Industrial process adaptive control system based on deep transfer learning

Abstract

The invention relates to the technical field of industrial process control, and discloses an industrial process adaptive control system for deep transfer learning, which comprises an offline identification and mapping unit, an online transfer and deduction unit, a transient fingerprint feature extraction module and a setting and execution unit, the system deduces an initial simplified model parameter through a mapping model to calculate an initial PID parameter; and generating a model parameter correction vector, calculating a secondary correction simplified model parameter and updating a PID (Proportion Integration Differentiation) parameter by utilizing a deviation correction model and combining the initial simplified model parameter and a transient fingerprint feature collected by a transient fingerprint feature extraction module. The initial deviation deduced by the static label is corrected by using the real transient response, the single dependence on the static label is avoided, the implicit disturbance during working condition switching can be rapidly corrected, and the accuracy of the PID parameter at the initial starting stage is improved.

Description

Technical Field

[0001] This invention relates to an adaptive control system for industrial processes using deep transfer learning, belonging to the field of industrial process control technology. Background Technology

[0002] Currently, the proportional-integral-derivative (PID) controller, due to its simple structure, robustness, and ease of implementation, constitutes the mainstream of industrial automatic control and is widely used for regulating process variables such as temperature, pressure, and flow. However, in modern batch production modes such as fine chemicals and pharmaceuticals, the production characteristics of small batches and multiple varieties have become the norm, causing the application of PID controllers to face long-term technical limitations. When production conditions change, such as product brand, raw material batch, or process target value, the dynamic characteristics of the controlled process, such as gain, time constant, and time delay, change accordingly. This causes the PID parameters, which were well tuned under the original conditions, to perform poorly under the new conditions, or even cause system oscillation. Those skilled in the art need to address this problem. The baseline approach involves switching the controller to manual mode after a change in operating conditions. Engineers then rely on experience and conservatively operate the system, spending tens of minutes or even hours performing trial and error or open-loop testing to retune the PID parameters. The time window required for this manual tuning and the amount of transition material generated during the tuning process increase dramatically in the total production batch, evolving into a systemic economic and efficiency problem. To shorten this time, the industry has tried various automatic tuning methods, such as relay feedback. However, most of these methods are invasive and still require injecting specific disturbance signals into the system after T>0. This also generates scrap in actual production, making it impossible to deliver parameters instantly at T=0 when the operating conditions change.

[0003] Meanwhile, in the field of intelligent manufacturing, although some technical solutions have attempted to use advanced models such as transfer learning to deal with multi-condition problems, the focus of application is mostly on equipment condition monitoring or diagnosis, without addressing the core challenge of instantaneous and predictive tuning of control parameters. For example, Chinese invention patent with authorization announcement number CN115351601B discloses a tool wear monitoring method based on transfer learning. The main purpose is to use existing monitoring data and wear labels to train a model, enabling it to quickly adapt to the tool wear prediction task under new conditions. Essentially, it is to build an identification model or diagnostic model from process signals to the current state. Its value lies in in-process evaluation, but it does not solve the fundamental bottleneck faced in this field when switching conditions, namely, the lack of a pre-emptive mechanism that cannot non-invasively and predictively generate a new set of near-optimal PID control parameters at T=0 based on the static formula or label of the target condition.

[0004] Therefore, the technical problem to be solved by this invention is how to provide a non-intrusive control method that can provide a set of near-optimal initial parameters to the PID controller at the T=0 time of the operating condition switch, so as to eliminate the time delay of manual tuning. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A deep transfer learning-based adaptive control system for industrial processes, the system comprising a data input interface, an offline identification and mapping unit, an online transfer and deduction unit, a transient fingerprint feature extraction module, and a tuning and execution unit: The data input interface is used to receive process data and operating condition tags from historical source operating conditions; The offline identification and mapping unit is connected to the data input interface and is used to identify the simplified model parameters {K,T,τ} of historical source working conditions based on process data and working condition labels, and to train a mapping model to establish the mapping relationship between working condition labels and simplified model parameters. The offline identification and mapping unit is also used to train the deviation correction model, which is used to establish the mapping from the initial simplified model parameters and transient fingerprint features to the model parameter correction vector. The online migration and simulation unit is used to receive the working condition label of the target working condition T. At that time, the mapping model is invoked to deduce the initial simplified model parameters of the target working condition T. The tuning and execution unit has a built-in PID tuning rule library. When switching operating conditions, the tuning and execution unit receives initial simplified model parameters, calls the PID tuning rule library, and calculates a set of initial PID parameters. }, and automatically load the initial PID parameters into the PID controller for execution; The transient fingerprint feature extraction module is used to collect the transient response of the process variable within a preset short time window after the initial PID parameters are loaded into the PID controller, extract a set of transient fingerprint features, and send them to the tuning and execution unit. The tuning and execution unit is used to receive transient fingerprint features at the end of a preset short time window, call the deviation correction model, and output a model parameter correction vector based on the initial simplified model parameters and transient fingerprint features; the initial simplified model parameters are superimposed with the model parameter correction vector to obtain the secondary corrected simplified model parameters. The PID tuning rule base is called again, and the parameters of the PID controller are calculated and updated based on the simplified model parameters after secondary correction.

[0006] Preferably, the simplified model parameters {K,T,τ} and the initial simplified model parameters { } and the second-corrected simplified model parameters { In the formula {K,T,τ}, all parameters include process gain K, time constant T, and time delay τ. The system also includes a simulation verification guardian unit, which is used to perform closed-loop simulation using the initial simplified model parameters as the controlled object and the initial PID parameters as the controller before the initial PID parameters are loaded into the PID controller. Based on the stability verification results of the closed-loop simulation, the guardian unit decides whether to load the initial PID parameters or a set of conservative PID parameters.

[0007] Preferably, the system also includes a control performance diagnosis module, which contains a preset expert rule base; the offline identification and mapping unit is also used to train a drift model, which is used to establish the mapping between operating condition labels and batch running time and simplified model parameters after drift; the control performance diagnosis module is used to monitor the control error signal of the PID controller when the PID controller is running during the T>0 period, calculate the statistical signature of the control error signal over a period of time, and diagnose whether there is a model-process mismatch based on the expert rule base; and when the control performance diagnosis module diagnoses a model-process mismatch, it obtains the current batch running time and operating condition label. It automatically triggers the drift model, derives the simplified model parameters after drift, and calls the tuning and execution unit to recalculate and update the PID controller parameters based on the simplified model parameters after drift.

[0008] Preferably, the statistical signature is the integral absolute error (IAE); the expert rule base diagnoses model-process mismatch by comparing the IAE with a preset threshold; the IAE is expressed by the following formula over a period of time. Internal calculation: ,in, To control the error signal, For the current time, The length of a period of time, It is the integral variable.

[0009] Preferably, the mapping model for offline identification and mapping unit training is a multilayer perceptron model.

[0010] The preferred mapping model for offline identification and mapping unit training is the K-nearest neighbor model.

[0011] Preferably, the PID tuning rule base includes at least one of the following: IMC rules, Cohen-Coon rules, and Ziegler-Nichols rules.

[0012] Preferably, transient fingerprint features include at least one of the following: initial response slope, overshoot, and first peak time.

[0013] Preferably, the system is also used to start an automatic tuning program after the PID controller updates its parameters, and to fine-tune the parameters of the PID controller online.

[0014] Preferably, when the simulation verification guardian unit decides to load conservative PID parameters, it is used to set the initial PID parameters { The proportional parameters in} Halve the parameters to generate conservative PID parameters.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This paper provides an implementation path for adaptive control. The system uses an online migration and deduction unit to instantly deduce the initial simplified model parameters of the target operating condition when switching operating conditions by utilizing the training mapping relationship. The tuning and execution unit then calls the fixed PID tuning rule library and calculates the initial PID parameters based on the model parameters. The parameter acquisition time is compressed from traditional manual trial and error or invasive testing to near zero, realizing non-invasive hot start of the controller. Before loading the parameters, the simulation verification guardian unit uses the deduced model parameters as the controlled object to perform closed-loop simulation verification on the loaded PID parameters. Based on the stability, the final execution parameters are determined. The intelligence of the deduction is placed under the constraint of the stability of the control system, avoiding the black-box runaway risk when artificial intelligence methods are directly applied to control. This allows the system to achieve adaptive speed while maintaining the safety and predictability of the control mechanism.

[0016] 2. By working in tandem with the transient feature extraction module and the deviation correction model, a secondary calibration mechanism at time T=0 is constructed. Hot start is achieved using the simulation results, immediately acquiring the true transient response of the process variables within a preset short time window. This response is extracted as a transient fingerprint feature and input into the deviation correction model along with the initial model parameters to deduce the model parameter correction vector. The tuning and execution units then use this correction vector to superimpose and correct the simplified model parameters, and again call the tuning rule base to calculate and update the PID parameters. Utilizing the true process feedback information at time T=0, the control system is freed from its sole reliance on static condition labels, enabling rapid correction of latent disturbances already present at time T=0, such as raw material batch fluctuations. This allows the PID parameters to approach the true optimal value in the initial startup phase, suppressing initial overshoot or slow response.

[0017] 3. By combining the control performance diagnostic module with the drift model, adaptive control is extended from the time T=0 to the entire batch operation process T>0. During T>0, the diagnostic module non-invasively monitors the control error signal of the PID controller and calculates its statistical signature over a period of time, such as integral absolute error. The module judges whether the statistical signature indicates a mismatch between the model and the process based on the expert rule base. By utilizing the existing error signal in the controller, complex online identification is avoided during T>0. When the diagnostic module confirms the existence of model drift, the system automatically triggers the drift model, deduces the drifted model parameters for the current batch time, and calls the tuning and execution unit to recalculate and update the PID parameters. This solves the problem of slow time-varying process characteristics during T>0 in a low-cost, on-demand triggering manner, ensuring the performance of the control system throughout the entire batch operation cycle. Attached Figure Description

[0018] Figure 1 This is a functional architecture and data flow diagram of the adaptive control system of the present invention; Figure 2 This is a comparison chart of the response curves of the simulation verification guardian unit of the present invention; Figure 3 This is a flowchart of the multi-stage adaptive control logic of the system of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0020] This invention provides a deep transfer learning-based adaptive control system for industrial processes. The system includes a data input interface for data interaction, an offline identification and mapping unit for offline modeling, an online transfer and deduction unit for prediction at time T=0, a transient fingerprint feature extraction module for instantaneous correction after T=0, and a tuning and execution unit for online calculation and execution. The system also includes a simulation verification and protection unit for safety verification before T=0 and a control performance diagnostic module for slow time-varying diagnosis during T>0, collectively forming an adaptive control architecture covering the entire cycle of intermittent processes. In specific engineering implementation, the system's working logic is divided into three time phases: an offline training phase, a T=0 instantaneous startup and correction phase, and a T>0 long-term operation diagnostic phase. In the offline training phase, the offline identification and mapping unit is used to build a knowledge base; the data input interface is used to retrieve process data and condition labels from historical source operating conditions from an industrial historical database; the process data consists of control variables actually collected by the control system in historical batches. With process variables Time series data, operating condition labels These are the static process formula parameters corresponding to that batch, such as product grade, key component ratio, or target temperature setpoint; the offline identification and mapping unit first processes the process data of each set of historical source conditions. To ensure identification accuracy, the system employs differentiated identification strategies based on data characteristics using standard system identification tools: for slow processes with large inertia or lag, step response analysis is used to extract model features by analyzing the response curves of process variables to control steps; while for scenarios with abundant historical samples, large data volumes, and random noise, least squares regression fitting is used to eliminate noise interference by utilizing the statistical characteristics of big data, thereby identifying and extracting the simplified model parameters {K,T,τ} corresponding to the operating condition.

[0021] In the specific working condition i, it is represented as { These three parameter sets correspond to a first-order plus pure time-delay FOPDT model, representing the process gain K, time constant T, and time delay τ, respectively; after obtaining all historical operating conditions... to After the mapping pairs are established, the unit trains three independent models: a mapping model, with input being N-dimensional working condition labels. The output is a simplified 3D model parameter. This model is used to establish the basic mapping relationship from static formulas to dynamic properties. A multilayer perceptron (MLP) model or a K-nearest neighbor (KNN) model can be selected. A bias correction model is trained, with input consisting of two parts of data, and initial simplified model parameters derived from the mapping model. A set of transient fingerprint features is used to output a 3D model parameter correction vector, which is used to learn the correction relationship between the initial response shape at time T=0 and the model prediction bias. The offline identification and mapping unit constructs the training dataset of the bias correction model in the offline stage. The procedure includes: the working condition label Li and the corresponding process data for any batch i in the historical database. The trained mapping model is invoked, and the initial simplified model parameters are derived based on the working condition label Li. (0), (0), (0)};Complete process data for batch i Using the application system identification tool, the actual simplified model parameters for this working condition are identified. (actual), (actual), (actual)}; Calculate the difference between these two to obtain { (actual)- (0), (actual)- (0), (actual)- (0)}, use this difference vector as the ground truth label required for training; from the process data of batch i Transient fingerprint features are extracted within a preset short time window, and these transient fingerprint features are compared with the initial simplified model parameters { (0), (0), The combination of (0)} serves as the input data for this training pair. The bias correction model is trained using the training set consisting of all historical batches of input data and ground truth labels. The drift model is trained using the working condition labels as input. and batch run time Its output is the simplified model parameters after drift. This model is used to learn the slow drift law of process characteristics with reaction time under specific operating conditions, such as catalyst deactivation and changes in material viscosity.

[0022] During the instantaneous startup and correction phase at T=0, i.e., the moment when the industrial site is preparing to switch operating conditions, the operator inputs the operating condition label of the target operating condition T through the data input interface. The online migration and simulation unit receives this. And immediately invoke the trained mapping model to instantly deduce the initial simplified model parameters of the target working condition T. The initial parameters are sent to the tuning and execution unit; in a preferred embodiment, to ensure the stable operation of the control system, the system activates the simulation verification guardian unit; at this time, the tuning and execution unit, based on its internally fixed PID tuning rule library, which may include at least one of IMC rules, Cohen-Coon rules, and Ziegler-Nichols rules, bases the following parameters on { } Calculate a set of temporary initial PID parameters { }; The simulation verification guardian unit will deduce { } as the controlled object model, the calculated { As controller parameters, a closed-loop simulation is instantaneously performed in the background of the control system. Based on preset stability criteria, such as damping rate or overshoot, the simulation results are checked. If the simulation is unstable or too aggressive, a set of conservative PID parameters is loaded. These conservative parameters can be generated using a gain reduction strategy, such as using the calculated proportional parameters. The initial PID parameters are halved; if the simulation is stable, the initial PID parameters are loaded. The tuning and execution unit will automatically load the PID parameters (initial or conservative) into the PID controller at T=0, immediately putting it into automatic operation, achieving a non-intrusive hot start. When the simulation verification and protection unit loads the conservative PID parameters, its internal gain reduction strategy can be preset according to different levels of stability requirements of the control loop. To achieve the fastest and most reliable stability recovery, the generation procedure for the conservative PID parameters is determined to adjust only the proportional term. While maintaining the integral parameter and differential parameters To ensure stability while seeking a faster response, a joint adjustment strategy can be adopted, adjusting the proportional parameter. While halving the points, the points accumulation time will be adjusted accordingly. and differential time Following the T=0 warm start, the system immediately enters the secondary calibration process; the transient fingerprint feature extraction module starts, and within a preset short time window (T=0 to T=10 seconds), it non-invasively acquires the true transient response curve of the process variable PV; from this curve, it extracts a set of transient fingerprint features, which are universal features characterizing the dynamic response, including at least one of the initial response slope, overshoot, or first peak time; at the end of the preset short time window, the tuning and execution unit receives this set of newly acquired transient fingerprint features; then it calls the offline-trained bias correction model and adjusts the initial simplified model parameters { The transient fingerprint feature is used as input; the deviation correction model outputs a model parameter correction vector; the tuning and execution unit superimposes this vector with the initial parameters to obtain the secondary correction simplified model parameters. The tuning and execution unit again calls its internal PID tuning rule base, and based on the second-corrected model parameters, calculates a new set of PID parameters. To ensure the security of online parameter updates, the system reactivates the simulation verification guardian unit, and performs a secondary correction of the simplified model parameters. As the controlled object, the closed-loop stability of the new PID parameters is verified. After the verification is passed, the tuning and execution units update the parameters of the PID controller in a non-disruptive switching manner. This mechanism uses the process feedback at T=0 to correct the initial deviation of the static label prediction and improve the control accuracy in the initial stage of startup.

[0023] The procedure for the transient fingerprint feature extraction module to extract features within a preset short time window includes: a preset short time window Tw length, and, during operating condition switching, based on the initial simplified model parameters {KT(0),T... T (0),τ T(0)} is dynamically determined. To prevent the acquisition window from closing prematurely due to model time delay inference deviations, its duration needs to be within the inference time delay τ. T (0) is based on an additional preset safety margin (e.g., 1.1 to 1.5 times τ). T (0)), and add a portion of the response time to ensure that a non-zero response can be collected, for example, it can be set to 1.2×τ. T (0) and 0.3×T T The sum of (0); within this Tw time window, this module prioritizes processing data at t=τ. T (0) Linear fitting is performed on the process variable data from t=Tw to calculate the initial response slope characteristics; for conditional characteristics such as overshoot or the first peak time, the first derivative of the process variable is monitored within the Tw time window. If no extreme point is detected at the end of Tw, the response is still in the monotonic change stage, and these conditional characteristic values ​​are set to preset default values, such as 0 or -1. The deviation correction model has learned samples containing this default value during offline training to ensure that the dimension of the model input is constant; during the long-term operation diagnosis stage of T>0, the control performance diagnosis module works continuously during the batch operation period of T>0 (which can last for several hours); non-invasively monitoring the control error signal of the PID controller, i.e. The system calculates the statistical signature of the error signal over a period of time in the background, with the preferred statistical signature being the integral absolute error (IAE). It contains a pre-defined expert rule base for diagnosing model-process mismatch. A typical diagnostic rule is to determine whether the rolling value of the IAE continuously increases and exceeds a certain pre-defined threshold. When a model-process mismatch is diagnosed based on the rule base, the current batch runtime is immediately obtained. and current operating condition labels Automatically triggers offline training of the drift model; the drift model is based on and Simplified model parameters after drift derivation The drifted parameters are sent to the tuning and execution unit, which reuses the PID tuning rule base to calculate the updated PID parameters. Subsequently, the system must call the simulation verification guardian unit to perform stability assessment using the simplified model parameters after drift. After ensuring parameter safety, the system executes a non-disruptive update of the PID controller parameters. The non-disruptive update is implemented as follows: at the moment the update command is triggered, the system first locks the current actual output value and current error of the PID controller; using the updated PID parameters, it calculates the current proportional and derivative outputs; then, by subtracting the new proportional and derivative terms from the locked actual output value, it reverses the calculation and resets the initial cumulative value of the integral term; this initial value is used as the integral reference when the new parameters take effect. This mechanism ensures that the total output value of the controller remains unchanged at the moment of parameter switching, thereby eliminating the step disturbance of the actuator caused by parameter changes, solving the slow time-varying problem of process characteristics in a low-cost way, and ensuring the control performance of the system throughout the entire batch operation cycle. In an optional implementation, after the PID controller updates its parameters, regardless of T= Whether it's the secondary correction at time point or the drift correction during the T>0 time period, the system can further initiate an automatic tuning program, such as the relay feedback method, to fine-tune the PID controller parameters online, achieving more refined optimization based on a hot start. When the system initiates the automatic tuning program to fine-tune the PID controller parameters online, the triggering time of this program is preset to be after the tuning and execution unit completes the PID parameter update of the simplified model parameters for the secondary correction (i.e., when T=10 seconds), and the process variable PV has entered a new steady state or a preset stable error band. The system can automatically inject non-destructive, small disturbance signals into the control loop or utilize natural disturbances that occur during the production process. The offline identification and mapping unit collects the process data of this stage online, performs a rapid online identification, obtains a more accurate {K,T,τ} model, and the tuning and execution unit calls the PID tuning rule base again to calculate and update the PID parameters, completing the final online fine-tuning.

[0024] Example 1: In a batch production scenario of a fine chemical polymerization reactor, the controlled object is the polymerization reactor. Operating conditions change frequently. The dynamic characteristics of the process, characterized by simplified model parameters {K,T,τ}, differ between different production products and between different batches of the same product using different operating conditions. The continuous increase in material viscosity during the reaction process also causes the model parameters to drift during batch operation when T>0. In a specific operating condition switching event, the system switches production from grade A (a low-viscosity product) to grade B (a high-viscosity product). Before the switching time T=0, the DCS operator inputs the operating condition label corresponding to grade B into the system. The formula parameters, including those for grade B, are received by the online migration and derivation unit. And by invoking the offline-trained mapping model, the initial simplified model parameters for grade B are instantly derived. The tuning and execution unit receives these initial parameters, calls its internally stored PID tuning rule base, and calculates a set of initial PID parameters. }; The simulation verification guardian unit then utilizes { As a simulation object, verify { The closed-loop stability of the system was verified, and the results showed that the system was stable. The tuning and execution units then... Load the PID controller and put it into automatic operation at time T=0.

[0025] Within this specific batch, a portion of Grade B raw materials contained no processing condition label. The low catalyst activity of the feedstock resulted in a slower actual temperature response of the reactor after a T=0 hot start, i.e., the process variable PV, than expected by the initial simplified model parameters, and a lower initial response slope. Within a preset short time window from T=0 to T=10 seconds, the transient fingerprint feature extraction module captured this response pattern and extracted it as a set of transient fingerprint features. At T=10 seconds, the tuning and execution unit invoked the deviation correction model, adjusting the initial simplified model parameters { The transient fingerprint features are input together to output a model parameter correction vector. This vector represents the model bias caused by low catalyst activity, i.e., a lower process gain K and a longer time constant T. The tuning and execution unit superimposes this correction vector with the initial parameters to obtain the secondary corrected simplified model parameters. Immediately, the PID tuning rule base is called again to calculate and update the parameters of the PID controller without disturbance, so that the controller adapts to the dynamic characteristics of this batch of raw materials to suppress the continuous low-temperature undertuning; after the batch runs for T=2 hours, the viscosity of the material in the reactor increases with the polymerization reaction, causing the dynamic characteristics of the process to drift slowly. The PID parameters calibrated at T=10 seconds are no longer suitable, which is manifested as a decrease in the ability to suppress upstream feed disturbances and an increase in control error. It begins to increase slowly; the control performance diagnostic module continuously monitors it. The signal is analyzed, and its rolling integral absolute error (IAE) over the past 30 minutes is calculated. Based on its internal expert rule base, it is determined that the continuously increasing trend of the IAE has triggered a model-process mismatch alarm; the current operating condition label is immediately obtained. That is, grade B and batch running time. That is, after 2 hours, the drift model is automatically triggered; the drift model instantaneously derives simplified model parameters after the drift at time T=2 hours. The tuning and execution unit restarts, and based on the drifted parameters, recalculates and updates the PID parameters to maintain control performance during batch operation with T>0.

[0026] Example 2: To objectively verify the effectiveness of the adaptive tuning at time T=0 and the secondary calibration mechanism at T=10 seconds of the present invention, the test scenario was set as a single operating condition switch. The PID controller was originally running in condition A, with model parameters {K=1.0, T=50s, τ=10s}. At time T=0, the system switched to condition B. To verify the system's ability to correct for implicit disturbances, a model mismatch was introduced, where the operating condition label did not match the actual process. The operating condition label entered by the operator at time T=0... The standard model parameters corresponding to working condition B are, i.e. ={2.0,30s,5s}; The actual model parameters for condition B running in the simulation platform are set to {K=1.8,T=35s,7s}. This deviation simulates unlabeled disturbances introduced by factors such as raw material batch fluctuations; In the experiment, the preset short time window is set to 10 seconds. This time window setting balances the requirements of rapid commissioning and the minimum amount of data required for feature extraction. The length of 10 seconds is greater than the time delay of 7 seconds in the actual model, which can ensure that the transient fingerprint feature extraction module can capture the initial dynamic response of the process variables; The experiment is set up with three test groups, respectively at T At time 0, a 20% setpoint step disturbance is applied to the actual model {1.8, 35s, 7s} of condition B. Control performance indicators are recorded, including overshoot, settling time (i.e., the time required for the process variable to enter the ±2% error band of the setpoint), and integral absolute error (IAE) (accumulated within 300 seconds after the start of the experiment). The first group is control group 1, where the controller keeps the PID parameters of condition A unchanged and simulates a non-adaptive condition switch. The second group is control group 2, whose controller uses the online migration and derivation unit of the system of this invention, based on condition labels. The standard model {2.0, 30s, 5s} calculates and loads the initial PID parameters, and disables the secondary calibration function at T=10 seconds; the third group is the experimental group. The controller uses the same initial PID parameters as the control group 2 to achieve a hot start at T=0. At T=10 seconds, the transient fingerprint feature extraction module and the deviation correction model are activated to perform secondary calibration and update of the PID parameters. The experimental data are summarized in Table 1.

[0027] Table 1: Performance Comparison of Different Control Schemes under Model Mismatch Conditions The experimental data summarized in Table 1 show that Control Group 1, due to severe model mismatch, experienced significant overshoot and continuous oscillations, resulting in an excessively long settling time, which is unacceptable in control practice. Control Group 2 achieved a T=0 hot start through online migration and deduction units, showing improved performance compared to Control Group 1. However, due to deviations between its operating condition labels and the actual process, its PID parameters were not optimal, resulting in a 24.5% overshoot and a settling time of 138.6 seconds. The experimental group performed similarly to Control Group 2 at T=0. At T=10 seconds, its transient fingerprint feature extraction module and deviation correction model worked together to deduce the model parameter correction vector based on the response data from T=0 to T=10 seconds, generating secondary correction simplified model parameters and updated PID parameters. This secondary calibration mechanism compensated for the initial model deviation introduced by implicit disturbances, reducing the overshoot of the control system to 3.8%, shortening the settling time to 45.1 seconds, and minimizing the integral absolute error (IAE).

[0028] Example 3: This example combines Figures 1 to 3 An explanation of adaptive control systems for industrial processes based on deep transfer learning, such as... Figure 1 As shown, the offline identification and mapping unit retrieves historical process data and operating condition labels from the historical database, trains and writes them into the mapping model, deviation correction model, and drift model. The operator inputs the target operating condition label into the online migration and deduction unit, reads the mapping model to deduce the initial simplified model parameters, and sends them to the tuning and execution unit. The tuning and execution unit then reads the PID tuning rule base, calculates the initial PID parameters, and sends them to the simulation verification and protection unit. Based on the verification result, it outputs the verified PID parameters or conservative PID parameters to the tuning and execution unit. The tuning and execution unit then sends the PID parameters to the industrial process DCS / PLC. The industrial process generates transient response process variables and control error signals. The transient response is captured by the transient fingerprint feature extraction module, extracted as transient fingerprint features, and sent to the tuning and execution unit. The deviation correction model is read to calculate the secondary corrected PID parameters, and the control error signal is sent to the control performance diagnosis module. When the module diagnoses model mismatch triggering, the tuning and execution unit reads the drift model to calculate the drifted PID parameters.

[0029] like Figure 2 As shown in the figure, the horizontal axis represents time in seconds, the vertical axis represents the process variable value, and a dotted line represents the setpoint. The figure also shows a stable response (as shown by the solid line) with initial PID parameters, exhibiting controllable overshoot and eventually converging to the setpoint, and an unstable response (as shown by the dashed line) with conservative PID parameters, exhibiting continuous large oscillations. Figure 3As shown, this process involves three parties: engineers, operators, and industrial process DCS / PLC. Engineers perform the task of training the system model, operators perform the task of starting a new operating condition and issuing PID parameters, and the industrial process DCS / PLC provides process data. In the internal logic of the system, starting a new operating condition will trigger the deduction of initial parameters. This action calls the mapping model. The deduced parameters are used for both safety verification and PID parameter calculation, ultimately leading to the issuance of PID parameters. The action of providing process data will trigger two parallel logics simultaneously. One is to perform secondary calibration, which will trigger the extraction of transient fingerprints and the invocation of the deviation correction model. The other is to diagnose long-term drift and then call the drift model. Both of these logics will trigger the calculation of PID parameters again to update the parameters.

[0030] Example 4: This example illustrates the specific procedures for the offline identification and mapping unit to construct a drift model and calibrate the internal expert rule base of the control performance diagnosis module during the offline phase of long-term operation (T>0). It further clarifies the adjudication logic of the simulation verification guardian unit at time T=0. The procedures for the offline identification and mapping unit to construct the drift model include: the system filters out complete historical source operating condition batches with long operating times from the historical database; each batch of data includes its operating condition label. and full-cycle process data For each batch Process data Batch run time Perform time slicing, with a window width of 1 hour, and obtain... , , Multiple data segments; for each time slice data segment, the system's identification tool is used independently to identify the average simplified model parameters within that time period. ,in This represents the midpoint or start point of the time period; the system combines the operating condition label, batch run time, and identified parameters into training samples, i.e. The training samples generated from all historical batches and all time slices are collected to form a complete dataset for training the drift model. This dataset is implemented using a multilayer perceptron and is used to establish a nonlinear mapping from [operating condition label, batch running time] to [simplified model parameters after drift].

[0031] The offline identification and mapping unit calibration control performance diagnosis module's expert rule base procedures are used for statistical signatures. Here, we take the integral absolute error (IAE) as an example to determine the trigger threshold. This procedure is used for online diagnosis of model-process mismatch during periods T>0; it includes: First, the system selects a baseline operating condition from historical data. and in Time-of-flight model and the corresponding initial PID parameters The second step is for the system to obtain the operating condition under these conditions. Model after time drift The third step is to use the system in a simulation environment. Initial PID parameter control The drift-out model is used to simulate the closed-loop response under this mismatch condition, and its performance index, namely overshoot, is recorded. The fourth step is for the system to calculate the IAE value of the simulation within a specific time window, such as 30 minutes. Fifth, the system repeats steps two through four, traversing combinations of different operating conditions and drift degrees in the database to generate a relational database of mismatch, IAE, and performance; Sixth, from this database, the boundary of control performance degradation is determined, such as... Identify all simulation conditions that cause this performance degradation. Values, these The statistical lower bound of the value, such as the 80th percentile, is set as global. The online diagnostic rules for the control performance diagnostic module are defined as follows: when IAE(last30min) > Furthermore, if the setpoint remains unchanged, a model drift alarm is triggered; when the simulation verification guardian unit performs a safety check at T=0, its internal decision logic is determined to extract the overshoot of the simulation response after performing closed-loop simulation. and attenuation rate ;when or If the condition is deemed unstable, a conservative PID parameter generation procedure is activated, employing one of several strategies implemented in this embodiment. In this embodiment, the procedure is determined to be to increase the proportional parameter in the initial PID parameters. Halve, that is ,Keep and Load the conservative parameters; otherwise, load the initial PID parameters. .

[0032] Example 5: When the system of the present invention is first deployed on a control loop without historical data, taking the temperature control loop of a newly built reactor as an example, the system executes a standardized initial calibration procedure; the historical database of the offline identification and mapping unit is empty, and the online migration and deduction unit cannot perform prediction; at this time, the system is triggered by the engineer to enter the initial calibration working mode and perform a controlled open-loop step test on the control loop; the offline identification and mapping unit collects the control variables of this test. With process variables Upon completion of the test, the system's built-in identification tool automatically identifies the baseline simplified model parameters of this loop from the set of response data. This set of parameters, along with its corresponding operating condition label, is recorded as "New Reactor - Initial Water Test Run" and stored in the historical database as the first anchor point data for subsequent transfer learning.

[0033] After the system has accumulated a certain amount of historical data, when the operator enters new operating condition labels that match all historical database entries... The target working condition labels all show significant differences. When prototyping a new grade, the system activates the cold start procedure for the new operating condition. When the mapping model of the online migration and derivation unit is implemented using the K-NN algorithm, it returns a low-confidence result during querying, failing to find a sufficient number of nearest neighbors in the historical data. At this time, the system does not generate prediction parameters. Instead, the tuning and execution unit automatically loads a set of general conservative PID parameters preset within the simulation verification guardian unit at T=0. These conservative parameters have low proportional gain and high integral time, ensuring closed-loop stability under the unknown model. After starting control at T=0, the system immediately activates the transient fingerprint feature extraction module. Within a preset short time window from T=0 to T=10 seconds, it captures the actual process response driven by these conservative parameters and extracts transient fingerprint features based on the following calculation procedure: For the initial response slope, the linear least squares method is used to fit the data of the effective response segment. The calculation eliminates the interference of process noise on differential operations; for overshoot, since steady state has not yet been reached within the short time window, it is defined as the percentage deviation of the maximum value of the process variable within the window from the current target setpoint; for the first peak time, the first derivative zero-crossing point determination method is adopted, and a preset noise suppression amplitude threshold (such as 0.5% of the sensor range) is introduced. Only when the peak amplitude exceeds this threshold is it confirmed as a valid extreme point to prevent misjudgment caused by signal jitter; the tuning and execution unit then sends this set of finely defined fingerprint features, together with the system's preset general reference model parameters, as input to the deviation correction model; this parameter serves as the physical center anchor point, helping the model establish an initial reference frame under unknown operating conditions, outputting the model parameter correction vector, thereby calculating the secondary correction simplified model parameters at T=10 seconds. The PID is updated to allow the system to quickly converge from a general safe state to a control state adapted to the new operating conditions.

[0034] Example 6: This example illustrates two alternative standardized engineering procedures used by the offline identification and mapping unit in the offline phase to construct the mapping model; the core task is to establish a system based on the working condition labels. (N-dimensional input) to simplified model parameters The nonlinear mapping relationship (3D output) is trained using data obtained from historical data identified in specific implementation methods. The mapping is applied to the database; in this implementation, the mapping model is constructed as a multilayer perceptron (MLP) model; the construction procedure includes: mapping all N-dimensional working condition labels in the database. Normalization is performed; a three-dimensional output layer is constructed, consisting of an N-dimensional input layer, at least one hidden layer, and three dimensions (corresponding to K, T, ..., ...). The neural network structure uses normalization. As input, the corresponding As labels, the MLP model is trained using the standard backpropagation algorithm. The training objective is to minimize the mean squared error between the model's output value and the label value. After training, the MLP model is embedded in the online transfer and inference unit, ready to be used upon receiving new target condition labels. At that time, through one forward calculation, the instantaneous output is obtained. To reduce model complexity, the mapping model is constructed as a K-Nearest Neighbor (K-NN) model; the training procedure involves using all historical data... Mapping pairs are all stored in a fast-retrieval N-dimensional feature space database; the online migration and inference unit receives the target working condition label. At that time, the unit executes the K-NN query procedure: calculate All historical operating condition tags in the database Find the Euclidean distance between them; find the one with the smallest distance. Neighbors (such as) ); calculate this Average distance of the nearest neighbors And compare it with a preset working condition similarity threshold. Compare; if ,show Similar to historical operating conditions, the system will then... The nearest neighbors correspond to The values ​​are weighted and averaged (weights are inversely proportional to distance) to calculate the result. ;like ,show This is a completely new operating condition that differs significantly from historical experience. At this point, the online migration and extrapolation unit determines it as a low-confidence prediction, automatically triggers the new operating condition cold start procedure, prioritizes loading general conservative PID parameters, and immediately starts the transient fingerprint feature extraction module for online calibration.

[0035] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep transfer learning-based adaptive control system for industrial processes, characterized in that, The system includes a data input interface, an offline identification and mapping unit, an online migration and deduction unit, a transient fingerprint feature extraction module, and a tuning and execution unit. The data input interface is used to receive process data and operating condition tags from historical source operating conditions; The offline identification and mapping unit is connected to the data input interface and is used to identify the simplified model parameters {K,T,τ} of historical source working conditions based on process data and working condition labels, and to train a mapping model to establish the mapping relationship between working condition labels and simplified model parameters. The offline identification and mapping unit is also used to train the deviation correction model, which is used to establish the mapping from the initial simplified model parameters and transient fingerprint features to the model parameter correction vector. The online migration and simulation unit is used to receive the working condition label of the target working condition T. At that time, the mapping model is invoked to deduce the initial simplified model parameters of the target working condition T. The tuning and execution unit has a built-in PID tuning rule library. When switching operating conditions, the tuning and execution unit receives initial simplified model parameters, calls the PID tuning rule library, and calculates a set of initial PID parameters. }, and automatically load the initial PID parameters into the PID controller for execution; The transient fingerprint feature extraction module is used to collect the transient response of the process variable within a preset short time window after the initial PID parameters are loaded into the PID controller, extract a set of transient fingerprint features, and send them to the tuning and execution unit. The tuning and execution unit is used to receive transient fingerprint features at the end of a preset short time window, call the deviation correction model, and output a model parameter correction vector based on the initial simplified model parameters and transient fingerprint features; the initial simplified model parameters are superimposed with the model parameter correction vector to obtain the secondary corrected simplified model parameters. The PID tuning rule base is called again, and the parameters of the PID controller are calculated and updated based on the simplified model parameters after secondary correction.

2. The deep transfer learning-based adaptive control system for industrial processes according to claim 1, characterized in that, Simplified model parameters {K,T,τ}, initial simplified model parameters { } and the second-corrected simplified model parameters { In the system, {K,T,τ} all include process gain K, time constant T, and time delay τ; the system also includes a simulation verification guardian unit, which is used to perform closed-loop simulation using the initial simplified model parameters as the controlled object and the initial PID parameters as the controller before the initial PID parameters are loaded into the PID controller. Based on the stability verification results of the closed-loop simulation, a decision is made as to whether to load the initial PID parameters or load a set of conservative PID parameters.

3. The deep transfer learning-based adaptive control system for industrial processes according to claim 1, characterized in that, The system also includes a control performance diagnostic module, which contains a pre-set expert rule base. The offline identification and mapping unit is used to train a drift model, which establishes a mapping between operating condition labels, batch running times, and simplified model parameters after drifting. The control performance diagnostic module monitors the control error signal of the PID controller when it operates during the T>0 time period, calculates the statistical signature of the control error signal over a period of time, and diagnoses whether model-process mismatch exists based on the expert rule base. When the control performance diagnostic module diagnoses model-process mismatch, it obtains the current batch running time and operating condition label. It automatically triggers the drift model, derives the simplified model parameters after drift, and calls the tuning and execution unit to recalculate and update the PID controller parameters based on the simplified model parameters after drift.

4. The deep transfer learning-based adaptive control system for industrial processes according to claim 3, characterized in that, The statistical signature is the integral absolute error (IAE); the expert rule base diagnoses model-process mismatch by comparing the IAE with a preset threshold; the IAE is expressed by the following formula over a period of time. Internal calculation: ,in, To control the error signal, For the current time, The length of a period of time, It is the integral variable.

5. The deep transfer learning-based adaptive control system for industrial processes according to claim 1, characterized in that, The mapping model trained by the offline identification and mapping unit is a multilayer perceptron model.

6. The deep transfer learning-based adaptive control system for industrial processes according to claim 1, characterized in that, The mapping model trained by the offline identification and mapping unit is a K-nearest neighbor model.

7. The deep transfer learning-based adaptive control system for industrial processes according to claim 1, characterized in that, The PID tuning rule base includes at least one of the following: IMC rules, Cohen-Coon rules, and Ziegler-Nichols rules.

8. The deep transfer learning-based adaptive control system for industrial processes according to claim 1, characterized in that, Transient fingerprint features include at least one of the following: initial response slope, overshoot, and first peak time.

9. The deep transfer learning-based adaptive control system for industrial processes according to claim 2, characterized in that, The system is also used to start an automatic tuning program after the PID controller updates its parameters, and to fine-tune the PID controller parameters online.

10. The deep transfer learning-based adaptive control system for industrial processes according to claim 2, characterized in that, The simulation verification guardian unit is used to load the initial PID parameters { when deciding on the conservative PID parameters. The proportional parameters in} Halve the parameters to generate conservative PID parameters.

Images