Apparatus and method for controlling a process using a trained parallel distributed processing network

• US 5,119,468 A
• Filed: 02/27/1991
• Issued: 06/02/1992
• Est. Priority Date: 02/28/1989
• Status: Expired due to Term

• Current Assignee: E. I. du Pont de Nemours and Company (Corteva, Inc.)
• Original Assignee: E. I. du Pont de Nemours and Company (Corteva, Inc.)
• Inventors: Owens, Aaron J.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/660,915
• Time in Patent Office: 461 Days
• Field of Search: 364/149, 364/150, 364/151, 364/164, 364/165, 364/513
• US Class Current: 706/23
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: Y10S 706/906 Process plant

Abstract

A parallel distributed processing network, previously trained to simulate a process, is responsive to input activations A to produce outputs Y. The activations are representative of the input parameters P of the process and the outputs are representative of the characteristics C of the product of the process. A set of goal signals G corresponding to the desired physical characteristics D of the product are produced and used to generate a signal representative of the error E between the goal signals G and the outputs Y of the network. The values of the updated inputs A_U of the network that are needed to drive the error E to a predetermined minimum are determined, and the actual physical input parameters P of the process changed to correspond to the updated inputs values A_U producing the minimum error. The present invention can be used in both a batch initialization situation and/or in an on-going process control situation.

Claims

1. A control system for both batch initialization and for on-going control of a physical process, the process being of the type in which inputs having physical parameters P are used to produce a product having physical characteristics C associated therewith, the control system comprising:

a parallel distributed processing network previously trained to simulate the process, the parallel distributed processing network being responsive to input activations A to produce outputs Y, the activations being representative of the input parameters P of the process and the outputs being representative of the characteristics C of the product of the process;

means for producing a set of goal signals G corresponding to the desired physical characteristics D of the product of the process;

means for generating a signal representative of the error E between the goal signals G and the outputs Y of the network;

means for determining the values of the updated inputs A^U of the parallel distributed processing network that are needed to drive the error E to a predetermined minimum, and

means for changing the actual physical input parameters P of the process to correspond to the updated inputs values A_U producing the minimum error.

2. A method of both batch initialization and on-going control of a physical process using a parallel distributed processing network previously trained to stimulate the process, the process being of the type in which inputs have predetermined physical parameters P produce a product having corresponding physical characteristics C, the method comprising the steps of:

(a) producing a set of goal signals G corresponding to the desired physical characteristics D of the product of the process

(b) determining the error E between the goal signals G and the outputs Y of the network

(c) determining the values of the updated inputs A_U of the parallel distributed processing network that are needed to drive the error E to a predetermined minimum, and thereafter

(d) changing the actual physical parameters P of the process to correspond to the updated inputs values A_U producing the minimum error.