Method and apparatus for analyzing a neural network within desired operating parameter constraints

• US 5,781,432 A
• Filed: 12/04/1996
• Issued: 07/14/1998
• Est. Priority Date: 03/02/1993
• Status: Expired due to Term

• Current Assignee: Rockwell Automation Technologies Incorporated (Rockwell Automation, Inc.)
• Original Assignee: Pavilion Technologies Incorporated (Rockwell Automation, Inc.)
• Inventors: Hartman, Eric Jon, Keeler, James David
• Primary Examiner(s): Elmore, Reba I.
• Assistant Examiner(s): Garland, Steven R.
• Application Number:
  US08/759,539
• Time in Patent Office: 587 Days
• Field of Search: 364/148-162, 395/2, 395/22, 395/24
• US Class Current: 700/44
• CPC Class Codes: F23N 2223/44 Optimum control
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: G05B 13/0285 using neural networks and fuzzy logic

Abstract

A distributed control system (14) receives on the input thereof the control inputs and then outputs control signals to a plant (10) for the operation thereof. The measured variables of the plant and the control inputs are input to a predictive model (34) that operates in conjunction with an inverse model (36) to generate predicted control inputs. The predicted control inputs are processed through a filter (46) to apply hard constraints, the values of which are received from a control parameter block (22). During operation, predetermined criterion stored in the control parameter block (22) are utilized by a cost minimization block (42) to generate an error control signal which is minimized by the inverse model (36) to generate the control signals. The system works in two modes, an analyze mode and a runtime mode. In the analyze mode, the predictive model (34) and the inverse model (36) are connected to either training data or simulated data from the analyzer (30) and the operation of the plant (10) evaluated. The values of the hard constraints in filter (46) and the criterion utilized for the cost minimization (42) can then be varied to change the constraints on the control signals input to the control network, the predicted output of the predictive model (34) and the hard constraints stored in the filter (46). Cost coefficients can be utilized as the criterion to set the input values in accordance with predetermined cost constraints.

Claims

1. A control system for controlling the operation of a plant, comprising:

a predictive network having an input layer for receiving control inputs for the plant, an output layer for outputting predicted outputs representing a prediction of the output of the plant and a mapping layer for mapping said input layer to said output layer through a stored representation of the plant;

an optimality device for receiving at least two of said predicted outputs or said control inputs as input properties as either the combination of at least one of said predicted outputs and one of said control inputs or the combination of at least two of said predicted outputs and operable to apply an optimality function to each of said input properties, the combination of said optimality functions defining a system optimality value, each of said optimality functions operating such that extremization thereof corresponds to a user-defined desired behavior of said associated input property, said user-defined behavior not the same for each of said optimality functions, said optimality function for each of said input properties having a user-defined behavior that is comprised of an associated weighting factor and a desired value, wherein said weighting factor is applied to a function of the difference between the predicted value of said associated input property and said desired value and said weighting factor is a variable value that is a function of predetermined limits, such that when said actual value of said input property exceeds said limit, the value of said weighting factor changes; and

a predictive system for generating updated control inputs that extremize said system optimality value in accordance with said optimality function.

2. The control network of claim 1, wherein said optimality device is operable to receive said predicted outputs as said input properties such that an associated optimality function can be applied to each of the received control inputs and predicted output.

3. The control network of claim 1, wherein said predictive system is an iterative system that is operable to iteratively change said received inputs to said optimality device to provide a new updated value of said control inputs in accordance with a predetermined procedure for extremizing said system optimality value.

4. The control network of claim 1, wherein said predictive system comprises an inverse model of the plant and substantially corresponding to said predictive network, and having:

an input layer;

an output layer;

a plurality of hidden layers for mapping said input layer and said output layer through a stored representation of said plant; and

said inverse model operable to operate in an inverse mode to receive on said output layer said system optimality value and back propagate said system optimality value through said plurality of hidden layers to said input layer in accordance with a search technique to generate said updated control inputs.

5. The control network of claim 1, and further comprising a filter device for constraining said control inputs in accordance with user-defined constraints, such that limits are placed on the values that can constitute said control inputs.

6. The control network of claim 5, wherein said user-defined constraints comprise hard limits that limit the value of said updated control inputs from exceeding a predetermined value.

7. The control network of claim 5, wherein said user-defined constraints comprise rate-of-change constraints that define the maximum incremental range that can be applied to the difference between said received inputs and said updated control inputs.

8. The control network of claim 5, wherein said user-defined constraints comprise combinatorial constraints, wherein a predetermined value is defined as the combination of selected ones of said control inputs, such that said selected ones of said control inputs have the values thereof limited so as not to violate said combinatorial constraints.

9. A control system for controlling the operation of a plant that receives control inputs and provides measurements of state variables associated with the operation of the plant and provides an output, comprising:

a distributed control system for distributing generated control inputs to the plant as the received control inputs;

a runtime predictive control system for receiving said generated control inputs and predicting updated control inputs in accordance with a predetermined desired plant behavior; and

a filter for placing predetermined constraints on said updated control inputs to provide constrained control inputs, such that limits are placed on the values that can constitute said control inputs, said constrained control inputs provided to said distributed control system for distribution to the plant as said received control inputs and also for distribution of said constrained inputs to said runtime predictive control network, wherein said predetermined constraints comprise combinatorial constraints, wherein a predetermined relationship or range of values is defined as the combination of selected ones of said constrained control inputs, such that said selected ones of said constrained control inputs have the values thereof limited so as not to violate said predetermined relationship.

10. The control system of claim 9, and further comprising an analysis system for determining said desired plant behavior and having:

a simulated predictive control system that is a substantial representation of said runtime predictive control system;

a system for generating analysis control inputs;

a behavior generation device for generating simulated desired plant behavior;

said analysis system for allowing a user to manipulate said simulated desired plant behavior and controlling said system for generating analysis control inputs to generate updated control inputs to cause said simulated predictive control system to operate in accordance with said desired behavior, said analysis system operable to allow the user to observe said desired behavior; and

a system for applying said manipulated desired plant behavior after manipulation thereof to said runtime predictive control system as said predetermined desired plant behavior.

11. The control system of claim 10, wherein said analysis system includes a predictive system model having an input layer for receiving said generated control inputs, an output layer for providing a predictive output representing a prediction of the output of the plant and a plurality of mapping layers for mapping said input layer to said output layer through a stored representation of the plant.

12. The control system of claim 9, wherein said runtime predictive control system comprises:

a predictive system model having an input layer for receiving said control inputs that are input to the plant, an output layer for outputting a predicted output representing a prediction of the output of the plant and a plurality of mapping layers for mapping said input layer to said output layer through a stored representation of the plant;

an optimality device for receiving at least two of said predicted outputs or said control inputs as input properties as either the combination of at least one of said predicted outputs and one of said control inputs or the combination of at least two of said predicted outputs and operable to apply an optimality function to each of said input properties, the combination of said optimality functions defining a system optimality value, each of said optimality functions operating such that extremization thereof corresponds to a user-defined desired behavior of said associated input property; and

a predictive system for generating updated control inputs that extremize said system optimality value.

13. The control system of claim 12, wherein said predictive system is an iterative system that is operable to iteratively change said received inputs to said optimality device to provide a new updated value of said control inputs in accordance with a predetermined procedure for extremizing said system optimality value.

14. The control system of claim 12, wherein said predictive system model comprises an inverse model of the plant and substantially corresponding to said predictive system model, and having:

an input layer;

an output layer;

a plurality of hidden layers for mapping said input layer and said output layer through a stored representation of said plant; and

said inverse model operable to operate in an inverse mode to receive on said output layer said system optimality value and back propagate said system optimality value through said hidden layer to said input layer in accordance with a search technique to generate said updated inputs.

15. The control system of claim 9, wherein said runtime predictive neural network is operable to receive as inputs the measured state variables of the plant in addition to said control inputs.

16. The control system of claim 15, wherein said runtime predictive control system comprises:

a predictive system model having an input layer for receiving said control inputs that are input to the plant and said measured state variables, an output layer for outputting a predicted output representing a prediction of the output of the plant and a plurality of mapping layers for mapping said input layer to said output layer through a stored representation of the plant;

an optimality device for receiving at least two of said predicted outputs as input properties and operable to apply an optimality function to each of said input properties, the combination of said optimality functions defining a system optimality value, each of said optimality functions operating such that the extremum thereof corresponds to a user-defined desired behavior of said associated input property; and

a predictive system for generating updated control inputs that extremize said system optimality value.

17. The control system of claim 9, wherein said predetermined constraints are hard constraints such that said constrained inputs have values that are limited relative to the value of said updated control inputs, said constrained inputs prevented from exceeding a predetermined value.

18. The control system of claim 9, wherein said predetermined constraints comprise rate-of-change constraints that define the maximum incremental change that can be applied to the difference between said constrained input and said received inputs received by said runtime predictive control network.

19. A control system for controlling the operation of a plant that receives control inputs and provides measurements of state variables associated with the operation of the plant and provides an output, comprising:

a distributed control system for distributing generated control inputs to the plant as the received control inputs;

a storage device for storing control system operating parameters;

a runtime predictive control system for receiving said control inputs and predicting updated control inputs for distribution by said distributed control system to the plant in accordance with said stored control system operating parameters to achieve a predetermined desired plant behavior;

an analysis system having a stored representation of said predictive control system and operable to simulate the operation of said runtime predictive control system with said stored representation thereof with internally generated simulated control inputs to determine a desired set of control system operating parameters; and

said analysis system having an output device for downloading said determined desired set of control system operating parameters to said storage device for storage as said stored control system operating parameters.

20. The control system of claim 19, wherein said analysis system comprises:

a simulated predictive control system that is a substantial representation of said runtime predictive control system;

a system for generating analysis control inputs;

a control system operating parameter generation device for generating simulated control system operating parameters; and

a manipulation device for allowing a user to manipulate said simulated predictive control network and said system for generating analysis control inputs to generate updated control inputs to cause said simulated predictive control system to operate in accordance with said simulated control system operating parameters, said manipulation device operable to allow the user to observe the desired plant behavior.

21. The control system of claim 20, wherein said manipulation device includes a predictive system model having an input layer for receiving said generated control inputs, an output layer for providing a predictive output representing a prediction of the output of the plant and a mapping layer for mapping said input layer to said output layer through a stored representation of the plant.

22. The control system of claim 19, wherein said runtime predictive control system comprises:

a predictive system model having an input layer for receiving said control inputs that are input to the plant, an output layer for outputting a predicted output representing a prediction of the output of the plant and a plurality of mapping layers for mapping said input layer to said output layer through a stored representation of the plant;

an optimality device for receiving at least two of said predicted outputs or said control inputs as input properties as either the combination of at least one of said predicted outputs and one of said control inputs or the combination of at least two of said predicted outputs and operable to apply an optimality function to each of said input properties, the combination of said optimality functions defining a system optimality value, each of said optimality functions operating such that the extremum thereof corresponds to a user-defined desired behavior of said associated input property; and

a predictive system for generating updated control inputs that extremize said system optimality value.

23. The control system of claim 19, wherein said predictive system model is an iterative system that is operable to iteratively change said received inputs to said optimality device to provide a new updated value of said control inputs in accordance with a predetermined procedure for extremizing said system optimality value.

24. The control system of claim 23, wherein said predictive system comprises an inverse model of the plant and substantially corresponding to said predictive network, and having:

an input layer;

an output layer;

a plurality of hidden layers for mapping said input layer and said output layer through a stored representation of said plant; and

said inverse model operable to operate in an inverse mode to receive on said output layer said system optimality value and back propagate said system optimality value through said hidden layers to said input layer in accordance with a search technique to generate said updated control inputs.

25. The control network of claim 19, and further comprising a filter device for constraining said control inputs in accordance with user-defined constraints, such that limits are placed on the values that can constitute said control inputs.

26. The control network of claim 25, wherein said user-defined constraints comprise hard limits that limit the value of said updated control inputs from exceeding a predetermined value.

27. The control system of claim 25, wherein said user-defined constraints comprise rate-of-change constraints that define the maximum incremental range that can be applied to the difference between said received inputs and said updated control inputs.

28. The control system of claim 25, wherein said user-defined constraints comprise combinatorial constraints, wherein a predetermined value is defined as the combination of selected ones of said control inputs, such that said selected ones of said control inputs have the values thereof limited so as not to violate said combinatorial constraints.

29. A method for generating updated control inputs for controlling the operation of a plant, comprising the steps of:

receiving control inputs for the plant, which control inputs are operable to control the plant;

processing the received control inputs through a predictive system model having a stored representation of the plant and providing predicted outputs;

receiving at least two of the predicted outputs or the control inputs as input properties as either the combination of at least one of the predicted outputs and one of the control inputs or the combination of at least two of the predicted outputs;

applying a separate optimality function to each of the input properties, the combination of the optimality functions defining a system optimality value, each of the optimality functions operating such that extremization thereof corresponds to a user-defined desired behavior of the associated input property, the user-defined behavior not the same for each of the optimality functions, which user-defined behavior is comprised of an associated weighting factor and a desired value, wherein the weighting factor is applied to a function of the difference between the predicted value of the associated input property and desired value; and

predicting the updated control inputs that extremize the system optimality value.

30. The method of claim 29, wherein the step of predicting the updated control inputs comprises iteratively changing the received control inputs to which the optimality function is applied to provide a new updated value of the control inputs in accordance with a predetermined procedure for extremizing the system optimality value.

31. The method of claim 29, wherein the step of predicting the updated control inputs comprises:

providing a predictive model having an input layer, an output layer and a plurality of hidden layers for mapping the input layer to the output layer through a stored representation of the plant;

operating the predictive model in an inverse mode;

receiving on the output layer the system optimality value when the system model is in the inverse mode; and

backpropagating the system optimality value through the hidden layer to the input layer in accordance with a search technique to generate the updated control inputs.

32. The method of claim 29, and further comprising constraining the control inputs in accordance with user-defined constraints.

33. The method of claim 32, wherein the step of constraining the control inputs comprises hard-limiting the control inputs to a predetermined value such that the updated control inputs do not exceed the predetermined value.

34. The method of claim 32, wherein the step of constraining the control inputs comprises limiting the rate of change of the updated control inputs to define the maximum incremental change that can be applied to the difference between the received inputs that are received by the plant and the updated control inputs.

35. The method of claim 32, wherein the step of constraining the control inputs comprises applying combinatorial constraints to the updated control inputs, wherein a predetermined value is defined as the combination of selected ones of the updated control inputs, such that selected ones of the updated control inputs have the value thereof limited so as not to violate the combinatorial constraints.

36. A method for generating updated control inputs for controlling the operation of a plant that receives control inputs and provides measurements of state variables associated with the operation of the plant and provides an output, comprising the steps of:

providing a distributed control system that is operable to distribute generated control inputs to the plant as the received control inputs;

processing the generated control inputs through a runtime predictive control system to predict updated control inputs in accordance with a predetermined desired plant behavior; and

placing predetermined constraints on the updated control inputs to provide constrained control inputs, such that limits are placed on the values that can constitute the control inputs, the constrained control inputs provided to the distributed control system for distribution to the plant as the received control inputs and also for distribution of the constrained inputs to the runtime predictive control network, wherein the predetermined constraints comprise combinatorial constraints, wherein a predetermined relationship or range of values is defined as the combination of selected ones of the constrained control inputs, such that the selected ones of the constrained control inputs have the values thereof limited so as not to violate the predetermined relationship.

37. The method of claim 36, and further comprising the step of determining the desired plant behavior by the steps of:

providing a simulated predictive control system that is a substantial representation of the runtime predictive control system;

generating analysis control inputs;

processing the analysis control inputs through the simulated predictive control system;

generating simulated desired behavior;

manipulating the simulated desired behavior and controlling the step of generating analysis control inputs to generate updated control inputs to cause the simulated predictive control system to operate in accordance with the simulated desired behavior, the step of manipulating allowing a user to observe the simulated desired behavior; and

applying a manipulated desired behavior after manipulation to the runtime predictive control system as the predetermined desired plant behavior.