Training a support vector machine with process constraints

• US 7,599,897 B2
• Filed: 05/05/2006
• Issued: 10/06/2009
• Est. Priority Date: 05/05/2006
• Status: Active Grant

• Current Assignee: IP3 2017, Series 200 of Allied Security Trust I (Allied Security Trust)
• Original Assignee: Rockwell Automation Technologies Incorporated (Allen-Bradley Co., Inc.)
• Inventors: Schweiger, Carl A., Hartman, Eric J., Johnson, W. Douglas, Sayyarrodsari, Bijan
• Primary Examiner(s): STARKS, WILBERT L
• Application Number:
  US11/418,971
• Publication Number:
  US 20070282766A1
• Time in Patent Office: 1,250 Days
• Field of Search: 701/4, 706/15, 706/52, 707/10
• US Class Current: 706/15
• CPC Class Codes: G05B 13/048 using a predictor
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]

Abstract

System and method for training a support vector machine (SVM) with process constraints. A model (primal or dual formulation) implemented with an SVM and representing a plant or process with one or more known attributes is provided. One or more process constraints that correspond to the one or more known attributes are specified, and the model trained subject to the one or more process constraints. The model includes one or more inputs and one or more outputs, as well as one or more gains, each a respective partial derivative of an output with respect to a respective input. The process constraints may include any of: one or more gain constraints, each corresponding to a respective gain; one or more Nth order gain constraints; one or more input constraints; and/or one or more output constraints. The trained model may then be used to control or manage the plant or process.

Claims

1. A method, comprising:

providing a model, wherein the model comprises a representation of a plant or process implemented with a support vector machine (SVM), wherein the model comprises one or more inputs and one or more outputs, and wherein the plant or process has one or more known attributes;

specifying one or more process constraints that correspond to the one or more known attributes; and

training the model subject to the one or more process constraints.

2. The method of claim 1, 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, and wherein the one or more process constraints comprise:

one or more gain constraints, wherein each of the one or more gain constraints corresponds to a respective gain.

3. The method of claim 1, 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, and wherein the one or more process constraints comprise:

one or more Nth order gain constraints, wherein N is a positive integer.

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

one or more inputs;

one or more outputs;

or

one or more Nth order gains.

5. The method of claim 1, wherein said providing the model comprises providing a primal formulation of the model.

6. The method of claim 1, wherein said providing the model comprises providing a dual formulation of the model.

7. 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) receiving training output data representative of outputs to the plant or process; and

c) optimizing a specified objective function subject to the one or more process constraints and one or more definitional constraints.

8. The method of claim 7, wherein said training the model comprises training the model with an optimizer utilizing a sequential minimal optimization (SMO) algorithm.

9. The method of claim 7, wherein said training the model comprises training the model with a non-linear programming (NLP) optimizer.

10. The method of claim 1, wherein the one or more attributes of the plant or process correspond to respective one or more parameters of the model, wherein the one or more process constraints comprise one or more of:

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

or

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

11. The method of claim 10, wherein the one or more process constraints comprise one or more soft constraints, and wherein said training the model comprises optimizing the model via gradient descent.

12. The method of claim 1, wherein at least one of the process constraints comprises a function of one or more model parameters.

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

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

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

training the model using one or more of;

one or more other constraints; and

one or more other optimizers.

16. 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.

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

specifying a desired behavior of the plant or process;

optimizing the model with an optimizer 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.

18. The method of claim 17, 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.

19. The method of claim 18, wherein the optimizer comprises a non-linear programming (NLP) optimizer.

20. The method of claim 18, wherein the optimizer utilizes a sequential minimal optimization (SMO) algorithm.

21. The method of claim 18, the method further comprising:

the optimizer optimizing a specified objective function with respect to the model.

22. The method of claim 21, wherein said optimizing the specified objective function comprises:

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

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

implement a model with a support vector machine (SVM), wherein the model comprises one or more inputs and one or more outputs, wherein the model comprises a representation of a plant or process, and wherein the plant or process has one or more known attributes;

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

training the model subject to the one or more process constraints.

24. A system, comprising:

means for providing a model, wherein the model comprises representation of a plant or process implemented with a support vector machine (SVM), wherein the model comprises one or more inputs and one or more outputs, and wherein the plant or process has one or more known attributes;

means for specifying one or more process constraints that correspond to the one or more known attributes; and

means for training the model subject to the one or more process constraints.

25. A system, comprising:

a model, representing a plant or process, wherein the model is implemented by a support vector machine (SVM), and wherein the plant or process has one or more known attributes, the model comprising;

one or more inputs;

one or more outputs; and

an optimizer, coupled to the model;

wherein the model is constrained by one or more process constraints that correspond to the one or more known attributes; and

wherein the optimizer is operable to train the model subject to the one or more process constraints.