CN-121995733-A - Industrial process self-adaptive control system for deep transfer learning

CN121995733ACN 121995733 ACN121995733 ACN 121995733ACN-121995733-A

Abstract

The invention relates to the technical field of industrial process control and discloses an industrial process self-adaptive control system for deep migration learning, which comprises an offline identification and mapping unit, an online migration and deduction unit, a transient fingerprint feature extraction module and a setting and execution unit, wherein the system deducts initial simplified model parameters through a mapping model to calculate initial PID parameters, and utilizes a deviation correction model to combine the initial simplified model parameters and the transient fingerprint features collected by the transient fingerprint feature extraction module to generate model parameter correction vectors, calculate secondary correction simplified model parameters and update PID parameters.

Inventors

LI XIANGLI
ZHOU LINCHENG

Assignees

苏州工学院

Dates

Publication Date: 20260508
Application Date: 20260408

Claims (10)

1. The industrial process self-adaptive control system for deep transfer learning is characterized by comprising a data input interface, an off-line identification and mapping unit, an on-line transfer and deduction unit, a transient fingerprint feature extraction module and a setting and executing unit: The data input interface is used for receiving process data of the history source working condition and the working condition label; The off-line identification and mapping unit is connected with the data input interface and is used for identifying simplified model parameters { K, T, tau } of the history source working condition and training a mapping model for establishing a mapping relation between the working condition label and the simplified model parameters based on the process data and the working condition label; the online migration and deduction unit is used for receiving the working condition label of the target working condition T When the method is used, a mapping model is called, and initial simplified model parameters { of the target working condition T are deduced The setting and executing unit is used for receiving initial simplified model parameters when working conditions are switched, calling the PID setting rule base and calculating a group of initial PID parameters { Automatically loading initial PID parameters into a PID controller for execution; The transient fingerprint feature extraction module is used for collecting transient response of a process variable in a preset short time window after the PID controller loads initial PID parameters, extracting a group of transient fingerprint features and sending the transient fingerprint features to the setting and executing unit; The setting and executing unit is used for receiving transient fingerprint characteristics when the preset short time window is over, calling a deviation correction model, outputting a model parameter correction vector based on the initial simplified model parameter and the transient fingerprint characteristics, and superposing the initial simplified model parameter and the model parameter correction vector to obtain a secondary corrected simplified model parameter { And calling the PID setting rule base again, and calculating and updating the parameters of the PID controller based on the parameters of the secondary correction simplified model.
2. The adaptive control system for an industrial process for deep learning of claim 1, wherein the reduced model parameters { K, T, τ }, the initial reduced model parameters { And a secondary correction of the simplified model parameters { The system comprises a PID controller, a simulation verification daemon, a closed loop simulation and a stability verification unit, wherein { K, T, τ } in the } comprises a process gain K, a time constant T and a time lag τ, the system further comprises the simulation verification daemon which is used for executing closed loop simulation by taking initial simplified model parameters as controlled objects and initial PID parameters as controllers before the initial PID parameters are loaded into the PID controller, and judging whether the initial PID parameters are loaded or a group of conservative PID parameters according to the stability verification result of the closed loop simulation.
3. The adaptive control system for an industrial process according to claim 1, further comprising a control performance diagnostic module including a predetermined expert rule base, wherein the offline identification and mapping unit is further configured to train a drift model for establishing a map of operating condition labels and run times of the batch and simplified model parameters after drift, wherein the control performance diagnostic module is configured to monitor control error signals of the PID controller when the PID controller is operating during a period of time T >0, calculate statistical signatures of the control error signals during a period of time, and diagnose whether a model-process mismatch exists based on the expert rule base, and acquire a current run time of the batch and the operating condition labels when the control performance diagnostic module diagnoses the model-process mismatch And calling the setting and executing unit to recalculate and update the parameters of the PID controller based on the simplified model parameters after drifting.
4. The adaptive control system for an industrial process for deep learning of claim 3 wherein the statistical signature is an integrated absolute error IAE, wherein the expert rules library diagnoses model-process mismatch by comparing the integrated absolute error IAE to a predetermined threshold, and wherein the integrated absolute error IAE is calculated over a period of time by the following formula And (3) internal calculation: , wherein, In order to control the error signal, For the current time period of time, For a length of time that is a period of time, Is an integral variable.
5. The adaptive control system for an industrial process of deep learning of claim 1, wherein the mapping model trained by the offline recognition and mapping unit is a multi-layer perceptron model.
6. The adaptive control system for an industrial process of deep learning of claim 1, wherein the mapping model trained by the offline recognition and mapping unit is a K-nearest neighbor model.
7. The adaptive control system for an industrial process for deep migration learning of claim 1, wherein the PID tuning rule base comprises at least one of IMC rules, cohen-Coon rules, and Ziegler-Nichols rules.
8. The adaptive control system for an industrial process for deep transfer learning of claim 1, wherein the transient fingerprint characteristics include at least one of an initial response slope, an overshoot, and a first peak time.
9. The adaptive control system for an industrial process for deep learning of claim 2, wherein the system is further configured to initiate an automatic tuning procedure to perform on-line fine tuning of the parameters of the PID controller after the PID controller updates the parameters.
10. The adaptive control system for industrial process of deep migration learning of claim 2, wherein the simulation verification daemon is configured to { the initial PID parameters when arbitrating loading the conservative PID parameters Proportional parameter in } Halving to generate conservative PID parameters.

Description

Industrial process self-adaptive control system for deep transfer learning Technical Field The invention relates to an industrial process self-adaptive control system for deep transfer learning, and belongs to the technical field of industrial process control. Background The current proportional integral derivative PID controller is simple in structure, good in robustness and easy to realize, forms a main stream of industrial automatic control, and is widely applied to adjustment of process variables such as temperature, pressure and flow; however, in modern intermittent production modes such as fine chemical industry and pharmacy, the production characteristics of small batches and multiple varieties become normal, so that the application of the PID controller faces long-term technical limitation, when the production working conditions such as product brands, raw material batches or process target values are switched, the dynamic characteristics of the controlled process such as gain, time constants and time lags are changed, so that the well-regulated PID parameters under the original working conditions perform poorly under the B working conditions and even cause system oscillation, and the baseline scheme for solving the problem is a standard scheme for a person skilled in the art, after the working conditions are switched, the controller is switched to a manual mode, an engineer takes tens of minutes or even hours through experience conservation operation, trial-and-error or open-loop tests are carried out to readjust the PID parameters, the time window required by the manual setting and the transition materials generated in the setting process are rapidly increased in the total production batch, the ratio in the total production batch is changed into systematic economic and efficiency problems, and various automatic setting methods such as a relay feedback method are tried in the industry for shortening the time, but the methods are still invasive, the method still needs to be switched to a specific disturbance signal T0 after the working conditions are still required to be injected, and the system is still difficult to finish in the transient condition when the actual working conditions are switched to be 0 = 0. Meanwhile, in the related field of intelligent manufacturing, although the prior art scheme tries to use advanced models such as transfer learning to process multiple working conditions, the application focus is more biased to state monitoring or diagnosis of equipment, and the core problem of instantaneous and predictive setting of control parameters is not touched, for example, the Chinese patent application with the authority of publication number CN115351601B discloses a tool wear monitoring method based on transfer learning, and the gist is to use monitoring data and wear label training models of existing working conditions to enable the tool wear monitoring method to be quickly adapted to tool wear prediction tasks under new working conditions, and the method is essentially to construct an identification model or a diagnosis model from a process signal to the current state. Therefore, how to provide a non-invasive control manner, a set of near-optimal initial parameters can be provided for the PID controller instantaneously at the time t=0 of the working condition switching, so as to eliminate the time delay of manual setting, which is a technical problem to be solved by the present invention. Disclosure of Invention In order to solve the problems in the background technology, the technical scheme of the invention is as follows, and the industrial process self-adaptive control system for deep transfer learning comprises a data input interface, an off-line identification and mapping unit, an on-line transfer and deduction unit, a transient fingerprint feature extraction module and a setting and executing unit: The data input interface is used for receiving process data of the history source working condition and the working condition label; The off-line identification and mapping unit is connected with the data input interface and is used for identifying simplified model parameters { K, T, tau } of the history source working condition and training a mapping model for establishing a mapping relation between the working condition label and the simplified model parameters based on the process data and the working condition label; the online migration and deduction unit is used for receiving the working condition label of the target working condition T When the method is used, a mapping model is called, and initial simplified model parameters { of the target working condition T are deducedThe setting and executing unit is used for receiving initial simplified model parameters when working conditions are switched, calling the PID setting rule base and calculating a group of initial PID parameters {Automatically loading initial PID parameters into a PID controller for execution; The transient f