Process control using neural network

• US 5,111,531 A
• Filed: 01/08/1990
• Issued: 05/05/1992
• Est. Priority Date: 01/08/1990
• Status: Expired due to Fees

• Current Assignee: Automation Technology, Inc. (California) (Intertek Group PLC)
• Original Assignee: Automation Technology, Inc. (California) (Intertek Group PLC)
• Inventors: Grayson, S. Keith, Rudd, John B.
• Primary Examiner(s): Fleming, Michael R.
• Assistant Examiner(s): Davis, George
• Application Number:
  US07/462,503
• Time in Patent Office: 848 Days
• Field of Search: 364/513, 395/23, 395/24
• US Class Current: 706/23
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: G06N 3/105 Shells for specifying net layout

Abstract

A control system and method for a continuous process in which a trained neural network predicts the value of an indirectly controlled process variable and the values of directly controlled process variables are changed to cause the predicted value to approach a desired value.

Claims

1. A control system for a continuous physical process characterized by directly controlled process variables and indirectly controlled process variables, said variables having current values and having had historical values that define the state of the physical process wherein the current values of directly controlled process variables are measured and wherein the current value of at least one indirectly controlled process variable is controlled indirectly by control of the current values of a plurality of directly controlled process variables comprising:

control means responsive to set point values for establishing the current values of the directly controlled process variables at said set point values applied to said control means,

means for implementing a trainable neural network having a plurality of input neurons for having input values applied thereto and at least one output neuron for providing an output value,

means for training the neural network to provide a predicted value for the at least one indirectly controlled process variable at an output neuron, said predicted value corresponding to the input values of the neural network,

means for measuring the values of directly controlled process variables,

means for establishing and continuously updating a computer database to store the current and historical values of measured process variables,

computer means for establishing the input values at the input neurons of the neural network based upon the values of process variables stored in the computer database,

computer means for establishing set point values to be applied to control means based upon values at an output neuron,

said control system so constructed and arranged that said computer means for establishing set point values, after the neural network has been trained to predict the value of the at least one indirectly controlled process variable, changes at least one set point value to cause the predicted value of the at least one indirectly controlled process variable to approach a desired value.

2. The control system according to claim 1 wherein the continuous physical process is characterized by an input stage, intermediate stages and an output stage, said process takes a material or workpieces at said input stage, processes said material or workpieces at the intermediate stages and delivers the material or workpieces at said output stage and wherein a period of time between the material or a workpiece being at the input stage and the material or workpiece being at the output stage is long compared to a period of time for the control means to establish directly controlled variables at set points.

3. The control system according to claim 2 wherein the indirectly controlled variable is measured at or near the output stage and other process variables are measured at or near all stages.

4. The control system according to claim 3 wherein control means are associated with a plurality of distinct process stages.

5. The control system according to claims 1, 2, 3 or 4 wherein said means for establishing and continuously updating the computer data base stores values of a plurality of measured variables at spaced time intervals in a data structure to maintain a process history.

6. The control system according to claim 5 wherein the set point values applied to each of the control means at a given process stage are established for those control means by presenting to the neural network inputs based upon the historical values of the measured process variables associated with prior process stages and based upon the current values of measured process variables at succeeding stages.

7. The control system according to claim 6 wherein the set point values of each control means are adjusted at spaced time intervals.

8. The control system according to claim 1, 2, 3 or 4 wherein the computer means for establishing set point values establishes new set point values during spaced adjustment time intervals and wherein the means responsive to set point values for establishing the values of the directly controlled variables bring the values of the respective controlled variables to set point values during settling time intervals, and wherein the means for establishing and continuously updating the computer database comprises means for storing values of measured variables in the database in a data structure immediately following the settling time intervals and before the next spaced adjustment time interval.

9. A control system for a process characterized by directly controlled process variables and indirectly controlled variables, all said process variables having values that define the state of the process wherein the values of a plurality of process variables are measured and wherein the value of at least one process variable is controlled indirectly by control of a plurality of directly controlled process variables comprising:

control means for establishing the value of the directly controlled variables at set point values applied to said control means,

means for implementing a neural network having a plurality of input neurons for having input values applied thereto and at least one output neuron for providing an output value,

computer means for storing an input value for input neurons of the neural network and at least one computer executable rule to establish the input value associated with an input neuron,

computer means for storing an output value for output neurons of the neural network and at least one computer executable rule for an output neuron,

computer means for establishing a process database defining the state of the process including the instantaneous values of the process variables,

means for continuously updating the process database to reflect the values of the measured process variables,

computer means for executing a rule associated with an input neuron to establish the value of the input neuron,

computer means for executing a rule associated with an output neuron for establishing set point values to be applied to control means,

means for applying the set point values to the control means,

means for training the neural network to generate a predicted value corresponding to the value of said indirectly controlled variable at an output neuron from values at input neurons corresponding to measured process variables,

said control system so constructed and arranged that after the neural network has been trained to generate the predicted value of said indirectly controlled variable a rule associated with said at least one output neuron changes at least one set point value to cause the predicted value of the indirectly controlled variable to approach a directed value.

10. The control system according to claim 9 wherein the control means include devices with feedback control that compare the values of the directly controlled variables with the set point values and generate error signals applied to the devices having a sign to drive the devices to diminish the error signals.

11. The control system according to claim 10 wherein the feedback control comprises proportional-integral-differential control.

12. The control system according to claim 9 wherein the means for implementing the neutral network is an integrated circuit defining the neural network and a teaching algorithm.

13. The control system according to claim 9 wherein the means for implementing the neural network is a computer program defining the neural network and a teaching algorithm.

14. The control system according to claim 9 wherein the computer means for storing the values and rule associated with an input neuron is a second computer database.

15. The control system according to claim 9 wherein the computer means for storing the values and rule associated with an output neuron is a second computer database.

16. The control system according to claim 9 wherein the means for continuously updating the process database to reflect the values of the directly controlled variables comprise computer input ports and a computer task for polling the computer input ports or fielding interrupts from the computer input ports to transfer data from the ports to the process database.

17. The control system according to claim 9 wherein the means for applying set point values to the control means comprise computer output ports and a computer task for transferring set point values to the computer output ports.

18. The control system according to claim 9 wherein the computer means for executing said computer executable rule associated with an input neuron and said computer executable rule associated with an output neuron comprise a computer evaluator task which loads said computer executable rules and loads values, if any, required to evaluate said computer executable rules from the process database and/or the input neurons and/or the output neurons and evaluates the rules one at a time.

19. The control system according to claim 9 further comprising means for storing values of measured process variables at spaced time intervals to maintain a process history.

20. The control system according to claim 19 wherein the process history is part of the process database.

21. The control system according to claim 19 wherein the rule associated with an input neuron for establishing the value of the input neuron takes as a parameter a value from the process history.

22. The control system according to claim 9 in which rules associated with input neurons comprise transfer, averaging, filtering, combining and/or switching rules.

23. The control system according to claim 9 in which rules associated with at least one output neuron comprise limit checking, weighing, scaling and/or ratioing rules.

24. The control system according to claim 9 in which rules associated with at least one output neuron comprise rules for selecting which set point should be modified in order to move the predicted value of the indirectly controlled variable to the desired value of the indirectly controlled variable.

25. The control system according to claim 9 further comprising means for storing values of measured process variables in a first-in-first-out buffer immediately following time spaced settling periods in which the rule associated with an input neuron and at least one output neuron are continuously updated to move the predicted value of the independently controlled variable to the desired value of the independently controlled variable so that a process history is maintained by the first-in-first-out buffer.

26. The control system according to claim 25, wherein a rule associated with at least one input neuron to establish the value of the input neuron takes as a parameter a value from the process history.

27. The control system according to claim 9 wherein an actual value of the indirectly controlled process variable cannot be determined until some period after it has been predicted.

28. The control system according to claim 9 wherein the means for training the neural network to generate a predicted value at an output neuron from the values at input neurons corresponding to directly controlled variables comprises a computer task for applying a plurality of training sets of input neuron and output neuron values captured from the process while it is in operation to the neural network during a training period.

29. The control system according to claim 28 wherein at least one input neuron value and at least one measured output neuron value are collected at different times.

30. The control system according to claim 9 wherein the value of the indirectly controlled variable is only determinable at a later time downstream from the control means which control directly controlled variables.

31. The control system according to claim 30 wherein the means for training the trainable neural network to generate a predicted value corresponding to the value of an indirectly controlled variable at an output neuron from the values at input neurons corresponding to measured process variables comprises a computer task for applying a plurality of training sets of input neuron values and output neuron values captured while the process is in operation to the input parts of the neural network during a training period said input neuron values and output neuron values being captured at different times.