Intelligent control method for chemical production equipment

Abstract

This invention discloses an intelligent control method for equipment used in chemical production, relating to the field of intelligent control technology. The method includes the following steps: collecting process variable data, automatic control commands, operator intervention behaviors, and equipment health status information; constructing a causal relationship model between operator intervention behaviors and equipment health status based on the collected process variable data; identifying anomalies and removing contamination data by combining sliding time window prediction and dynamic tolerance thresholds; constructing a multivariate causal graph based on process variable deviations to distinguish between control error paths caused by equipment degradation and intervention; deploying multi-version control strategy models in the control system, evaluating and selecting the optimal strategy; guiding the operator to record the purpose and method of intervention on the human-machine interface; continuously evaluating the performance of the control strategy and triggering a model retraining and replacement mechanism; this invention ensures that the control logic always closely approximates the real state.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to an intelligent control method for equipment used in chemical production. Background Technology

[0002] In continuous chemical and other process industries, equipment will experience performance degradation as it ages. For example, valves may become sluggish, sensors may become less sensitive, and signal noise may increase. These aging effects cause the control system to become distorted in its perception and response to the process state, causing the originally optimized control algorithm to deviate from the optimal operating point.

[0003] At the same time, based on experience, operators may make "compensatory interventions" to the system through manual operations (such as temporarily opening and closing valves, fine-tuning temperature / pressure, etc.). This behavior may seem to improve the operating effect, but because it is highly subjective and fluctuates with human preferences, it is easy for the control system to learn it as "standard behavior", which further aggravates the model drift of the control strategy.

[0004] The end result is a synergistic drift caused by "aging + human intervention": the control system is affected by equipment aging and learns from contaminated data, mistakenly believing that human micro-operations are ideal control, thus continuously deviating from the original control target, and even making it difficult to determine the root cause of the deviation, which is detrimental to production control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent control method for chemical production equipment.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an intelligent control method for chemical production equipment, comprising the following steps:

[0008] Process data acquisition phase: Deploy sensors, actuator interface modules, controllers and human-machine interaction terminals at the equipment level to collect process variable data, automatic control commands, operator intervention behaviors and equipment health status information. All data are timestamped and the acquisition frequency meets the dynamic response requirements.

[0009] In the joint modeling stage of intervention behavior and equipment health status, a causal relationship model between operator intervention behavior and equipment health status is constructed based on the collected process variable data. Characteristic parameters of control links are introduced as part of the model. By analyzing the frequency, magnitude and process variable response of intervention, control deviation trends and compensatory intervention behaviors are identified.

[0010] In the process signal purification and contamination data identification stage, a process signal prediction model is deployed at the edge node. By combining sliding time window prediction and dynamic tolerance threshold, anomalies are identified and contamination data is removed.

[0011] In the control deviation causal tracing stage, a multivariate causal graph is constructed based on process variable deviations to distinguish between control error paths caused by equipment degradation and intervention, and to output an interpretable deviation causal chain.

[0012] In the control strategy version integration and optimization stage, multiple versions of control strategy models are deployed in the control system. Based on process variable deviation, control offset trend and intervention records, the optimal strategy is evaluated and screened, executed and fed back to the model optimization process, forming a dynamic closed-loop optimization mechanism.

[0013] During the human-computer interaction recording and feedback phase, the operator is guided to record the purpose and method of intervention through the human-computer interface. The data is synchronously bound to the process data and control strategy information for causal modeling and expert review.

[0014] During the control strategy evolution management phase, the performance of the control strategy is continuously evaluated, triggering model retraining and replacement mechanisms.

[0015] As a preferred technical solution of the present invention, the joint probabilistic graphical model is adopted in the joint modeling stage of intervention behavior and equipment health status. The model variables include equipment health status H, operator intervention behavior A, process control response Y, and control deviation E.

[0016] The structure of the joint probabilistic graphical model is as follows:

[0017] ;

[0018] Wherein, P(H,A,Y,E) is the joint probability distribution of variables H, A, Y, and E, that is, the probability that the four variables appear simultaneously at a certain time point (or in multiple samples), P(H) represents the prior probability of the equipment health status, P(A|H) represents the probability of a specific operator intervention behavior occurring under a specific equipment health status, P(Y|A,H) represents the system response probability given the operator intervention behavior and the equipment health status, and P(E|Y) represents the probability of control deviation when a given process response is given;

[0019] Using the maximum a posteriori estimation formula:

[0020] ;

[0021] in, This represents the estimated equipment health status under the maximum a posteriori probability. Let P(H,A|E) represent the operator intervention behavior estimated under the maximum posterior probability, and let P(H,A|E) represent the joint posterior probability of equipment health status and operator intervention behavior under a given control deviation.

[0022] As a preferred embodiment of the present invention, the process signal purification and contamination data identification stage uses an edge-deployed process signal prediction model function f. edge For process signal S t Perform point-by-point prediction:

[0023] ;

[0024] Among them, S t Represented as the observed value of the process signal at time t. Represented as the process signal value predicted by the model;

[0025] And determine whether the process signal is contaminated data using an indicator function:

[0026] ;

[0027] Among them, L clean (S t () represents an indicator function that takes the value 0 or 1, determining the process signal observation value S. t Whether the data is contaminated or not, ε is the tolerance threshold.

[0028] As a preferred embodiment of the present invention, the process signal prediction model is trained using a labeled sample set:

[0029] ;

[0030] Where, x i Let y represent the input sequence of i samples. i L represents the target output for i samples. clean (x i ) represents sample x i Is the data contaminated?

[0031] As a preferred embodiment of the present invention, the control strategy version fusion and optimization stage includes a strategy model fusion unit, which is used to process multiple control strategy models M1, M2, ..., M... n Perform a weighted set and determine the optimal strategy M based on the scoring mechanism. ＊ ;

[0032] The control strategy model set is as follows:

[0033] ;

[0034] Each model in the sliding time window W t The average error within is:

[0035] ;

[0036] Among them, W t Let x represent the set of samples within the current sliding time window. j Let y represent the input of j samples. j M represents the target output for j samples. i (x j ) represents the control strategy model M i For sample x j The prediction results, δ i For the control strategy model M i In the sliding time window W t The average control error within;

[0037] The optimal model is selected using the following formula:

[0038] ;

[0039] Among them, M ＊ This represents the currently selected optimal control strategy model, M0 is the baseline control strategy model, and λ represents the balance coefficient between prediction error and model parameter drift. DriftPenalty(M i M0) represents the drift penalty function, used to measure the control policy model M. i The degree of difference in parameter space between the baseline control strategy model M0 and the baseline control strategy model M0.

[0040] As a preferred embodiment of the present invention, the drift penalty function is defined as follows:

[0041] ;

[0042] Where, θ i Represents the control strategy model M i The parameter vector, θ j Represents the control strategy model M j The parameter vector.

[0043] As a preferred technical solution of the present invention, the human-machine interface records the purpose, method, duration and other details of each operator intervention operation through a form template, and binds it to the process control data to form an intervention behavior database. This data participates in sample selection and model label correction during the model retraining stage.

[0044] As a preferred embodiment of the present invention, the control method further includes constructing a feature vector of the operational intervention behavior:

[0045] ;

[0046] Where τ represents the duration of the intervention, f represents the frequency of the intervention, and α represents the magnitude of the intervention. A The feature vector representing the intervention behavior;

[0047] Clustering algorithms are used to classify intervention behaviors, and compensatory intervention judgments are made in conjunction with the equipment health status H.

[0048] As a preferred technical solution of the present invention, the strategy model fusion device has a built-in drift detection module. When any model in the model set experiences parameter drift exceeding a set threshold during continuous operation, the system automatically switches to the baseline control strategy model M0 and activates the stability monitoring module until the new strategy converges and stabilizes before replacing the baseline strategy.

[0049] The beneficial effects of this invention are:

[0050] 1. In this invention, by introducing a joint probabilistic graphical model in the joint modeling stage of intervention behavior and equipment health status, the system can identify which human interventions are compensatory adjustments to equipment performance degradation and which are accidental or biased operations. This decoupling mechanism effectively prevents the control system from "learning" contaminated human behavior during the model training stage, ensuring the cleanliness and representativeness of the training samples from the source, which is beneficial to maintaining the long-term stability and generalization ability of the control model.

[0051] 2. In this invention, by deploying a process signal prediction model on edge computing nodes, the process variables within the sliding window are dynamically predicted and compared. Combined with the causal chain tracing mechanism of control error, the system can not only identify data pollution points caused by sensor drift, signal noise, etc. in real time, but also trace the root cause of deviations to whether they come from equipment degradation or human intervention. This capability is equivalent to giving the control system an "immune mechanism", enabling it to have the functions of anomaly self-identification, self-purification and deviation self-discrimination, ensuring that the control logic is always close to the real state. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart illustrating the intelligent control method for chemical production equipment according to the present invention. Detailed Implementation

[0054] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Example 1

[0055] like Figure 1 As shown, an intelligent control method for chemical production equipment includes the following steps:

[0056] Process data acquisition phase: Deploy sensors, actuator interface modules, controllers and human-machine interaction terminals at the equipment level to collect process variable data, automatic control commands, operator intervention behaviors and equipment health status information. All data are timestamped to ensure horizontal data alignment and the integrity of time series analysis. The acquisition frequency meets the dynamic response requirements. For different equipment response characteristics and control rhythms, a minimum sampling frequency is set to avoid signal aliasing and loss of key dynamic information.

[0057] Specifically, process variable data such as temperature, pressure, flow rate, and liquid level are recorded in real time by distributed sensors;

[0058] Automatic control commands are controller command signals, including output commands, electrical signals, current / voltage changes, actuator feedback, etc. of the automatic control system;

[0059] Operator intervention behaviors such as setpoint changes, valve opening and closing, and manual adjustment of PID parameters;

[0060] Equipment health status information is mainly obtained through the equipment status monitoring module. The equipment health status information is quantified using a status classification mechanism, including three categories: "healthy", "degraded", and "faulty", and is accompanied by diagnostic parameters characterizing each status (such as signal-to-noise ratio, hysteresis coefficient, frequency response change, etc.). Specifically, the equipment health status information is determined by an SVM classifier based on degradation feature vectors, mainly based on motor temperature, bearing vibration frequency domain characteristics, speed deviation trend, etc.

[0061] In addition, the collected data also includes the alarm trigger status and frequency that occur during system operation;

[0062] The collection of the above process data helps to form a complete, unified, and traceable holographic data system for process operation, providing data support for subsequent behavior modeling and control optimization.

[0063] In the joint modeling stage of intervention behavior and equipment health status, a causal relationship model between operator intervention behavior and equipment health status is constructed based on the collected process variable data. Characteristic parameters of control links are introduced as part of the model. By analyzing the frequency, magnitude and process variable response of intervention, control deviation trends and compensatory intervention behaviors are identified.

[0064] The core of this phase lies in clarifying the underlying causes of operator behavior, determining whether it constitutes a "compensatory intervention," and improving the quality of modeling data by using joint modeling to isolate the intervention's impact. Specific methods include:

[0065] Process decoupling modeling: control links are divided according to process units, and a structured system modeling framework is established;

[0066] Introducing characteristic parameters of the control loop: such as system gain, time delay, control bandwidth, etc., for model representation to better describe the dynamic characteristics of the system;

[0067] Causal analysis of intervention and health status: Compare changes in equipment status and differences in control response before and after the intervention to identify whether the intervention behavior stemmed from control errors caused by equipment degradation;

[0068] Behavioral determination mechanism: If there is a significant causal relationship between the operational behavior and the equipment degradation, and it has periodic characteristics and stability, it is marked as "compensatory behavior";

[0069] Training data filter: This judgment result is used to filter and remove contaminated data for subsequent model training data, thereby improving the scientific accuracy of model fitting.

[0070] In the process signal purification and contamination data identification stage, a process signal prediction model is deployed at the edge node. By combining sliding time window prediction and dynamic tolerance threshold, anomalies are identified and contamination data is removed.

[0071] Specifically, a lightweight process signal prediction model can be deployed on edge computing nodes, along with a real-time anomaly detection module. This module performs sliding time window prediction on the currently collected process variable data, calculates the deviation between the predicted and actual observed values, identifies pollution signal data based on dynamic tolerance thresholds, marks anomalies, and removes them from the subsequent modeling process. This process signal prediction model adaptively adjusts model parameters in conjunction with different operating conditions, achieving generalization capability for process signal prediction and improving the reliability and consistency of data input.

[0072] In the control deviation causal tracking stage, a multivariate causal graph is constructed based on process variable deviations. Starting from the variable deviations, the source of error is traced using structural causal models (SCM) or graph neural networks. The control error paths caused by equipment degradation and intervention are distinguished, and systematic control errors caused by equipment degradation or sudden deviations caused by human operation are differentiated. An interpretable deviation causal chain is output. Combined with equipment health status and operational intervention information, a basis is provided for adjusting the control strategy.

[0073] In the control strategy version fusion and optimization stage, multiple versions of control strategy models are deployed in the control system. Multiple versions are used for different equipment states. The control strategy model can be selected as MPC, reinforcement learning controller, fuzzy controller, etc. Based on indicators such as process variable deviation, control offset trend and intervention record, the optimal strategy is evaluated and screened, executed and fed back to the model optimization process to form a dynamic closed-loop optimization mechanism, that is, the optimal strategy is selected to issue control commands and used as a reference benchmark for the next cycle.

[0074] This step ensures that the system can still dynamically select the optimal control model to maintain the best response, even under conditions of fluctuating equipment status and external interference.

[0075] During the human-computer interaction recording and feedback phase, the operator is guided to record the purpose and method of intervention through the human-computer interface. The data is synchronously bound to the process data and control strategy information for causal modeling and expert review.

[0076] Specifically, operation guidance and data recording functions are embedded in the human-machine interface in the control room. When the operator performs an intervention operation, the system automatically pops up the intervention record interface, which can guide the operator to fill in the operation purpose, expected effect and execution method, and bind this information to process variable data, equipment health status and control strategy information, supporting subsequent causal modeling and expert review. At the same time, the system provides real-time feedback to the operator on the deviation between the current process status and the system's expected behavior through the early warning module.

[0077] During the control strategy evolution management phase, the performance of the control strategy is continuously evaluated, triggering model retraining and replacement mechanisms.

[0078] The system can continuously evaluate the performance and analyze trends of multiple deployed control strategy models. When a control strategy model shows characteristics such as an increase in cumulative control error or an increase in target offset rate over multiple consecutive periods, the control system automatically lists the strategy model as a candidate for elimination and starts the model retraining module to generate a replacement strategy model and add it to the model pool. The control system retains the baseline control strategy for the control logic fallback mechanism to ensure the stability and continuity of system control during the strategy model replacement process.

[0079] It should be noted that the control system mentioned above refers to an intelligent control system used in chemical production equipment. This system includes components such as edge computing nodes, control strategy models, drift detection modules, and human-machine interaction terminals, which work together to complete data acquisition, strategy decision-making, and optimized control tasks.

[0080] Furthermore, the joint probabilistic graphical model is used in the joint modeling stage of intervention behavior and equipment health status. The model variables include equipment health status H, operator intervention behavior A, process control response Y, and control deviation E.

[0081] The structure of the joint probabilistic graphical model is as follows:

[0082] ;

[0083] Wherein, P(H,A,Y,E) is the joint probability distribution of variables H, A, Y, and E, that is, the probability that the four variables appear simultaneously at a certain time point (or in multiple samples); P(H) is the prior probability of the equipment health status, which can be obtained through on-site collection of historical operation statistics; P(A|H) represents the probability of a specific operator intervention behavior occurring under a specific equipment health status; P(Y|A,H) represents the system response probability given the operator intervention behavior and the equipment health status, reflecting the driving influence of equipment degradation on operation behavior, and is used to describe the process control response being affected by the combined effect of equipment health status and operator intervention behavior; P(E|Y) represents the probability of control deviation given a process response, reflecting the propagation path of system error.

[0084] To determine whether "compensatory intervention" exists in the currently collected data, the following maximum a posteriori (MAP) estimation method is proposed, using the MAP estimation formula:

[0085] ;

[0086] in, This represents the estimated equipment health status under the maximum a posteriori probability. Let P(H,A|E) represent the operator intervention behavior estimated under the maximum posterior probability, and let P(H,A|E) represent the joint posterior probability of equipment health status and operator intervention behavior under a given control deviation.

[0087] The specific derivation logic is as follows:

[0088] Transformed using Bayes' theorem:

[0089] ;

[0090] By observing the known control deviation E, summing (marginalizing) all possible process response variables Y, we can infer the equipment health status most likely to cause this deviation. and operational behavior P(H,A,E) represents the joint probability after marginalizing Y;

[0091] If the operation behavior In the health status of the equipment The probability of a degradation scenario is much higher than that of a healthy device. If the probability is the same as that under healthy conditions, then the intervention may be an operation to compensate for equipment degradation, i.e., a "compensatory intervention." Preferably, when the equipment is in a healthy state... The probability of degradation exceeds that of the device's healthy state. Only when the probability is 2 to 4 times higher than that under healthy conditions will it be considered a "compensatory intervention".

[0092] Using this discrimination mechanism, data corresponding to "compensatory intervention" can be marked as "contaminated data" and removed from subsequent model training, thus preventing the control system from "mislearning" human behavior and ensuring the objectivity of modeling and the stability of control strategies.

[0093] Furthermore, the process signal purification and contamination data identification stage utilizes an edge-deployed process signal prediction model function f. edge For process signal S t Perform point-by-point prediction:

[0094] ;

[0095] Among them, S t Represented as the observed value of the process signal at time t. f represents the process signal value predicted by the model. edge Lightweight regression models, such as moving average models, exponential smoothing, and LSTM variants, should possess characteristics such as low latency, strong generalization ability, and edge deployment capability. For example, a prediction model f... edge A two-layer LSTM structure is used, with the input being a sequence of historical variables (x) of length 10. t-9 ,...,x t The output is the predicted value of the variable at the next time step. ;

[0096] And determine whether the process signal is contaminated data using an indicator function:

[0097] ;

[0098] Among them, L clean (S t () represents an indicator function that takes the value 0 or 1, determining the process signal observation value S. t Whether it is contaminated data, ε is the tolerance threshold, 1[·] represents the indicator function, which takes the value 1 when the internal condition is met, and considers the data to be clean; otherwise, it takes the value 0, which is contaminated data. Contaminated data is marked and removed or downweighted from subsequent training samples.

[0099] use Measuring prediction bias is a means of directly reflecting whether the current signal deviates significantly from the model's "expected value";

[0100] The tolerance threshold ε is usually set based on the following considerations:

[0101] Process requirements: Different processes or equipment have different tolerances for signal fluctuations, and the tolerance threshold needs to be selected in combination with the stability requirements of the specific process.

[0102] Historical data fluctuations: By analyzing the fluctuation range and error of historical data, a suitable tolerance value is set;

[0103] Signal type: Different signals (such as temperature, pressure, flow rate, etc.) may have different tolerance ranges.

[0104] For continuous variables such as temperature, pressure, and flow rate, the tolerance threshold ε is usually set as a proportion of the signal fluctuation range. The tolerance threshold ε for pressure signals can be set to 0.5% to 2.5% of the full scale.

[0105] For example: assuming the full scale of temperature is 100°C, the tolerance threshold can be set to [0.5°C, 2.5°C]; assuming the full scale of pressure is 10MPa, the tolerance threshold can be set to [0.05MPa, 0.25MPa]; assuming the full scale of flow rate is 1000L / h, the tolerance threshold can be set to [5L / h, 25L / h].

[0106] This setup ensures that the prediction errors of different process parameters will not have an excessive impact on subsequent modeling and control decisions, thereby maintaining system stability and data quality, and thus ensuring the data quality of the input joint probabilistic graphical model P(H,A,Y,E).

[0107] Furthermore, the process signal prediction model is trained using a labeled sample set:

[0108] ;

[0109] Where, x i Let y represent the input sequence of i samples, with dimensions n×d, where n is the time window length, d is the dimension of the process variables, and y i L represents the target output for i samples. clean (x i ) represents sample x i Whether the data is contaminated or not, the initial training data comes from real running data and historical annotation records, while the contaminated data is annotated by experts and automatically generated based on the sliding prediction error, and semi-supervised learning is used to optimize the label accuracy.

[0110] This edge prediction model is not only used for signal reconstruction; its key task lies in utilizing historical sequences x. i To pre-determine the presence of potential contamination, during the training phase, the model uses triples (x... i ,y i ,L cleanLearning pollution characteristics; in the inference stage, the residuals between predicted and observed values. Used again for dynamic cleaning;

[0111] The core mechanism is as follows:

[0112] Data preprocessing: Detrending, normalizing, and sliding window slicing are performed on the collected raw sequences;

[0113] Pollution label generation: Initial generation of L using an empirical model clean Labels serve as supervision signals for the first phase of training;

[0114] Model training: Supervised learning structures with contaminated labels (such as BiLSTM+MLP structure) are used to optimize the dual objectives of prediction and recognition;

[0115] Edge deployment: After training, the model is deployed on edge nodes to achieve online prediction and cleaning;

[0116] Prediction update: Clean samples are fed back into the model input to further optimize the prediction path and form a closed loop.

[0117] The aforementioned prediction model focuses on pollution characteristics caused by "human intervention + equipment degradation" through labeled learning, and its identification ability is more in line with actual problems.

[0118] Furthermore, the control strategy version fusion and optimization phase includes a strategy model fusion unit, which is used to integrate multiple control strategy models M1, M2, ..., M n Perform a weighted set and determine the optimal strategy M based on the scoring mechanism. ＊ ;

[0119] The control strategy model set is as follows:

[0120] ;

[0121] Each M i This represents a control strategy model (which includes a first-order inertial delay model based on system identification combined with an MPC controller, and a policy network trained using a self-reinforcement learning framework), applicable to specific equipment health states (e.g., healthy, degraded, minor faults) and operating load levels (e.g., low load, high dynamic fluctuations). Each model contains a parameter vector θ. i .

[0122] The design intent of this multi-model structure is to establish a model redundancy pool to avoid affecting the overall control performance due to the inadequacy of a certain model to adapt to specific operating conditions.

[0123] To evaluate the current effectiveness of each control strategy, a sliding time window W is introduced. tIt is defined as the data sequence corresponding to the most recent continuous time period during system operation. The window length is set according to the process response time. Preferably, W t ∈[10 minutes, 100 minutes], the specific value is configured according to the actual dynamic characteristics of the equipment, and each model is within the sliding time window W t The average control error δ within i for:

[0124] ;

[0125] Among them, W t Let x represent the set of samples within the current sliding time window. j Let y represent the input of j samples. j M represents the target output for j samples. i (x j ) represents the control strategy model M i For sample x j The prediction results, δ i For the control strategy model M i In the sliding time window W t The average control error within the current window is a metric that measures the real-time control accuracy of the control strategy within the current window.

[0126] The final optimal model is selected using the following formula:

[0127] ;

[0128] Among them, M ＊ M0 represents the currently selected optimal control strategy model, which is the baseline control strategy model, i.e., the reference model. It is usually a previously trained model with stable performance and reasonable behavior. λ represents the balance coefficient between prediction error and model parameter drift. Preferably, λ∈[0.1,1.0]. When λ is low (e.g., λ is less than 0.4), it is biased towards control accuracy. When λ is high (e.g., λ is greater than 0.7), it is focused on model stability. When the control system is operating under stable conditions, small error differences will not have a significant impact on production, but the drift risk of the model may significantly affect subsequent operation. Therefore, it is recommended to set λ to a higher value, such as 0.8. When the system is in a period of high dynamic fluctuation (e.g., the start-up and shutdown phase or a period of large adjustment), it is necessary to ensure control accuracy first. It is recommended to reduce λ to 0.2~0.4.

[0129] DriftPenalty(M i M0) represents the drift penalty function, used to measure the control policy model M. i The degree of difference in parameter space between the baseline control strategy model M0 and the baseline control strategy model M0.

[0130] Furthermore, considering that even if a model exhibits a small error, a significant drift in its structural parameters (indicating that the model has deviated from its original design or is overfitting) could lead to a decrease in future control stability, a drift penalty function is introduced as one of the policy selection factors. The drift penalty function is defined as follows:

[0131] ;

[0132] Where, θ i Represents the control strategy model M i The parameter vector, θ j Represents the control strategy model M j The parameter vector.

[0133] In a typical continuous reactor control system, several control strategies are set as follows: M1 is applicable to the initial equipment state, M2 is for parameter adjustment in partial aging conditions, and M3 is a non-standard strategy optimized by the operator based on experience.

[0134] During the model fusion stage:

[0135] Although M3 has the smallest average control error δ3 in recent cycles, its parameter drift value is much higher than that of M2. After system evaluation, it is determined that M3 has the risk of overfitting. Therefore, under the condition of λ=0.15, M2 has the lowest comprehensive score and is thus used as the new control strategy.

[0136] Furthermore, the human-machine interface records detailed information such as the purpose, method, and duration of each operator intervention through form templates, and synchronously binds it to the process control data to form an intervention behavior database. This data participates in sample selection and model label correction during the model retraining stage. The human-machine interface is deployed on the web and uses Vue.js to implement a dynamic form recording module. All input fields are synchronously bound to process variable data over time.

[0137] Traditional systems only record operational behaviors (such as adjusting a valve) but cannot know the underlying intentions. Through form templates, operators can be guided to fill in the intervention purpose (such as "controlling temperature fluctuations"), execution method (such as "fine-tuning the set value"), and expected effect (such as "stabilizing outlet pressure"). This provides interpretability factors for subsequent model analysis, enabling the machine learning module to model the intervention behavior as a "conditional dependent variable".

[0138] Each operator intervention is structured and saved as a labeled data item, including: operation timestamp, controlled object, action range, intervention duration, operator identity, expected goal, and actual feedback.

[0139] With standardized record templates, the system can establish causal chains between operational actions and changes in process variables, such as determining whether a temperature anomaly was caused by a human-caused change in a setpoint. It also supports subsequent expert review to determine whether there were any operational errors or whether interventions were appropriate.

[0140] To quantify and analyze operator intervention behaviors, this invention introduces a mechanism for constructing intervention behavior feature vectors based on intervention records:

[0141] ;

[0142] Where τ represents the intervention duration, that is, the duration of each manual intervention action. Too short or too frequent adjustments indicate that the system is unstable or the operation is immature.

[0143] f represents the intervention frequency, reflecting the intervention density. Frequent adjustments may mean a decrease in system control accuracy or model mismatch.

[0144] α represents the magnitude of intervention, which measures the degree of deviation of each intervention from the process variable or set point, and helps to distinguish between fine adjustments and drastic interventions;

[0145] f A The feature vector representing the intervention behavior, specifically, can be processed using a clustering algorithm. A Classification is used to discover behavioral patterns. Commonly used algorithms include K-Means (suitable for a defined number of clusters, with fast convergence), DBSCAN (robust to outliers, suitable for detecting boundary behaviors), and GMM (Gaussian Mixture Model, capable of handling behavioral partitioning under fuzzy boundaries). Taking K-Means as an example, the clustering objective function is:

[0146] ;

[0147] Where K is the number of clusters, representing how many categories the intervention behavior should be divided into. This is usually a pre-defined parameter, with a recommended range of 3-6 clusters. C i Let μ be the i-th cluster, representing all samples included in the i-th clustering result. i Let be the center vector of the i-th cluster, and let represent the average feature vector of all samples in the cluster. The objective function is to calculate the distance between the feature vector of each intervention behavior and its cluster center, and minimize the sum of squared distances to determine the optimal cluster partition.

[0148] Clustering algorithms are used to classify intervention behaviors, and compensatory intervention is judged in conjunction with the equipment health status H. Specifically, after clustering, the average τ, f, α combination characteristics of each cluster are analyzed together with the equipment health status H∈{healthy, degraded, faulty} to determine whether it is a compensatory intervention.

[0149] If τ>τ th (τ) th The threshold for the duration of the intervention (usually between 3 and 10 seconds) and f > f th (f) th The threshold for intervention frequency is typically between 2 times / hour and 10 times / hour, and α > α. th (α) th If the threshold for the intervention magnitude is typically between 5% and 20% relative to the range of the control variable, and the corresponding device H is degraded or faulty, then it is marked as "compensatory intervention". That is, when all conditions are met, the intervention time, frequency, and magnitude all exceed the set threshold, and the device health status is degraded or faulty, the behavior is judged as "compensatory intervention".

[0150] Furthermore, the strategy model fusion unit has a built-in drift detection module. When any model in the model set experiences parameter drift exceeding a set threshold during continuous operation, the system automatically switches to the baseline control strategy model M0 and activates the stability monitoring module until the new strategy converges and stabilizes before replacing the baseline strategy.

[0151] To identify drift states, the system periodically calculates the Euclidean distance between the current parameters of the control strategy model and its initial / reference parameters:

[0152] ;

[0153] Where, θ i Represents the control strategy model M i The parameter vector, θ0 represents the parameter vector of the baseline control strategy model M0;

[0154] Once this value exceeds the system-set threshold δ drift ,Right now:

[0155] ;

[0156] The system immediately recognizes that the model has experienced severe parameter drift and triggers a fallback mechanism. For low-dimensional linear strategy models, the threshold δ drift The preferred range is [0.05, 0.2]; for medium-complexity nonlinear models (such as tree models, shallow neural networks), the threshold δ drift The preferred range is [0.2, 1.0]; for deep neural networks, the threshold δ drift The preferred range is [0.1, 0.3].

[0157] Specifically, when a certain control strategy model M iIf a severe drift is detected, the system immediately interrupts the control output of the model and automatically switches to the pre-deployed baseline control strategy model M0. The baseline strategy M0 has the following characteristics: it has a certain fault tolerance for the health status of all devices, it has shown stable control effects in historical verification, and it can maintain the safe boundary operation of the system in complex and abnormal scenarios.

[0158] The role of the rollback mechanism is to quickly remove the effects of the failure model, preventing it from continuing to contaminate the control logic or trigger cascading failures.

[0159] After switching to M0, the system does not immediately abandon candidate control strategy model M. i Instead, it enters a phase called the "strategy recovery period," with the following specific steps:

[0160] Retrain / fine-tune the control policy model M i : Update the model parameters by calling historical data and the cleaned new data to obtain the new parameters θ. i And the i-th control strategy model M after fine-tuning and restoration. i ′, θ i ′ is M i The parameter vector corresponding to ', compared to the original θ i It has better control and stability;

[0161] Stability assessment: Running M in an isolated environment i Tested in parallel with M0 strategy, and the average control error δ was compared. i and deviation trends;

[0162] The policy evaluation function is as follows:

[0163] ;

[0164] Where, δ i For the control strategy model M i In the sliding time window W t The average control error within the range, and the weighting of control error and parameter changes;

[0165] The final step is the recovery judgment: if Score(M) i If M0 is significantly better than M0, then the system will revert to using M. i ' is the primary strategy; otherwise, continue to retain M0.

[0166] Taking the temperature control system of a continuous distillation column as an example, after long-term operation, the heat exchange efficiency of the top condenser decreases, causing an imbalance in the condensate-reflux ratio. The original control strategy M2 of the system experiences parameter drift, resulting in temperature control lag and amplified fluctuations. The process at this time is as follows:

[0167] First, system calculation =0.27, while the preset threshold is δ drift =0.2;

[0168] Secondly, if the M2 drift is determined to be excessive, the system switches to the baseline control strategy model M0;

[0169] Simultaneously, the model recovery module is activated to fine-tune M2, forming M2′;

[0170] Finally, in the sliding time window W t The error and drift cost are evaluated. If Score(M2′) < Score(M0), then M2′ is restored. If the effect is not achieved, M0 is maintained and M2 is marked as a candidate for elimination.

[0171] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent control of equipment used in chemical production, characterized in that, Includes the following steps: Process data acquisition phase: Deploy sensors, actuator interface modules, controllers and human-machine interaction terminals at the equipment level to collect process variable data, automatic control commands, operator intervention behaviors and equipment health status information. All data are timestamped and the acquisition frequency meets the dynamic response requirements. In the joint modeling stage of intervention behavior and equipment health status, a causal relationship model between operator intervention behavior and equipment health status is constructed based on the collected process variable data. Characteristic parameters of control links are introduced as part of the causal relationship model. By analyzing the frequency, magnitude and process variable response of intervention, control deviation trends and compensatory intervention behaviors are identified. In the process signal purification and contamination data identification stage, a process signal prediction model is deployed at the edge node. By combining sliding time window prediction and dynamic tolerance threshold, anomalies are identified and contamination data is removed. In the control deviation causal tracing stage, a multivariate causal graph is constructed based on process variable deviations to distinguish between control error paths caused by equipment degradation and intervention, and to output an interpretable deviation causal chain. In the control strategy version integration and optimization stage, multiple versions of control strategy models are deployed in the control system. Based on process variable deviation, control offset trend and intervention records, the optimal strategy is evaluated and screened, executed and fed back to the model optimization process, forming a dynamic closed-loop optimization mechanism. During the human-computer interaction recording and feedback phase, the operator is guided to record the purpose and method of intervention through the human-computer interface. The data is synchronously bound to the process data and control strategy information for causal modeling and expert review. During the control strategy evolution management phase, the performance of the control strategy is continuously evaluated, triggering the control strategy model retraining and replacement mechanism. The joint modeling phase of intervention behavior and equipment health status adopts a joint probabilistic graphical model, and the model variables include equipment health status H, operator intervention behavior A, process control response Y, and control deviation E. The structure of the joint probabilistic graphical model is as follows: ； Wherein, P(H) represents the prior probability of the equipment health state, P(A|H) represents the probability of a specific operator intervention behavior occurring under a specific equipment health state, P(Y|A,H) represents the system response probability given the operator intervention behavior and the equipment health state, and P(E|Y) represents the probability of control deviation when a given process response occurs. Using the maximum a posteriori estimation formula: ； in, This represents the estimated equipment health status under the maximum a posteriori probability. Let P(H,A|E) represent the operator intervention behavior estimated under the maximum posterior probability, and let P(H,A|E) represent the joint posterior probability of equipment health status and operator intervention behavior under a given control deviation.

2. The intelligent control method for chemical production equipment according to claim 1, characterized in that, The process signal purification and contamination data identification stage uses a process signal prediction model function f deployed at edge nodes. edge For process signal S t Perform point-by-point prediction: ； Among them, S t Represented as the observed value of the process signal at time t. Represented as the process signal value predicted by the model; And determine whether the process signal is contaminated data using an indicator function: ； Among them, L clean (S t () represents an indicator function that takes the value 0 or 1, determining the process signal observation value S. t Whether the data is contaminated or not, ε is the tolerance threshold.

3. The intelligent control method for chemical production equipment according to claim 2, characterized in that, The process signal prediction model is trained using a labeled sample set: ； Where, x i Let y represent the input sequence of i samples. i L represents the target output for i samples. clean (x i ) represents sample x i Is the data contaminated? 4. The intelligent control method for chemical production equipment according to claim 1, characterized in that, The control strategy version fusion and optimization phase includes a strategy model fusion unit, which is used to fuse multiple control strategy models M1, M2, ..., M n Perform a weighted set and determine the optimal strategy M based on the scoring mechanism. ＊ ; The control strategy model set is as follows: ； Each model in W t The average error within is: ； Where j represents the current sliding time window W t Samples within, x j Let y represent the input of j samples. j M represents the target output for j samples. i (x j ) represents the control strategy model M i For sample x j The prediction results, δ i For the control strategy model M i In the sliding time window W t The average control error within; The optimal model is selected using the following formula: ； Among them, M ＊ This represents the currently selected optimal control strategy model, M0 is the baseline control strategy model, and λ represents the balance coefficient between prediction error and model parameter drift. DriftPenalty(M i M0) represents the drift penalty function, used to measure the control policy model M. i The degree of difference in parameter space between the baseline control strategy model M0 and the baseline control strategy model M0.

5. The intelligent control method for chemical production equipment according to claim 4, characterized in that, The drift penalty function is defined as follows: ； Where, θ i Represents the control strategy model M i The parameter vector, θ j Represents the control strategy model M j The parameter vector.

6. The intelligent control method for chemical production equipment according to claim 1, characterized in that, The human-machine interface records the purpose, method, and duration of each operator intervention through a form template, and simultaneously binds it to the process control data to form an intervention behavior database. This data participates in sample selection and model label correction during the retraining phase of the control strategy model.

7. The intelligent control method for chemical production equipment according to claim 1, characterized in that, The control method further includes constructing a feature vector for operational intervention behaviors: ； Where τ represents the duration of the intervention, f represents the frequency of the intervention, and α represents the magnitude of the intervention. A The feature vector representing the intervention behavior; Clustering algorithms are used to classify intervention behaviors, and compensatory intervention judgments are made in conjunction with the equipment health status H.

8. The intelligent control method for chemical production equipment according to claim 4, characterized in that, The strategy fusion unit has a built-in drift detection module. When any control strategy model in the set of control strategy models experiences parameter drift exceeding a set threshold during continuous operation, the system automatically switches to the baseline control strategy model M0 and activates the stability monitoring module until the new strategy converges and stabilizes before replacing the baseline control strategy model.

Images