Method for steady-state identification based upon identified dynamics

• US 6,047,221 A
• Filed: 10/03/1997
• Issued: 04/04/2000
• Est. Priority Date: 10/03/1997
• Status: Expired due to Term

• Current Assignee: Rockwell Automation Technologies Incorporated (Rockwell Automation, Inc.)
• Original Assignee: Pavilion Technologies Incorporated (Rockwell Automation, Inc.)
• Inventors: Johnson, William D., Hartman, Eric, Piche, Stephen, Liano, Kadir, Gerules, Mark, Keeler, James David
• Primary Examiner(s): Gordon, Paul P.
• Assistant Examiner(s): Marc, McDieunel, Marc, McDieunel
• Application Number:
  US08/943,489
• Time in Patent Office: 914 Days
• Field of Search: 327/198, 327/185, 399/77, 700/44, 700/46, 700/71, 700/29
• US Class Current: 700/44
• CPC Class Codes: G05B 13/048 using a predictor
• CPC Class Codes: G05B 17/02 electric

Abstract

A method for modeling a steady-state network in the absence of steady-state historical data. A steady-state neural network can be tied by impressing the dynamics of the system onto the input data during the training operation by first determining the dynamics in a local region of the input space, this providing a set of dynamic training data. This dynamic training data is then utilized to train a dynamic model, gain thereof then set equal to unity such that the dynamic model is now valid over the entire input space. This is a linear model, and the historical data over the entire input space is then processed through this model prior to input to the neural network during training thereof to remove the dynamic component from the data, leaving the steady-state component for the purpose of training. This provides a valid model in the presence of historical data that has a large content of dynamic behavior. A single dynamic model is required for each output variable in a multi-input multi-output steady-state model such that for each output there is a separate dynamic model required for pre-filtering. They are combined in a single network made up of multiple individual steady-state models for each output. The dynamic model can be identified utilizing a weighting factor for the gain to force the dynamic gain of the dynamic model to the steady-state gain by weighting the difference thereof during optimization of the dynamic model. The steady-state model is optimized utilizing gain constraints during the optimization procedure such that the gain of the network is prevented from exceeding the gain constraints.

Claims

1. A method for training a steady-state non-linear model of a plant, comprising the steps of:

generating an input set of historical training data representative of the historical operation of the plant and having input data u(t) and output data y(t);

identifying a linear dynamic model of the plant representing the dynamics of the plant for a given output variable of the plant;

pre-filtering the historical input data through the dynamic model to impress the dynamics of the plant as defined by the linear dynamic model on the input data u(t); and

training a non-linear model of the plant with the pre-filtered input data u(t) to provide a stored representation of the steady-state operation in the non-linear model associated with the given output variable.

2. The method of claim 1 wherein the step of training the non-linear model comprises training a non-linear neural network having an input layer, an output layer and a hidden layer for storing a learned representation of the plant and for mapping the input layer to the output layer through the learned representation of the plant, which learned representation of the plant comprises the steady-state operation of the plant.

3. The method of claim 1 wherein the step of identifying comprises identifying a linear dynamic model of the plant representing the dynamics of the plant in a defined portion of the input space.

4. The method of claim 3 wherein the step of training the non-linear model comprises training a non-linear model of the plant with the pre-filtered data to provide a stored representation of the steady-state operation in the non-linear model over a substantial portion of the input space.

5. The method of claim 1 wherein the step of identifying a linear model of a plant comprises:

generating a set of dynamic data defining the dynamics of the plant over the defined portion of the input space over which the operation of the plant is linear;

defining a linear dynamic model; and

identifying the linear dynamic model utilizing the generated dynamic data over the defined portion of the input space such that the dynamics embodied within the identified linear dynamic model represent the dynamics of the plant over the entire input space.

6. The method of claim 5 and further comprising the step of normalizing the gain of the identified linear dynamic model to unity.

7. The method of claim 5 wherein the step of generating the set of dynamic data comprises performing a step test on the plant in a localized portion of input space from an initial steady-state value to a second value and then back to the initial steady-state value.

8. The method of claim 1 wherein the input set of historical training data has dynamics of the plant associated therewith.

9. A method for training a steady-state, non-linear model of a plant having multiple input variables and multiple output variables, comprised of the steps of:

generating an input of historical training data representative of the historical operations of the plant and having input data u(t) and output data y(t);

identifying a plurality of linear dynamic models, one for each output variable of the plant, each of which represents the dynamics of the plant for the given output variables of the plant;

pre-filtering the historical input data for each of the input variables through each of the dynamic models to impress the dynamics of the plant for the associated output variable with the associated dynamic model on the input data u(t); and

training a plurality of non-linear models in the plant, each non-linear model associated with one of the output variables, and each of the non-linear models filtered with the pre-filtered input data u(t) for the associated output variable to provide a stored representation of the steady-state operation of each of the output variables in the associated non-linear model.

10. The method of claim 9 and further comprising combining all of the non-linear models into a single combined model, such that a single multi-input multi-output model is provided, with each of the outputs having associated therewith one of the non-linear models.

11. A method 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 plant, comprising:

providing a training data set having a set of input data u(t) and target output data y(t) representative of the operation of a plant;

training the model with a predetermined training algorithm; and

constraining the training algorithm to maintain the sensitivity of the output with respect to the input substantially within user defined bounds.

12. The method of claim 11, wherein the model comprises a neural network having an input layer, an output layer and a hidden layer that provides the mapping layer operation.

13. The method of claim 11, wherein the training algorithm comprises a gradient based training algorithm for generating an error term during the operation of the training algorithm and minimizing this error term and wherein the step of constraining the training algorithm comprises minimizing the error term substantially within the user defined bounds.

14. The method of claim 13, wherein the training algorithm comprises a back propagation algorithm.

15. The method of claim 13, wherein the training algorithm minimizes an error E in accordance with the following:

##EQU32## where P is the number of training patterns, y^p (t) comprises the output of the model and y(t) comprises the training data set.

16. The method of claim 15, wherein the sensitivity of the outputs with respect to the inputs during training are constrained between a first lower sensitivity limit and a second upper sensitivity limit.

17. The method of claim 16, wherein the upper and lower sensitivity limits are hard constraints.

18. The method of claim 16, wherein the constraints are soft limits wherein a penalty factor is attributed to the training algorithm when the upper and lower limits of sensitivity are violated.

19. The method of claim 18, wherein the back propagation algorithm with soft constraints is in accordance with the following relationship:

##EQU33## where H represents a predetermined non-negative penalty function for violating the constraints, λ

represents a user defined weighting factor, G₁ represents a lower sensitivity limit, G_h represents an upper sensitivity limit, G represents the sensitivity of the outputs with respect to the inputs.

20. The method of claim 15, wherein the sensitivity comprises the gain of the model.

21. The method of claim 11, wherein the sensitivity is defined as a sensitivity of each output with respect to each input.

22. A method for developing a linear dynamic model of a plant, comprising the steps of:

creating a linear model for predicting a dynamic response y(t) from an input u(t) and having a dynamic gain k_d, which linear model can define the dynamic response of the plant;

generating a training data set of dynamic input data u_d (t) and output data y_d (t);

providing a user defined sensitivity value representing the sensitivity of the output of the plant with respect to the input of the plant at steady-state;

training the linear model on the training data set with a predetermined training algorithm; and

constraining the training algorithm with a scaled level of the difference between the dynamic gain and the user defined sensitivity such that an error generated with the training algorithm will increase as the difference between the dynamic gain and the user defined sensitivity increases.

23. The method of claim 22, wherein the training data set is generated over a localized region of the input data space.

24. The method of claim 23, wherein the step of generating the training data comprises step testing the plant by stepping the data from a steady-state value at u_i (t) for each value of the input.

25. The method of claim 22, wherein the step of providing a user defined sensitivity comprises providing a user defined sensitivity over the entire input space.

26. The method of claim 25, wherein the step of providing the user defined sensitivity comprises providing a non-linear predictive network that represents a predictive model of the plant and which has a steady-state gain K_ss, and wherein the average value of the gain K_ss, K_ss '"'"', represents the user defined sensitivity.

27. The method of claim 22, wherein the predetermined training algorithm comprises a gradient based training algorithm.

28. The method of claim 27, wherein the step of providing the user defined sensitivity comprises providing a non-linear network that represents a predictive model of the plant and which has a steady-state gain K_ss '"'"', and wherein the average value of the gain K_ss, K_ss '"'"', represents the user defined sensitivity.

29. The method of claim 28, wherein the training algorithm and the step of constraining operate in accordance with the following relationship:

##EQU34## where J=the error that is to be minimized and λ

=a user defined weighting factor where y(t) is the predicted value and y(t) is y(t).

30. The method of claim 22, wherein the step of creating the linear model creates a linear model in accordance with the following relationship:

space="preserve" listing-type="equation">y(t)=-α

.sub.1 y(t-1)+-α

.sub.2 y(t-2)+b.sub.1 u(t-d-1)+b.sub.2 u(t-d-2)

where the dynamic gain k_d is defined by the relationship;

##EQU35##

31. The method of claim 30, wherein the gain k_d for the linear network is set equal to unity after training by varying the values b₁ and b₂ in the gain relationship.

32. A method for determining the dead-time d between an input variable u(t) to a plant and an output variable y(t) from the plant, comprising the steps of:

providing a historical data set of input variables u(t) for a given output variable y(t);

defining a minimum dead-time d_i,min and a maximum dead-time d_i,max;

varying the value of d_i from d_i,min to d_i,max for each input variable u_i (t-d_i);

determining the strength of the statistical relationship between the input variable u(t) and the output variable y(t) for each value of i as d_i varies from d_imin to d_imax ; and

select the value of j for d_i as the d_i value having the strongest statistical relationship for the given input variable.

33. The method of claim 32 wherein the step of determining determines the statistical independence of y(t) for a given value of u_i (t-d_i) as d_i varies from d_i,min--d_i,max, with the one of the u_i (t-d_i) values for d_i corresponding to the lowest statistical dependent comprising the strongest relationship value of u_i (t-d_i).

34. The method of claim 33 wherein the step of determining utilizes mutual information (MI) in accordance with the following relationship:

##EQU36## wherein x₁ (t) represents u_i (t-d_j) and x₂ (t) represents y(t).

35. The method of claim 34, wherein MI is computed by estimating the probability distribution of a given data set of u(t) and y(t) utilizing a binning method wherein a grid with two dimensions, u_i (t-d_i) and y(t), is divided into a fixed number of divisions.

36. The method of claim 32 wherein a plurality of output variables y(t) and u_i (t) and the step of determining the strength of the statistical relationship comprises determining the strength and the statistical between each of the input variables u(t) and each of the output variables y(t).