Process optimization using a neural network

• US 5,671,335 A
• Filed: 01/03/1994
• Issued: 09/23/1997
• Est. Priority Date: 05/23/1991
• Status: Expired due to Fees

• Current Assignee: Allen-Bradley Company (Rockwell Automation, Inc.)
• Original Assignee: Allen-Bradley Company (Rockwell Automation, Inc.)
• Inventors: Gasperi, Michael L., Davis, Gerald Wesley
• Primary Examiner(s): Davis, George B.
• Application Number:
  US08/707,191
• Time in Patent Office: 1,359 Days
• Field of Search: 395/23, 395/24
• US Class Current: 706/25
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent

Abstract

An input to a complex multi-input process, such as injection molding, is optimized to produce a target output from that process through the use of a neural network trained to that process. A trial input is forward-propagated through the neural network and the output of the network compared to the target output. The difference is back-propagated through the network to determine an input error value in the network. This error value is used to correct the trial input. This correction process is repeated until the trial input produces the target output to within a predetermined degree of accuracy.

Claims

1. A method of determining an optimized input to produce a target output in a complex multi-input process, the process having a teaching set of historical inputs and outputs, the method using a neural network having input, hidden, and output neurons, each neuron having an activation and having weighted interconnections with other neurons, where the neural network has been trained according to the teaching set to establish the weights of the interconnections between neurons and to produce a trained neural network comprising the steps of:

a) presenting a trial input to the input neurons of the trained neural network;

b) forward-propagating the trial input to determine the activations of the output neurons of the trained neural network;

c) presenting the target output to the output neurons of the trained neural network;

d) back-propagating the difference between the activations of the output neurons and the target output to compute an input error value for the input neurons of the trained neural network;

e) adding a factor of the input error value to the trial input to create a modified trial input; and

f) inputting the modified trial input to the complex multi-input process.

2. The method recited in claim 1 wherein at step (d) the input error value is also multiplied by a correction rate ε

'"'"'.

3. The method recited in claim 1 wherein the neural network is a feed forward network.

4. The method recited in claim 1 including the step of:

repeating steps (c) through (f) while the difference of step (d) is greater than a predetermined value.

5. An apparatus for optimizing a trial input to produce a target output in a complex multi-input process, the apparatus comprising:

a neural network having an input and output layer of neurons having activation and, for receiving the trial input at the input layer and forward-propagating the trial input to produce an output value at the output layer;

a comparing means communicating with the output layer for producing an output error value dependant on the difference between the output value and the target output;

the neural network also for receiving the output error value at the output layer and back-propagating the output error value to produce an input error value at the input layer; and

a trial input means communicating with the input layer for scaling the trial input by adding a factor of the input error value to the trial input.

6. The apparatus recited in claim 5 wherein the modification means modifies the trial input, tr(i);

by adding a value dtr(i) to the trial input, where;

dtr(i)=ε

'"'"'δ

(1,i)

where ε

'"'"' is a correction rate and δ

(1,i) is the input error value.