DCS operation and maintenance method based on fusion of dynamic weight distribution and fault chain deduction

Abstract

The application discloses a DCS operation and maintenance method based on fusion of dynamic weight distribution and fault chain deduction, and relates to the technical field of distributed control system operation and maintenance.The method comprises the following steps: obtaining multiple groups of source data of a monitoring area; determining corresponding monitoring periods according to the multiple groups of source data, and combining the monitoring period information; the technical key points are as follows: a multi-source data fusion module is used to associate the data of a DCS system, an associated production system and a cross-factory area collaborative system, then through adjustment of time-space alignment and multi-dimensional verification, fragmented data such as equipment operation data, production load fluctuation and cross-factory area interaction information are woven into a complete data network; such global perception capability not only solves the one-sidedness of traditional single data judgment, such as misjudgment of environmental interference by relying only on equipment temperature data, but also can capture implicit associated faults across systems; secondly, when the multiple groups of source data are mutually verified, an evidence chain is formed through logical consistency verification, so that the misjudgment rate of suspected faults is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of distributed control system operation and maintenance technology, specifically a DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain deduction. Background Technology

[0002] In modern industrial production, Distributed Control Systems (DCS), as the core hub connecting the equipment and management layers, are like the nervous brain of the production system; their stable operation directly affects the efficiency and safety of the entire production line. However, current traditional operation and maintenance models are facing multiple scenario-based challenges:

[0003] 1. In continuous production workshops, the 24-hour uninterrupted operation of equipment makes it difficult for traditional manual inspections to cover critical periods such as the early morning. For example, if the temperature sensor of the reactor experiences a slight drift and it is not detected in time, it will cause an imbalance in the raw material ratio after 6 hours, resulting in the scrapping of the entire batch of products. This mode of relying on timed inspections is difficult to capture potential faults that change dynamically.

[0004] 2. In cross-plant collaborative production scenarios, control system data from different areas are stored in isolation. When the raw material conveying pump in Plant A experiences pressure fluctuations, maintenance personnel cannot correlate the liquid level changes in the storage tanks of Plant B in real time, leading to misjudgments of local equipment failures when in fact it is a systemic pipeline pressure imbalance. This makes fault location more difficult, with an average troubleshooting time exceeding 4 hours.

[0005] 3. When new equipment is used in combination with old systems, maintenance experience is difficult to reuse. After a chemical plant introduced imported smart valves, their high-frequency vibration characteristics were significantly different from those of traditional valves. The original threshold alarm mechanism gave false alarms for three consecutive days, forcing maintenance personnel to turn off the alarm function. Ultimately, the production line was shut down because the real fault was not discovered. It is precisely because of this experience barrier that the handling of new faults can only be done by trial and error.

[0006] Furthermore, fixed operation and maintenance strategies cannot adapt to dynamic production scenarios; the normal operating parameters of the same compressor can differ by up to 30% under low load in winter and full load in summer, but traditional systems still use the same set of judgment criteria, which leads to frequent triggering of unnecessary load reduction protection in summer, resulting in an annual loss of about 2,000 tons of production capacity.

[0007] The above-mentioned common issues expose the core pain points of existing technologies: fragmented data utilization, delayed fault response, closed experience inheritance, and rigid decision-making logic. Therefore, there is an urgent need for a new operation and maintenance method that can achieve multi-source collaborative perception, cross-domain knowledge transfer, and dynamic intelligent decision-making to adapt to the increasingly complex and intelligent development needs of industrial production. Summary of the Invention

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The DCS operation and maintenance method based on the fusion of dynamic weight allocation and fault chain inference includes the following steps:

[0010] The system acquires system operation information, correlates production data, and cross-system interaction information. System operation information includes real-time records of various parameters during equipment operation, status records during data transmission, relevant records of control commands from issuance to execution, and records of the relevant impacts of the equipment's environment. Correlated production data includes records of changes in production load per unit time, records of the proportion settings of various raw materials during production, and records of quantitative results obtained after testing products according to preset standards. Cross-system interaction information includes records of the time difference between data transmission and reception, records of the proportion of successfully executed collaborative commands to the total number of commands, and records of the time from the issuance of resource scheduling commands to the start of resource allocation.

[0011] The acquired system operation information, associated production data, and cross-system interaction information are preprocessed. Preprocessing includes removing obviously erroneous data, converting data of different magnitudes to the same numerical range, and filling in missing data based on reasonable trends of change before and after the time intervals to generate preprocessed fused data.

[0012] The preprocessed fused data is spatiotemporally aligned and multidimensionally verified to generate a multi-source fused dataset. Spatiotemporal alignment ensures that data collected from different sources and at different times are matched in time and space. Multidimensional verification includes checking whether the data is within the valid time range, verifying whether the logical relationship between different data is reasonable, and confirming whether the association between data from different systems conforms to normal patterns.

[0013] The multi-source fusion dataset is input into a feature learning framework with cross-domain knowledge application capabilities, and the output is a multi-class feature sequence reflecting the system state. The multi-class feature sequence includes features reflecting the overall health status of the equipment in chronological order, features reflecting the stability of the data transmission link in chronological order, features reflecting the effect of control command execution in chronological order, features reflecting the degree of environmental influence on the system in chronological order, features reflecting the degree of correlation between the production system and the control system in chronological order, and features reflecting the cross-system collaborative effect in chronological order.

[0014] The causal reasoning and scenario adaptation mechanism is used to perform weighted fusion of multiple feature sequences to generate a comprehensive evaluation result. The causal reasoning and scenario adaptation mechanism is based on the causal correlation strength between various feature sequences and system faults and the priority requirements of the current production scenario.

[0015] By analyzing and comprehensively evaluating the results through anomaly identification models, the system outputs operational indicators, anomaly probabilities, fault trends, and associated impacts. Operational indicators are overall quantitative indicators of the current operational status of the system; anomaly probabilities are the likelihood of an anomaly occurring in the system; fault trends are the possible development of the fault in the future; and associated impacts are other effects that the fault may cause.

[0016] Based on operational and maintenance metrics, anomaly probability, failure trends, and associated impacts, a multi-level progressive response mechanism is triggered.

[0017] When the system is in a stable operating state, it initiates self-evolutionary learning to continuously optimize operation and maintenance strategies.

[0018] Preferably, the triggering conditions and execution content of the first-level response (automatic repair) in the multi-level progressive response mechanism are as follows:

[0019] When the operation and maintenance indicators are within the normal range, the probability of anomalies is in the preset low level range, the slope of the fault development trend curve is less than 0.02, and there are no associated fault risks, the system automatically executes preset repair actions, including parameter fine-tuning, backup link switching, and load distribution adjustment, while recording the execution effect and time of each action; parameter fine-tuning is to make minor adjustments to some settings during equipment operation; backup link switching is to switch data transmission to a backup transmission path; load distribution adjustment is to redistribute the workload of each device;

[0020] If the system status fails to recover to a reasonable range after multiple actions are performed consecutively, it will automatically escalate to a Level 2 response.

[0021] Preferably, the triggering conditions and execution content of the second-level response (remote intervention) in the multi-level progressive response mechanism are as follows:

[0022] When operational indicators exceed normal ranges, anomaly probabilities are within a preset moderate range, the slope of the fault development trend curve is 0.02-0.05, and there are associated fault risks, an intervention decision package is pushed to the remote operation and maintenance platform. This package includes fault cause analysis, recommended solutions, expected effect simulation, and alternative solutions. Fault cause analysis is an analysis of the causes of the fault and their interrelationships. Recommended solutions are specific solutions proposed for the fault. Expected effect simulation simulates the possible effects after the solution is implemented. Alternative solutions are alternatives when the recommended solution is not feasible.

[0023] It supports maintenance personnel to configure parameters and provide operation guidance through a remote operation interface. The execution progress is fed back in real time during the operation. If the execution effect does not meet expectations or the fault shows signs of aggravation, it will automatically escalate to a level three response.

[0024] Preferably, the triggering conditions and execution content of the third-level response (on-site emergency repair) in the multi-level progressive response mechanism are as follows:

[0025] When maintenance indicators exceed normal ranges, the probability of anomalies is in a preset high-level range, the slope of the fault development trend curve is 0.05-0.1, and there is associated fault risk, the audible and visual alarm device will be activated immediately. Simultaneously, a field handling plan will be pushed out, including the specific location of the fault, the safety isolation procedure, the required spare parts list, the emergency repair route plan, and the personnel qualification requirements. The specific location of the fault is the specific equipment and part where the fault occurred; the safety isolation procedure is the operational steps for isolating the faulty area from the normal area; the required spare parts list is the name and quantity of the parts that need to be replaced to repair the fault; the emergency repair route plan is the optimal route from the current location of the emergency repair personnel to the fault location; and the personnel qualification requirements are the skills and qualifications that the personnel participating in the emergency repair must possess.

[0026] It also links with the production scheduling system to adjust the production load in advance. If the problem is not resolved after the repair exceeds the preset time, or if the scope of the fault shows an expanding trend, it will automatically escalate to a level four response.

[0027] Preferably, the triggering conditions and execution content of the fourth-level response (cross-system collaborative emergency response) in the multi-level progressive response mechanism are as follows:

[0028] When operational indicators significantly exceed normal ranges, the probability of anomalies is at a preset extremely high level, the slope of the fault development trend curve is greater than 0.1, and there are associated fault risks, a cross-system emergency command mode is activated. This mode pushes a collaborative emergency plan that includes multi-system collaborative isolation strategies, emergency resource allocation plans, production plan adjustment measures, and emergency team dispatch instructions. The multi-system collaborative isolation strategy is a method for multiple systems to cooperate in fault isolation; the emergency resource allocation plan is an arrangement for the emergency allocation of human and material resources; the production plan adjustment measures are modifications to the production plan to address the fault; and the emergency team dispatch instructions are mobilization orders for the teams involved in emergency response. Simultaneously, a cross-regional emergency communication line is established to continuously monitor the progress of fault handling until the fault is completely resolved and the system returns to normal operation.

[0029] Preferably, the construction and updating process of a feature learning framework with cross-domain knowledge application capabilities includes:

[0030] Historical operation and maintenance data of industrial control systems from multiple industries are collected. The data for each industry includes sufficient normal operation records and fault handling records, and covers data related to new types of faults that have appeared in recent years. Normal operation records are data records of various aspects when the system is working normally. Fault handling records are data records of system faults and the handling process. Data related to new types of faults are information related to faults that have not occurred in the past.

[0031] Historical operation and maintenance data is standardized to generate a feature sample set that can be used for model training; the standardization process is to convert feature data of different formats and scales into a uniform format and scale.

[0032] An analytical model with a multi-layered processing structure is constructed, and the model is trained using a feature sample set. By continuously optimizing the model parameters, the model's accuracy in identifying fault types reaches the preset standard.

[0033] Preferably, the weighted fusion mechanism based on causal reasoning and scenario adaptation includes the following steps:

[0034] The system acquires key parameters such as the importance of equipment, data transmission priority, command response timeliness requirements, and the degree of environmental impact. Equipment importance is the level of importance set according to the role of the equipment in the system. Data transmission priority is the order in which different data are transmitted. Command response timeliness requirements are the requirements for the speed of response to control commands. The degree of environmental impact is the magnitude of the impact of environmental factors on system operation.

[0035] The basic weights of equipment status characteristics are calculated based on the importance of the equipment; the dynamic adjustment coefficients of transmission characteristics are calculated by combining transmission priority and real-time transmission delay; the timeliness correction coefficients of instruction execution characteristics are calculated based on instruction response timeliness requirements and historical response deviations; the interference compensation coefficients of environmental interference characteristics are calculated based on the degree of environmental impact and interference duration; historical response deviation is the difference between historical instruction response time and expected time; interference duration is the duration for which environmental factors have an adverse impact on the system.

[0036] The basic weights, dynamic adjustment coefficients, time-based correction coefficients, and interference compensation coefficients mentioned above are normalized to obtain the final weight allocation result; the normalization process converts each coefficient into a value whose sum is 1.

[0037] Preferably, the workflow of the anomaly detection model includes:

[0038] A pattern recognition benchmark library is constructed, comprising a normal pattern library and a fault pattern library. The normal pattern library is a set of characteristic patterns when the system is running normally; the fault pattern library is a set of characteristic patterns when the system malfunctions. The fault pattern library is divided into three levels according to the scope of impact: local fault, regional fault, and system-level fault. Local fault is a fault that only affects a single device or a small area; regional fault is a fault that affects multiple devices within a certain area; and system-level fault is a fault that affects the operation of the entire system.

[0039] The similarity between the comprehensive evaluation result and the features of various patterns in the benchmark library is calculated. The pattern with the highest similarity is identified as the recognition result, and the anomaly probability of this pattern is calculated.

[0040] When the identification result is a fault mode, the corresponding fault type, possible impact range assessment, and preliminary response plan are output simultaneously.

[0041] Preferably, the self-evolutionary learning process includes:

[0042] Set a fixed learning cycle, and collect sufficient execution result data of automatic decision-making and manual intervention decision-making in each cycle. This data covers the improvement of system status after decision implementation, the time spent on execution and the resource consumption.

[0043] The parameters to be optimized are selected by combining specific strategies with knowledge extraction. Among them, the parameters used to balance the exploration of new strategies and the application of mature strategies are dynamically adjusted based on the system stability score and the decision success rate. The system stability score is a score that measures the degree of stable operation of the system. The decision success rate is the proportion of the number of decisions that successfully solve problems out of the total number of decisions.

[0044] By analyzing the correlation between decision results and changes in system energy consumption, fault interval duration, and production efficiency fluctuations, the criteria used to evaluate the merits of decisions are updated, and an incentive mechanism is introduced to encourage the experimentation with new fault handling strategies. Changes in system energy consumption refer to the increase or decrease in system energy consumption after the decision is implemented; fault interval duration is the length of time between two faults; and production efficiency fluctuations refer to the changes in production efficiency after the decision is implemented.

[0045] When the improvement of the decision evaluation criteria exceeds the preset value for multiple consecutive cycles, the current optimization parameters are determined as the baseline configuration, while the baseline configurations of the previous two cycles are retained as alternatives.

[0046] A DCS operation and maintenance system based on integrated dynamic weight allocation and fault chain simulation, comprising:

[0047] The information collection unit is used to acquire system operation information, correlated production data, and cross-system interaction information through distributed sensing devices, and to perform time stamping and integrity verification on the information. Distributed sensing devices are devices distributed throughout the system for sensing data; time stamping records the specific time of information collection; integrity verification checks whether the information is complete. This unit includes multiple types of interfaces for connecting various sensing devices, a preprocessing subunit for processing raw information, and an encrypted transmission subunit for encrypting and transmitting the processed information. The preprocessing subunit is used to filter noise, fill in missing values, and convert the format of the collected raw information; the encrypted transmission subunit encrypts the preprocessed information using a specific encryption method and transmits it to the feature learning module through a dedicated communication path.

[0048] The feature learning module includes a multi-faceted feature extraction tool for extracting features from data and a tool optimization unit that periodically updates feature extraction parameters based on newly generated operational information.

[0049] The fusion unit has built-in dynamic weight allocation rules, which are used to perform weighted fusion of various features to generate a comprehensive evaluation result;

[0050] The decision center includes an anomaly pattern recognition tool for identifying abnormal patterns in the system, a fault level assessment unit for evaluating the severity of faults, and a decision suggestion generation unit for generating operation and maintenance decision suggestions. It is used to output comprehensive operation and maintenance indicators, anomaly probabilities, and corresponding handling solutions.

[0051] The response execution system executes corresponding level operation and maintenance operations based on the results output by the decision center. It includes an automatic adjustment module for automatically executing adjustment operations, a remote collaboration module to support remote collaborative operation and maintenance, and an emergency response module to deal with emergencies.

[0052] The optimization module is used to dynamically adjust operating parameters during normal system operation to achieve continuous optimization of system performance.

[0053] This invention provides a DCS operation and maintenance method based on the fusion of dynamic weight allocation and fault chain inference, which has the following beneficial effects:

[0054] 1. By using a multi-source data fusion module to link data from the DCS system, related production systems, and cross-plant collaborative systems, and then adjusting through spatiotemporal alignment and multi-dimensional verification, fragmented data such as equipment operation data, production load fluctuations, and cross-plant interaction information are woven into a complete data network. This all-domain perception capability not only solves the one-sidedness of traditional single-data judgment, such as the easy misjudgment of environmental interference by relying solely on equipment temperature data, but also captures hidden cross-system related faults. Secondly, when multi-source data corroborate each other, the accuracy of fault identification is no longer a simple superposition of individual data judgments, but rather a chain of evidence formed through logical consistency verification, which greatly reduces the misjudgment rate of suspected faults.

[0055] 2. The cross-domain knowledge transfer technology module breaks down industry and scenario barriers, allowing the operation and maintenance experience of reactors in the chemical industry to provide a reference for boiler systems in the power industry, and the clean environment control logic in the pharmaceutical industry to empower production workshops in the food industry. This knowledge flow is not a simple replication of experience, but a migration of application scenarios through the dynamic adaptation mechanism of the feature learning framework. When faced with new types of faults, the system does not need to accumulate experience from scratch, but can quickly call similar cases in the cross-domain knowledge base and generate solutions by combining them with local scenario features. This significantly improves the response speed to new faults compared to the traditional model that relies on internal experts, and its processing capacity grows exponentially as the knowledge base continues to expand.

[0056] 3. The dynamic weight allocation mechanism avoids the constraints of traditional fixed thresholds and can adjust the weight of evaluation dimensions in real time according to the importance of equipment and the priority of production scenarios. During high-load production, the weight of production-related features is automatically increased to ensure that operation and maintenance decisions prioritize production capacity. During equipment start-up and shutdown, the weight of environmental interference features is increased to avoid misjudging faults due to environmental fluctuations. This dynamic adjustment is not a simple increase or decrease of the weight of each dimension, but rather establishes a correlation model through causal reasoning, prioritizing the feature changes of core equipment, which avoids resource waste and ensures the stability of key links.

[0057] 4. The combination of a four-level response mechanism and fault trend prediction constructs a full-chain prevention and control system from prevention to handling; Level 1 automatic repair intervenes quickly in the nascent stage of a fault to prevent small problems from escalating; Level 2 remote intervention shortens the response cycle through human-machine collaboration; Level 3 on-site emergency repair links production scheduling to reduce downtime losses; and Level 4 cross-system emergency response constructs a comprehensive defense line. This tiered response is not a simple superposition of steps. When the system predicts that a fault may escalate within 120 minutes, it will initiate resource preparation for a higher-level response in advance, reducing the spare parts allocation time for on-site emergency repair by more than half and significantly reducing the overall impact of the fault compared to the traditional passive response mode.

[0058] 5. The self-evolving learning module enables the system to grow on its own. It does not simply record operation and maintenance data, but rather uses reinforcement learning to discover correlation patterns so that the system can automatically increase its priority. When changes in the production scenario cause the original strategy to fail, it will quickly adapt to the new rules. This combination of self-evolution and multi-source data and cross-domain knowledge enables the system's operation and maintenance efficiency to continuously improve over time, realizing the long-term value of becoming smarter the more it is used.

[0059] In summary, the various technical modules of this invention achieve comprehensive efficiency far exceeding the sum of individual technologies through deep collaboration in data exchange, knowledge flow, decision linkage, and capability iteration. This not only solves the problems of one-sidedness and lag in traditional operation and maintenance, but also constructs a dynamic operation and maintenance system that can adapt to the increasing complexity of industrial scenarios and the trend of intelligent production, providing comprehensive and adaptive intelligent protection for the stable operation of distributed control systems. Attached Figure Description

[0060] Figure 1 This is a system flowchart of an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0062] Example: Please refer to Figure 1 This embodiment provides a DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference. The operation and maintenance method includes the following steps:

[0063] S1. Acquire real-time operation and maintenance information of the distributed control system, key operating data of the associated production system, and interactive information of the cross-plant collaborative system. Real-time operation and maintenance information includes equipment operating values, i.e., real-time records of specific parameters such as temperature, pressure, and speed during equipment operation; communication link status values, i.e., data transmission speed, number of lost data, and link connection time; control command response values, i.e., the time interval from command issuance to execution and the degree of command completion; environmental impact values, i.e., the temperature, humidity, and dust content of the environment in which the equipment is located; key operating data of the associated production system includes production load fluctuation values, i.e., the amount of change in production load per unit time; raw material ratio parameters, i.e., the proportion settings of various raw materials in the production process; product quality inspection values, i.e., the quantitative results obtained after testing the product according to preset standards; interactive information of the cross-plant collaborative system includes data transmission delay time, i.e., the time difference from data transmission to reception; command collaboration completion rate, i.e., the proportion of successfully executed collaborative commands to the total number; and resource scheduling response speed, i.e., the time from the issuance of a resource scheduling command to the start of resource allocation.

[0064] S2. Preprocess the acquired real-time operation and maintenance information, key operational data, and interactive information. Preprocessing includes data cleaning, which removes obviously erroneous data such as values ​​that are outside the reasonable range; data standardization, which converts data of different magnitudes to the same numerical range; and data completion, which fills in missing data based on reasonable trends of change over time to generate preprocessed multi-source data.

[0065] S3. Perform spatiotemporal correlation alignment and multi-dimensional verification on the preprocessed multi-source data to generate a multi-source fusion dataset. Spatiotemporal correlation alignment ensures that data collected from different sources and at different times are matched in time and space. Multi-dimensional verification includes data timeliness verification, which checks whether the data is within the valid time range; logical consistency verification, which verifies whether the logical relationship between different data is reasonable, for example, the corresponding operating parameters of the heat dissipation system should change accordingly when the equipment temperature rises; and cross-boundary correlation verification, which confirms whether the correlation between data in different systems conforms to normal rules.

[0066] S4. Input the multi-source fusion dataset into a feature learning framework with cross-domain knowledge transfer and real-time knowledge update capabilities. This framework absorbs the operation and maintenance experience of distributed control systems in different industries through a dynamic adaptation mechanism, such as the experience of chemical and power industries in handling similar equipment failures. It can also autonomously update feature extraction rules according to newly emerging failure types and output equipment health feature sequences, link transmission reliability feature sequences, command execution performance feature sequences, environmental interference intensity feature sequences, production association feature sequences, and cross-system collaboration feature sequences. The equipment health feature sequence is a series of feature values ​​reflecting the overall health status of the equipment, arranged in chronological order. The link transmission reliability feature sequence is a series of feature values ​​reflecting the stability of communication link transmission, arranged in chronological order. The command execution performance feature sequence is a series of feature values ​​reflecting the effect of control command execution, arranged in chronological order. The environmental interference intensity feature sequence is a series of feature values ​​reflecting the degree of environmental impact on the system, arranged in chronological order. The production association feature sequence is a series of feature values ​​reflecting the degree of association between the production system and the distributed control system, arranged in chronological order. The cross-system collaboration feature sequence is a series of feature values ​​reflecting the collaborative effect of cross-plant system, arranged in chronological order.

[0067] S5. A weight allocation mechanism based on causal reasoning and dynamic scenario adaptation is adopted. The six types of feature sequences are weighted and fused according to the causal correlation strength between each feature sequence and system failure and the priority requirements of the current production scenario to generate a comprehensive operation and maintenance evaluation matrix. The causal correlation strength between each feature sequence and system failure is determined by analyzing historical data to determine the probability of failure caused by changes in each feature sequence. The priority requirements of the current production scenario, such as the relatively high priority of production-related feature sequences during high-load production, are considered. The comprehensive operation and maintenance evaluation matrix is ​​a comprehensive evaluation result of the system operation and maintenance status presented in matrix form. Each element in the matrix corresponds to the evaluation value of a specific dimension.

[0068] S6. Input the comprehensive operation and maintenance assessment matrix into the anomaly identification model that integrates fault evolution trajectory prediction and cascading fault deduction. This model analyzes the historical fault development process and the correlation between faults, and outputs the current comprehensive operation and maintenance indicators, anomaly probability values, fault development trend curves within the next 120 minutes, and possible associated fault chains. The current comprehensive operation and maintenance indicators are overall quantitative indicators of the current operation and maintenance status of the system. The anomaly probability value is the probability of the system experiencing an anomaly, expressed as a value between 0 and 100%. The fault development trend curve within the next 120 minutes is a curve plotted with time on the horizontal axis and fault severity on the vertical axis, used to show the possible development of the fault. The possible associated fault chains are the sequence of other faults that may be triggered in sequence after a certain fault occurs.

[0069] S7. Based on whether the comprehensive operation and maintenance indicators are within the preset normal range, whether the abnormal probability value is lower than the threshold, the slope change rate of the fault development trend curve, and the impact range of the associated fault chain, an intelligent operation and maintenance response mechanism with four-level progressive decision suggestions is triggered. The preset normal range is a reasonable interval set according to system design standards and historical normal operation data. The threshold is the upper limit of the abnormal probability set according to system safety requirements, such as 10%. The slope change rate of the fault development trend curve is the rate of change of the slope of the curve, reflecting the change in the speed of fault development. The impact range of the associated fault chain is the number of devices, areas, or systems that may be involved in the associated fault. The four-level progressive decision suggestions include Level 1 response (automatic repair), Level 2 response (remote intervention), Level 3 response (on-site emergency repair), and Level 4 response (cross-system collaborative emergency response).

[0070] When the system is in a stable operating state, the self-evolutionary learning mode is activated. Through a combination of reinforcement learning and transfer learning, the feature extraction weights, anomaly detection thresholds, decision suggestion generation rules, and cross-system collaboration strategies are continuously optimized. Feature extraction weights represent the importance of various features in the evaluation; anomaly detection thresholds are the critical values ​​for determining whether something is an anomaly; decision suggestion generation rules are the logic and basis for generating decision suggestions; and cross-system collaboration strategies are the ways and methods for different systems to work together.

[0071] As shown in steps S1-S7 above, the system operation and maintenance can be automated and precise through the synergy of multi-dimensional data processing and intelligent decision-making.

[0072] The data acquisition process clearly covers three types of information:

[0073] First, there is the real-time operation and maintenance information of the distributed control system, which specifically includes equipment operating values, which are records of directly measurable parameters such as temperature, pressure, and speed during equipment operation; communication link status values, which are status parameters such as data transmission speed, amount of lost data, and connection duration; control command response values, which are response parameters such as the time interval from issuance to execution of control commands and the percentage of execution completed; and environmental impact values, which are environmental parameters such as temperature, humidity, and dust content of the environment in which the equipment is located.

[0074] Second, key operational data related to the production system, including production load fluctuation values, which are the changes in production load per unit time; raw material ratio parameters, which are the ratio settings of various raw materials during the production process; and product quality test values, which are the quantitative results of products tested according to preset quality standards.

[0075] Thirdly, the interactive information of the cross-plant collaborative system includes data transmission delay, which is the time difference between sending and receiving data; instruction collaboration completion rate, which is the proportion of successfully executed collaborative instructions to the total number of instructions; and resource scheduling response speed, which is the time from the issuance of a resource scheduling instruction to the start of resource allocation.

[0076] The data preprocessing stage includes three distinct operations: data cleaning removes erroneous values ​​that are clearly outside the reasonable range, such as sudden temperature spikes that far exceed physical limits; data standardization converts parameters of different magnitudes to the same numerical range (e.g., 0-1), using the difference between the actual and minimum values ​​of a parameter, divided by the difference between the maximum and minimum values, i.e., (actual value - minimum value) ÷ (maximum value - minimum value), ensuring data comparability; and data completion fills in missing parameter values ​​based on reasonable trends in their changes over time. For example, if pressure data is missing for a certain period, the rate of pressure change at adjacent times is calculated first, and then the pressure value at the missing time is estimated based on that rate of change.

[0077] The multi-source data fusion process is achieved through two core operations: spatiotemporal correlation alignment, which matches data from different sources and at different collection times according to a unified timestamp and spatial location to ensure the temporal and spatial consistency of the data; and multi-dimensional verification, which includes data timeliness verification, retaining only valid data within a set time range, such as only retaining real-time data within 1 hour; logical consistency verification, verifying whether changes in correlation parameters conform to physical laws, such as the flow rate of the heat dissipation system should increase synchronously when the equipment temperature rises, and if the flow rate does not change when the temperature rises by 5°C, it is judged as logical inconsistency; and cross-boundary correlation verification, confirming whether the correlation between data in different systems conforms to production logic, such as when component A in the raw material ratio increases by 10%, component A in the product should increase by 8%-12%, and if it exceeds this range, it is judged as correlation anomaly, ultimately forming a complete multi-source fusion dataset.

[0078] The feature learning process is implemented through a framework with cross-domain knowledge transfer capabilities: this framework can absorb the operation and maintenance experience of different industries such as chemical and power, such as transferring the temperature control experience of chemical reactors to power boiler systems, and can automatically update the feature extraction rules when new faults occur. The six output feature sequences are all sets of feature values ​​arranged in chronological order: Equipment health feature sequence reflects the overall operating status of the equipment, and its feature value is calculated as the weighted sum of the normal rates of various equipment parameters, such as temperature normal rate × 0.4 + pressure normal rate × 0.3 + speed normal rate × 0.3; Link transmission reliability feature sequence reflects the stability of the communication link, and its feature value is (1 - packet loss rate) × 0.6 + connection duration ÷ total monitoring time × 0.4; Command execution efficiency feature sequence reflects the execution effect of control commands, and its feature value is (1 - command delay rate) × 0.5 + command completion rate × 0.5; Environmental interference intensity feature sequence reflects the degree of influence of the environment on the system, and its feature value is the duration of environmental parameter exceedance ÷ total monitoring time; Production correlation feature sequence reflects the closeness of the correlation between the system and the production process, and its feature value is the matching degree between production load fluctuations and equipment parameter fluctuations; Cross-system collaboration feature sequence reflects the collaboration efficiency of cross-plant systems, and its feature value is command collaboration completion rate × 0.6 + (1 - data transmission delay rate) × 0.4.

[0079] The dynamic weight allocation process is based on causal reasoning and scenario adaptation: It calculates the probability of each feature sequence causing a failure by analyzing historical data and the current production scenario's priority (e.g., production-related features have higher weights during high-load production). The six feature sequences are then weighted and fused. The fusion result is a comprehensive operation and maintenance evaluation matrix, where each element corresponds to a specific evaluation value for a particular dimension. The calculation process involves summing the products of the feature values ​​of each feature sequence and their corresponding weights.

[0080] The anomaly identification and response process is achieved through model prediction and tiered response: The anomaly identification model includes a normal operation mode library and a fault mode library categorized by impact range. By calculating the similarity between the current comprehensive operation and maintenance assessment matrix and various modes in the mode library (the similarity is calculated as the complement of the average of the absolute differences between corresponding elements of the two matrices, i.e., 1 - the absolute value of the average difference), the fault type is determined and an anomaly probability of 0-100% is output. Simultaneously, the model predicts the fault development trend within the next 120 minutes, displayed as a time-severity curve, and the potential associated fault chains, such as pump failure leading to raw material delivery interruption, and then production line shutdown. Based on whether the comprehensive operation and maintenance indicators are within the preset normal range, whether the anomaly probability is below a threshold (e.g., 10%), the fault trend slope, and the impact range of associated faults, a four-level progressive response is triggered: automatic repair, remote intervention, on-site emergency repair, and cross-system collaborative emergency response.

[0081] The self-evolutionary learning process ensures continuous improvement in system capabilities: When the system is running stably, at least 200 decision execution results are collected every 12 hours, including decision effectiveness, execution time, and resource consumption. Reinforcement learning is used to optimize feature extraction weights (the importance of various features), anomaly identification thresholds (the critical values ​​for judging anomalies), decision generation rules (the logic for generating response plans), and cross-system collaboration strategies (the way multiple systems collaborate), so that the system's operation and maintenance capabilities are continuously optimized over time.

[0082] In a specific implementation process, the triggering conditions and execution content of Level 1 response (automatic repair) are as follows: When the comprehensive operation and maintenance indicators fluctuate at the edge of the normal range, the probability of anomalies is 3%-8%, the slope of the fault development trend curve is less than 0.02, and there is no associated fault risk, the system automatically executes 20 preset repair actions, including parameter fine-tuning, redundant link switching, and load balancing adjustment, and records the quantitative value of the effect of each action and the execution time; parameter fine-tuning is, for example, fine-tuning the set value of the equipment operating temperature; redundant link switching is switching data transmission to the backup link; load balancing adjustment is adjusting the load ratio of each device; the quantitative value of the effect is, for example, the degree of improvement in the stability of equipment operation after parameter adjustment; if the effect does not meet expectations after the execution of 3 consecutive actions, such as the equipment operating parameters not returning to a reasonable range, the system automatically upgrades to Level 2 response.

[0083] An autonomous handling mechanism for minor anomalies is defined, which is suitable for scenarios with low failure risk and no possibility of propagation.

[0084] The triggering conditions are clearly defined as four items: the comprehensive operation and maintenance indicators fluctuate at the edge of the preset normal range without exceeding the normal range, the abnormality probability value is in the range of 3%-8%, the slope of the fault development trend curve is less than 0.02, the slope is calculated as the difference in fault severity between adjacent moments ÷ time interval, and there is no related fault risk, which will not cause abnormalities in other equipment or systems.

[0085] The execution includes: the system automatically performing 20 preset repair actions, specifically covering parameter fine-tuning (e.g., adjusting the device operating temperature setpoint within a small range, with an adjustment margin of ±5% of the normal range), redundant link switching (switching the data transmission path to a backup link, with a switching time not exceeding 1 second), and load balancing adjustment (redistributing the task proportions of each device so that the load difference between devices does not exceed 10%). The effect of each action, such as the improvement in device stability after parameter adjustment, is calculated as the fluctuation range of the adjusted parameters ÷ the fluctuation range of the parameters before adjustment, and the execution time is recorded. If the system status still does not return to the normal range after three consecutive actions are executed, such as if the device operating parameters have not recovered to a reasonable range, a secondary response is automatically triggered.

[0086] In a specific implementation process, the triggering conditions and execution content of Level 2 response (remote intervention) are as follows: When the comprehensive operation and maintenance indicators exceed the normal range by 10%-30%, the abnormal probability value is 8%-20%, the slope of the fault development trend curve is 0.02-0.05, and the impact range of the associated fault chain is small, an intervention decision package containing fault causal analysis, recommended repair solutions, expected effect simulation, and alternative solutions is pushed to the remote operation and maintenance platform. This supports operation and maintenance personnel to remotely configure parameters and provide operation guidance through an augmented reality interface. Fault causal analysis is an analysis of the causes of the fault and their interrelationships. Recommended repair solutions are specific solutions proposed for the fault. Expected effect simulation is a simulation of the possible effects after the repair solution is implemented. During the operation, the execution progress and effects are fed back to the operation and maintenance personnel in real time. If the execution effect does not meet expectations or the fault worsens, it is automatically upgraded to Level 3 response.

[0087] A human-machine collaborative processing mechanism for the system under moderate abnormal scenarios is defined, which is applicable to scenarios that require remote human intervention but do not require on-site emergency repair.

[0088] The triggering conditions are clearly defined as follows: the comprehensive operation and maintenance indicators exceed the preset normal range by 10%-30%, and the excess ratio is calculated as (actual value - upper limit of normal range) ÷ upper limit of normal range × 100%; the abnormal probability value is in the range of 8%-20%; the slope of the fault development trend curve is in the range of 0.02-0.05, and the slope is calculated as the difference in fault severity between adjacent moments ÷ time interval; and the impact range of the associated fault chain is small, affecting only a single device or local area.

[0089] The execution includes: pushing an intervention decision package to the remote operation and maintenance platform. This package contains a causal analysis of the fault, identifying the direct cause and related factors leading to the fault, such as filter clogging causing pressure drop, with the degree of clogging being positively correlated with the pressure drop; recommended repair solutions: specific steps for resolving the fault, such as remotely initiating a backwashing procedure for 5 minutes; simulated expected results: predicting the system state after the solution is implemented using historical data, such as the pressure recovering to 0.32 MPa, with the recovery rate calculated as the pressure before backwashing × 1.2; and at least one alternative solution; and supports operation and maintenance. Through an augmented reality interface, personnel can intuitively view equipment status and operation instructions; remotely configure parameters and provide operation guidance; during operation, the system provides real-time feedback to maintenance personnel on execution progress, such as when the backwashing procedure has been 50% completed (the progress is calculated as executed time ÷ total planned time × 100%); and on the effect, such as when the current pressure is 0.29 MPa, the recovery rate is calculated as (current pressure - minimum pressure) ÷ (normal pressure - minimum pressure) × 100%; if the execution effect does not meet expectations, such as when the pressure does not rise as predicted; or if the fault worsens, such as when the pressure continues to drop, a level three response is automatically triggered.

[0090] In a specific implementation process, the triggering conditions and execution content of Level 3 response (on-site emergency repair) are as follows: When the comprehensive operation and maintenance indicators exceed the normal range by more than 30%, the abnormal probability value exceeds 20%, the slope of the fault development trend curve is 0.05-0.1, and the impact range of the associated fault chain is moderate, an audible and visual alarm is immediately activated, and an on-site handling plan including precise fault location, safety isolation steps, spare parts replacement list, emergency repair route planning, and personnel qualification requirements is pushed out. Simultaneously, linkage with the production scheduling system is triggered to adjust the production load in advance. Precise fault location is to determine the specific equipment and location where the fault occurred. Safety isolation steps are the operation procedures for isolating the fault area from the normal area. Spare parts replacement list is the name, model, and quantity of the parts that need to be replaced. Emergency repair route planning is the optimal route from the current location of the emergency repair personnel to the fault location. Personnel qualification requirements are the skills and qualifications that the personnel participating in the emergency repair must possess. If the emergency repair is not resolved after a preset time, such as 1 hour, or the scope of the fault impact expands, it will automatically escalate to Level 4 response.

[0091] The system defines an on-site handling mechanism for severe abnormal scenarios, applicable to scenarios requiring on-site handling by professional personnel and potentially affecting production.

[0092] The triggering conditions are clearly defined as four items: the comprehensive operation and maintenance indicators exceed the preset normal range by more than 30%, the excess ratio is calculated as (actual value - upper limit of normal range) ÷ upper limit of normal range × 100%; the abnormal probability value is above 20%; the slope of the fault development trend curve is in the range of 0.05-0.1, the slope is calculated as the difference in fault severity between adjacent moments ÷ time interval; and the impact range of the associated fault chain is moderate, which may affect multiple production lines or multiple devices.

[0093] The execution includes: immediately activating the audible and visual alarms to alert on-site personnel via sound and light signals, with the alarm frequency increasing as the severity of the fault increases; sending an on-site handling plan to the repair personnel, which includes precise fault location, specifying the equipment model and location, such as the temperature sensor interface of reactor A, with a positioning accuracy of ±0.5 meters; safety isolation steps, listing operations such as cutting off power and isolating the area in sequence, such as 1. closing the reactor feed valve, with a closing time ≤3 seconds; 2. cutting off the heating power supply, with a power-off time ≤5 seconds; and a spare parts replacement list, clearly specifying the name, model, and quantity of the parts to be replaced, such as the temperature sensor, model, etc. PT100, Quantity 1, Emergency Repair Path Planning: Calculates the shortest safe route from the current location of the repair personnel to the fault point. The path length is the sum of the distances of each segment. Personnel qualification requirements: Certificates and skills required for participating repair personnel, such as holding a special equipment operation certificate and having ≥3 years of operating experience. Simultaneously, a signal is sent to the production scheduling system to adjust the production load in advance, such as reducing it from 100% to 80%. The adjustment range is determined based on the fault impact assessment to reduce the impact of the fault. If the repair time exceeds a preset threshold, such as if the problem is not resolved within 1 hour, or if the fault impact expands, such as affecting the entire production line from a single piece of equipment, a Level 4 response is automatically triggered.

[0094] In a specific implementation process, the triggering conditions and execution content of the Level 4 response (cross-system collaborative emergency response) are as follows: When the comprehensive operation and maintenance indicators exceed the normal range by more than 50%, the probability of anomalies exceeds 50%, the slope of the fault development trend curve is greater than 0.1, and the associated fault chain may affect the cross-plant collaborative system, the cross-system emergency command mode is activated. A collaborative emergency plan including a multi-system collaborative isolation scheme, an emergency resource allocation plan, a comprehensive production adjustment strategy, and emergency team dispatch instructions is pushed out. At the same time, a cross-plant emergency communication line is established. The multi-system collaborative isolation scheme is a strategy for multiple systems to cooperate in fault isolation. The emergency resource allocation plan is an arrangement for the emergency allocation of human, material, and other resources. The comprehensive production adjustment strategy is an overall adjustment of the production plan to cope with the fault. The emergency team dispatch instructions are the mobilization orders for the teams participating in the emergency response. The cross-plant emergency communication line is an independent communication line dedicated to cross-plant emergency communication. The progress of fault handling is continuously monitored until the fault is resolved and the system returns to normal operation.

[0095] A comprehensive emergency response mechanism for the system under severe abnormal scenarios is defined, which is applicable to scenarios that may affect cross-plant systems and are extremely risky.

[0096] The triggering conditions are clearly defined as four items: the comprehensive operation and maintenance indicators seriously exceed the preset normal range by more than 50%, the excess ratio is calculated as (actual value - upper limit of normal range) ÷ upper limit of normal range × 100%; the abnormal probability value is above 50%; the slope of the fault development trend curve is greater than 0.1, the slope is calculated as the difference in fault severity between adjacent moments ÷ time interval; and the associated fault chain may affect cross-plant collaborative systems, such as raw material supply and product transportation.

[0097] The execution includes: activating a cross-system emergency command mode, integrating monitoring and control permissions across various plants and systems, with the highest priority given to these permissions; pushing out a collaborative emergency plan, which includes multi-system collaborative isolation schemes, specifying the isolation operations for each system, such as Plant A closing the main raw material conveying valve with a closing time ≤ 10 seconds; Plant B activating the backup raw material tank with a switchover time ≤ 30 seconds; an emergency resource allocation plan, listing the manpower, equipment, spare parts, and other resources to be mobilized and their allocation routes, such as transporting two backup pumps from Plant C to Plant A, with a route length of 20 kilometers and an estimated transportation time of 30 minutes; and a comprehensive production adjustment strategy, targeting the production plans of each plant. Adjustments will be made, such as suspending production in Plant A and switching Plant B to emergency production mode, maintaining 30% of normal capacity. Emergency team dispatch instructions will be issued, clearly defining the tasks and arrival times of each emergency team. For example, the first repair team will arrive at Plant A within 30 minutes, carrying a tool list including 2 sets of wrenches and 5 gaskets. A dedicated emergency communication line will be established across plant areas, independent of the regular communication network, with a bandwidth of ≥100Mbps and a latency of ≤50ms, to ensure smooth emergency commands. The progress of fault handling will be continuously monitored. For example, if 80% of the pipeline repair is completed, the progress will be calculated as the length repaired ÷ the total fault length × 100%, until the fault is resolved and the system returns to normal operation.

[0098] In a specific implementation process, the construction and updating process of a feature learning framework with cross-domain knowledge transfer and real-time knowledge update capabilities includes:

[0099] S401. Collect historical operation and maintenance data of distributed control systems in four industries: chemical, power, metallurgy, and pharmaceutical. Each industry's data includes at least 8,000 normal operation records and 5,000 fault handling records, and covers records of new fault types that have appeared in the past five years. Normal operation records are data records of various aspects when the system is working normally. Fault handling records are data records of system faults and the handling process. New fault type records are relevant data of faults that have not occurred in the past.

[0100] S402. Perform feature standardization processing on historical operation and maintenance data to generate a standardized feature sample set. Feature standardization processing is to convert feature data of different formats and magnitudes into a unified format and magnitude. The standardized feature sample set is a set of feature data that has been standardized and can be used for model training.

[0101] S403. Construct a neural network model that includes convolutional layers, recurrent layers, and an attention mechanism. Convolutional layers are used to extract local features, i.e., features within a local range of data, such as the features of equipment operating parameters within a certain time period. Recurrent layers are used to capture temporal correlation features, i.e., the correlation features between data at different time points, such as the relationship between changes in equipment operating parameters over time. The attention mechanism is used to strengthen the weights of key features, i.e., to increase the influence of features that play an important role in system evaluation and fault diagnosis.

[0102] S404. Train the neural network model using a standardized feature sample set, and iteratively optimize the model parameters through the loss function until the model's accuracy in identifying fault labels reaches a preset standard. The loss function is a function that measures the difference between the model's prediction results and the actual results. The model parameters are adjustable variables in the model. The identification accuracy is the proportion of the number of fault types correctly identified by the model to the total number of identifications. The preset standard is, for example, above 95%.

[0103] As shown in steps S401-S404 above, the implementation method of the feature learning framework with cross-domain knowledge transfer and real-time update capabilities is defined to ensure that the system can adapt to different industries and new types of faults.

[0104] The data collection is defined as follows: collecting historical operation and maintenance data of distributed control systems in four industries: chemical, power, metallurgy, and pharmaceutical. The data for each industry must include at least 8,000 records of various parameters during normal system operation and 5,000 complete records of system failures and handling processes. It must also cover relevant data on new types of failures that have not been recorded in the past five years, such as new communication failures of intelligent devices.

[0105] The feature standardization process is defined as follows: convert feature data in different formats, such as temperature units ℃ and ℉, and different magnitudes, such as pressure 0-1MPa and 0-10MPa, into a unified format and magnitude, such as converting them all into values ​​in the 0-1 range. The conversion formula is (feature value - industry minimum value) ÷ (industry maximum value - industry minimum value), generating a standardized feature sample set that can be directly used for model training.

[0106] The model construction is clearly defined as follows: a neural network model containing convolutional layers, recurrent layers, and an attention mechanism. Convolutional layers are used to extract features within local ranges of data, such as temperature fluctuation patterns over a 10-minute period, calculating local feature values ​​through sliding 3×3 convolutional kernels. Recurrent layers are used to capture correlation features between data points at different times, such as the temporal relationship of a simultaneous decrease in flow rate after a 10-minute pressure drop, calculating temporal correlation values ​​through Long Short-Term Memory (LSTM) units. The attention mechanism is used to increase the weights of features that play a crucial role in system evaluation and fault diagnosis; the weight is calculated as the mutual information value between the feature and the fault label divided by the sum of the mutual information values ​​of all features.

[0107] Model training and updating are defined as follows: The neural network model is trained using a standardized feature sample set. The loss function, mean squared error loss, is calculated as the average of the squared differences between the predicted fault probability and the actual fault label. Iterative adjustments are made to optimizable parameters in the model, such as neuron connection weights, with 1000 iterations and a learning rate of 0.001, until the model achieves a certain accuracy rate in fault type identification, representing the proportion of correctly identified faults to the total number of identified faults. This reaches a preset standard, such as 95%. When new faults appear, the model updates its parameters through incremental learning. New fault data accounts for 10% of the training set. After training, the accuracy rate for identifying existing faults remains above 92%, and the update time does not exceed 48 hours.

[0108] In a specific implementation process, the execution steps of the weight allocation mechanism based on causal reasoning and dynamic scenario adaptation include:

[0109] S501. Obtain the device importance level, link transmission priority, command response timeliness requirement, and environmental impact coefficient. The device importance level is a level set according to the device's role in the system, such as core device, important device, and general device. The link transmission priority is the priority order of communication links in data transmission, such as critical data transmission links having higher priority than ordinary data transmission links. The command response timeliness requirement is the requirement for the speed of response to control commands, such as emergency commands needing to be responded to within 1 second. The environmental impact coefficient is the degree of influence of environmental factors on system operation, such as a higher coefficient for environments with a greater impact on device operation.

[0110] S502. Calculate the basic weight coefficient of the equipment status feature vector according to the equipment importance level; the basic weight coefficient is a value that reflects the basic importance of the equipment status feature in the comprehensive evaluation. For example, the basic weight coefficient for core equipment is 0.3, for important equipment it is 0.2, and for general equipment it is 0.1.

[0111] S503. Calculate the dynamic adjustment coefficient of the link transmission feature vector based on the link transmission priority and real-time transmission delay. The dynamic adjustment coefficient is a value that adjusts the basic weight according to the actual transmission situation of the link. The dynamic adjustment coefficient is higher for links with high transmission priority and low real-time transmission delay, such as 1.2, and lower for links with low transmission priority and low real-time transmission delay, such as 0.8.

[0112] S504. Calculate the timeliness correction coefficient of the instruction execution feature vector based on the instruction response timeliness requirements and the historical response deviation rate; the historical response deviation rate is the ratio of the deviation between the historical instruction response time and the expected time; the timeliness correction coefficient is a value that corrects the basic weight based on the instruction response timeliness. Instructions with high timeliness requirements and low historical response deviation rates have higher timeliness correction coefficients, such as 1.1, and vice versa, they have lower timeliness correction coefficients, such as 0.9.

[0113] S505. Calculate the interference compensation coefficient of the environmental interference feature vector based on the environmental impact coefficient and the interference duration. The interference duration is the duration for which environmental factors have an adverse effect on the system. The interference compensation coefficient is a value used to compensate the basic weights based on the environmental interference situation. The interference compensation coefficient is higher when the environmental impact coefficient is large and the interference duration is long, for example, 1.3. Conversely, it is lower when the environmental impact coefficient is small, for example, 0.7.

[0114] S506. Normalize the basic weight coefficients, dynamic adjustment coefficients, time-based correction coefficients, and interference compensation coefficients to obtain the final weight allocation matrix. Normalization converts each coefficient into a value that sums to 1. The final weight allocation matrix is ​​a matrix representation of the final weight values ​​of each feature.

[0115] As shown in steps S501-S506 above, a weight calculation method based on causal reasoning and scenario adaptation is defined to ensure that the importance assessment of each feature is in line with actual needs.

[0116] The basic parameters are clearly defined as follows: Equipment importance level, categorized into core equipment, important equipment, and general equipment based on their role in the system (e.g., a reactor is a core equipment); Link transmission priority, categorized into critical links and ordinary links based on the importance of the transmitted data (e.g., a control command transmission link is a critical link); Command response time requirements, setting the response time based on the urgency of the command (e.g., an urgent command must respond within 1 second); and Environmental impact coefficient, set according to the degree of influence of environmental factors on the system (e.g., a high-temperature environmental coefficient of 0.8 and a normal-temperature coefficient of 0.3).

[0117] The coefficient calculation is as follows: The basic weight coefficient is set according to the importance level of the equipment: 0.3 for core equipment, 0.2 for important equipment, and 0.1 for general equipment; the dynamic adjustment coefficient is calculated based on the link transmission priority and real-time transmission delay, with the following formula: if it is a critical link and the delay is ≤100ms, then the coefficient = 1.2; if it is a critical link and the delay is >100ms, then the coefficient = 1.0; if it is a normal link and the delay is ≤100ms, then the coefficient = 0.9; if it is a normal link and the delay is >100ms, then the coefficient = 0.8; the timeliness correction coefficient is based on the instruction response timeliness requirements and the historical response deviation rate, historical response... The ratio of the difference between the required time and the expected time is calculated using the following formula: If the required time is ≤1 second and the deviation rate is ≤5%, then the coefficient = 1.1; if the required time is ≤1 second and the deviation rate is >5%, then the coefficient = 1.0; if the required time is >1 second and the deviation rate is ≤5%, then the coefficient = 1.0; if the required time is >1 second and the deviation rate is >5%, then the coefficient = 0.9. The interference compensation coefficient is calculated based on the environmental impact coefficient and the duration of interference, which is the duration for which environmental factors have an adverse impact on the system. The formula is: Coefficient = Environmental Impact Coefficient × (1 + Interference Duration ÷ 30 minutes), and the result is controlled between 0.7 and 1.3.

[0118] The normalization process is defined as follows: calculate the sum of the basic weight coefficient, dynamic adjustment coefficient, timeliness correction coefficient, and interference compensation coefficient, and set it as S. Then divide each coefficient by the sum S to obtain the final weight of each feature. For example, the weight of core equipment = 0.3 ÷ S, forming the final weight allocation matrix to ensure that the sum of all weights is 1.

[0119] In a specific implementation process, the workflow of the anomaly identification model that integrates fault evolution trajectory prediction and cascading fault deduction includes:

[0120] S601. Construct a pattern recognition benchmark library that includes a normal pattern library and a fault pattern library. The normal pattern library is a set of characteristic patterns when the system is running normally. The fault pattern library is a set of characteristic patterns when the system fails. The fault pattern library is divided into three levels according to the scope of the fault's impact: local fault, regional fault, and system-level fault. A local fault is a fault that only affects a single device or a small area. A regional fault is a fault that affects multiple devices within a certain area. A system-level fault is a fault that affects the operation of the entire system.

[0121] S602. Calculate the similarity between the integrated operation and maintenance feature matrix and each pattern feature in the pattern recognition benchmark library; the similarity is an index that measures the similarity between two feature vectors, with a value range of 0-1, and the closer to 1, the more similar they are.

[0122] S603. Take the pattern category with the highest similarity as the recognition result, and calculate the anomaly probability value of the pattern; the anomaly probability value is the probability that the pattern belongs to an abnormal situation.

[0123] When the identification result is a fault mode, the corresponding fault type, impact range assessment, and preliminary handling plan are output. Fault type includes, for example, equipment overload fault, link interruption fault, etc. Impact range assessment includes the number of devices and the size of the area that the fault may affect. The preliminary handling plan is the initial solution for the fault.

[0124] As shown in steps S601-S603 above, an identification method that integrates fault prediction and cascading fault analysis is defined to ensure the comprehensiveness and foresight of fault judgment.

[0125] The pattern library construction is defined as follows: a pattern recognition benchmark library is constructed, which includes a normal pattern library and a fault pattern library; the normal pattern library stores characteristic patterns when the system is running normally, such as a flow rate of 20-30L / min corresponding to a pulp concentration of 3%-5%, and the pattern characteristic value is the average value ± standard deviation of various parameters during normal operation; the fault pattern library is divided into local faults, which only affect a single device, such as sensor drift, and regional faults, which affect multiple devices, such as agitator faults, and system-level faults, which affect the entire system, such as main power supply faults, according to the scope of impact. The characteristic value of each fault mode is the typical value of various parameters when the fault occurs.

[0126] The similarity calculation is defined as follows: calculate the degree of similarity between the current integrated operation and maintenance feature matrix and the features of each mode in the mode library. The steps are as follows: 1. Calculate the absolute value of the difference between corresponding elements of the two matrices; 2. Calculate the average value of all absolute differences, denoted as M; 3. Similarity = 1 - M, where M takes the value 0-1, and the closer it is to 1, the more similar it is; for example, if the average absolute value of the difference between the current matrix and the stirrer fault mode matrix is ​​0.12, then the similarity = 0.88.

[0127] The identification results are clearly output as follows: the pattern category with the highest similarity is taken as the identification result, and the probability that the pattern belongs to an anomaly is calculated. The anomaly probability value is calculated as similarity × 100% (if it is a fault mode). When a fault mode is identified, the specific fault type is output, such as insufficient stirrer speed, impact range assessment, such as it may cause the paper thickness deviation to exceed 0.1mm in the next 2 hours, the deviation is calculated as current concentration deviation × 0.2 × time, and preliminary treatment plan, such as increasing the stirrer speed to 110% of the rated value, the adjustment range is determined according to the difference between the current speed and the rated value.

[0128] In a specific implementation process, the operation of the self-evolutionary learning model includes:

[0129] S701. Set a learning cycle of 12 hours. Collect more than 200 execution result data of automatic decision-making and manual intervention decision-making in each cycle. The execution result data includes decision effect, execution efficiency and resource consumption data. The decision effect is the degree of improvement of the system status after the decision is executed. The execution efficiency is the time from decision-making to execution completion. The resource consumption data is the manpower, material resources, energy and other resources consumed in the decision execution process.

[0130] S702. Select parameters to be optimized by combining specific strategies with knowledge extraction; among them, the parameters used to balance exploration and utilization are dynamically adjusted according to the system stability score and decision success rate; the system stability score is a score that measures the stability of the system's operation, ranging from 0 to 100 points; the decision success rate is the proportion of the number of decisions that successfully solve the problem to the total number of decisions; the higher the stability score and the higher the decision success rate, the smaller the value of this parameter, ranging from 0.05 to 0.4. The smaller the value of this parameter, the more the system tends to select parameters that have been verified to be effective for optimization.

[0131] S703. By calculating the correlation between the decision results and the system energy consumption reduction rate, fault interval extension rate, and production efficiency improvement rate, the function used to evaluate the merits of the decision is updated, and an exploration reward mechanism is introduced to encourage the attempt of new fault handling strategies. The system energy consumption reduction rate is the proportion of system energy consumption reduction after the decision is implemented; the fault interval extension rate is the proportion of time extension between two faults after the decision is implemented; the production efficiency improvement rate is the proportion of production efficiency improvement after the decision is implemented; the exploration reward mechanism is to reward the attempt of new and effective decision strategies.

[0132] When the improvement rate of the function used to evaluate the merits of the decision exceeds 8% for three consecutive cycles, the current optimization parameters are fixed as the baseline configuration, while the previous two baseline configurations are retained as alternatives. The improvement rate of the function used to evaluate the merits of the decision is the growth rate of the function in adjacent cycles. The baseline configuration is the default parameter setting of the system. The alternatives can be switched to when the current baseline configuration has problems.

[0133] As shown in steps S701-S703 above, a learning mechanism for system self-optimization is defined to ensure that operation and maintenance capabilities continue to improve over time.

[0134] The learning cycle and data collection are defined as follows: each learning cycle is set at 12 hours. In each cycle, at least 200 execution result data of automatic system decision-making and manual intervention decision-making are collected, including decision effect, such as the degree of improvement in system stability after the decision, calculated as the failure rate after the decision ÷ the failure rate before the decision; execution efficiency, such as the time from decision-making to completion, accurate to the second; and resource consumption, such as the electricity and manpower consumed in the execution of the decision, with electricity in kWh and manpower in man-hours.

[0135] The parameter optimization is defined as follows: A specific strategy is adopted: balancing the exploration of new strategies with the utilization of effective strategies; and knowledge extraction: extracting effective rules from historical decisions; these are combined to select the parameters to be optimized. The parameter used to balance exploration and utilization is denoted as ε. It is dynamically adjusted based on the system stability score (0-100 points, higher scores indicate greater stability) and the decision success rate (the proportion of decisions that successfully solve the problem out of the total number of decisions). The calculation formula is: ε = 0.4 - (stability score ÷ 100 + decision success rate) × 0.35, with the result controlled between 0.05 and 0.4. A smaller ε value indicates that the system is more inclined to select parameters that have been verified to be effective.

[0136] The function update and strategy fixation are defined as follows: The correlation between the decision result and the system energy consumption reduction rate is calculated as follows: the percentage reduction in energy consumption after the decision is calculated as (energy consumption before decision - energy consumption after decision) ÷ energy consumption before decision; the fault interval extension rate is calculated as the percentage increase in time between two faults after the decision is calculated as (interval after decision - interval before decision) ÷ interval before decision; the production efficiency improvement rate is calculated as the percentage increase in production efficiency after the decision is calculated as (efficiency after decision - efficiency before decision) ÷ efficiency before decision. The correlation is the weighted sum of the three indicators: energy consumption reduction rate × 0.3 + fault interval extension rate × 0.4 + production efficiency improvement rate × 0.3. This is used to update the function used to evaluate the merits of the decision. When the improvement rate of this function for three consecutive cycles, the percentage increase in adjacent cycles is calculated as (current cycle function value - previous cycle function value) ÷ previous cycle function value × 100%. If both exceed 8%, the current optimized parameters are determined as the baseline configuration and the system default parameters. The previous two baseline configurations are retained as alternatives to be used when the current configuration fails.

[0137] A DCS operation and maintenance system based on integrated dynamic weight allocation and fault chain inference, the system comprising:

[0138] The intelligent information collection unit is used to acquire real-time operation and maintenance information of the distributed control system through distributed sensing nodes, and to perform time stamping and integrity verification on the information. Distributed sensing nodes are devices located throughout the system for sensing data. Time stamping records the specific time of information acquisition. Integrity verification checks whether the information is complete and free of missing information. This unit includes multiple types of sensing interfaces, an information preprocessing subunit, and an information encryption and transmission subunit. The multiple types of sensing interfaces are used to connect temperature sensors, pressure sensors, flow sensors, vibration sensors, and environmental sensors. Temperature sensors are devices that measure temperature; pressure sensors are devices that measure pressure; flow sensors are devices that measure... The device includes a fluid flow meter; a vibration sensor measures equipment vibration; an environmental sensor measures environmental parameters; an information preprocessing subunit performs noise filtering, missing value supplementation, and format conversion on the collected raw information; noise filtering removes interference signals from the information; missing value supplementation appropriately fills in missing information; format conversion converts the information into a unified format; the information encryption and transmission subunit uses symmetric encryption to encrypt the preprocessed information and transmits it to the feature learning module through a dedicated communication path; symmetric encryption uses the same key for both encryption and decryption; a dedicated communication path is a communication line or channel specifically used for transmitting this information.

[0139] The feature learning module includes a variety of feature extraction tools and a tool optimization unit. The feature extraction tools are tools used to extract features from data. The tool optimization unit is a unit used to optimize the performance of the feature extraction tools. The tool optimization unit updates the feature extraction tool parameters regularly based on newly generated operation and maintenance information. The feature extraction tool parameters are adjustable variables in the feature extraction tools.

[0140] The dynamic fusion unit has built-in dynamic weight allocation rules, which are used to perform weighted fusion of various feature combinations and generate a comprehensive operation and maintenance feature set; the dynamic weight allocation rules are the rules for calculating the weight of each feature; the comprehensive operation and maintenance feature set is the feature set obtained after fusion, which is used to comprehensively evaluate the system operation and maintenance status;

[0141] The intelligent decision center includes an anomaly pattern recognition tool, a fault level assessment unit, and a decision suggestion generation unit. It is used to output comprehensive operation and maintenance indicators, anomaly probability values, and corresponding handling methods. The anomaly pattern recognition tool is used to identify abnormal patterns in the system. The fault level assessment unit is used to assess the severity level of faults. The decision suggestion generation unit is used to generate operation and maintenance decision suggestions.

[0142] The tiered response execution system executes corresponding level operation and maintenance operations based on the output results of the intelligent decision-making center. It includes an automatic adjustment module, a remote collaboration module, and an emergency response module. The automatic adjustment module is used to automatically execute adjustment operations; the remote collaboration module is used to support remote collaborative operation and maintenance; and the emergency response module is used to deal with emergency situations.

[0143] The adaptive optimization module is used to dynamically adjust operating parameters during normal system operation to achieve continuous performance optimization.

[0144] Implement the system modules and functions of each module to ensure the successful implementation of the above methods.

[0145] The intelligent information collection unit is defined as follows: It utilizes distributed sensing nodes, sensors, and data acquisition devices located throughout the system, with a sampling frequency ≥10Hz; it acquires real-time operation and maintenance information, simultaneously recording the specific time of information collection with a timestamp accurate to milliseconds; and checks the completeness of the information, ensuring no missing fields and a missing rate ≤0.1%. This unit includes multiple types of sensing interfaces, connecting to sensors for temperature, pressure, flow, vibration, and environment, supporting signal types such as 4-20mA and RS485. An information preprocessing subunit filters noise from the raw information, removing high-frequency interference signals using a moving average method with a window size of 5 sampling points; it fills in missing values ​​according to trends using linear interpolation; it performs format conversion, unifying the data format to JSON; and an information encryption transmission subunit uses symmetric encryption, such as AES-256, with the same key for encryption and decryption, transmitting data to the feature learning module via a dedicated communication line at a transmission rate ≥1Mbps.

[0146] The feature learning module is defined as follows: it includes multiple feature extraction tools, a software module that extracts six types of feature sequences from the data, a processing latency of ≤1 second, and a tool optimization unit. It regularly adjusts the adjustable parameters in the feature extraction tools, such as feature weight thresholds, based on new operation and maintenance data, with an adjustment cycle of 24 hours.

[0147] The dynamic fusion unit is defined as follows: it has built-in dynamic weight allocation rules for calculating the weights of each feature, performs weighted fusion on various feature combinations, and generates a comprehensive operation and maintenance feature set for comprehensively evaluating the system operation and maintenance status, with a fusion time of ≤2 seconds.

[0148] The intelligent decision center is defined as follows: it includes an anomaly pattern recognition tool for identifying abnormal patterns in the system, a fault level assessment unit for evaluating the severity of faults and assigning levels 1-4 based on the scope of the fault's impact and its development speed, and a decision suggestion generation unit for generating operational and maintenance decision suggestions based on the fault level and historical handling solutions. The suggestion generation time is ≤5 seconds, and it outputs comprehensive operational and maintenance indicators, anomaly probability values, and corresponding handling methods.

[0149] The hierarchical response execution system is defined as follows: based on the output of the intelligent decision center, it executes corresponding level operation and maintenance operations, including an automatic adjustment module that automatically performs adjustment operations with a response time of ≤1 second, a remote collaboration module that supports remote collaborative operation and maintenance with a data transmission latency of ≤100ms, and an emergency response module that responds to emergencies with a startup time of ≤3 seconds.

[0150] The adaptive optimization module is defined as follows: during normal system operation, it dynamically adjusts operating parameters, such as feature extraction weights and recognition thresholds, at a frequency of once per hour, to achieve continuous optimization of system performance, resulting in an improvement of fault handling efficiency of ≥5% per week after optimization.

[0151] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0152] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0153] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A DCS operation and maintenance method based on the fusion of dynamic weight allocation and fault chain inference, characterized in that, This operation and maintenance method includes: acquiring multi-source operation and maintenance data, which includes system operation data, related production data, and cross-system interaction data; The acquired multi-source operation and maintenance data is preprocessed. The preprocessing includes removing erroneous data, converting data of different magnitudes to the same numerical range, filling in missing data based on reasonable trends of change before and after the time interval, and generating preprocessed fused data. Spatiotemporal correlation alignment and multi-dimensional verification are performed on the preprocessed fused data to generate a multi-source fused dataset. Input the multi-source fusion dataset into a feature learning framework with cross-domain knowledge application capabilities, and output a multi-class feature sequence that reflects the system state. A weight allocation mechanism based on causal reasoning and dynamic scenario adaptation is adopted. Based on the causal correlation strength between each feature sequence and system faults, and the priority requirements of the current production scenario, the aforementioned multiple feature sequences are weighted and fused to generate a comprehensive operation and maintenance evaluation matrix. The execution steps of the weight allocation mechanism based on causal reasoning and dynamic scenario adaptation include: The system equipment is assigned a high priority level in the system, different data transmission priorities are set, the response speed requirements for control commands are set, and the impact of environmental factors on system operation is determined. The basic weights of equipment status characteristics are calculated based on the importance of the equipment; the dynamic adjustment coefficients of transmission characteristics are calculated by combining transmission priority and real-time transmission delay; the timeliness correction coefficients of instruction execution characteristics are calculated based on instruction response timeliness requirements and historical response deviations; and the interference compensation coefficients of environmental interference characteristics are calculated based on the degree of environmental impact and interference duration. The basic weights, dynamic adjustment coefficients, time-based correction coefficients, and interference compensation coefficients mentioned above are normalized to obtain the final weight allocation result; the normalization process converts each coefficient into a value whose sum is 1. The comprehensive operation and maintenance assessment matrix is ​​input into the anomaly identification model that integrates fault evolution trajectory prediction and chain fault inference. This model analyzes the historical fault development process and the correlation between faults, and outputs the current comprehensive operation and maintenance indicators, anomaly probability values, fault development trend curves and possible associated fault chains. Based on the current comprehensive operation and maintenance indicators, anomaly probability values, fault development trend curves and possible associated fault chains, a multi-level progressive response mechanism is triggered. When the system is in a stable operating state, it initiates self-evolutionary learning to continuously optimize operation and maintenance strategies.

2. The DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference according to claim 1, characterized in that, The triggering conditions and execution content of the first-level response in the multi-level progressive response mechanism are as follows: When the operation and maintenance indicators are within the normal range, the probability of anomalies is in the preset low level range, the slope of the fault development trend curve is less than 0.02 and there is no associated fault risk, the system automatically executes preset repair actions including parameter fine-tuning, backup link switching and load distribution adjustment, and records the execution effect and time of each action. If the system status fails to recover to a reasonable range after multiple actions are performed consecutively, it will automatically escalate to a Level 2 response.

3. The DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference according to claim 2, characterized in that, The triggering conditions and execution content of the second-level response in the multi-level progressive response mechanism are as follows: When the operation and maintenance indicators exceed the normal range, the probability of anomalies is in the preset medium level range, the slope of the fault development trend curve is 0.02-0.05 and there is associated fault risk, an intervention decision package containing fault cause analysis, recommended solutions, expected effect simulation and alternative solutions is pushed to the remote operation and maintenance platform. It supports maintenance personnel to configure parameters and provide operation guidance through a remote operation interface. The execution progress is fed back in real time during the operation. If the execution effect does not meet expectations or the fault shows signs of aggravation, it will automatically escalate to a level three response.

4. The DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference according to claim 3, characterized in that, The triggering conditions and execution content of the three-level response in the multi-level progressive response mechanism are as follows: When the operation and maintenance indicators exceed the normal range, the probability of abnormality is in the preset high level range, the slope of the fault development trend curve is 0.05-0.1 and there is a related fault risk, the audible and visual alarm device will be activated immediately, and at the same time, an on-site handling plan including the specific location of the fault, the safety isolation process, the list of required spare parts, the emergency repair route plan and the personnel qualification requirements will be pushed. It also links with the production scheduling system to adjust the production load in advance. If the problem is not resolved after the repair exceeds the preset time, or if the scope of the fault shows an expanding trend, it will automatically escalate to a level four response.

5. The DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference according to claim 4, characterized in that, The triggering conditions and execution content of the four-level response in the multi-level progressive response mechanism are as follows: When the operation and maintenance indicators seriously exceed the normal range, the probability of anomalies is in the preset extremely high level range, the slope of the fault development trend curve is greater than 0.1 and there is a related fault risk, the cross-system emergency command mode is activated, and a collaborative emergency plan including multi-system collaborative isolation strategy, emergency resource allocation plan, production plan adjustment measures and emergency team scheduling instructions is pushed. At the same time, a cross-regional emergency communication line was established to continuously monitor the progress of fault handling until the fault was completely resolved and the system returned to normal operation.

6. The DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference according to claim 1, characterized in that, The process of building and updating a feature learning framework with cross-domain knowledge application capabilities includes: Collect historical operation and maintenance data of industrial control systems from multiple industries; Historical operation and maintenance data are standardized to generate a feature sample set that can be used for model training; An analytical model with a multi-layered processing structure is constructed, and the model is trained using a feature sample set. By continuously optimizing the model parameters, the model's accuracy in identifying fault types reaches the preset standard.

7. The DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference according to claim 1, characterized in that, The workflow of the anomaly detection model includes: Construct a pattern recognition benchmark library that includes a normal pattern library and a fault pattern library; the normal pattern library is a set of feature patterns when the system is running normally; the fault pattern library is a set of feature patterns when the system fails, and the fault pattern library is divided into three levels according to the scope of impact: local fault, regional fault, and system-level fault. Calculate the similarity between the comprehensive operation and maintenance evaluation matrix and the features of various patterns in the benchmark library, determine the pattern with the highest similarity as the identification result, and calculate the anomaly probability under this pattern; When the identification result is a fault mode, the corresponding fault type, possible impact range assessment, and preliminary response plan are output simultaneously.

8. The DCS operation and maintenance method based on fusion dynamic weight allocation and fault chain inference according to claim 1, characterized in that, The process of self-evolutionary learning includes: Set a fixed learning cycle, and collect sufficient execution result data of automatic decision-making and manual intervention decision-making in each cycle, covering the improvement of system status after decision implementation, execution time and resource consumption; The parameters to be optimized are selected by combining specific strategies with knowledge extraction. Among them, the parameters used to balance the exploration of new strategies and the application of mature strategies are dynamically adjusted based on the system stability score and the decision success rate. The system stability score is a score that measures the degree of stable operation of the system. The decision success rate is the proportion of the number of decisions that successfully solve problems out of the total number of decisions. By analyzing the correlation between decision results and changes in system energy consumption, fault interval duration, and fluctuations in production efficiency, the criteria used to evaluate the merits of decisions are updated, and incentive mechanisms are introduced to encourage attempts at new fault handling strategies. When the improvement of the decision evaluation criteria exceeds the preset value for multiple consecutive cycles, the current optimization parameters are determined as the baseline configuration, while the baseline configurations of the previous two cycles are retained as alternatives.

Images