Residual activation neural network

• US 5,353,207 A
• Filed: 06/10/1992
• Issued: 10/04/1994
• Est. Priority Date: 06/10/1992
• Status: Expired due to Term

• Current Assignee: Rockwell Automation Technologies Incorporated (Allen-Bradley Co., Inc.)
• Original Assignee: Pavilion Technologies Incorporated (Rockwell Automation, Inc.)
• Inventors: Keeler, James D., Liano, Kadir, Hartman, Eric J., Ferguson, Ralph B.
• Primary Examiner(s): Smith, Jerry
• Assistant Examiner(s): BROWN, THOMAS
• Application Number:
  US07/896,755
• Time in Patent Office: 846 Days
• Field of Search: 364/148, 364/162, 364/160, 364/161, 364/163-165, 364/149-151, 395/21, 395/23, 395/24, 395/22
• US Class Current: 700/44
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: Y10S 706/906 Process plant

Abstract

A plant (72) is operable to receive control inputs c(t) and provide an output y(t). The plant (72) has associated therewith state variables s(t) that are not variable. A control network (74) is provided that accurately models the plant (72). The output of the control network (74) provides a predicted output which is combined with a desired output to generate an error. This error is back propagated through an inverse control network (76), which is the inverse of the control network (74) to generate a control error signal that is input to a distributed control system (73) to vary the control inputs to the plant (72) in order to change the output y(t) to meet the desired output. The control network (74) is comprised of a first network NET 1 that is operable to store a representation of the dependency of the control variables on the state variables. The predicted result is subtracted from the actual state variable input and stored as a residual in a residual layer (102). The output of the residual layer (102) is input to a hidden layer (108) which also receives the control inputs to generate a predicted output in an output layer (106). During back propagation of error, the residual values in the residual layer (102) are latched and only the control inputs allowed to vary.

Claims

1. A control network for controlling a plant having plant control inputs for receiving plant control variables and desired plant outputs, the plant outputs being a function of the plant control variables and external influences on the plant, comprising:

a control network input for receiving as network inputs the current plant control variables and desired plant outputs;

a control network output for outputting predicted plant control variables necessary to achieve the desired plant outputs;

a processing system for processing the received plant control variables through an inverse representation of the plant that represents the dependencies of the plant output on the plant control variables parameterized by an estimation of the external influences to provide the predicted plant control variables to achieve the desired plant outputs, said processing system having;

an estimation network for estimating the external influences on the plant and output estimated external influences,

means for parameterizing the inverse representation of the plant with the estimated influences.

a first intermediate output for providing a predicted plant output,

a first intermediate processing system for receiving the plant control variables from said control network input and the estimated external influences from said estimation network for processing through a predictive model of the plant to generate the predicted plant outputs for output from said intermediate output,

an error generation device for comparing the predicted plant outputs to the desired plant outputs and generating an error representing the difference therebetween,

a second intermediate processing system for processing the error through the inverse representation of the plant that represents the dependencies of the plant output on the plant control variables parameterized by the estimated external influences to output predicted control variable change values,

a control system for inputting said predicted control variable change values to the input of said first intermediate processing system for summing with the control variable input to provide a summed control variable value, and processing the summed control variable through said first processing system to minimize said error and output the summed control variable value as the predicted control variables; and

an interface device for inputting the predicted plant control variables that are output by said control network output to the plant as plant control variables to achieve the desired plant outputs.

2. The control network of claim 1, wherein the inverse representation of said processing system is a general non-linear inverse representation.

3. The control network of claim 1, wherein the control variable inputs are variables that can be manipulated.

4. The control network of claim 1, wherein said second intermediate processing system comprises:

a neural network having an input layer for receiving said error;

an output layer for providing the predicted output of the plant;

a hidden layer for mapping said input layer to said output layer through an inverse representation of the plant that represents the dependencies of the plant output on the plant control variables parameterized by the external influences to provide as an output from the output layer the control variable change values.

5. The control network of claim 1 wherein said control system utilizes a gradient descent procedure to minimize said error.

6. The control network of claim 1 wherein said control system utilizes a Monte Carlo technique to minimize said error.

7. The control network of claim 1, wherein said second intermediate processing system and said estimation network comprise:

a residual activation neural network having;

a residual neural network for receiving as inputs in an input layer the plant control variables and non-manipulatable plant state variables dependant on the plant control variables, and mapping the received plant control variables through a hidden layer to an output layer, the hidden layer having a representation of the dependencies of the plant state variables on the plant control variables to provide as an output from said output layer predicted state variables,

a residual layer for determining as a residual the difference between the plant state variables and the predicted state variables as an estimation of the external influences on the plant, and

a latch for latching said residual determined in said residual layer after determination thereof; and

a main neural network having;

an input layer for receiving the plant control variables and said latched residual,

an output layer for outputting a predicted plant output, and

a hidden layer for mapping said input layer to said output layer through a representation of the plant as a function of the plant control variable inputs and said latched residual, said main neural network operating in an inverse mode to receive at the output layer said error and back propagate said error through said hidden layer to said input layer with said residual latched in said latch to output from said input layer said predicted control variable change values.

8. A method for controlling a plant having plant outputs and plant control inputs for receiving plant control variables and desired plant outputs, the plant outputs being a function of the plant control variables and external influences on the plant, comprising the steps of:

receiving the current plant control variables and desired plant outputs;

estimating the external influences on the plant as estimated external influences;

parameterizing the inverse representation of the plant with the estimated external influences;

processing the received plant control variables through an inverse representation of the plant that represents the dependencies of the plant output on the plant control variables parameterized by an estimation of the external influences to provide the predicted plant control variables necessary to achieve the desired plant outputs, the step of processing including the steps of;

processing in a first intermediate processing step the plant control variables and the estimated external influences through a predictive model of the plant to generate the predicted plant outputs for output from an intermediate output,

comparing the predicted plant outputs to the desired plant outputs and generating an error representing the difference therebetween,

processing in a second intermediate processing step the error through the inverse representation of the plant that represents the dependencies of the plant output on the plant control variables parameterized by the estimated external influences to output predicted control variable change values, and

changing the input control variables to the first intermediate step by the control variable change values to provide the predicted plant control variables;

outputting on an output the predicted plant control variables necessary to achieve the desired plant outputs; and

controlling the plant with the predicted plant control variables.

9. The method of claim 8 wherein the inverse representation of the processing system is a general non-linear inverse representation.

10. The method of claim 8 wherein the control variables are variables that can be manipulated.

11. The method of claim 8 wherein the second intermediate processing step comprises:

receiving on input layer of a neural network the error;

mapping the neural network input layer to a neural network output layer through a neural network hidden layer having stored therein a local representation of the plant parameterized by the external influences; and

operating the neural network in an inverse relationship wherein the error is received as an input in the output layer and propagated through the hidden layer having a local inverse representation of the plant that represents the dependencies of the plant output on the plant control variables parameterized by the estimate external influences to provide as an output from the neural network input layer the predicted plant control variable change values, wherein the error is back propagated through the neural network hidden layer to the neural network input layer.

12. The method of claim 8 wherein the first intermediate processing step includes the step of estimating and comprises:

receiving the plant control variables on an input layer to a residual neural network and mapping the received plant control variables to a residual neural network output layer through a hidden layer, the hidden layer having a representation of the dependencies of non-manipulatable plant state variables on the plant control variables to provide from the output layer predicted state variables as a function of the plant control variables, the residual comprising the estimation of the external influences;

determining as a residual the difference between the plant state variables and the predicted state variables;

latching the determined residual after determination thereof;

receiving the plant control variables and the latched residual on an input layer of a main neural network; and

mapping the input layer of the main neural network to an output layer of the main neural network through a main neural network hidden layer having stored therein a representation of the plant as a function of the plant control variable inputs and the residual, to output from the output layer the predicted plant outputs.

13. The method of claim 12 wherein the step of changing the input control variables comprises iteratively changing the input control variables by summing with the predicted control variable change values to minimize the error in accordance with a gradient descent technique.