Method and apparatus for training a system model with gain constraints using a non-linear programming optimizer

• US 20060224534A1
• Filed: 04/03/2006
• Published: 10/05/2006
• Est. Priority Date: 05/06/1996
• Status: Active Grant

• Current Assignee: Rockwell Automation Technologies Incorporated (Rockwell Automation, Inc.)
• Original Assignee: Rockwell Automation, Inc.
• Inventors: Piche, Stephen, Hartman, Eric Jon, Gerules, Mark
• Granted Patent: US 7,624,079 B2
• Time in Patent Office: Days
• US Class Current: 706/15
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: G05B 13/042 in which a parameter or coefficient is automatically adjusted to optimise the performance
• CPC Class Codes: G05B 13/048 using a predictor
• CPC Class Codes: G05B 17/02 electric
• CPC Class Codes: G06N 3/02 Neural networks

Abstract

Method and apparatus for training a system model with gain constraints. A method is disclosed for training a steady-state model, the model having an input and an output and a mapping layer for mapping the input to the output through a stored representation of a system. A training data set is provided having a set of input data u(t) and target output data y(t) representative of the operation of a system. The model is trained with a predetermined training algorithm which is constrained to maintain the sensitivity of the output with respect to the input substantially within user defined constraint bounds by iteratively minimizing an objective function as a function of a data objective and a constraint objective. The data objective has a data fitting learning rate and the constraint objective has constraint learning rate that are varied as a function of the values of the data objective and the constraint objective after selective iterative steps.

Claims

1. A method, comprising:

providing a model, wherein the model comprises a representation of a plant or process, wherein the model comprises one or more inputs and one or more outputs, wherein the model comprises one or more gains, and wherein each gain comprises a respective partial derivative of an output with respect to a respective input;

specifying one or more gain constraints, wherein each of the one or more gain constraints corresponds to a respective gain; and

training the model with a non-linear programming optimizer subject to the one or more gain constraints.

2. The method of claim 1, wherein said training the model comprises:

a) receiving training input data representative of inputs to the plant or process;

b) executing the model to generate predicted training output data for the plant or process corresponding to the training input data;

c) comparing the predicted training output data to training output data representative of output of the plant or process to determine an error;

d) updating the model in accordance with the determined error; and

e) performing a)-d) in an iterative manner to minimize the error.

3. The method of claim 2, wherein said training the model further comprises optimizing a specified objective function.

4. The method of claim 1, wherein the one or more gain constraints comprise one or more of:

at least one hard constraint, comprising a strict limitation on a corresponding at least one gain used in said training the model;

or

at least one soft constraint, comprising a weighted penalty with respect to a corresponding at least one gain used in said training the model.

5. The method of claim 1, wherein at least one of the gain constraints comprises one or more of:

a lower bound on the corresponding gain;

or

an upper bound on the corresponding gain.

6. The method of claim 1, wherein the model comprises a neural network.

7. The method of claim 1, the method further comprising:

providing plant or process input data to the model;

executing the model to generate predicted plant or process output based on the plant or process input data; and

wherein the predicted plant or process output is useable to manage the plant or process.

8. The method of claim 1, the method further comprising:

specifying a desired behavior of the plant or process;

optimizing the model with one or more other optimizers to determine one or more input values that result in the desired behavior of the plant or process; and

wherein the one or more input values are usable to operate the plant or process to produce the desired behavior.

9. The method of claim 8, further comprising:

providing the one or more input values as input to the plant or process;

operating the plant or process in accordance with the one or more input values to produce the desired behavior.

10. The method of claim 8, wherein the one or more other optimizers comprise the non-linear programming optimizer.

11. The method of claim 8, the method further comprising:

the one or more other optimizers optimizing a specified objective function with respect to the model.

12. The method of claim 11, wherein said optimizing the specified objective function comprises:

optimizing the specified objective function subject to one or more constraints.

13. The method of claim 12, wherein the one or more constraints comprise one or more of:

at least one hard constraint, wherein the at least one hard constraint is strictly satisfied during said optimization;

or

at least one soft constraint, comprising a weighted penalty used in said optimizing the specified objective function.

14. The method of claim 12, wherein the one or more constraints comprise one or more soft constraints, and wherein the one or more other optimizers are operable to optimize the plant or process via gradient descent.

15. The method of claim 1, wherein the model comprises a steady state model.

16. The method of claim 1, wherein the model comprises a dynamic model.

17. The method of claim 1, wherein said training comprises:

training the model using one or more of;

one or more other constraints;

or one or more other optimizers.

18. A computer-accessible memory medium, wherein the memory medium comprises program instructions executable to:

implement a model, wherein the model comprises one or more inputs and one or more outputs, wherein the model comprises a representation of a plant or process, wherein the model comprises one or more gains, and wherein each gain comprises a respective partial derivative of an output with respect to a respective input;

specify one or more gain constraints, wherein each of the one or more gain constraints corresponds to a respective gain; and

training the model with a non-linear programming optimizer subject to the one or more gain constraints.

19. A system, comprising:

means for providing a model, wherein the model comprises one or more inputs and one or more outputs, wherein the model comprises a representation of a plant or process, wherein the model comprises one or more gains, and wherein each gain comprises a respective partial derivative of an output with respect to a respective input;

means for specifying one or more gain constraints, wherein each of the one or more gain constraints corresponds to a respective gain; and

training the model with a non-linear programming optimizer subject to the one or more gain constraints.

20. A system, comprising:

a model, representing a plant or process, the model comprising;

one or more inputs;

one or more outputs; and

one or more gains, wherein each gain comprises a respective partial derivative of an output with respect to a respective input; and

a non-linear programming optimizer, coupled to the model;

wherein the model is constrained by one or more gain constraints, wherein each of the one or more gain constraints corresponds to a respective gain; and

wherein the non-linear programming optimizer is operable to train the model subject to the one or more gain constraints.