Computer neural network regulatory process control system and method

• US 5,197,114 A
• Filed: 08/03/1990
• Issued: 03/23/1993
• Est. Priority Date: 08/03/1990
• Status: Expired due to Term

• Current Assignee: Rockwell Automation Technologies Incorporated (Allen-Bradley Co., Inc.)
• Original Assignee: E. I. du Pont de Nemours and Company (Corteva, Inc.)
• Inventors: Skeirik, Richard D.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/562,462
• Time in Patent Office: 963 Days
• Field of Search: 364/513, 364/148, 364/449, 364/150, 364/151, 364/164, 364/165, 364/807, 364/200, 364/474.15, 364/500, 364/501, 364/502, 364/503, 395/22, 395/23, 395/68, 395/906
• US Class Current: 706/23
• CPC Class Codes: G06N 3/042 Knowledge-based neural networks; Logical representations of neural networks
• CPC Class Codes: Y10S 706/906 Process plant

Abstract

A computer neural network regulatory process control system and method allows for the elimination of a human operator from real time control of the process. The present invention operates in three modes: training, operation (prediction), and retraining. In the training mode, training input data is produced by the control adjustment made to the process by the human operator. The neural network of the present invention is trained by producing output data using input data for prediction. The output data is compared with the training input data to produce error data, which is used to adjust the weight(s) of the neural network. When the error data is less than a preselected criterion, training has been completed. In the operation mode, the neutral network of the present invention provides output data based upon predictions using the input data. The output data is used to control a state of the process via an actuator. In the retraining mode, retraining data is supplied by monitoring the supplemental actions of the human operator. The retraining data is used by the neural network for adjusting the weight(s) of the neural network.

Claims

1. A computer neural network process control method adapted for predicting output data provided to an actuator used to control a process for producing a product having at least one product property, said method allowing the process to be controlled without a human operator, the computer neural network process control method comprising the steps of:

(1) configuring the neural network by specifying at least one input, at least one output, at least one training input, and at least one specified interval;

(2) training the neural network to produce a trained neural network comprising the substeps of;

(a) retrieving a first raw training input data;

(b) retrieving a second raw training input data;

(c) computing a corresponding first training input data based on said first raw training input data and said second raw training input data, said first training input data indicative of the action of a human operator of the process;

(d) retrieving a first input data;

(e) predicting a first output data using said first input data;

(f) computing a first error data in accordance with said first training input data and said first output data; and

(g) training the neural network to produce said trained neural network in accordance with said first error data;

(3) at said at least one specified interval, retrieving a second input data and predicting, with said trained neural network weights, second output data using said second input data;

(4) retrieving said second output data; and

(5) changing a state of the actuator in response to said second output data of step (4).

2. The computer neural network process control method of claim 1, wherein step (1) comprises a step of configuring the neural network by having at least two layers and/or up to full connectivity and/or at least one weight.

3. The computer neural network process control method of claim 1, wherein step (1) comprises a step of configuring the neural network to have feedback and/or feedforward connectivity.

4. The computer neural network process control method of claim 1, wherein said step (2)(e) comprises a step of predicting said first output data using said first input data and a transfer function of said configured neural network of step (1).

5. The computer neural network process control method of claim 1, wherein said step (2)(f) comprises a step of computing said first error data by subtracting said first training input data from said first output data.

6. The computer neural network process control method of claim 1, wherein said step (2)(g) comprises a step of training said neural network by using said first error data to adjust a first weight data.

7. The computer neural network process control method of claim 6, wherein said training step comprises a step of utilizing back propagation method.

8. The computer neural network process control method of claim 1 wherein said step (3) comprises a step of using said at least one specified interval selected in accordance with a time constant for control of the process.

9. The computer neural network process control method of claim 1, further comprising a step of using a sensor responsive to the actuator, for outputting said first raw training input data used in step (2)(a) and said second raw training input data used in step (2)(b).

10. The computer neural network process control method of claim 1, wherein said retrieving steps 2(a), 2(b) and/or 2(d) further comprise the step of retrieving from an historical database.

11. The computer neural network process control method of claim 10, wherein step (2)(b) comprises a step of retrieving said second raw training input data from said historical database with an associated first timestamp.

12. The computer neural network process control method of claim 11, wherein step (2)(d) comprises a step of retrieving said first input data from said historical database using said associated first timestamp.

13. The computer neural network process control method of claim 11, further comprising a step of using a clock for producing said associated first timestamp of step (2)(b).

14. The computer neural network process control method of claim 11, wherein step (2)(d) further comprises a step of retrieving said first input data with an associated second timestamp selected in accordance with said associated first timestamp.

15. The computer neural network process control method of claim 1, wherein step (2) further comprises a step

(a)(i) of using a human operator to make an adjustment to a sate of the actuator,

wherein said first training input data is indicative of said adjustment.

16. The computer neural network process control method of claim 1, further comprising,

a step (5) of retrieving a third raw training input data with a corresponding second input data;

and

a step (6) of retraining said trained neural network in accordance with a second training input data computed using said second raw training input data, a second error data computed using said second training input data and a second output data predicted using said corresponding second input data.

17. The computer neural network process control method of claim 16, wherein said step (6) further comprises a step of stopping the use of step (4) for controlling the process when said second error data exceeds a metric.

18. The computer neural network process control method of claim 16, wherein step (5) of retrieving further comprises a step of retrieving past output data.

19. The computer neural network process control method of claim 1, wherein step (2)(a) comprises a step of retrieving the state of an actuator as said first raw training input data.

20. The computer neural network process control method of claim 1 wherein step (2)(b) comprises a step of retrieving the state of an actuator as said second raw training input data.

21. The computer neural network process control method of claim 1, wherein step (2)(d) comprises a step of using an aim value for a product property value as said first input data.

22. The computer neural network process control method of claim 21, wherein step (1) of configuring further comprises a step of specifying said at least one specified interval to expire whenever a change in said aim value occurs.

23. The computer neural network process control method of claim 1, wherein step (2)(d) comprises a step of using an aim value for a process condition as said first input data.

24. The computer neural network process control method of claim 23, wherein step (1) of configuring further comprises a step of specifying said at least one specified interval to expire whenever a change in said aim value occurs.

25. The computer neural network process control method of claim 1, wherein said first input data of step (2)(d) of retrieving comprises the step of retrieving an associated second timestamp.

26. The computer neural network process control method of claim 25, further comprising a step of using a lab to provide said first input data and to provide said associated second timestamp indicating the time at which a lab sample was taken from the process and was used to produce said first input data.

27. The computer neural network process control method of claim 1, further comprising the steps:

(1)(a) operating the process with one or more sensors connected to sense process conditions and produce process condition measurements, and an actuator connected to change a controllable process state;

(1)(b) sensing the state of said actuator to produce an actuator state measurement; and

wherein step (2) further comprises the step of retrieving said actuator state measurement as said first raw training input data and/or said second raw training input data.

28. A computer neural network process control system adapted for predicting output data provided to an actuator used to control a process for producing a product having at least one product property, said system allowing the process to be controlled without a human operator, the computer neural network process control system comprising:

(a) a neural network, responsive to a first input data indicative of sensor or aim or lab data, comprising;

(1) training means, responsive to a first training input data indicative of the action of a human operator of the process, for training said neural network in accordance with a training step having said first input data and said first training input data to produce trained neural network weights, and

(2) predicting means using said trained neural network weights for predicting the output data in accordance with a second input data indicative of sensor or aim or lab data; and

(b) an actuator, responsive to said predicting means, comprising,

(1) sending means for adjusting a state of the actuator in accordance with the output data without a setpoint adjustment by a controller, and

(2) adjustment means for accepting a change made by said human operator to said state of the actuator.

29. The computer neural network process control system of claim 28, wherein said neural network further comprises computing means for computing error data in accordance with said first training input data and said first input data, said error data used by said training means to train said neural network.

30. The computer neural network process control system of claim 29, wherein said predicting means is connected to retrieve said weights for predicting the output data, and

wherein said training means is connected to modify said weights in accordance with said error data of said computing means.

31. The computer neural network process control system of claim 29, wherein said neural network comprises a software system adapted to run on a computer platform.

32. The computer neural network process control system of claim 28, further comprising lab means for receiving a sample from the process, which said lab means provides said first input data based on said sample.

33. The computer neural network process control system of claim 28, wherein said training means further comprises:

(i) storage means for saving a past raw training input data value indicative of a past value of said state of the actuator;

(ii) retrieving means for retrieving a current raw training input value indicative of a current value of said state of the actuator;

(iii) computing means for computing a first training input data indicative of a change between said current raw training input value and said past raw training input data value and for saving said current raw training input value in said storage means; and

(iv) weight adjusting means, responsive to said first training input data, for training said neural network weights in accordance with a training set having said first input data and said first training input data.

34. The computer neural network process control system of claim 33, wherein said predicting means further comprises:

(a) output data storage means for storing an output data value;

wherein said predicting means is connected to store the output data in said output data storage means; and

wherein said training means further comprises,

(v) comparing means for retrieving the output data in said output data storage means, and connected to execute said weight adjusting means when said first training input data is different from the output data.

35. The computer neural network process control system of claim 28, further comprising storage means for storing said first training input data as a past operator action value, and

wherein said training means is connected to retrieve said past operator action value as said first input data.

36. The computer neural network process control system of claim 28, further comprising the storage means for storing the output data as a control action value, and

wherein said training means is connected to retrieve said control action value as said first input data.

37. The computer neural network process control system of claim 28, further comprising:

(c) a sensor for sensing said state of the actuator and generating an actuator state measurement; and

wherein said training means is connected to retrieve said actuator state measurement to compute said first training input data.

38. The computer neural network process control system adapted for predicting output data provided to an actuator used to control a process of producing a product having at least one product property, said system allowing the process to be controlled without a human operator, the computer neural network process control system comprising:

(a) a neural network, responsive to a first input data indicative of sensor or aim or lab data, comprising,

(1) training means, responsive to a first training input data indicative of the action of a human operator of the process, for training said neural network in accordance with a first training set having said first input data and said first training input data to produce trained neural weights, and

(2) predicting means using said trained neural network weights for predicting the output data in accordance with a second input data indicative of sensor or aim or lab data;

(b) an actuator, responsive to said predicting means, comprising,

(1) sending means for adjusting a state of the actuator in accordance with the output data without a setpoint adjustment by a controller, and

(2) adjustment means for accepting a change made by said human operator to said state of the actuator; and

(c) an historical database comprising,

(1) storing means for storing said first training input data with an associated first timestamp, and for storing said first input data indicated by said associated first timestamp, and

(2) retrieving means, responsive to said storing means, for retrieving said first training set comprising said first training input data and said first input data indicated by said associated first timestamp, and connected to provide said first training set to said training means.

39. The computer neural network process control system of claim 38, further comprising lab means for receiving a sample from the process, which said lab means provides said first input data based on said sample.

40. The computer neural network process control system of claim 39, further comprising clock means, associated with said lab means, for providing said associated first timestamp to said storing means.

41. The computer neural network process control system of claim 40, further comprising sensor means, responsive to said neural network means, for providing said first input data.

42. The computer neural network process control system of claim 38, further comprising sensor means, responsive to said neural network means, for providing said first input data.

43. The computer neural network process control system of claim 42, further comprising clock means for providing an associated second timestamp of said first input data to said historical database, and

wherein said retrieving means selects said first input data indicated by said associated first timestamp using said associated second timestamp.

44. The computer neural network process control system of claim 38, further comprising output means for storing the output data in said historical database.