the integral time of the secondary PID, modifying the upper limit of the desuperheating valve position, enabling / disabling a certain feedforward signal, etc. The virtual environment is based on a high-precision boiler-turbine model built in MATLAB / Simulink.

[0212] Total number of migration cycles According to the large inertia characteristics of the temperature object, it is set to 40, and the S-type function parameter . The safety boundary strictly follows the temperature and pressure rate of change limits in the boiler operation procedures.

[0213] During the trial operation, the system successfully handled multiple typical abnormalities automatically:

[0214] 1. Case one: temperature sensor drift causes control oscillation;

[0215] 5# Boiler A-side main steam temperature display value appears low-frequency periodic fluctuations, causing the desuperheating water regulating valve to act frequently. The system identifies a short path starting with the faulty temperature measurement point TT0101.PV, with a path strength , appearing for three consecutive cycles. The matching action switches the input of the main PID from TT0101 to the redundant measurement point TT0102 at the same location, and verifies in the virtual environment that the switching process will not cause a sudden temperature change. After approval, S-type migration is performed. After migration is completed, the temperature fluctuations disappear and the valve action returns to normal frequency. The single-point sensor switching strategy is marked and stored in the case library.

[0216] 2. Case two: slight sticking of desuperheating water regulating valve causes cascade loop mismatch;

[0217] 2# Boiler B-side desuperheating water regulating valve CV0202 appears slight sticking, causing the secondary loop to respond slowly and the main steam temperature to drift slowly. The causal path points to the abnormal relationship between CV0202.OP and CV0202.PV, with a path strength , and is strongly related to the load change condition. The system combines two actions: a) temporarily change the sub-loop PID control mode to manual with dead zone compensation; b) slightly increase the output bias of the main control PID in the DCS logic to manually compensate for the flow. The virtual preview shows that the combination can stabilize the temperature. After the strategy is deployed, the temperature drift is suppressed. The system also generates a maintenance work order to check the CV0202 valve. This manual compensation + mode switching combination strategy is stored as a composite action.

[0218] 3. Case three: insufficient feedforward compensation under fuel heat value fluctuation;

[0219] During the replacement of the coal batch, the dynamic response of the 3# boiler main steam temperature to load changes deteriorated, and the overshoot increased. The system identified multiple paths with the boiler master control command as the starting point and the main steam temperature deviation as the ending point, and the path intensity significantly increased in a specific load interval, forming a sensitive condition cluster. The matching action dynamically enhances the feedforward compensation coefficient related to the main steam flow. The virtual environment optimizes the coefficient for pre-performance. When executing the S-type migration to the 15th cycle, the monitoring thread detects a short-term temperature disturbance caused by other reasons, and the newly born path intensity . The system immediately freezes the migration. After returning to S200 for re-evaluation, it is determined that the newly born disturbance is caused by known short-term operations, and the disturbance disappears after 10 minutes. The system unfreezes and continues to complete the remaining migration. The migration is finally completed, and the temperature control quality is improved under the new fuel condition. This case verifies the effectiveness of the migration monitoring and freezing mechanism.

[0220] Implementation effect statistics:

[0221] The accuracy rate of identifying driving abnormal paths reaches 96.8%.

[0222] The average time from abnormality highlighting to the readiness of the reconstructed strategy package is 65 seconds.

[0223] The virtual preview successfully intercepted two aggressive reconstruction strategies that could have triggered the over-limit of boiler metal thermal stress.

[0224] The S-type function migration reduces the maximum instantaneous change rate of the main steam temperature during reconstruction by about 70% compared to the traditional direct switching method, significantly improving the safety of unit operation.

[0225] Through the accumulation of case libraries, the system shortens the diagnosis and strategy generation time for similar abnormalities by more than 50%.

[0226] During the project period, the cumulative avoidance of main protection actions or unplanned shutdowns due to delayed abnormal handling is 3 times, which reflects the significant value of the invention in improving the resilience and autonomy of industrial control systems.

[0227] In particular, in case three, the migration freeze and re-evaluation mechanism successfully avoids over-reaction to short-term operation disturbances, embodying the robustness and practicality of the method in complex industrial environments. The system accumulates optimized feedforward coefficient strategies, which are automatically recommended and applied during subsequent replacement of the same coal batch, achieving effective knowledge inheritance and reuse.

[0228] Embodiment three:

[0229] An abnormal event stream causal inference control loop dynamic reconstruction method system, comprising:

[0230] A multi-source event collection and causal inference module is used to collect multi-source abnormal event streams generated during the operation of an industrial control system in real time, perform online lightweight causal inference on event sequences within a recent time window, and identify driving causal paths with controller input or output variables as endpoints.

[0231] A multi-layer path evaluation and decision module is used to start a four-layer nested judgment mechanism for the identified driving causal paths, and only when all four layers are true, the reconstruction preparation state is activated.

[0232] A reconstruction strategy generation and verification module is used to match and combine multiple basic reconstruction actions from a pre-set atomic action library based on the structure, variables and timing characteristics of the current driving causal path in the reconstruction preparation state, forming a reconstruction strategy package; and perform a rehearsal verification in a virtual environment, and only when the verification is successful, the execution is approved.

[0233] A safe migration deployment module is used to set safety boundary constraints for each action in the reconstruction strategy package, deploy the strategy package to the real control loop under the constraints; control the new and old control structure to migrate in a S-shaped function over multiple cycles; if a new high-intensity driving causal path is detected during migration, the transition migration is immediately frozen, triggering the multi-layer path evaluation and decision module to re-evaluate.

[0234] A monitoring feedback and knowledge learning module is used to monitor abnormal event streams within a predetermined time scale after reconstruction deployment; if the abnormal event stream disappears, the verified reconstruction strategy package is marked as a steady state, and after being labeled by a scene, it is included in the atomic action library; if the abnormality still exists or a new abnormal pattern appears, trigger strategy rollback, restart the multi-source event collection and causal inference module, and form an adaptive reconstruction cycle.

[0235] Those skilled in the art will appreciate that embodiments of the application described herein are provided as methods, systems, or computer program products. Accordingly, embodiments of the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, embodiments of the application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage devices, etc.) embodying computer program code thereon for use by or in connection with an instruction execution system. Program code embodied on one or more computer readable storage media can be transmitted using any apparatus adapted to transfer a set of data or instructions for execution by a machine.

[0236] The present application is described in reference to the flow diagrams and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flow diagrams and / or block diagrams, and combinations of blocks in the flow diagrams and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing system or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flow diagrams and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks.

[0237] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow diagrams and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks.

[0238] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow diagrams and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks. Figure 1 one or more functions specified in the flow diagram and / or block diagram block or blocks.

[0239] While preferred embodiments of the application have been described, modifications and alterations thereto will occur to those skilled in the art upon reading the preceding description. It is intended to include all such modifications and alterations insofar as they come within the scope of the basic inventive concepts disclosed herein.

[0240] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover the modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

Claims

1. A method for dynamically reconstructing control loops for causal inference of abnormal event flows, characterized in that, include: Real-time acquisition of multi-source abnormal event streams generated during the operation of industrial control systems; online lightweight causal inference of event sequences from multi-source abnormal event streams; identification of driving causal paths ending at controller input or output variables. For the identified driving causal path, a four-layer nested judgment mechanism is initiated. The reconstruction preparation state is activated only when all four judgments are true. The first-level judgment mechanism is to screen the key causal path of the endpoint-related controlled variable. The second-level judgment mechanism is to calculate the comprehensive causal strength of each path; The third-level judgment mechanism is to determine whether the path strength exceeds a preset safety threshold. The fourth-layer judgment mechanism is to verify whether the path exceeding the threshold has been repeated twice or more within a consecutive number of cycles; In the pre-reconstruction state, based on the structure, variables and temporal characteristics of the current driving causal path, multiple basic reconstruction actions are matched and combined from the pre-set atomic action library to form a reconstruction strategy package; pre-rehearsal verification is performed in the virtual environment, and execution is approved only if the verification is successful; Set safety boundary constraints for each action in the reconstructed strategy package, and deploy the strategy package to the real control loop under the constraints; the old and new control structures are gradually migrated over multiple cycles using an S-shaped function; if a new high-intensity driving causal path is detected during the migration, the transition migration is immediately frozen and the four-layer nested judgment mechanism is returned for re-evaluation. After refactoring and deployment, monitor the abnormal event flow within a predetermined time scale; if the abnormal event flow disappears, mark the verified refactoring strategy package as steady state, and incorporate it into the atomic action library after scenario tagging; if the abnormality still exists or a new abnormal pattern appears, trigger strategy rollback, restart the causal inference and evaluation loop, and form an adaptive refactoring mechanism.

2. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 1, characterized in that, The online lightweight causal inference performed on the event sequence of the multi-source anomalous event stream includes: Deploy data acquisition agents at key locations in the system to collect controller process variables, setpoints, control outputs, network packets, and system log data; Based on preset rules or models, the collected data is anomaly detected, and a uniformly formatted anomaly event triplet is generated; the anomaly event triplet includes a timestamp, variable identifier, and confidence level; The abnormal event stream is pushed to the central processing engine. Within the sliding time window, the continuous time axis is discretized into equally spaced time slots, and the confidence scores of multiple events for each variable within the time slots are aggregated to form a multivariate time series. A causal discovery algorithm is used to analyze multivariate time series data, generate a directed causal graph, and extract all candidate sets of driving causal paths ending with key control variables.

3. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 1, characterized in that, The first-level judgment mechanism specifically includes: Define and maintain a list of key controlled variables based on process safety information, key product quality parameters, and equipment protection interlock lists; For each path in the set of driving causal paths, obtain the endpoint variable; If the endpoint variable is a controller output variable, then query the control loop configuration library to determine whether the final execution object manipulated by the controller output variable is used to adjust the variables in the list of key controlled variables. If the endpoint variable is a process variable, trace the control loop to which the endpoint variable belongs and determine whether the controller output of the control loop is ultimately used to adjust the variables in the list of key controlled variables. The first-level judgment mechanism filters the paths into a subset of key causal paths.

4. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 3, characterized in that, In the second-layer judgment mechanism, the comprehensive causal strength of each path is calculated. The calculation formula is: ;in: It is the path length factor. , It is the path length; It is the average confidence factor. It is the confidence level of all events along the path. The arithmetic mean, It is the time coupling tightness factor. , It is a path The maximum time interval between adjacent events, It is the maximum time interval between adjacent events on the path. The attenuation coefficient is... These are the weight coefficients of the corresponding factors, and satisfy the following conditions: .

5. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 4, characterized in that, The third-level judgment includes: A strength safety threshold is set by analyzing the intensity distribution of causal paths under normal operating conditions in the system's long-term historical operating data. ; Traverse the subset of key causal paths and select all that satisfy the condition. The paths constitute a subset of high-intensity paths; If the high-intensity path subset is empty, it indicates that no path intensity exceeds the safety line, and the system determines that immediate reconstruction is not necessary, and the process continues to monitor. If the high-intensity path subset is not empty, the process proceeds to the fourth level of judgment.

6. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 5, characterized in that, The fourth-level judgment mechanism includes: The system maintains a circular buffer for high-intensity path subsets identified within several control cycles; For each path in the high-intensity path subset of the current period, a match is performed in the historical buffer. The matching criteria include identical variable sequences and the fact that the intensity of each path exceeds the intensity safety threshold in each historical period. ; The number of times a statistical path appears within a consecutive number of periods; If there exists at least one path that appears more than or equal to 2 times, then the fourth-level judgment is valid, the reconstruction preparation state is activated, and the set of paths that meet the condition is recorded as the trigger path set.

7. The method for dynamic reconstruction of control loops for causal inference of abnormal event flows according to claim 1, characterized in that, In the pre-reconstruction state, based on the structure, variables, and temporal characteristics of the current driving causal path, multiple basic reconstruction actions are matched and combined from a pre-set atomic action library to form a reconstruction strategy package, including: Deep feature extraction is performed on the paths in the trigger path set, including identifying the variable types of the start and end points, analyzing the event timestamp sequence to determine the abnormal pattern, and combining the factory information model to analyze the physical equipment and control loop to which the variable belongs; Based on the extracted features, the anomalies are categorized into at least one of the following reconstruction dimensions: signal source correction, control law adjustment, execution path switching, and logical association optimization. From a pre-built library of atomic actions, matching atomic actions are retrieved based on reconstruction dimension labels, semantic similarity, or correlation with historical success cases; the atomic actions are independently executable and reversible control system modification instructions. Based on path strength, scope of influence, and dependencies between actions, multiple atomic actions are combined into an ordered refactoring strategy package, and static conflict and dependency analysis is performed on the refactoring strategy package.

8. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 7, characterized in that, The pre-performance verification in the virtual environment, with approval for execution only upon successful verification, includes: Call upon a digital twin model that matches the actual control system; Load the current real system state snapshot into the virtual environment as the initial condition; The actions in the reconstruction strategy package are executed sequentially in the virtual environment, and the simulation runs for a predetermined duration at accelerated or real-time speed. Monitor whether the original abnormal event flow disappears during and after the simulation, whether the dynamic response of the key controlled variables meets the standards, whether all process variables are within the safe operating range, and whether new high-intensity abnormal event flows or causal paths are generated. The refactoring strategy package is approved to proceed to the actual deployment phase only if all predefined success criteria are met.

9. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 1, characterized in that, For each action in the reconfiguration strategy package, safety boundary constraints are set, and the strategy package is deployed to the real control loop under these constraints. The old and new control structures are gradually transitioned over multiple cycles using a sigmoid function, including: Define safety boundary constraints for each atomic action in the refactoring strategy package, including upper and lower limits for parameter adjustment, preconditions for mode switching, execution timeout, number of retries for failure, and rollback trigger conditions, to form a refactoring strategy package with safety constraints; Determine the total number of migration cycles, and calculate the mixing coefficient for each migration cycle based on the sigmoid function. : ,in, For shape parameters; This represents the total number of migration cycles. This is the current migration cycle; Calculate the actual effective control action in each migration cycle. and will When applied to the actual control loop, when hour, , The migration is complete, and the system is now fully operational under the new architecture; among them, ,in The output is from the old control structure. Output for the new control structure; During the migration process, lightweight abnormal event flow monitoring and causal strength calculation are performed in parallel. If the comprehensive causal strength of the new path is detected to exceed the safety threshold, the migration is immediately frozen, the current mixing coefficient and control effect are maintained, and the process returns to the four-layer nested judgment mechanism for re-evaluation.

10. The control loop dynamic reconstruction method for causal inference of abnormal event flow according to claim 1, characterized in that, After the reconstruction is deployed, the abnormal event stream is monitored within a predetermined time scale; if the abnormal event stream disappears, the verified reconstruction strategy package is marked as steady state and incorporated into the atomic action library after scene tagging. If the anomaly persists or a new anomaly pattern emerges, a policy rollback is triggered, restarting the causal inference and evaluation cycle, including: After the migration is successfully completed, the predetermined observation period will be entered to continuously monitor whether the original abnormal event flow disappears, whether new high-intensity paths with the same or similar patterns as the original trigger path set appear, and whether the key controlled variable operating indicators have recovered. If all monitoring conditions are met during the observation period, the reconstruction is considered successful. The topological features of the trigger path set, the variable process semantic labels, and the system steady-state operating points are encoded into scene feature vectors. The successful reconstruction strategy package with safety constraints is associated with the scene feature vectors and stored in the reconstruction case knowledge base. If the original anomaly persists, occurs periodically, or a new high-intensity anomaly pattern appears during the observation period, the reconstruction is deemed to have failed, and an automatic rollback procedure is initiated to restore the control structure to its pre-reconstruction state using a smooth transition curve. After the rollback is completed, a failure log is recorded, and a new round of diagnosis and reconstruction cycle is started from the real-time data collection and causal inference steps. In subsequent action matching, failed strategy combinations are avoided, so as to realize the system's continuous adaptive learning.

11. A control loop dynamic reconstruction system for causal inference of abnormal event flows according to any one of claims 1-10, characterized in that the system include: The multi-source event acquisition and causal inference module is used to acquire multi-source abnormal event streams generated during the operation of industrial control systems in real time, perform online lightweight causal inference on event sequences within the most recent time window, and identify driving causal paths ending with controller input or output variables. The multi-layer path evaluation and decision-making module is used to initiate a four-layer nested judgment mechanism for the identified driving causal path. The reconstruction preparation state is activated only when all four judgments are true. The refactoring strategy generation and verification module is used to match and combine multiple basic refactoring actions from a pre-set atomic action library to form a refactoring strategy package based on the structure, variables and temporal characteristics of the current driving causal path in the refactoring preparation state; it performs pre-rehearsal verification in a virtual environment and approves execution only when the verification is successful; The secure migration deployment module is used to set security boundary constraints for each action in the refactoring strategy package, and deploy the strategy package to the real control loop under the constraints; it controls the new and old control structures to migrate slowly over multiple cycles according to an S-shaped function; If a new high-intensity driving causal path is detected during the migration process, the transition migration is immediately frozen, triggering a re-evaluation by the multi-layer path evaluation and decision-making module; The monitoring, feedback, and knowledge learning module is used to monitor the flow of abnormal events within a predetermined time scale after the refactoring and deployment. If the abnormal event stream disappears, the verified refactoring strategy package will be marked as steady state and incorporated into the atomic action library after scene tagging; If the anomaly persists or a new anomaly pattern emerges, a policy rollback is triggered, and the multi-source event collection and causal inference module is restarted, forming an adaptive reconstruction loop.

Images