Control system using an adaptive neural network for target and path optimization for a multivariable, nonlinear process

• US 5,640,491 A
• Filed: 12/18/1995
• Issued: 06/17/1997
• Est. Priority Date: 09/14/1992
• Status: Expired due to Fees

• Current Assignee: Aspentech Corporation, Texaco Incorporated (Chevron Corp.)
• Original Assignee: NeuralWare, Inc. (Aspentech Corporation), Texaco Incorporated (Chevron Corp.)
• Inventors: Graettinger, Timothy J., Federowicz, Alexander J., Bhat, Naveen V., Dubose, Paul A., Braden, William B., Heckendoorn, Kent E.
• Primary Examiner(s): Davis, George B.
• Application Number:
  US08/574,250
• Time in Patent Office: 547 Days
• Field of Search: 395/22, 395/23, 395/903, 395/24, 395/906
• US Class Current: 706/25
• CPC Class Codes: B01D 3/425 Head-, bottom- and feed stream
• CPC Class Codes: B01J 19/0033 Optimalisation processes, i.e. processes with adaptive control systems adaptive control systems per se G05B13/00
• CPC Class Codes: G05B 13/027 using neural networks only

Abstract

A control system having four major components: a target optimizer, a path optimizer, a neural network adaptation controller and a neural network. In the target optimizer, the controlled variables are optimized to provide the most economically desirable outputs, subject to operating constraints. Various manipulated variable and disturbance values are provided for modeling purposes. The neural network receives as inputs a plurality of settings for each manipulated and disturbance variable. For target optimization all the neural network input values are set equal to produce a steady state controlled variable value. The entire process is repeated with differing manipulated variable values until an optimal solution develops. The resulting target controlled and manipulated variable values are provided to the path optimizer to allow the manipulated variables to be adjusted to obtain the target output. Various manipulated variable values are developed to model moves from current to desired values. In this case trend indicating values of the manipulated and disturbance variables are provided to produce time varying values of the controlled variables. The process is repeated until an optimal path is obtained, at which time the manipulated variable values are applied to the actual process. On a periodic basis all of the disturbance, manipulated and controlled variables are sampled to find areas where the training of the neural network is sparse or where high dynamic conditions are indicated. These values are added to the set of values used to train the neural network.

Claims

1. A method for developing a sample set for training a neural network, the method comprising the steps of:

obtaining values of various inputs and outputs of the neural network at a specific time to form a new sample;

developing an n-dimensional matrix of cells, wherein n is a total number of inputs and outputs of the neural network, an axis has a total range corresponding to a value range of respective inputs or outputs of the neural network, each axis total range being subdivided into cell ranges to result in a plurality of cell ranges for each axis, whereby a total number of cells in said matrix is a product of number of cell ranges in the total range for each axis;

determining a corresponding cell in said n-dimensional matrix based on obtained input and output values in said new sample;

determining a number of previously stored samples in said corresponding cell; and

adding said new sample to previously stored samples if said determined number is below a predetermined list.

2. The method of claim 1, further comprising the steps of:

if said determined number of previously stored samples exceeds said predetermined number, further determining if said new sample includes certain values indicating other predetermined conditions desirable for training the neural network; and

removing a previously stored sample in said corresponding cell if said new sample includes said certain values and adding said new sample to the previously stored samples.

3. The method of claim 2, wherein a previously stored sample which is removed is an oldest sample.

4. The method of claim 1, wherein a previously stored sample which is removed is an oldest sample which is not marked as being a permanent sample.