CN112379601A - MFA control system design method based on industrial process - Google Patents

MFA control system design method based on industrial process Download PDF

Info

Publication number
CN112379601A
CN112379601A CN202011387575.5A CN202011387575A CN112379601A CN 112379601 A CN112379601 A CN 112379601A CN 202011387575 A CN202011387575 A CN 202011387575A CN 112379601 A CN112379601 A CN 112379601A
Authority
CN
China
Prior art keywords
mfa
controller
control system
delay
output
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011387575.5A
Other languages
Chinese (zh)
Inventor
孙京诰
陈显锋
张晨阳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
East China University of Science and Technology
Original Assignee
East China University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by East China University of Science and Technology filed Critical East China University of Science and Technology
Priority to CN202011387575.5A priority Critical patent/CN112379601A/en
Publication of CN112379601A publication Critical patent/CN112379601A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G05B13/027—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance

Landscapes

  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Feedback Control In General (AREA)

Abstract

The invention adopts a universal Model Free Adaptive (MFA) controller, the controller structure is based on an LSTM dynamic circulation neural network, different single-input single-output processes can be controlled by simple configuration, and a precise mathematical Model of the process is not needed. On the basis of the controller, the invention mainly discloses a general method for designing an MFA control system of an industrial process, which comprises the design of a multivariable MFA control system and the design of an anti-delay MFA control system, and the problems of multivariable coupling, large delay and the like in the industrial process are solved by designing a feedforward and feedback MFA control system to compensate process measurement.

Description

MFA control system design method based on industrial process
Technical Field
The invention provides a general MFA control system design method of an industrial process based on an MFA controller designed by an LSTM recurrent neural network, wherein the MFA control system design method comprises a multivariable MFA control system and an anti-delay MFA control system.
Background
Increasingly complex modern industrial processes place increasing demands on product quality and production efficiency, and indirectly place higher demands on the accuracy and adaptability of controllers. In the face of different complex industrial processes, the optimal adjustment of the controller parameters will directly affect the process control effect, however, the optimal controller parameters are usually difficult to obtain in the practical application process. Therefore, the newly designed control method needs to have certain adaptivity to different processes and does not need a complex parameter adjustment process, and can have good control performance.
In modern industrial processes, PID controllers are widely used due to their simple and practical nature. There are still significant difficulties with the control of complex industrial control systems (e.g., multivariable, strongly coupled, large hysteresis, non-linearity, etc.).
Therefore, a corresponding simple and universal industrial process MFA control system needs to be designed for industrial process systems with different characteristics, the system can be suitable for different processes without considering a process accurate mathematical model, and a good control effect can be obtained only through strong self-learning and self-adaptive capacity of the controller system.
Disclosure of Invention
The invention provides a general MFA control system design method of an industrial process based on an MFA controller designed by an LSTM recurrent neural network. First, a multivariable feedback control system based on a general MFA controller is proposed for use in industrial processes. Then, a general anti-delay MFA controller system is proposed for large hysteresis processes.
The MFA controller designs of two complex systems proposed by the invention are based on a universal single-input single-output MFA controller. Therefore, first, consider a single-input single-output MFA controller system design process, which is implemented as follows:
considering the actual industrial process transfer function
Figure 220672DEST_PATH_IMAGE001
Comprises the following steps:
Figure 436889DEST_PATH_IMAGE002
。
wherein the content of the first and second substances,
Figure 912870DEST_PATH_IMAGE003
,
Figure 265354DEST_PATH_IMAGE004
are respectively a measured variable
Figure 411164DEST_PATH_IMAGE005
And controller output
Figure 849099DEST_PATH_IMAGE006
Is performed by the laplace transform.
The MFA controller design framework is based on the LSTM recurrent neural network architecture. This is because the LSTM loop network structure has a certain memory in processing time series information, and it can make full use of historical time error series information to obtain an appropriate output control amount.
The main task of an LSTM-based MFA controller is to make the currently controlled variable follow a desired trajectory, i.e. to make the error signal
Figure 128771DEST_PATH_IMAGE007
Tending towards zero.
The single-input single-output general LSTM-MFA controller of the invention comprises the following specific steps:
the method comprises the following steps: selecting the length of the input error time series
Figure 335761DEST_PATH_IMAGE008
Then inputting the error sequence
Figure 918052DEST_PATH_IMAGE009
,
Figure 938223DEST_PATH_IMAGE010
,…,
Figure 162531DEST_PATH_IMAGE011
By standardising units
Figure 958449DEST_PATH_IMAGE012
Conversion to
Figure 242799DEST_PATH_IMAGE013
As an input signal to the controller.
Wherein the standardization unit
Figure 983222DEST_PATH_IMAGE012
As tan h function:
Figure 11221DEST_PATH_IMAGE014
。
step two: designing a controller network structure, taking the standardized error signal as network input and obtaining network output through the forward propagation process of the LSTM unit
Figure 661645DEST_PATH_IMAGE015
。
Wherein, the forward propagation calculation formula of the LSTM unit is as follows:
Figure 116897DEST_PATH_IMAGE016
Figure 610196DEST_PATH_IMAGE017
Figure 176306DEST_PATH_IMAGE018
Figure 681237DEST_PATH_IMAGE019
Figure 307390DEST_PATH_IMAGE020
Figure 22405DEST_PATH_IMAGE021
Figure 392207DEST_PATH_IMAGE022
。
wherein the content of the first and second substances,
Figure 17223DEST_PATH_IMAGE023
represents an operation matrix multiplication;
Figure 814278DEST_PATH_IMAGE024
then represents a matrix addition;
Figure 515124DEST_PATH_IMAGE025
for memory-losing gates, for deciding the state of the last cell
Figure 423038DEST_PATH_IMAGE026
The information removed in (1);
Figure 902560DEST_PATH_IMAGE027
for a candidate vector, this value is added to the cell state;
Figure 870516DEST_PATH_IMAGE028
is an input gate for updating the cell state
Figure 560124DEST_PATH_IMAGE027
；
Figure 271728DEST_PATH_IMAGE029
An output section for determining a state of the cell as an output gate; a total of 8 sets of weight vectors to be adjusted are contained in each LSTM
Figure 871336DEST_PATH_IMAGE030
Unlike a normal neural network, all LSTM units in a single recursive hidden layer share this set of weights and do not increase the computational effort.
The activation function adopted by the gate control unit is a sigmoid function, and the expression is as follows:
Figure 10194DEST_PATH_IMAGE031
。
step three: current sampling instant
Figure 921518DEST_PATH_IMAGE032
MFA controller output
Figure 436813DEST_PATH_IMAGE033
Output from the network
Figure 890928DEST_PATH_IMAGE015
And the error value, the expression is
Figure 599307DEST_PATH_IMAGE035
。
Wherein
Figure 652714DEST_PATH_IMAGE036
For controller gains, which are usually greater than zero, the controller performance can be improved by fine tuning, which is generally taken during practical industrial applications
Figure 961335DEST_PATH_IMAGE037
。
Step four: after each sampling moment, a set of data is used to update the network weight through a back propagation algorithm over time so as to obtain a more appropriate control quantity to act on the actual process.
Wherein the input data is of length
Figure 441995DEST_PATH_IMAGE008
Time error sequence of
Figure 94956DEST_PATH_IMAGE009
,
Figure 686474DEST_PATH_IMAGE010
,…,
Figure 115181DEST_PATH_IMAGE011
The output data may be of length
Figure 766742DEST_PATH_IMAGE008
Time error sequence and current network output
Figure 139955DEST_PATH_IMAGE015
Is constructed to output a structural value
Figure 535164DEST_PATH_IMAGE038
The expression of (a) is as follows:
Figure 818378DEST_PATH_IMAGE039
。
wherein the content of the first and second substances,
Figure 640840DEST_PATH_IMAGE040
is of length of
Figure 501349DEST_PATH_IMAGE008
Is passed through a normalization unit
Figure 700249DEST_PATH_IMAGE012
The value obtained is used to determine the value of,
Figure 369128DEST_PATH_IMAGE041
for the adaptive rate, usually take the value of
Figure 831333DEST_PATH_IMAGE042
It can be seen that the error values that are closer to the current time have a greater effect.
Because the MFA controller has strong self-adaptive capacity, the MFA controller has good control effect on a single-input single-output system and does not need a mathematical model with accurate process.
The above process is a single-input single-output MFA controller design process. However, in actual industrial processes, the system is usually a multivariable process or a large hysteresis process, and in order to adapt the MFA controller to a more complex industrial process, a more rational design of the MFA control system is required.
Therefore, the invention designs and develops a multivariable MFA control system and an anti-delay MFA control system on the basis of the original single-input single-output MFA controller.
Firstly, a multivariable feedback control system based on a general MFA controller is provided, which comprises the following specific implementation steps:
the method comprises the following steps: the number of input-output variables, and thus the number of primary MFA controllers and compensators, of an actual industrial process is determined, and then the desired value or desired trajectory for each sub-process is determined.
Step two: by means of a signal comparator device, a deviation signal is generated between the desired value and the actual measured signal for each sub-process, the length of which is truncated
Figure 585663DEST_PATH_IMAGE008
As an input signal to the controller.
Step three: the MFA controller and compensator based on the LSTM recurrent neural network obtains the actual output of the network according to the forward propagation of the input error sequence.
Step four: the actual outputs of the controller and compensator are generated using the network output and the current actual error to derive a control signal for the actual process.
Step five: in each sampling period, according to the input length
Figure 181729DEST_PATH_IMAGE008
And the constructed expected output signal as training data, and updating the network weights on-line by a back propagation algorithm over time, so that the parameters of the controller and compensator are updated on-line to adapt to the actual industrial process.
Step six: and continuously executing the second step to the fifth step until the multivariable industrial process is stable, and finishing the iteration.
Then, aiming at the industrial process with large lag, a general anti-delay MFA controller system is provided, which comprises the following specific steps:
the method comprises the following steps: design delay predictor
Figure 705114DEST_PATH_IMAGE043
Usually in the form of first order plus hysteresis.
Step two: selecting predictors according to certain rules or experience
Figure 338221DEST_PATH_IMAGE044
,
Figure 579846DEST_PATH_IMAGE045
,
Figure 36778DEST_PATH_IMAGE046
And (4) parameters. Gain in general
Figure 414670DEST_PATH_IMAGE044
Can be set close to 1, delay
Figure 484257DEST_PATH_IMAGE046
It can be estimated on the basis of the actual process,
Figure 947600DEST_PATH_IMAGE045
the value may be in terms of a sampling time
Figure 885469DEST_PATH_IMAGE047
And input error sequence length
Figure 117867DEST_PATH_IMAGE008
The product of (a). Because the MFA controller has strong self-adaptive capacity, the MFA controller can also have good control effect when certain deviation exists in the parameters.
Output of delay predictor
Figure 358355DEST_PATH_IMAGE048
The transfer function of (a) is of the form:
Figure 308994DEST_PATH_IMAGE049
。
wherein the content of the first and second substances,
Figure 784975DEST_PATH_IMAGE050
,
Figure 137459DEST_PATH_IMAGE048
,
Figure 548848DEST_PATH_IMAGE043
and
Figure 986783DEST_PATH_IMAGE004
are respectively a signal
Figure 266454DEST_PATH_IMAGE051
,
Figure 942286DEST_PATH_IMAGE052
,
Figure 55736DEST_PATH_IMAGE053
And
Figure 75907DEST_PATH_IMAGE054
is performed by the laplace transform.
Step three: generating dynamic signals from delay predictors
Figure 300215DEST_PATH_IMAGE052
Instead of measuring variables
Figure 96132DEST_PATH_IMAGE051
As a feedback signal, then
Figure 114904DEST_PATH_IMAGE055
。
Step four: in each sampling period, according to the input length
Figure 120906DEST_PATH_IMAGE008
And the constructed expected output signal as training data, and updating the network weights online by a back-propagation algorithm over time, thereby updating the controller online to adapt to the actual industrial process.
Step five: the process is trained in conjunction with standard MFA controllers until the process stabilizes.
The invention develops an MFA control system design method based on an industrial process based on a single-input single-output MFA controller design, wherein a multivariable MFA control system and an anti-delay MFA control system. Compared with the prior control technology, the invention has the advantages that:
(1) the MFA controller of the present invention is a modeless controller, i.e., there is no need to know the exact data model of the process, and there is no need to readjust the controller structure for different processes.
(2) The multivariable MFA control system provided by the invention aims at the problem that the complex industrial process is difficult to decouple and control, provides a multivariable control system based on an MFA controller, can easily control the multivariable process, and has a simple structure.
(3) The delay-resistant MFA control system effectively solves the large delay process by designing the delay-resistant predictor structure based on the MFA controller, and the existing PID control is difficult to control and needs to consume a large amount of parameter adjustment processes.
(4) The design method of the general MFA controller based on the LSTM network is relatively simple in structure and can be easily applied to the actual process.
Drawings
Fig. 1 is a block diagram of a general MFA control system based on an industrial process according to the present invention.
FIG. 2 is a multivariable MFA industrial control system of the present invention.
FIG. 3 is a schematic diagram of a multivariable MFA control system of the present invention.
Fig. 4 is an MFA controller architecture of the present invention.
Fig. 5 is an MFA compensator structure of the present invention.
FIG. 6 is a comparison of two-input two-output system LSTM-MFA and NN-MFA control simulations.
Fig. 7 is an industrial control system for delay tolerant MFA of the present invention.
FIG. 8 is a comparison of LSTM-MFA and NN-MFA control simulations for a delay system.
Detailed Description
The invention will be further explained with reference to the drawings.
The invention provides a design method of a general MFA control system based on an industrial process, the structure of which is shown in figure 1 and comprises basic components of a control process loop such as a controller, the industrial process, a signal adder and the like. The design of the controllers may also differ, as the characteristics of the actual industrial process may include various characteristics such as multivariable, hysteresis, etc.
Considering first the multivariable MFA control system architecture of the present invention as shown in fig. 2, it comprises a set of controllers for each control loop, a multiple-input multiple-output process, a set of signal summers, and a set of back propagation modules over time. Since the system is a multivariable control system, all signal variables are represented as vectors in bold, as follows:
Figure 148905DEST_PATH_IMAGE056
is a set of process settings.
Figure 799329DEST_PATH_IMAGE057
Is a set of process control quantities.
Figure 254581DEST_PATH_IMAGE058
Is a set of process disturbances.
Figure 482300DEST_PATH_IMAGE059
Is a set of process measurements.
Figure 517252DEST_PATH_IMAGE060
Is a set of error signals.
As shown in FIG. 3, without loss of generality, we take as an example a 2-input-2-output multivariable MFA-controlled industrial system, where the MFA controller is
Figure 818921DEST_PATH_IMAGE061
The compensator is
Figure 38550DEST_PATH_IMAGE062
The process comprises four sub-processes, respectively
Figure 628931DEST_PATH_IMAGE063
。
Figure 90743DEST_PATH_IMAGE064
Are respectively controllers
Figure 715759DEST_PATH_IMAGE061
The set value of (2).
Figure 512814DEST_PATH_IMAGE065
Respectively representing sub-processes
Figure 715125DEST_PATH_IMAGE066
To output of (c).
Figure 91880DEST_PATH_IMAGE067
Respectively represent measured values of a process, wherein
Figure 102561DEST_PATH_IMAGE068
，
Figure 663993DEST_PATH_IMAGE069
。
Figure 760125DEST_PATH_IMAGE070
Respectively, error signals between set value and measured variable, wherein
Figure 940570DEST_PATH_IMAGE071
，
Figure 805758DEST_PATH_IMAGE072
。
Figure 538091DEST_PATH_IMAGE073
Are respectively controllers
Figure 121519DEST_PATH_IMAGE061
The output value of (1).
Figure 105655DEST_PATH_IMAGE074
Are respectively compensators
Figure 825350DEST_PATH_IMAGE075
The output value of (1).
Figure 230048DEST_PATH_IMAGE076
Respectively, a control variable of the controller, i.e. an input of a process, wherein
Figure 300772DEST_PATH_IMAGE077
，
Figure 823021DEST_PATH_IMAGE078
。
Figure 662801DEST_PATH_IMAGE079
Respectively represent
Figure 736936DEST_PATH_IMAGE067
The perturbation value of (1).
Controller
Figure 294956DEST_PATH_IMAGE061
Having a single-input single-output MFA controller structure, compensator as shown in FIG. 4
Figure 355316DEST_PATH_IMAGE062
Is shown in FIG. 5, wherein
Figure 49603DEST_PATH_IMAGE009
,
Figure 560218DEST_PATH_IMAGE010
,…,
Figure 808797DEST_PATH_IMAGE011
Respectively for each miningThe input error signal at the time of the sample,
Figure 797482DEST_PATH_IMAGE080
,
Figure 80696DEST_PATH_IMAGE081
,…,
Figure 995169DEST_PATH_IMAGE082
respectively for the network output signal at each sampling instant,
Figure 262202DEST_PATH_IMAGE083
,
Figure 929944DEST_PATH_IMAGE084
,…,
Figure 192298DEST_PATH_IMAGE085
a signal is output for the controller at each sampling instant.
The specific implementation steps of the controller are as follows:
the method comprises the following steps: selecting the length of the input error time series
Figure 654503DEST_PATH_IMAGE008
Then inputting the error sequence
Figure 2308DEST_PATH_IMAGE086
By standardising units
Figure 739320DEST_PATH_IMAGE012
Conversion to
Figure 731546DEST_PATH_IMAGE013
As an input signal to the controller.
Wherein the standardization unit
Figure 161391DEST_PATH_IMAGE012
As tan h function:
Figure 996491DEST_PATH_IMAGE014
。
step two: constructing a controller network structure, selecting an LSTM recurrent neural network as a controller basic unit framework as an input error signal is a time sequence signal, and acquiring network output through forward propagation of the network
Figure 271615DEST_PATH_IMAGE087
。
Step three: current sampling instant
Figure 383928DEST_PATH_IMAGE032
Controller
Figure 719094DEST_PATH_IMAGE061
Output of (2)
Figure 542956DEST_PATH_IMAGE088
Output from the network
Figure 621770DEST_PATH_IMAGE087
And the error value, the expression is
Figure 360236DEST_PATH_IMAGE089
，
Figure 904350DEST_PATH_IMAGE090
。
Compensator
Figure 786855DEST_PATH_IMAGE075
Slightly different from the controller, the output of the controller does not contain an error signal, and the expression is
Figure 139339DEST_PATH_IMAGE091
，
Figure 550729DEST_PATH_IMAGE092
。
Wherein
Figure 988663DEST_PATH_IMAGE093
For controller gains, which are usually greater than zero, the controller performance can be improved by fine tuning, which is generally taken during practical industrial applications
Figure 2756DEST_PATH_IMAGE037
。
Figure 475326DEST_PATH_IMAGE094
For the compensator type of action, the value is 1 for positive action and-1 for negative action.
Step four: after each sampling moment, a set of data is used to update the network weight through a back propagation algorithm over time so as to obtain a more appropriate control quantity to act on the actual process.
Wherein the input data is of length
Figure 57617DEST_PATH_IMAGE008
Time error sequence of
Figure 982847DEST_PATH_IMAGE009
,
Figure 800631DEST_PATH_IMAGE010
,…,
Figure 862127DEST_PATH_IMAGE011
The output data may be of length
Figure 615320DEST_PATH_IMAGE008
Time error sequence and current network output
Figure 27847DEST_PATH_IMAGE087
Is constructed to output a structural value
Figure 147856DEST_PATH_IMAGE038
The expression of (a) is as follows:
Figure 329439DEST_PATH_IMAGE039
。
wherein the content of the first and second substances,
Figure 987953DEST_PATH_IMAGE040
is of length of
Figure 543568DEST_PATH_IMAGE008
Is passed through a normalization unit
Figure 578520DEST_PATH_IMAGE012
The value obtained is used to determine the value of,
Figure 880189DEST_PATH_IMAGE041
for the adaptive rate, usually take the value of
Figure 834238DEST_PATH_IMAGE042
It can be seen that the error values that are closer to the current time have a greater effect.
The MFA controller has strong self-adaptive capacity, so that the MFA controller has certain decoupling capacity on a multivariable system and can work well.
Consider a two-input two-output industrial process system as follows:
Figure 221357DEST_PATH_IMAGE095
Figure 60000DEST_PATH_IMAGE096
Figure 950596DEST_PATH_IMAGE097
Figure 842591DEST_PATH_IMAGE098
wherein the LSTM-MFA controller parameters are all set to
Figure 451427DEST_PATH_IMAGE099
NN-MFA controller parameter set to
Figure 93760DEST_PATH_IMAGE100
. The sampling times are all
Figure 104442DEST_PATH_IMAGE101
。
As shown in fig. 6, it can be seen that the LSTM-based multivariable MFA controller has more stable control performance and higher adaptability to changes in the set value. The invention provides a general anti-delay MFA controller applied to an industrial process aiming at a large-lag process, and the general anti-delay MFA controller is specifically described as follows:
the delay tolerant MFA controller system contemplated by the present invention is shown in FIG. 7 and comprises an MFA controller, a large-lag industrial process, a delay predictor, and a back-propagation over time module.
The delay predictor is designed to generate a dynamic signal
Figure 665873DEST_PATH_IMAGE052
Instead of measuring variables
Figure 762005DEST_PATH_IMAGE051
As a feedback signal, therefore
Figure 942451DEST_PATH_IMAGE055
。
By regenerating the error signal for the controller, the control effect of the process is obtained with less delay, thereby generating a more appropriate control signal.
Because of the strong adaptability of the MFA controller, the actual process output can be predicted by simply designing a delay predictor, which can be designed in a first-order plus lag fashion, as follows:
Figure 807639DEST_PATH_IMAGE049
。
wherein the content of the first and second substances,
Figure 539971DEST_PATH_IMAGE050
,
Figure 123399DEST_PATH_IMAGE048
,
Figure 841957DEST_PATH_IMAGE043
and
Figure 827230DEST_PATH_IMAGE004
are respectively a signal
Figure 730464DEST_PATH_IMAGE051
,
Figure 801188DEST_PATH_IMAGE052
,
Figure 323436DEST_PATH_IMAGE053
And
Figure 163217DEST_PATH_IMAGE054
is performed by the laplace transform.
Figure 747606DEST_PATH_IMAGE053
Is a prediction signal that is a function of the signal,
Figure 40047DEST_PATH_IMAGE052
is the predictor output.
Figure 365986DEST_PATH_IMAGE044
,
Figure 60272DEST_PATH_IMAGE045
,
Figure 305309DEST_PATH_IMAGE046
Approximate model parameters based on a first order plus lag form for an industrial process.
In the practical application process, the gain
Figure 350625DEST_PATH_IMAGE044
May be set close to 1.
Estimating in MFA predictor by rough estimation of actual process delay time
Figure 214676DEST_PATH_IMAGE046
The value of (a).
Figure 763469DEST_PATH_IMAGE045
Can be estimated by the user, or set to
Figure 851511DEST_PATH_IMAGE008
*
Figure 712020DEST_PATH_IMAGE047
Wherein
Figure 379761DEST_PATH_IMAGE008
In order to input the length of the error sequence,
Figure 783061DEST_PATH_IMAGE047
is the sampling time.
The delay-resistant MFA control system has strong adaptive capacity, so that the delay-resistant MFA controller has parameter pair
Figure 42004DEST_PATH_IMAGE044
,
Figure 389808DEST_PATH_IMAGE045
,
Figure 126820DEST_PATH_IMAGE046
Is not very sensitive.
Consider a delayed industrial process system as follows:
Figure 119047DEST_PATH_IMAGE102
the simulation results are shown in fig. 8, and it can be seen that the LSTM-based anti-delay MFA controller has a lower overshoot, a shorter settling time, and a smoother performance when dealing with the delay process.

Claims (10)

1. A multivariable MFA control system for an open-loop stabilized, controllable industrial process, comprising four points:
a) firstly, the multivariable MFA control system consists of a plurality of MFA controllers and compensators, wherein each MFA controller and each compensator are constructed on the basis of an LSTM recurrent neural network;
b) the actual industrial process consists of a plurality of sub-processes, the values of the corresponding control variables of the sub-processes consist of the output of the main MFA controller and the output of the compensator, wherein the compensator can be positive or negative;
c) the control target is that the controlled variable corresponding to the process follows the expected track (set value), and an error signal is calculated according to the difference between the controlled variable and the expected track;
d) the controller and the compensator input historical error sequence signals of corresponding sub-processes, and the output is an expected prediction signal; the parameters of the controller and compensator are continuously iteratively updated by the back-propagation algorithm of the LSTM network to change the control values to reduce the error values.
2. The multivariable MFA control system of claim 1, wherein the controller and compensator are each constructed based on an LSTM recurrent neural network, with a historical time error series signal corresponding to the controlled variable as an input.
3. The multivariable MFA control system of claim 1, wherein the compensators are acting positively as an addition and are acting negatively as a subtraction.
4. The multivariable MFA control system of claim 1, wherein process control values are jointly determined by corresponding primary MFA controller values and compensator values.
5. An anti-delay MFA control system for an open-loop stable, controllable industrial process, comprising four points:
a) the input of the controller is a time error sequence signal, the output is a control value, and the target is a minimized error value;
b) for large lag processes, a delay predictor is designed with inputs for measured process variable and control value outputs, the delay predictor output defined as:
Figure 900661DEST_PATH_IMAGE001
wherein
Figure 806300DEST_PATH_IMAGE002
,
Figure 902432DEST_PATH_IMAGE003
And
Figure 551719DEST_PATH_IMAGE004
the laplace transform of the measured variable, the controlled variable and the delay predictor output respectively,
Figure 853125DEST_PATH_IMAGE005
,
Figure 991982DEST_PATH_IMAGE006
,
Figure 513094DEST_PATH_IMAGE007
parameters that approximate a first order plus lag process for the predictor;
c) the control objective is to minimize the error value between the output of the delay predictor and the set point;
d) the error value is reduced by continuously iteratively updating the controller parameters to change the control values via a back-propagation algorithm of the LSTM network based on the controller input error sequence signal and the desired output control value.
6. The anti-delay MFA control system of claim 5, wherein the structure of the predictor is generally designed in the form of first order plus hysteresis.
7. The delay-tolerant MFA control system of claim 5, wherein the purpose of the delay predictor is to simulate the actual output of the process.
8. The anti-delay MFA control system of claim 5, wherein the parameter
Figure 762809DEST_PATH_IMAGE005
To predict the process gain, it may be set close to 1.
9. The anti-delay MFA control system of claim 5, wherein the parameter
Figure 685766DEST_PATH_IMAGE007
To predict the process delay, it may be obtained by a user's estimation of the actual process.
10. The anti-delay MFA control system of claim 5, wherein the parameter
Figure 995525DEST_PATH_IMAGE006
Can be estimated by the user, or set to
Figure 3932DEST_PATH_IMAGE008
*
Figure 57338DEST_PATH_IMAGE009
Wherein
Figure 333337DEST_PATH_IMAGE008
In order to input the length of the error sequence,
Figure 751680DEST_PATH_IMAGE009
is the sampling time.
CN202011387575.5A 2020-12-01 2020-12-01 MFA control system design method based on industrial process Pending CN112379601A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011387575.5A CN112379601A (en) 2020-12-01 2020-12-01 MFA control system design method based on industrial process

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011387575.5A CN112379601A (en) 2020-12-01 2020-12-01 MFA control system design method based on industrial process

Publications (1)

Publication Number Publication Date
CN112379601A true CN112379601A (en) 2021-02-19

Family

ID=74590279

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011387575.5A Pending CN112379601A (en) 2020-12-01 2020-12-01 MFA control system design method based on industrial process

Country Status (1)

Country Link
CN (1) CN112379601A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115576205A (en) * 2022-11-03 2023-01-06 南栖仙策(南京)科技有限公司 Feedback control method, general feedback controller, training method, readable storage medium, computer program product, and system

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1121196A (en) * 1994-09-08 1996-04-24 侯忠生 Model-less control technology and controller for industrial control
CA2308541A1 (en) * 1997-10-06 1999-04-15 General Cybernation Group, Inc. Model-free adaptive process control
CN101957598A (en) * 2010-09-26 2011-01-26 上海电力学院 Gray model-free control method for large time lag system
WO2020065024A1 (en) * 2018-09-27 2020-04-02 Deepmind Technologies Limited Stacked convolutional long short-term memory for model-free reinforcement learning
CN111930010A (en) * 2020-06-29 2020-11-13 华东理工大学 LSTM network-based general MFA controller design method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1121196A (en) * 1994-09-08 1996-04-24 侯忠生 Model-less control technology and controller for industrial control
CA2308541A1 (en) * 1997-10-06 1999-04-15 General Cybernation Group, Inc. Model-free adaptive process control
CN1274435A (en) * 1997-10-06 2000-11-22 美国通控集团公司 Model-free adaptive process control
CN101957598A (en) * 2010-09-26 2011-01-26 上海电力学院 Gray model-free control method for large time lag system
WO2020065024A1 (en) * 2018-09-27 2020-04-02 Deepmind Technologies Limited Stacked convolutional long short-term memory for model-free reinforcement learning
CN111930010A (en) * 2020-06-29 2020-11-13 华东理工大学 LSTM network-based general MFA controller design method

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115576205A (en) * 2022-11-03 2023-01-06 南栖仙策(南京)科技有限公司 Feedback control method, general feedback controller, training method, readable storage medium, computer program product, and system
CN115576205B (en) * 2022-11-03 2023-08-29 南栖仙策(南京)高新技术有限公司 Feedback control method, universal feedback controller, training method, readable storage medium, computer program product and system

Similar Documents

Publication Publication Date Title
Zribi et al. A new PID neural network controller design for nonlinear processes
Akpan et al. Nonlinear model identification and adaptive model predictive control using neural networks
Liu et al. Adaptive neural output feedback tracking control for a class of uncertain discrete-time nonlinear systems
Zhang et al. Adaptive neural tracking control of pure-feedback nonlinear systems with unknown gain signs and unmodeled dynamics
CN110609476B (en) Multivariate nonlinear dynamic system model prediction control method based on Gaussian process model
CN111930010A (en) LSTM network-based general MFA controller design method
CN105867138A (en) Stable platform control method and device based on PID controller
Lakshmanan et al. Estimation and model predictive control of non-linear batch processes using linear parameter varying models
Sakr et al. Improving the performance of networked control systems with time delay and data dropouts based on fuzzy model predictive control
Georgieva et al. Neural network-based control strategies applied to a fed-batch crystallization process
Xu et al. Terminal sliding mode control using adaptive fuzzy-neural observer
Azmi et al. Robust adaptive fault detection and diagnosis observer design for a class of nonlinear systems with uncertainty and unknown time-varying internal delay
CN112379601A (en) MFA control system design method based on industrial process
CN111240201B (en) Disturbance suppression control method
Antão et al. Model-based control using interval type-2 fuzzy logic systems
Yan et al. Robust model predictive control of nonlinear affine systems based on a two-layer recurrent neural network
Sun et al. GRU-based model-free adaptive control for industrial processes
Uçak et al. Model-free MIMO self-tuning controller based on support vector regression for nonlinear systems
Fatehi et al. Disturbance rejection in neural network model predictive control
CN112731801A (en) Symmetric dead zone nonlinear self-adaptive dynamic surface output feedback control method
Sadeghassadi et al. Dissolved oxygen control of BSM1 benchmark using generalized predictive control
Romasevych et al. Identification and optimal control of a dynamical system via ANN-based approaches
Chu IMC‐PID tuning method based on maximum sensitivity for uncertain multivariable systems
Wang et al. Distributed Double‐Layered Dynamic Matrix Control for Large‐Scale System
Yu et al. Predictive control based on neural networks of the chemical process

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210219