Universal process control using artificial neural networks

• US 5,159,660 A
• Filed: 10/11/1990
• Issued: 10/27/1992
• Est. Priority Date: 08/09/1990
• Status: Expired due to Term

• Current Assignee: Western Thunder
• Original Assignee: Western Thunder
• Inventors: Lu, Yong-Zai, Cheng, George S., Manoff, Michael
• Primary Examiner(s): Fleming, Michael R.
• Assistant Examiner(s): Downs, Robert W.
• Application Number:
  US07/596,056
• Time in Patent Office: 747 Days
• Field of Search: 364/513, 364/148-166, 364/178, 395/22
• US Class Current: 706/23
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: Y10S 706/906 Process plant

Abstract

Adaptive control for a wide variety of complex processes is provided by an ANN controller with input and hidden layers having a plurality of neurons, and an output layer with a single neuron. The inputs to the ANN are a time sequence of error values, and the neuron paths are weighted as a function of these error values and the present-time process output. The present-time error value may be added to the output layer of the ANN to provide faster response to sudden input changes. The controller of this invention can efficiently handle processes with nonlinear, time-varying, coupled and variable-structure behaviors as well as process parameter and/or structure uncertainties. Large steady-state gains in the process can be compensated by attenuating the ANN block output.

Claims

1. A control system for controlling an open-loop stable process, said process having a process output represented by a process output signal, and being responsive to a control signal for changing said process output as a function of said control signal, said system comprising:

a) setpoint selection means for generating a setpoint signal representative of a desired value of said process output;

b) comparator means connected to said setpoint selection means and said process output for comparing said process output signal to said setpoint signal and for deriving from said comparison an error signal representative of the difference between said process output signal and said setpoint signal;

c) sampling means connected to said comparator means for producing time-spaced samples of said error signal at selected time intervals; and

d) control means connected to said sampling means and said process, said control means being responsive to said error signal samples for generating said control signal,

e) said control means being an artificial neural network, said network including a plurality of layers of neurons connected so as to interact with each other in accordance with variable weights, a first of said layers having a plurality of neuron inputs, and another of said layers having a network output; and

f) said sampling means being so connected to said network as to simultaneously apply to individual ones of said neuron input signals representative of individual ones of said time-spaced error signal samples, and also to simultaneously vary individual ones of said weights as a function of individual ones of said time-spaced error signal samples.

2. The system of claim 1, in which said artificial neural network includes;

i) an input layer composed of a first plurality of neurons each having an input representative of one of said error signal samples;

ii) at least one hidden layer composed of a second plurality of neurons, each of said first plurality of neurons having a separate output additively connected to the input of each of said second plurality of neurons; and

iii) an output layer composed of only one neuron, each of said second plurality of neurons having a separate output additively connected to the input of said output-layer neuron;

iv) each output of each of said neurons being an individually weighted function of that neuron'"'"'s input.

3. The system of claim 1, in which said network inputs to include functions of said error signal taken at discrete points in time, and in which one of said discrete points in time is present time, said system further comprising means for adding to the output of said network a function of said present-time error signal.

4. The system of claim 1, in which the inputs and weights of said artificial neural network are functions of said error signal at discrete times, and in which said weights are also a function of a signal representative of the present-time value of said process output.

5. The system of claim 4, in which said functions of said error signal and said process-output-representing signal are the respective normalization factors of said signals.

6. The system of claim 1, in which said process has a dominant time constant, and said time intervals, in seconds, are selected to be substantially equal to one twentieth of said dominant time constant.

7. The system of claim 1, further comprising means for iteratively updating said weights for each successive one of said time intervals, the weights associated with the network paths from said hidden layer to said output layer being first updated according to the formula

8. The system of claim 7, in which the constant η

is a selectable learning rate.

9. The system of claim 8, in which said learning rate is selected from a range of substantially 1 to substantially 3.

10. The system of claim 9, in which said learning rate is selected to be substantially 2.

11. The system of claim 1, further comprising means for attenuating the output of said network to compensate for high DC gain in said process.